Skip to content

P
PaddleSeg

PaddlePaddle / PaddleSeg

通知 285
Star 8
Fork 1
P
PaddleSeg
提交 7835c042 编写于 作者: C chenguowei01
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleSeg into humanseg

上级 7781902e f0dba857
展开全部 隐藏空白更改
内联 并排
Showing with 3422 addition and 1780 deletion
README.md
浏览文件 @ 7835c042
......@@ -35,7 +35,7 @@ PaddleSeg是基于[PaddlePaddle](https://www.paddlepaddle.org.cn)开发的端到
- **高性能**
PaddleSeg支持多进程I/O、多卡并行、跨卡Batch Norm同步等训练加速策略，结合飞桨核心框架的显存优化功能，可大幅度减少分割模型的显存开销，让开发者更低成本、更高效地完成图像分割训练。
PaddleSeg支持多进程I/O、多卡并行等训练加速策略，结合飞桨核心框架的显存优化功能，可大幅度减少分割模型的显存开销，让开发者更低成本、更高效地完成图像分割训练。
- **工业级部署**
......
contrib/ACE2P/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
contrib/ACE2P/config.py
浏览文件 @ 7835c042
# -*- coding: utf-8 -*-
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from utils.util import AttrDict, merge_cfg_from_args, get_arguments
import os
......@@ -19,10 +33,10 @@ cfg.class_num = 20
# 均值, 图像预处理减去的均值
cfg.MEAN = 0.406, 0.456, 0.485
# 标准差，图像预处理除以标准差
cfg.STD = 0.225, 0.224, 0.229
cfg.STD = 0.225, 0.224, 0.229
# 多尺度预测时图像尺寸
cfg.multi_scales = (377,377), (473,473), (567,567)
cfg.multi_scales = (377, 377), (473, 473), (567, 567)
# 多尺度预测时图像是否水平翻转
cfg.flip = True
......
contrib/ACE2P/download_ACE2P.py
浏览文件 @ 7835c042
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/ACE2P/infer.py
浏览文件 @ 7835c042
# -*- coding: utf-8 -*-
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import cv2
import numpy as np
......@@ -12,18 +26,19 @@ config = importlib.import_module('config')
cfg = getattr(config, 'cfg')
# paddle垃圾回收策略FLAG，ACE2P模型较大，当显存不够时建议开启
os.environ['FLAGS_eager_delete_tensor_gb']='0.0'
os.environ['FLAGS_eager_delete_tensor_gb'] = '0.0'
import paddle.fluid as fluid
# 预测数据集类
class TestDataSet():
def __init__(self):
self.data_dir = cfg.data_dir
self.data_dir = cfg.data_dir
self.data_list_file = cfg.data_list_file
self.data_list = self.get_data_list()
self.data_num = len(self.data_list)
def get_data_list(self):
# 获取预测图像路径列表
data_list = []
......@@ -56,10 +71,10 @@ class TestDataSet():
img_path = self.data_list[index]
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
if img is None:
return img, img,img_path, None
return img, img, img_path, None
img_name = img_path.split(os.sep)[-1]
name_prefix = img_name.replace('.'+img_name.split('.')[-1],'')
name_prefix = img_name.replace('.' + img_name.split('.')[-1], '')
img_shape = img.shape[:2]
img_process = self.preprocess(img)
......@@ -90,39 +105,44 @@ def infer():
if image is None:
print(im_name, 'is None')
continue
# 预测
if cfg.example == 'ACE2P':
# ACE2P模型使用多尺度预测
reader = importlib.import_module('reader')
multi_scale_test = getattr(reader, 'multi_scale_test')
parsing, logits = multi_scale_test(exe, test_prog, feed_name, fetch_list, image, im_shape)
parsing, logits = multi_scale_test(exe, test_prog, feed_name,
fetch_list, image, im_shape)
else:
# HumanSeg,RoadLine模型单尺度预测
result = exe.run(program=test_prog, feed={feed_name[0]: image}, fetch_list=fetch_list)
result = exe.run(
program=test_prog,
feed={feed_name[0]: image},
fetch_list=fetch_list)
parsing = np.argmax(result[0][0], axis=0)
parsing = cv2.resize(parsing.astype(np.uint8), im_shape[::-1])
# 预测结果保存
result_path = os.path.join(cfg.vis_dir, im_name + '.png')
if cfg.example == 'HumanSeg':
logits = result[0][0][1]*255
logits = result[0][0][1] * 255
logits = cv2.resize(logits, im_shape[::-1])
ret, logits = cv2.threshold(logits, thresh, 0, cv2.THRESH_TOZERO)
logits = 255 *(logits - thresh)/(255 - thresh)
logits = 255 * (logits - thresh) / (255 - thresh)
# 将分割结果添加到alpha通道
rgba = np.concatenate((ori_img, np.expand_dims(logits, axis=2)), axis=2)
rgba = np.concatenate((ori_img, np.expand_dims(logits, axis=2)),
axis=2)
cv2.imwrite(result_path, rgba)
else:
else:
output_im = PILImage.fromarray(np.asarray(parsing, dtype=np.uint8))
output_im.putpalette(palette)
output_im.save(result_path)
if (idx + 1) % 100 == 0:
print('%d processd' % (idx + 1))
print('%d processd done' % (idx + 1))
print('%d processd done' % (idx + 1))
return 0
......
contrib/ACE2P/reader.py
浏览文件 @ 7835c042
# -*- coding: utf-8 -*-
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle.fluid as fluid
from config import cfg
import cv2
def get_affine_points(src_shape, dst_shape, rot_grad=0):
# 获取图像和仿射后图像的三组对应点坐标
# 三组点为仿射变换后图像的中心点, [w/2,0], [0,0]，及对应原始图像的点
......@@ -23,7 +38,7 @@ def get_affine_points(src_shape, dst_shape, rot_grad=0):
# 原始图像三组点
points = [[0, 0]] * 3
points[0] = (np.array([w, h]) - 1) * 0.5
points[0] = (np.array([w, h]) - 1) * 0.5
points[1] = points[0] + 0.5 * affine_shape[0] * np.array([sin_v, -cos_v])
points[2] = points[1] - 0.5 * affine_shape[1] * np.array([cos_v, sin_v])
......@@ -34,6 +49,7 @@ def get_affine_points(src_shape, dst_shape, rot_grad=0):
return points, points_trans
def preprocess(im):
# ACE2P模型数据预处理
im_shape = im.shape[:2]
......@@ -42,13 +58,10 @@ def preprocess(im):
# 获取图像和仿射变换后图像的对应点坐标
points, points_trans = get_affine_points(im_shape, scale)
# 根据对应点集获得仿射矩阵
trans = cv2.getAffineTransform(np.float32(points),
np.float32(points_trans))
trans = cv2.getAffineTransform(
np.float32(points), np.float32(points_trans))
# 根据仿射矩阵对图像进行仿射
input = cv2.warpAffine(im,
trans,
scale[::-1],
flags=cv2.INTER_LINEAR)
input = cv2.warpAffine(im, trans, scale[::-1], flags=cv2.INTER_LINEAR)
# 减均值测，除以方差，转换数据格式为NCHW
input = input.astype(np.float32)
......@@ -66,19 +79,20 @@ def preprocess(im):
return input_images
def multi_scale_test(exe, test_prog, feed_name, fetch_list,
input_ims, im_shape):
def multi_scale_test(exe, test_prog, feed_name, fetch_list, input_ims,
im_shape):
# 由于部分类别分左右部位, flipped_idx为其水平翻转后对应的标签
flipped_idx = (15, 14, 17, 16, 19, 18)
ms_outputs = []
# 多尺度预测
for idx, scale in enumerate(cfg.multi_scales):
input_im = input_ims[idx]
parsing_output = exe.run(program=test_prog,
feed={feed_name[0]: input_im},
fetch_list=fetch_list)
parsing_output = exe.run(
program=test_prog,
feed={feed_name[0]: input_im},
fetch_list=fetch_list)
output = parsing_output[0][0]
if cfg.flip:
# 若水平翻转，对部分类别进行翻转，与原始预测结果取均值
......@@ -92,7 +106,8 @@ def multi_scale_test(exe, test_prog, feed_name, fetch_list,
# 仿射变换回图像原始尺寸
points, points_trans = get_affine_points(im_shape, scale)
M = cv2.getAffineTransform(np.float32(points_trans), np.float32(points))
logits_result = cv2.warpAffine(output, M, im_shape[::-1], flags=cv2.INTER_LINEAR)
logits_result = cv2.warpAffine(
output, M, im_shape[::-1], flags=cv2.INTER_LINEAR)
ms_outputs.append(logits_result)
# 多尺度预测结果求均值，求预测概率最大的类别
......@@ -100,4 +115,3 @@ def multi_scale_test(exe, test_prog, feed_name, fetch_list,
ms_fused_parsing_output = np.mean(ms_fused_parsing_output, axis=0)
parsing = np.argmax(ms_fused_parsing_output, axis=2)
return parsing, ms_fused_parsing_output
contrib/ACE2P/utils/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
contrib/ACE2P/utils/palette.py
浏览文件 @ 7835c042
......@@ -7,6 +7,7 @@
## This source code is licensed under the MIT-style license found in the
## LICENSE file in the root directory of this source tree
##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
contrib/ACE2P/utils/util.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import argparse
import os
def get_arguments():
parser = argparse.ArgumentParser()
parser.add_argument("--use_gpu",
action="store_true",
help="Use gpu or cpu to test.")
parser.add_argument('--example',
type=str,
help='RoadLine, HumanSeg or ACE2P')
parser.add_argument(
"--use_gpu", action="store_true", help="Use gpu or cpu to test.")
parser.add_argument(
'--example', type=str, help='RoadLine, HumanSeg or ACE2P')
return parser.parse_args()
......@@ -34,6 +48,7 @@ class AttrDict(dict):
else:
self[name] = value
def merge_cfg_from_args(args, cfg):
"""Merge config keys, values in args into the global config."""
for k, v in vars(args).items():
......@@ -44,4 +59,3 @@ def merge_cfg_from_args(args, cfg):
value = v
if value is not None:
cfg[k] = value
contrib/HumanSeg/datasets/__init__.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/HumanSeg/datasets/dataset.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/datasets/shared_queue/__init__.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/datasets/shared_queue/queue.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/datasets/shared_queue/sharedmemory.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -12,9 +13,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# utils for memory management which is allocated on sharedmemory,
# note that these structures may not be thread-safe
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
contrib/HumanSeg/export.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import models
import argparse
......
contrib/HumanSeg/infer.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
import os.path as osp
......
contrib/HumanSeg/models/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .humanseg import HumanSegMobile
from .humanseg import HumanSegServer
from .humanseg import HumanSegLite
......
contrib/HumanSeg/models/humanseg.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......@@ -24,6 +25,7 @@ import time
import tqdm
import cv2
import yaml
import shutil
import paddleslim as slim
import utils
......@@ -102,7 +104,7 @@ class SegModel(object):
# 当前模型状态
self.status = 'Normal'
def _get_single_car_bs(self, batch_size):
def _get_single_card_bs(self, batch_size):
if batch_size % len(self.places) == 0:
return int(batch_size // len(self.places))
else:
......@@ -144,7 +146,7 @@ class SegModel(object):
capacity=64,
use_double_buffer=True,
iterable=True)
batch_size_each_gpu = self._get_single_car_bs(batch_size)
batch_size_each_gpu = self._get_single_card_bs(batch_size)
self.train_data_loader.set_sample_list_generator(
dataset.generator(batch_size=batch_size_each_gpu),
places=self.places)
......@@ -242,30 +244,11 @@ class SegModel(object):
if self.status == 'Normal':
fluid.save(self.train_prog, osp.join(save_dir, 'model'))
model_info['status'] = 'Normal'
elif self.status == 'Quant':
float_prog, _ = slim.quant.convert(
self.test_prog, self.exe.place, save_int8=True)
test_input_names = [
var.name for var in list(self.test_inputs.values())
]
test_outputs = list(self.test_outputs.values())
fluid.io.save_inference_model(
dirname=save_dir,
executor=self.exe,
params_filename='__params__',
feeded_var_names=test_input_names,
target_vars=test_outputs,
main_program=float_prog)
fluid.save(self.test_prog, osp.join(save_dir, 'model'))
model_info['status'] = 'QuantOnline'
model_info['_ModelInputsOutputs'] = dict()
model_info['_ModelInputsOutputs']['test_inputs'] = [
[k, v.name] for k, v in self.test_inputs.items()
]
model_info['_ModelInputsOutputs']['test_outputs'] = [
[k, v.name] for k, v in self.test_outputs.items()
]
model_info['status'] = self.status
with open(
osp.join(save_dir, 'model.yml'), encoding='utf-8',
mode='w') as f:
......@@ -307,40 +290,57 @@ class SegModel(object):
logging.info("Model for inference deploy saved in {}.".format(save_dir))
def export_quant_model(self,
dataset,
save_dir,
dataset=None,
save_dir=None,
batch_size=1,
batch_nums=10,
cache_dir="./.temp"):
self.arrange_transform(transforms=dataset.transforms, mode='quant')
dataset.num_samples = batch_size * batch_nums
try:
from utils import HumanSegPostTrainingQuantization
except:
raise Exception(
"Model Quantization is not available, try to upgrade your paddlepaddle>=1.7.0"
)
is_use_cache_file = True
if cache_dir is None:
is_use_cache_file = False
post_training_quantization = HumanSegPostTrainingQuantization(
executor=self.exe,
dataset=dataset,
program=self.test_prog,
inputs=self.test_inputs,
outputs=self.test_outputs,
batch_size=batch_size,
batch_nums=batch_nums,
scope=None,
algo='KL',
quantizable_op_type=["conv2d", "depthwise_conv2d", "mul"],
is_full_quantize=False,
is_use_cache_file=is_use_cache_file,
cache_dir=cache_dir)
post_training_quantization.quantize()
post_training_quantization.save_quantized_model(save_dir)
if cache_dir is not None:
os.system('rm -r' + cache_dir)
cache_dir=".temp",
quant_type="offline"):
if quant_type == "offline":
self.arrange_transform(transforms=dataset.transforms, mode='quant')
dataset.num_samples = batch_size * batch_nums
try:
from utils import HumanSegPostTrainingQuantization
except:
raise Exception(
"Model Quantization is not available, try to upgrade your paddlepaddle>=1.8.1"
)
is_use_cache_file = True
if cache_dir is None:
is_use_cache_file = False
post_training_quantization = HumanSegPostTrainingQuantization(
executor=self.exe,
dataset=dataset,
program=self.test_prog,
inputs=self.test_inputs,
outputs=self.test_outputs,
batch_size=batch_size,
batch_nums=batch_nums,
scope=None,
algo='KL',
quantizable_op_type=["conv2d", "depthwise_conv2d", "mul"],
is_full_quantize=False,
is_use_cache_file=is_use_cache_file,
cache_dir=cache_dir)
post_training_quantization.quantize()
post_training_quantization.save_quantized_model(save_dir)
if cache_dir is not None:
shutil.rmtree(cache_dir)
else:
float_prog, _ = slim.quant.convert(
self.test_prog, self.exe.place, save_int8=True)
test_input_names = [
var.name for var in list(self.test_inputs.values())
]
test_outputs = list(self.test_outputs.values())
fluid.io.save_inference_model(
dirname=save_dir,
executor=self.exe,
params_filename='__params__',
feeded_var_names=test_input_names,
target_vars=test_outputs,
main_program=float_prog)
model_info = self.get_model_info()
model_info['status'] = 'Quant'
......@@ -592,6 +592,16 @@ class SegModel(object):
'Current evaluated best model in eval_dataset is epoch_{}, miou={}'
.format(best_model_epoch, best_miou))
if quant:
if osp.exists(osp.join(save_dir, "best_model")):
fluid.load(
program=self.test_prog,
model_path=osp.join(save_dir, "best_model"),
executor=self.exe)
self.export_quant_model(
save_dir=osp.join(save_dir, "best_model_export"),
quant_type="online")
def evaluate(self, eval_dataset, batch_size=1, epoch_id=None):
"""评估。
......
contrib/HumanSeg/models/load_model.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -24,7 +25,7 @@ import models
def load_model(model_dir):
if not osp.exists(osp.join(model_dir, "model.yml")):
raise Exception("There's not model.yml in {}".format(model_dir))
raise Exception("There's no model.yml in {}".format(model_dir))
with open(osp.join(model_dir, "model.yml")) as f:
info = yaml.load(f.read(), Loader=yaml.Loader)
status = info['status']
......@@ -33,7 +34,7 @@ def load_model(model_dir):
raise Exception("There's no attribute {} in models".format(
info['Model']))
model = getattr(models, info['Model'])(**info['_init_params'])
if status == "Normal":
if status in ["Normal", "QuantOnline"]:
startup_prog = fluid.Program()
model.test_prog = fluid.Program()
with fluid.program_guard(model.test_prog, startup_prog):
......@@ -41,11 +42,16 @@ def load_model(model_dir):
model.test_inputs, model.test_outputs = model.build_net(
mode='test')
model.test_prog = model.test_prog.clone(for_test=True)
if status == "QuantOnline":
print('test quant online')
import paddleslim as slim
model.test_prog = slim.quant.quant_aware(
model.test_prog, model.exe.place, for_test=True)
model.exe.run(startup_prog)
import pickle
with open(osp.join(model_dir, 'model.pdparams'), 'rb') as f:
load_dict = pickle.load(f)
fluid.io.set_program_state(model.test_prog, load_dict)
fluid.load(model.test_prog, osp.join(model_dir, 'model'))
if status == "QuantOnline":
model.test_prog = slim.quant.convert(model.test_prog,
model.exe.place)
elif status in ['Infer', 'Quant']:
[prog, input_names, outputs] = fluid.io.load_inference_model(
......
contrib/HumanSeg/nets/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .backbone import mobilenet_v2
from .backbone import xception
from .deeplabv3p import DeepLabv3p
......
contrib/HumanSeg/nets/backbone/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .mobilenet_v2 import MobileNetV2
from .xception import Xception
contrib/HumanSeg/nets/backbone/mobilenet_v2.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -10,6 +11,7 @@
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
......
contrib/HumanSeg/nets/backbone/xception.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/nets/deeplabv3p.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/nets/hrnet.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/nets/libs.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/nets/seg_modules.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/nets/shufflenet_slim.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
contrib/HumanSeg/pretrained_weights/download_pretrained_weights.py
浏览文件 @ 7835c042
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/HumanSeg/quant_offline.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
from datasets.dataset import Dataset
import transforms
......
contrib/HumanSeg/quant_online.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
from datasets.dataset import Dataset
from models import HumanSegMobile, HumanSegLite, HumanSegServer
......
contrib/HumanSeg/requirements.txt
浏览文件 @ 7835c042
visualdl == 2.0.0b1
visualdl >= 2.0.0b1
paddleslim
contrib/HumanSeg/train.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
from datasets.dataset import Dataset
from models import HumanSegMobile, HumanSegLite, HumanSegServer
......
contrib/HumanSeg/transforms/__init__.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/HumanSeg/transforms/functional.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/HumanSeg/transforms/transforms.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/HumanSeg/utils/__init__.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/HumanSeg/utils/humanseg_postprocess.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import cv2
import os
......
contrib/HumanSeg/utils/logging.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/utils/metrics.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/utils/post_quantization.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/HumanSeg/utils/utils.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -205,11 +206,9 @@ def load_pretrained_weights(exe, main_prog, weights_dir, fuse_bn=False):
vars_to_load.append(var)
logging.debug("Weight {} will be load".format(var.name))
fluid.io.load_vars(
executor=exe,
dirname=weights_dir,
main_program=main_prog,
vars=vars_to_load)
params_dict = fluid.io.load_program_state(
weights_dir, var_list=vars_to_load)
fluid.io.set_program_state(main_prog, params_dict)
if len(vars_to_load) == 0:
logging.warning(
"There is no pretrain weights loaded, maybe you should check you pretrain model!"
......
contrib/HumanSeg/val.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
from datasets.dataset import Dataset
import transforms
......
contrib/HumanSeg/video_infer.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
import os.path as osp
......
contrib/LaneNet/data_aug.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -21,6 +21,7 @@ from models.model_builder import ModelPhase
from pdseg.data_aug import get_random_scale, randomly_scale_image_and_label, random_rotation, \
rand_scale_aspect, hsv_color_jitter, rand_crop
def resize(img, grt=None, grt_instance=None, mode=ModelPhase.TRAIN):
"""
改变图像及标签图像尺寸
......@@ -44,7 +45,8 @@ def resize(img, grt=None, grt_instance=None, mode=ModelPhase.TRAIN):
if grt is not None:
grt = cv2.resize(grt, target_size, interpolation=cv2.INTER_NEAREST)
if grt_instance is not None:
grt_instance = cv2.resize(grt_instance, target_size, interpolation=cv2.INTER_NEAREST)
grt_instance = cv2.resize(
grt_instance, target_size, interpolation=cv2.INTER_NEAREST)
elif cfg.AUG.AUG_METHOD == 'stepscaling':
if mode == ModelPhase.TRAIN:
min_scale_factor = cfg.AUG.MIN_SCALE_FACTOR
......
contrib/LaneNet/dataset/download_tusimple.py
浏览文件 @ 7835c042
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/LaneNet/eval.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -122,7 +122,10 @@ def evaluate(cfg, ckpt_dir=None, use_gpu=False, use_mpio=False, **kwargs):
if ckpt_dir is not None:
print('load test model:', ckpt_dir)
fluid.io.load_params(exe, ckpt_dir, main_program=test_prog)
try:
fluid.load(test_prog, os.path.join(ckpt_dir, 'model'), exe)
except:
fluid.io.load_params(exe, ckpt_dir, main_program=test_prog)
# Use streaming confusion matrix to calculate mean_iou
np.set_printoptions(
......
contrib/LaneNet/loss.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -19,8 +19,9 @@ from utils.config import cfg
def unsorted_segment_sum(data, segment_ids, unique_labels, feature_dims):
zeros = fluid.layers.fill_constant_batch_size_like(unique_labels, shape=[1, feature_dims],
dtype='float32', value=0)
unique_labels_shape = fluid.layers.shape(unique_labels)
zeros = fluid.layers.fill_constant(
shape=[unique_labels_shape[0], feature_dims], dtype='float32', value=0)
segment_ids = fluid.layers.unsqueeze(segment_ids, axes=[1])
segment_ids.stop_gradient = True
segment_sum = fluid.layers.scatter_nd_add(zeros, segment_ids, data)
......@@ -30,29 +31,23 @@ def unsorted_segment_sum(data, segment_ids, unique_labels, feature_dims):
def norm(x, axis=-1):
distance = fluid.layers.reduce_sum(fluid.layers.abs(x), dim=axis, keep_dim=True)
distance = fluid.layers.reduce_sum(
fluid.layers.abs(x), dim=axis, keep_dim=True)
return distance
def discriminative_loss_single(
prediction,
correct_label,
feature_dim,
label_shape,
delta_v,
delta_d,
param_var,
param_dist,
param_reg):
correct_label = fluid.layers.reshape(
correct_label, [
label_shape[1] * label_shape[0]])
def discriminative_loss_single(prediction, correct_label, feature_dim,
label_shape, delta_v, delta_d, param_var,
param_dist, param_reg):
correct_label = fluid.layers.reshape(correct_label,
[label_shape[1] * label_shape[0]])
prediction = fluid.layers.transpose(prediction, [1, 2, 0])
reshaped_pred = fluid.layers.reshape(
prediction, [
label_shape[1] * label_shape[0], feature_dim])
prediction, [label_shape[1] * label_shape[0], feature_dim])
unique_labels, unique_id, counts = fluid.layers.unique_with_counts(correct_label)
unique_labels, unique_id, counts = fluid.layers.unique_with_counts(
correct_label)
correct_label.stop_gradient = True
counts = fluid.layers.cast(counts, 'float32')
num_instances = fluid.layers.shape(unique_labels)
......@@ -69,24 +64,29 @@ def discriminative_loss_single(
distance = norm(tmp)
distance = distance - delta_v
distance_pos = fluid.layers.greater_equal(distance, fluid.layers.zeros_like(distance))
distance_pos = fluid.layers.greater_equal(distance,
fluid.layers.zeros_like(distance))
distance_pos = fluid.layers.cast(distance_pos, 'float32')
distance = distance * distance_pos
distance = fluid.layers.square(distance)
l_var = unsorted_segment_sum(distance, unique_id, unique_labels, feature_dims=1)
l_var = unsorted_segment_sum(
distance, unique_id, unique_labels, feature_dims=1)
l_var = fluid.layers.elementwise_div(l_var, counts_rsp)
l_var = fluid.layers.reduce_sum(l_var)
l_var = l_var / fluid.layers.cast(num_instances * (num_instances - 1), 'float32')
l_var = l_var / fluid.layers.cast(num_instances * (num_instances - 1),
'float32')
mu_interleaved_rep = fluid.layers.expand(mu, [num_instances, 1])
mu_band_rep = fluid.layers.expand(mu, [1, num_instances])
mu_band_rep = fluid.layers.reshape(mu_band_rep, (num_instances * num_instances, feature_dim))
mu_band_rep = fluid.layers.reshape(
mu_band_rep, (num_instances * num_instances, feature_dim))
mu_diff = fluid.layers.elementwise_sub(mu_band_rep, mu_interleaved_rep)
intermediate_tensor = fluid.layers.reduce_sum(fluid.layers.abs(mu_diff), dim=1)
intermediate_tensor = fluid.layers.reduce_sum(
fluid.layers.abs(mu_diff), dim=1)
intermediate_tensor.stop_gradient = True
zero_vector = fluid.layers.zeros([1], 'float32')
bool_mask = fluid.layers.not_equal(intermediate_tensor, zero_vector)
......@@ -95,7 +95,8 @@ def discriminative_loss_single(
mu_norm = norm(mu_diff_bool)
mu_norm = 2. * delta_d - mu_norm
mu_norm_pos = fluid.layers.greater_equal(mu_norm, fluid.layers.zeros_like(mu_norm))
mu_norm_pos = fluid.layers.greater_equal(mu_norm,
fluid.layers.zeros_like(mu_norm))
mu_norm_pos = fluid.layers.cast(mu_norm_pos, 'float32')
mu_norm = mu_norm * mu_norm_pos
mu_norm_pos.stop_gradient = True
......@@ -122,8 +123,8 @@ def discriminative_loss(prediction, correct_label, feature_dim, image_shape,
output_ta_reg = 0.
for i in range(batch_size):
disc_loss_single, l_var_single, l_dist_single, l_reg_single = discriminative_loss_single(
prediction[i], correct_label[i], feature_dim, image_shape, delta_v, delta_d, param_var, param_dist,
param_reg)
prediction[i], correct_label[i], feature_dim, image_shape, delta_v,
delta_d, param_var, param_dist, param_reg)
output_ta_loss += disc_loss_single
output_ta_var += l_var_single
output_ta_dist += l_dist_single
......@@ -134,5 +135,3 @@ def discriminative_loss(prediction, correct_label, feature_dim, image_shape,
l_dist = output_ta_dist / batch_size
l_reg = output_ta_reg / batch_size
return disc_loss, l_var, l_dist, l_reg
contrib/LaneNet/models/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/LaneNet/models/model_builder.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/LaneNet/models/modeling/lanenet.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -18,7 +18,6 @@ from __future__ import print_function
import paddle.fluid as fluid
from utils.config import cfg
from pdseg.models.libs.model_libs import scope, name_scope
from pdseg.models.libs.model_libs import bn, bn_relu, relu
......@@ -86,7 +85,12 @@ def bottleneck(inputs,
with scope('down_sample'):
inputs_shape = inputs.shape
with scope('main_max_pool'):
net_main = fluid.layers.conv2d(inputs, inputs_shape[1], filter_size=3, stride=2, padding='SAME')
net_main = fluid.layers.conv2d(
inputs,
inputs_shape[1],
filter_size=3,
stride=2,
padding='SAME')
#First get the difference in depth to pad, then pad with zeros only on the last dimension.
depth_to_pad = abs(inputs_shape[1] - output_depth)
......@@ -95,12 +99,16 @@ def bottleneck(inputs,
net_main = fluid.layers.pad(net_main, paddings=paddings)
with scope('block1'):
net = conv(inputs, reduced_depth, [2, 2], stride=2, padding='same')
net = conv(
inputs, reduced_depth, [2, 2], stride=2, padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
with scope('block2'):
net = conv(net, reduced_depth, [filter_size, filter_size], padding='same')
net = conv(
net,
reduced_depth, [filter_size, filter_size],
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
......@@ -137,13 +145,18 @@ def bottleneck(inputs,
# Second conv block --- apply dilated convolution here
with scope('block2'):
net = conv(net, reduced_depth, filter_size, padding='SAME', dilation=dilation_rate)
net = conv(
net,
reduced_depth,
filter_size,
padding='SAME',
dilation=dilation_rate)
net = bn(net)
net = prelu(net, decoder=decoder)
# Final projection with 1x1 kernel (Expansion)
with scope('block3'):
net = conv(net, output_depth, [1,1])
net = conv(net, output_depth, [1, 1])
net = bn(net)
net = prelu(net, decoder=decoder)
......@@ -172,9 +185,11 @@ def bottleneck(inputs,
# Second conv block --- apply asymmetric conv here
with scope('block2'):
with scope('asymmetric_conv2a'):
net = conv(net, reduced_depth, [filter_size, 1], padding='same')
net = conv(
net, reduced_depth, [filter_size, 1], padding='same')
with scope('asymmetric_conv2b'):
net = conv(net, reduced_depth, [1, filter_size], padding='same')
net = conv(
net, reduced_depth, [1, filter_size], padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
......@@ -211,7 +226,8 @@ def bottleneck(inputs,
with scope('unpool'):
net_unpool = conv(inputs, output_depth, [1, 1])
net_unpool = bn(net_unpool)
net_unpool = fluid.layers.resize_bilinear(net_unpool, out_shape=output_shape[2:])
net_unpool = fluid.layers.resize_bilinear(
net_unpool, out_shape=output_shape[2:])
# First 1x1 projection to reduce depth
with scope('block1'):
......@@ -220,7 +236,12 @@ def bottleneck(inputs,
net = prelu(net, decoder=decoder)
with scope('block2'):
net = deconv(net, reduced_depth, filter_size=filter_size, stride=2, padding='same')
net = deconv(
net,
reduced_depth,
filter_size=filter_size,
stride=2,
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
......@@ -253,7 +274,10 @@ def bottleneck(inputs,
# Second conv block
with scope('block2'):
net = conv(net, reduced_depth, [filter_size, filter_size], padding='same')
net = conv(
net,
reduced_depth, [filter_size, filter_size],
padding='same')
net = bn(net)
net = prelu(net, decoder=decoder)
......@@ -281,17 +305,33 @@ def ENet_stage1(inputs, name_scope='stage1_block'):
= bottleneck(inputs, output_depth=64, filter_size=3, regularizer_prob=0.01, type=DOWNSAMPLING,
name_scope='bottleneck1_0')
with scope('bottleneck1_1'):
net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01,
name_scope='bottleneck1_1')
net = bottleneck(
net,
output_depth=64,
filter_size=3,
regularizer_prob=0.01,
name_scope='bottleneck1_1')
with scope('bottleneck1_2'):
net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01,
name_scope='bottleneck1_2')
net = bottleneck(
net,
output_depth=64,
filter_size=3,
regularizer_prob=0.01,
name_scope='bottleneck1_2')
with scope('bottleneck1_3'):
net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01,
name_scope='bottleneck1_3')
net = bottleneck(
net,
output_depth=64,
filter_size=3,
regularizer_prob=0.01,
name_scope='bottleneck1_3')
with scope('bottleneck1_4'):
net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.01,
name_scope='bottleneck1_4')
net = bottleneck(
net,
output_depth=64,
filter_size=3,
regularizer_prob=0.01,
name_scope='bottleneck1_4')
return net, inputs_shape_1
......@@ -302,17 +342,38 @@ def ENet_stage2(inputs, name_scope='stage2_block'):
name_scope='bottleneck2_0')
for i in range(2):
with scope('bottleneck2_{}'.format(str(4 * i + 1))):
net = bottleneck(net, output_depth=128, filter_size=3, regularizer_prob=0.1,
name_scope='bottleneck2_{}'.format(str(4 * i + 1)))
net = bottleneck(
net,
output_depth=128,
filter_size=3,
regularizer_prob=0.1,
name_scope='bottleneck2_{}'.format(str(4 * i + 1)))
with scope('bottleneck2_{}'.format(str(4 * i + 2))):
net = bottleneck(net, output_depth=128, filter_size=3, regularizer_prob=0.1, type=DILATED, dilation_rate=(2 ** (2*i+1)),
name_scope='bottleneck2_{}'.format(str(4 * i + 2)))
net = bottleneck(
net,
output_depth=128,
filter_size=3,
regularizer_prob=0.1,
type=DILATED,
dilation_rate=(2**(2 * i + 1)),
name_scope='bottleneck2_{}'.format(str(4 * i + 2)))
with scope('bottleneck2_{}'.format(str(4 * i + 3))):
net = bottleneck(net, output_depth=128, filter_size=5, regularizer_prob=0.1, type=ASYMMETRIC,
name_scope='bottleneck2_{}'.format(str(4 * i + 3)))
net = bottleneck(
net,
output_depth=128,
filter_size=5,
regularizer_prob=0.1,
type=ASYMMETRIC,
name_scope='bottleneck2_{}'.format(str(4 * i + 3)))
with scope('bottleneck2_{}'.format(str(4 * i + 4))):
net = bottleneck(net, output_depth=128, filter_size=3, regularizer_prob=0.1, type=DILATED, dilation_rate=(2 ** (2*i+2)),
name_scope='bottleneck2_{}'.format(str(4 * i + 4)))
net = bottleneck(
net,
output_depth=128,
filter_size=3,
regularizer_prob=0.1,
type=DILATED,
dilation_rate=(2**(2 * i + 2)),
name_scope='bottleneck2_{}'.format(str(4 * i + 4)))
return net, inputs_shape_2
......@@ -320,52 +381,106 @@ def ENet_stage3(inputs, name_scope='stage3_block'):
with scope(name_scope):
for i in range(2):
with scope('bottleneck3_{}'.format(str(4 * i + 0))):
net = bottleneck(inputs, output_depth=128, filter_size=3, regularizer_prob=0.1,
name_scope='bottleneck3_{}'.format(str(4 * i + 0)))
net = bottleneck(
inputs,
output_depth=128,
filter_size=3,
regularizer_prob=0.1,
name_scope='bottleneck3_{}'.format(str(4 * i + 0)))
with scope('bottleneck3_{}'.format(str(4 * i + 1))):
net = bottleneck(net, output_depth=128, filter_size=3, regularizer_prob=0.1, type=DILATED, dilation_rate=(2 ** (2*i+1)),
name_scope='bottleneck3_{}'.format(str(4 * i + 1)))
net = bottleneck(
net,
output_depth=128,
filter_size=3,
regularizer_prob=0.1,
type=DILATED,
dilation_rate=(2**(2 * i + 1)),
name_scope='bottleneck3_{}'.format(str(4 * i + 1)))
with scope('bottleneck3_{}'.format(str(4 * i + 2))):
net = bottleneck(net, output_depth=128, filter_size=5, regularizer_prob=0.1, type=ASYMMETRIC,
name_scope='bottleneck3_{}'.format(str(4 * i + 2)))
net = bottleneck(
net,
output_depth=128,
filter_size=5,
regularizer_prob=0.1,
type=ASYMMETRIC,
name_scope='bottleneck3_{}'.format(str(4 * i + 2)))
with scope('bottleneck3_{}'.format(str(4 * i + 3))):
net = bottleneck(net, output_depth=128, filter_size=3, regularizer_prob=0.1, type=DILATED, dilation_rate=(2 ** (2*i+2)),
name_scope='bottleneck3_{}'.format(str(4 * i + 3)))
net = bottleneck(
net,
output_depth=128,
filter_size=3,
regularizer_prob=0.1,
type=DILATED,
dilation_rate=(2**(2 * i + 2)),
name_scope='bottleneck3_{}'.format(str(4 * i + 3)))
return net
def ENet_stage4(inputs, inputs_shape, connect_tensor,
skip_connections=True, name_scope='stage4_block'):
def ENet_stage4(inputs,
inputs_shape,
connect_tensor,
skip_connections=True,
name_scope='stage4_block'):
with scope(name_scope):
with scope('bottleneck4_0'):
net = bottleneck(inputs, output_depth=64, filter_size=3, regularizer_prob=0.1,
type=UPSAMPLING, decoder=True, output_shape=inputs_shape,
name_scope='bottleneck4_0')
net = bottleneck(
inputs,
output_depth=64,
filter_size=3,
regularizer_prob=0.1,
type=UPSAMPLING,
decoder=True,
output_shape=inputs_shape,
name_scope='bottleneck4_0')
if skip_connections:
net = fluid.layers.elementwise_add(net, connect_tensor)
with scope('bottleneck4_1'):
net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.1, decoder=True,
name_scope='bottleneck4_1')
net = bottleneck(
net,
output_depth=64,
filter_size=3,
regularizer_prob=0.1,
decoder=True,
name_scope='bottleneck4_1')
with scope('bottleneck4_2'):
net = bottleneck(net, output_depth=64, filter_size=3, regularizer_prob=0.1, decoder=True,
name_scope='bottleneck4_2')
net = bottleneck(
net,
output_depth=64,
filter_size=3,
regularizer_prob=0.1,
decoder=True,
name_scope='bottleneck4_2')
return net
def ENet_stage5(inputs, inputs_shape, connect_tensor, skip_connections=True,
def ENet_stage5(inputs,
inputs_shape,
connect_tensor,
skip_connections=True,
name_scope='stage5_block'):
with scope(name_scope):
net = bottleneck(inputs, output_depth=16, filter_size=3, regularizer_prob=0.1, type=UPSAMPLING,
decoder=True, output_shape=inputs_shape,
name_scope='bottleneck5_0')
net = bottleneck(
inputs,
output_depth=16,
filter_size=3,
regularizer_prob=0.1,
type=UPSAMPLING,
decoder=True,
output_shape=inputs_shape,
name_scope='bottleneck5_0')
if skip_connections:
net = fluid.layers.elementwise_add(net, connect_tensor)
with scope('bottleneck5_1'):
net = bottleneck(net, output_depth=16, filter_size=3, regularizer_prob=0.1, decoder=True,
name_scope='bottleneck5_1')
net = bottleneck(
net,
output_depth=16,
filter_size=3,
regularizer_prob=0.1,
decoder=True,
name_scope='bottleneck5_1')
return net
......@@ -378,14 +493,16 @@ def decoder(input, num_classes):
segStage3 = ENet_stage3(stage2)
segStage4 = ENet_stage4(segStage3, inputs_shape_2, stage1)
segStage5 = ENet_stage5(segStage4, inputs_shape_1, initial)
segLogits = deconv(segStage5, num_classes, filter_size=2, stride=2, padding='SAME')
segLogits = deconv(
segStage5, num_classes, filter_size=2, stride=2, padding='SAME')
# Embedding branch
with scope('LaneNetEm'):
emStage3 = ENet_stage3(stage2)
emStage4 = ENet_stage4(emStage3, inputs_shape_2, stage1)
emStage5 = ENet_stage5(emStage4, inputs_shape_1, initial)
emLogits = deconv(emStage5, 4, filter_size=2, stride=2, padding='SAME')
emLogits = deconv(
emStage5, 4, filter_size=2, stride=2, padding='SAME')
elif 'vgg' in cfg.MODEL.LANENET.BACKBONE:
encoder_list = ['pool5', 'pool4', 'pool3']
......@@ -396,14 +513,16 @@ def decoder(input, num_classes):
encoder_list = encoder_list[1:]
for i in range(len(encoder_list)):
with scope('deconv_{:d}'.format(i + 1)):
deconv_out = deconv(score, 64, filter_size=4, stride=2, padding='SAME')
deconv_out = deconv(
score, 64, filter_size=4, stride=2, padding='SAME')
input_tensor = input[encoder_list[i]]
with scope('score_{:d}'.format(i + 1)):
score = conv(input_tensor, 64, 1)
score = fluid.layers.elementwise_add(deconv_out, score)
with scope('deconv_final'):
emLogits = deconv(score, 64, filter_size=16, stride=8, padding='SAME')
emLogits = deconv(
score, 64, filter_size=16, stride=8, padding='SAME')
with scope('score_final'):
segLogits = conv(emLogits, num_classes, 1)
emLogits = relu(conv(emLogits, 4, 1))
......@@ -415,7 +534,8 @@ def encoder(input):
model = vgg_backbone(layers=16)
#output = model.net(input)
_, encode_feature_dict = model.net(input, end_points=13, decode_points=[7, 10, 13])
_, encode_feature_dict = model.net(
input, end_points=13, decode_points=[7, 10, 13])
output = {}
output['pool3'] = encode_feature_dict[7]
output['pool4'] = encode_feature_dict[10]
......@@ -427,8 +547,9 @@ def encoder(input):
stage2, inputs_shape_2 = ENet_stage2(stage1)
output = (initial, stage1, stage2, inputs_shape_1, inputs_shape_2)
else:
raise Exception("LaneNet expect enet and vgg backbone, but received {}".
format(cfg.MODEL.LANENET.BACKBONE))
raise Exception(
"LaneNet expect enet and vgg backbone, but received {}".format(
cfg.MODEL.LANENET.BACKBONE))
return output
......
contrib/LaneNet/reader.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -58,7 +58,8 @@ class LaneNetDataset():
if self.shuffle and cfg.NUM_TRAINERS > 1:
np.random.RandomState(self.shuffle_seed).shuffle(self.all_lines)
num_lines = len(self.all_lines) // cfg.NUM_TRAINERS
self.lines = self.all_lines[num_lines * cfg.TRAINER_ID: num_lines * (cfg.TRAINER_ID + 1)]
self.lines = self.all_lines[num_lines * cfg.TRAINER_ID:num_lines *
(cfg.TRAINER_ID + 1)]
self.shuffle_seed += 1
elif self.shuffle:
np.random.shuffle(self.lines)
......@@ -86,7 +87,8 @@ class LaneNetDataset():
if self.shuffle and cfg.NUM_TRAINERS > 1:
np.random.RandomState(self.shuffle_seed).shuffle(self.all_lines)
num_lines = len(self.all_lines) // self.num_trainers
self.lines = self.all_lines[num_lines * self.trainer_id: num_lines * (self.trainer_id + 1)]
self.lines = self.all_lines[num_lines * self.trainer_id:num_lines *
(self.trainer_id + 1)]
self.shuffle_seed += 1
elif self.shuffle:
np.random.shuffle(self.lines)
......@@ -118,7 +120,8 @@ class LaneNetDataset():
def batch_reader(is_test=False, drop_last=drop_last):
if is_test:
imgs, grts, grts_instance, img_names, valid_shapes, org_shapes = [], [], [], [], [], []
for img, grt, grt_instance, img_name, valid_shape, org_shape in reader():
for img, grt, grt_instance, img_name, valid_shape, org_shape in reader(
):
imgs.append(img)
grts.append(grt)
grts_instance.append(grt_instance)
......@@ -126,14 +129,15 @@ class LaneNetDataset():
valid_shapes.append(valid_shape)
org_shapes.append(org_shape)
if len(imgs) == batch_size:
yield np.array(imgs), np.array(
grts), np.array(grts_instance), img_names, np.array(valid_shapes), np.array(
org_shapes)
yield np.array(imgs), np.array(grts), np.array(
grts_instance), img_names, np.array(
valid_shapes), np.array(org_shapes)
imgs, grts, grts_instance, img_names, valid_shapes, org_shapes = [], [], [], [], [], []
if not drop_last and len(imgs) > 0:
yield np.array(imgs), np.array(grts), np.array(grts_instance), img_names, np.array(
valid_shapes), np.array(org_shapes)
yield np.array(imgs), np.array(grts), np.array(
grts_instance), img_names, np.array(
valid_shapes), np.array(org_shapes)
else:
imgs, labs, labs_instance, ignore = [], [], [], []
bs = 0
......@@ -144,12 +148,14 @@ class LaneNetDataset():
ignore.append(ig)
bs += 1
if bs == batch_size:
yield np.array(imgs), np.array(labs), np.array(labs_instance), np.array(ignore)
yield np.array(imgs), np.array(labs), np.array(
labs_instance), np.array(ignore)
bs = 0
imgs, labs, labs_instance, ignore = [], [], [], []
if not drop_last and bs > 0:
yield np.array(imgs), np.array(labs), np.array(labs_instance), np.array(ignore)
yield np.array(imgs), np.array(labs), np.array(
labs_instance), np.array(ignore)
return batch_reader(is_test, drop_last)
......@@ -299,10 +305,12 @@ class LaneNetDataset():
img, grt = aug.rand_crop(img, grt, mode=mode)
elif ModelPhase.is_eval(mode):
img, grt, grt_instance = aug.resize(img, grt, grt_instance, mode=mode)
img, grt, grt_instance = aug.resize(
img, grt, grt_instance, mode=mode)
elif ModelPhase.is_visual(mode):
ori_img = img.copy()
img, grt, grt_instance = aug.resize(img, grt, grt_instance, mode=mode)
img, grt, grt_instance = aug.resize(
img, grt, grt_instance, mode=mode)
valid_shape = [img.shape[0], img.shape[1]]
else:
raise ValueError("Dataset mode={} Error!".format(mode))
......
contrib/LaneNet/train.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -40,10 +40,10 @@ from pdseg.utils.timer import Timer, calculate_eta
from reader import LaneNetDataset
from models.model_builder import build_model
from models.model_builder import ModelPhase
from models.model_builder import parse_shape_from_file
from eval import evaluate
from vis import visualize
from utils import dist_utils
from utils.load_model_utils import load_pretrained_weights
def parse_args():
......@@ -101,37 +101,6 @@ def parse_args():
return parser.parse_args()
def save_vars(executor, dirname, program=None, vars=None):
"""
Temporary resolution for Win save variables compatability.
Will fix in PaddlePaddle v1.5.2
"""
save_program = fluid.Program()
save_block = save_program.global_block()
for each_var in vars:
# NOTE: don't save the variable which type is RAW
if each_var.type == fluid.core.VarDesc.VarType.RAW:
continue
new_var = save_block.create_var(
name=each_var.name,
shape=each_var.shape,
dtype=each_var.dtype,
type=each_var.type,
lod_level=each_var.lod_level,
persistable=True)
file_path = os.path.join(dirname, new_var.name)
file_path = os.path.normpath(file_path)
save_block.append_op(
type='save',
inputs={'X': [new_var]},
outputs={},
attrs={'file_path': file_path})
executor.run(save_program)
def save_checkpoint(exe, program, ckpt_name):
"""
Save checkpoint for evaluation or resume training
......@@ -141,29 +110,22 @@ def save_checkpoint(exe, program, ckpt_name):
if not os.path.isdir(ckpt_dir):
os.makedirs(ckpt_dir)
save_vars(
exe,
ckpt_dir,
program,
vars=list(filter(fluid.io.is_persistable, program.list_vars())))
fluid.save(program, os.path.join(ckpt_dir, 'model'))
return ckpt_dir
def load_checkpoint(exe, program):
"""
Load checkpoiont from pretrained model directory for resume training
Load checkpoiont for resuming training
"""
print('Resume model training from:', cfg.TRAIN.RESUME_MODEL_DIR)
if not os.path.exists(cfg.TRAIN.RESUME_MODEL_DIR):
raise ValueError("TRAIN.PRETRAIN_MODEL {} not exist!".format(
cfg.TRAIN.RESUME_MODEL_DIR))
fluid.io.load_persistables(
exe, cfg.TRAIN.RESUME_MODEL_DIR, main_program=program)
model_path = cfg.TRAIN.RESUME_MODEL_DIR
print('Resume model training from:', model_path)
if not os.path.exists(model_path):
raise ValueError(
"TRAIN.PRETRAIN_MODEL {} not exist!".format(model_path))
fluid.load(program, os.path.join(model_path, 'model'), exe)
# Check is path ended by path spearator
if model_path[-1] == os.sep:
model_path = model_path[0:-1]
......@@ -178,7 +140,6 @@ def load_checkpoint(exe, program):
else:
raise ValueError("Resume model path is not valid!")
print("Model checkpoint loaded successfully!")
return begin_epoch
......@@ -271,44 +232,7 @@ def train(cfg):
begin_epoch = load_checkpoint(exe, train_prog)
# Load pretrained model
elif os.path.exists(cfg.TRAIN.PRETRAINED_MODEL_DIR):
print_info('Pretrained model dir: ', cfg.TRAIN.PRETRAINED_MODEL_DIR)
load_vars = []
load_fail_vars = []
def var_shape_matched(var, shape):
"""
Check whehter persitable variable shape is match with current network
"""
var_exist = os.path.exists(
os.path.join(cfg.TRAIN.PRETRAINED_MODEL_DIR, var.name))
if var_exist:
var_shape = parse_shape_from_file(
os.path.join(cfg.TRAIN.PRETRAINED_MODEL_DIR, var.name))
if var_shape != shape:
print(var.name, var_shape, shape)
return var_shape == shape
return False
for x in train_prog.list_vars():
if isinstance(x, fluid.framework.Parameter):
shape = tuple(fluid.global_scope().find_var(
x.name).get_tensor().shape())
if var_shape_matched(x, shape):
load_vars.append(x)
else:
load_fail_vars.append(x)
fluid.io.load_vars(
exe, dirname=cfg.TRAIN.PRETRAINED_MODEL_DIR, vars=load_vars)
for var in load_vars:
print_info("Parameter[{}] loaded sucessfully!".format(var.name))
for var in load_fail_vars:
print_info(
"Parameter[{}] don't exist or shape does not match current network, skip"
" to load it.".format(var.name))
print_info("{}/{} pretrained parameters loaded successfully!".format(
len(load_vars),
len(load_vars) + len(load_fail_vars)))
load_pretrained_weights(exe, train_prog, cfg.TRAIN.PRETRAINED_MODEL_DIR)
else:
print_info(
'Pretrained model dir {} not exists, training from scratch...'.
......@@ -393,8 +317,7 @@ def train(cfg):
avg_emb_loss, avg_acc, avg_fp, avg_fn, speed,
calculate_eta(all_step - step, speed)))
if args.use_vdl:
log_writer.add_scalar('Train/loss', avg_loss,
step)
log_writer.add_scalar('Train/loss', avg_loss, step)
log_writer.add_scalar('Train/lr', lr[0], step)
log_writer.add_scalar('Train/speed', speed, step)
sys.stdout.flush()
......@@ -423,8 +346,7 @@ def train(cfg):
use_gpu=args.use_gpu,
use_mpio=args.use_mpio)
if args.use_vdl:
log_writer.add_scalar('Evaluate/accuracy', accuracy,
step)
log_writer.add_scalar('Evaluate/accuracy', accuracy, step)
log_writer.add_scalar('Evaluate/fp', fp, step)
log_writer.add_scalar('Evaluate/fn', fn, step)
......
contrib/LaneNet/utils/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
contrib/LaneNet/utils/config.py
浏览文件 @ 7835c042
# -*- coding: utf-8 -*-
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......@@ -80,8 +80,8 @@ cfg.DATASET.DATA_DIM = 3
cfg.DATASET.SEPARATOR = ' '
# 忽略的像素标签值, 默认为255，一般无需改动
cfg.DATASET.IGNORE_INDEX = 255
# 数据增强是图像的padding值
cfg.DATASET.PADDING_VALUE = [127.5,127.5,127.5]
# 数据增强是图像的padding值
cfg.DATASET.PADDING_VALUE = [127.5, 127.5, 127.5]
########################### 数据增强配置 ######################################
# 图像镜像左右翻转
......
contrib/LaneNet/utils/dist_utils.py
浏览文件 @ 7835c042
#copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
......
contrib/LaneNet/utils/generate_tusimple_dataset.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
generate tusimple training dataset
"""
......@@ -14,12 +28,16 @@ import numpy as np
def init_args():
parser = argparse.ArgumentParser()
parser.add_argument('--src_dir', type=str, help='The origin path of unzipped tusimple dataset')
parser.add_argument(
'--src_dir',
type=str,
help='The origin path of unzipped tusimple dataset')
return parser.parse_args()
def process_json_file(json_file_path, src_dir, ori_dst_dir, binary_dst_dir, instance_dst_dir):
def process_json_file(json_file_path, src_dir, ori_dst_dir, binary_dst_dir,
instance_dst_dir):
assert ops.exists(json_file_path), '{:s} not exist'.format(json_file_path)
......@@ -39,11 +57,14 @@ def process_json_file(json_file_path, src_dir, ori_dst_dir, binary_dst_dir, inst
h_samples = info_dict['h_samples']
lanes = info_dict['lanes']
image_name_new = '{:s}.png'.format('{:d}'.format(line_index + image_nums).zfill(4))
image_name_new = '{:s}.png'.format(
'{:d}'.format(line_index + image_nums).zfill(4))
src_image = cv2.imread(image_path, cv2.IMREAD_COLOR)
dst_binary_image = np.zeros([src_image.shape[0], src_image.shape[1]], np.uint8)
dst_instance_image = np.zeros([src_image.shape[0], src_image.shape[1]], np.uint8)
dst_binary_image = np.zeros(
[src_image.shape[0], src_image.shape[1]], np.uint8)
dst_instance_image = np.zeros(
[src_image.shape[0], src_image.shape[1]], np.uint8)
for lane_index, lane in enumerate(lanes):
assert len(h_samples) == len(lane)
......@@ -62,13 +83,23 @@ def process_json_file(json_file_path, src_dir, ori_dst_dir, binary_dst_dir, inst
lane_pts = np.vstack((lane_x, lane_y)).transpose()
lane_pts = np.array([lane_pts], np.int64)
cv2.polylines(dst_binary_image, lane_pts, isClosed=False,
color=255, thickness=5)
cv2.polylines(dst_instance_image, lane_pts, isClosed=False,
color=lane_index * 50 + 20, thickness=5)
dst_binary_image_path = ops.join(src_dir, binary_dst_dir, image_name_new)
dst_instance_image_path = ops.join(src_dir, instance_dst_dir, image_name_new)
cv2.polylines(
dst_binary_image,
lane_pts,
isClosed=False,
color=255,
thickness=5)
cv2.polylines(
dst_instance_image,
lane_pts,
isClosed=False,
color=lane_index * 50 + 20,
thickness=5)
dst_binary_image_path = ops.join(src_dir, binary_dst_dir,
image_name_new)
dst_instance_image_path = ops.join(src_dir, instance_dst_dir,
image_name_new)
dst_rgb_image_path = ops.join(src_dir, ori_dst_dir, image_name_new)
cv2.imwrite(dst_binary_image_path, dst_binary_image)
......@@ -78,7 +109,12 @@ def process_json_file(json_file_path, src_dir, ori_dst_dir, binary_dst_dir, inst
print('Process {:s} success'.format(image_name))
def gen_sample(src_dir, b_gt_image_dir, i_gt_image_dir, image_dir, phase='train', split=False):
def gen_sample(src_dir,
b_gt_image_dir,
i_gt_image_dir,
image_dir,
phase='train',
split=False):
label_list = []
with open('{:s}/{}ing/{}.txt'.format(src_dir, phase, phase), 'w') as file:
......@@ -92,7 +128,8 @@ def gen_sample(src_dir, b_gt_image_dir, i_gt_image_dir, image_dir, phase='train'
image_path = ops.join(image_dir, image_name)
assert ops.exists(image_path), '{:s} not exist'.format(image_path)
assert ops.exists(instance_gt_image_path), '{:s} not exist'.format(instance_gt_image_path)
assert ops.exists(instance_gt_image_path), '{:s} not exist'.format(
instance_gt_image_path)
b_gt_image = cv2.imread(binary_gt_image_path, cv2.IMREAD_COLOR)
i_gt_image = cv2.imread(instance_gt_image_path, cv2.IMREAD_COLOR)
......@@ -102,7 +139,8 @@ def gen_sample(src_dir, b_gt_image_dir, i_gt_image_dir, image_dir, phase='train'
print('image: {:s} corrupt'.format(image_name))
continue
else:
info = '{:s} {:s} {:s}'.format(image_path, binary_gt_image_path, instance_gt_image_path)
info = '{:s} {:s} {:s}'.format(image_path, binary_gt_image_path,
instance_gt_image_path)
file.write(info + '\n')
label_list.append(info)
if phase == 'train' and split:
......@@ -110,10 +148,12 @@ def gen_sample(src_dir, b_gt_image_dir, i_gt_image_dir, image_dir, phase='train'
val_list_len = len(label_list) // 10
val_label_list = label_list[:val_list_len]
train_label_list = label_list[val_list_len:]
with open('{:s}/{}ing/train_part.txt'.format(src_dir, phase, phase), 'w') as file:
with open('{:s}/{}ing/train_part.txt'.format(src_dir, phase, phase),
'w') as file:
for info in train_label_list:
file.write(info + '\n')
with open('{:s}/{}ing/val_part.txt'.format(src_dir, phase, phase), 'w') as file:
with open('{:s}/{}ing/val_part.txt'.format(src_dir, phase, phase),
'w') as file:
for info in val_label_list:
file.write(info + '\n')
return
......@@ -130,12 +170,14 @@ def process_tusimple_dataset(src_dir):
for json_label_path in glob.glob('{:s}/label*.json'.format(src_dir)):
json_label_name = ops.split(json_label_path)[1]
shutil.copyfile(json_label_path, ops.join(traing_folder_path, json_label_name))
shutil.copyfile(json_label_path,
ops.join(traing_folder_path, json_label_name))
for json_label_path in glob.glob('{:s}/test_label.json'.format(src_dir)):
json_label_name = ops.split(json_label_path)[1]
shutil.copyfile(json_label_path, ops.join(testing_folder_path, json_label_name))
shutil.copyfile(json_label_path,
ops.join(testing_folder_path, json_label_name))
train_gt_image_dir = ops.join('training', 'gt_image')
train_gt_binary_dir = ops.join('training', 'gt_binary_image')
......@@ -154,9 +196,11 @@ def process_tusimple_dataset(src_dir):
os.makedirs(os.path.join(src_dir, test_gt_instance_dir), exist_ok=True)
for json_label_path in glob.glob('{:s}/*.json'.format(traing_folder_path)):
process_json_file(json_label_path, src_dir, train_gt_image_dir, train_gt_binary_dir, train_gt_instance_dir)
process_json_file(json_label_path, src_dir, train_gt_image_dir,
train_gt_binary_dir, train_gt_instance_dir)
gen_sample(src_dir, train_gt_binary_dir, train_gt_instance_dir, train_gt_image_dir, 'train', True)
gen_sample(src_dir, train_gt_binary_dir, train_gt_instance_dir,
train_gt_image_dir, 'train', True)
if __name__ == '__main__':
......
contrib/LaneNet/utils/lanenet_postprocess.py
浏览文件 @ 7835c042
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# this code heavily base on https://github.com/MaybeShewill-CV/lanenet-lane-detection/blob/master/lanenet_model/lanenet_postprocess.py
"""
LaneNet model post process
......@@ -22,12 +35,14 @@ def _morphological_process(image, kernel_size=5):
:return:
"""
if len(image.shape) == 3:
raise ValueError('Binary segmentation result image should be a single channel image')
raise ValueError(
'Binary segmentation result image should be a single channel image')
if image.dtype is not np.uint8:
image = np.array(image, np.uint8)
kernel = cv2.getStructuringElement(shape=cv2.MORPH_ELLIPSE, ksize=(kernel_size, kernel_size))
kernel = cv2.getStructuringElement(
shape=cv2.MORPH_ELLIPSE, ksize=(kernel_size, kernel_size))
# close operation fille hole
closing = cv2.morphologyEx(image, cv2.MORPH_CLOSE, kernel, iterations=1)
......@@ -46,13 +61,15 @@ def _connect_components_analysis(image):
else:
gray_image = image
return cv2.connectedComponentsWithStats(gray_image, connectivity=8, ltype=cv2.CV_32S)
return cv2.connectedComponentsWithStats(
gray_image, connectivity=8, ltype=cv2.CV_32S)
class _LaneFeat(object):
"""
"""
def __init__(self, feat, coord, class_id=-1):
"""
lane feat object
......@@ -108,18 +125,21 @@ class _LaneNetCluster(object):
"""
Instance segmentation result cluster
"""
def __init__(self):
"""
"""
self._color_map = [np.array([255, 0, 0]),
np.array([0, 255, 0]),
np.array([0, 0, 255]),
np.array([125, 125, 0]),
np.array([0, 125, 125]),
np.array([125, 0, 125]),
np.array([50, 100, 50]),
np.array([100, 50, 100])]
self._color_map = [
np.array([255, 0, 0]),
np.array([0, 255, 0]),
np.array([0, 0, 255]),
np.array([125, 125, 0]),
np.array([0, 125, 125]),
np.array([125, 0, 125]),
np.array([50, 100, 50]),
np.array([100, 50, 100])
]
@staticmethod
def _embedding_feats_dbscan_cluster(embedding_image_feats):
......@@ -186,15 +206,16 @@ class _LaneNetCluster(object):
# get embedding feats and coords
get_lane_embedding_feats_result = self._get_lane_embedding_feats(
binary_seg_ret=binary_seg_result,
instance_seg_ret=instance_seg_result
)
instance_seg_ret=instance_seg_result)
# dbscan cluster
dbscan_cluster_result = self._embedding_feats_dbscan_cluster(
embedding_image_feats=get_lane_embedding_feats_result['lane_embedding_feats']
)
embedding_image_feats=get_lane_embedding_feats_result[
'lane_embedding_feats'])
mask = np.zeros(shape=[binary_seg_result.shape[0], binary_seg_result.shape[1], 3], dtype=np.uint8)
mask = np.zeros(
shape=[binary_seg_result.shape[0], binary_seg_result.shape[1], 3],
dtype=np.uint8)
db_labels = dbscan_cluster_result['db_labels']
unique_labels = dbscan_cluster_result['unique_labels']
coord = get_lane_embedding_feats_result['lane_coordinates']
......@@ -219,11 +240,13 @@ class LaneNetPostProcessor(object):
"""
lanenet post process for lane generation
"""
def __init__(self, ipm_remap_file_path='./utils/tusimple_ipm_remap.yml'):
"""
convert front car view to bird view
"""
assert ops.exists(ipm_remap_file_path), '{:s} not exist'.format(ipm_remap_file_path)
assert ops.exists(ipm_remap_file_path), '{:s} not exist'.format(
ipm_remap_file_path)
self._cluster = _LaneNetCluster()
self._ipm_remap_file_path = ipm_remap_file_path
......@@ -232,14 +255,16 @@ class LaneNetPostProcessor(object):
self._remap_to_ipm_x = remap_file_load_ret['remap_to_ipm_x']
self._remap_to_ipm_y = remap_file_load_ret['remap_to_ipm_y']
self._color_map = [np.array([255, 0, 0]),
np.array([0, 255, 0]),
np.array([0, 0, 255]),
np.array([125, 125, 0]),
np.array([0, 125, 125]),
np.array([125, 0, 125]),
np.array([50, 100, 50]),
np.array([100, 50, 100])]
self._color_map = [
np.array([255, 0, 0]),
np.array([0, 255, 0]),
np.array([0, 0, 255]),
np.array([125, 125, 0]),
np.array([0, 125, 125]),
np.array([125, 0, 125]),
np.array([50, 100, 50]),
np.array([100, 50, 100])
]
def _load_remap_matrix(self):
fs = cv2.FileStorage(self._ipm_remap_file_path, cv2.FILE_STORAGE_READ)
......@@ -256,15 +281,20 @@ class LaneNetPostProcessor(object):
return ret
def postprocess(self, binary_seg_result, instance_seg_result=None,
min_area_threshold=100, source_image=None,
def postprocess(self,
binary_seg_result,
instance_seg_result=None,
min_area_threshold=100,
source_image=None,
data_source='tusimple'):
# convert binary_seg_result
binary_seg_result = np.array(binary_seg_result * 255, dtype=np.uint8)
# apply image morphology operation to fill in the hold and reduce the small area
morphological_ret = _morphological_process(binary_seg_result, kernel_size=5)
connect_components_analysis_ret = _connect_components_analysis(image=morphological_ret)
morphological_ret = _morphological_process(
binary_seg_result, kernel_size=5)
connect_components_analysis_ret = _connect_components_analysis(
image=morphological_ret)
labels = connect_components_analysis_ret[1]
stats = connect_components_analysis_ret[2]
......@@ -276,8 +306,7 @@ class LaneNetPostProcessor(object):
# apply embedding features cluster
mask_image, lane_coords = self._cluster.apply_lane_feats_cluster(
binary_seg_result=morphological_ret,
instance_seg_result=instance_seg_result
)
instance_seg_result=instance_seg_result)
if mask_image is None:
return {
......@@ -292,15 +321,15 @@ class LaneNetPostProcessor(object):
for lane_index, coords in enumerate(lane_coords):
if data_source == 'tusimple':
tmp_mask = np.zeros(shape=(720, 1280), dtype=np.uint8)
tmp_mask[tuple((np.int_(coords[:, 1] * 720 / 256), np.int_(coords[:, 0] * 1280 / 512)))] = 255
tmp_mask[tuple((np.int_(coords[:, 1] * 720 / 256),
np.int_(coords[:, 0] * 1280 / 512)))] = 255
else:
raise ValueError('Wrong data source now only support tusimple')
tmp_ipm_mask = cv2.remap(
tmp_mask,
self._remap_to_ipm_x,
self._remap_to_ipm_y,
interpolation=cv2.INTER_NEAREST
)
interpolation=cv2.INTER_NEAREST)
nonzero_y = np.array(tmp_ipm_mask.nonzero()[0])
nonzero_x = np.array(tmp_ipm_mask.nonzero()[1])
......@@ -309,16 +338,19 @@ class LaneNetPostProcessor(object):
[ipm_image_height, ipm_image_width] = tmp_ipm_mask.shape
plot_y = np.linspace(10, ipm_image_height, ipm_image_height - 10)
fit_x = fit_param[0] * plot_y ** 2 + fit_param[1] * plot_y + fit_param[2]
fit_x = fit_param[0] * plot_y**2 + fit_param[
1] * plot_y + fit_param[2]
lane_pts = []
for index in range(0, plot_y.shape[0], 5):
src_x = self._remap_to_ipm_x[
int(plot_y[index]), int(np.clip(fit_x[index], 0, ipm_image_width - 1))]
int(plot_y[index]),
int(np.clip(fit_x[index], 0, ipm_image_width - 1))]
if src_x <= 0:
continue
src_y = self._remap_to_ipm_y[
int(plot_y[index]), int(np.clip(fit_x[index], 0, ipm_image_width - 1))]
int(plot_y[index]),
int(np.clip(fit_x[index], 0, ipm_image_width - 1))]
src_y = src_y if src_y > 0 else 0
lane_pts.append([src_x, src_y])
......@@ -366,8 +398,10 @@ class LaneNetPostProcessor(object):
continue
lane_color = self._color_map[index].tolist()
cv2.circle(source_image, (int(interpolation_src_pt_x),
int(interpolation_src_pt_y)), 5, lane_color, -1)
cv2.circle(
source_image,
(int(interpolation_src_pt_x), int(interpolation_src_pt_y)),
5, lane_color, -1)
ret = {
'mask_image': mask_image,
'fit_params': fit_params,
......
contrib/LaneNet/utils/load_model_utils.py 0 → 100644
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import os.path as osp
import six
import numpy as np
def parse_param_file(param_file, return_shape=True):
from paddle.fluid.proto.framework_pb2 import VarType
f = open(param_file, 'rb')
version = np.fromstring(f.read(4), dtype='int32')
lod_level = np.fromstring(f.read(8), dtype='int64')
for i in range(int(lod_level)):
_size = np.fromstring(f.read(8), dtype='int64')
_ = f.read(_size)
version = np.fromstring(f.read(4), dtype='int32')
tensor_desc = VarType.TensorDesc()
tensor_desc_size = np.fromstring(f.read(4), dtype='int32')
tensor_desc.ParseFromString(f.read(int(tensor_desc_size)))
tensor_shape = tuple(tensor_desc.dims)
if return_shape:
f.close()
return tuple(tensor_desc.dims)
if tensor_desc.data_type != 5:
raise Exception(
"Unexpected data type while parse {}".format(param_file))
data_size = 4
for i in range(len(tensor_shape)):
data_size *= tensor_shape[i]
weight = np.fromstring(f.read(data_size), dtype='float32')
f.close()
return np.reshape(weight, tensor_shape)
def load_pdparams(exe, main_prog, model_dir):
import paddle.fluid as fluid
from paddle.fluid.proto.framework_pb2 import VarType
from paddle.fluid.framework import Program
vars_to_load = list()
vars_not_load = list()
import pickle
with open(osp.join(model_dir, 'model.pdparams'), 'rb') as f:
params_dict = pickle.load(f) if six.PY2 else pickle.load(
f, encoding='latin1')
unused_vars = list()
for var in main_prog.list_vars():
if not isinstance(var, fluid.framework.Parameter):
continue
if var.name not in params_dict:
print("{} is not in saved model".format(var.name))
vars_not_load.append(var.name)
continue
if var.shape != params_dict[var.name].shape:
unused_vars.append(var.name)
vars_not_load.append(var.name)
print(
"[SKIP] Shape of pretrained weight {} doesn't match.(Pretrained: {}, Actual: {})"
.format(var.name, params_dict[var.name].shape, var.shape))
continue
vars_to_load.append(var)
for var_name in unused_vars:
del params_dict[var_name]
fluid.io.set_program_state(main_prog, params_dict)
if len(vars_to_load) == 0:
print(
"There is no pretrain weights loaded, maybe you should check you pretrain model!"
)
else:
print("There are {}/{} varaibles in {} are loaded.".format(
len(vars_to_load),
len(vars_to_load) + len(vars_not_load), model_dir))
def load_pretrained_weights(exe, main_prog, weights_dir):
if not osp.exists(weights_dir):
raise Exception("Path {} not exists.".format(weights_dir))
if osp.exists(osp.join(weights_dir, "model.pdparams")):
return load_pdparams(exe, main_prog, weights_dir)
import paddle.fluid as fluid
vars_to_load = list()
vars_not_load = list()
for var in main_prog.list_vars():
if not isinstance(var, fluid.framework.Parameter):
continue
if not osp.exists(osp.join(weights_dir, var.name)):
print("[SKIP] Pretrained weight {}/{} doesn't exist".format(
weights_dir, var.name))
vars_not_load.append(var)
continue
pretrained_shape = parse_param_file(osp.join(weights_dir, var.name))
actual_shape = tuple(var.shape)
if pretrained_shape != actual_shape:
print(
"[SKIP] Shape of pretrained weight {}/{} doesn't match.(Pretrained: {}, Actual: {})"
.format(weights_dir, var.name, pretrained_shape, actual_shape))
vars_not_load.append(var)
continue
vars_to_load.append(var)
params_dict = fluid.io.load_program_state(
weights_dir, var_list=vars_to_load)
fluid.io.set_program_state(main_prog, params_dict)
if len(vars_to_load) == 0:
print(
"There is no pretrain weights loaded, maybe you should check you pretrain model!"
)
else:
print("There are {}/{} varaibles in {} are loaded.".format(
len(vars_to_load),
len(vars_to_load) + len(vars_not_load), weights_dir))
contrib/LaneNet/vis.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -45,6 +45,7 @@ from models.model_builder import ModelPhase
from utils import lanenet_postprocess
import matplotlib.pyplot as plt
def parse_args():
parser = argparse.ArgumentParser(description='PaddeSeg visualization tools')
parser.add_argument(
......@@ -106,7 +107,6 @@ def minmax_scale(input_arr):
return output_arr
def visualize(cfg,
vis_file_list=None,
use_gpu=False,
......@@ -119,7 +119,6 @@ def visualize(cfg,
if vis_file_list is None:
vis_file_list = cfg.DATASET.TEST_FILE_LIST
dataset = LaneNetDataset(
file_list=vis_file_list,
mode=ModelPhase.VISUAL,
......@@ -139,7 +138,12 @@ def visualize(cfg,
ckpt_dir = cfg.TEST.TEST_MODEL if not ckpt_dir else ckpt_dir
fluid.io.load_params(exe, ckpt_dir, main_program=test_prog)
if ckpt_dir is not None:
print('load test model:', ckpt_dir)
try:
fluid.load(test_prog, os.path.join(ckpt_dir, 'model'), exe)
except:
fluid.io.load_params(exe, ckpt_dir, main_program=test_prog)
save_dir = os.path.join(vis_dir, 'visual_results')
makedirs(save_dir)
......@@ -161,22 +165,26 @@ def visualize(cfg,
for i in range(num_imgs):
gt_image = org_imgs[i]
binary_seg_image, instance_seg_image = segLogits[i].squeeze(-1), emLogits[i].transpose((1,2,0))
binary_seg_image, instance_seg_image = segLogits[i].squeeze(
-1), emLogits[i].transpose((1, 2, 0))
postprocess_result = postprocessor.postprocess(
binary_seg_result=binary_seg_image,
instance_seg_result=instance_seg_image,
source_image=gt_image
)
pred_binary_fn = os.path.join(save_dir, to_png_fn(img_names[i], name='_pred_binary'))
pred_lane_fn = os.path.join(save_dir, to_png_fn(img_names[i], name='_pred_lane'))
pred_instance_fn = os.path.join(save_dir, to_png_fn(img_names[i], name='_pred_instance'))
source_image=gt_image)
pred_binary_fn = os.path.join(
save_dir, to_png_fn(img_names[i], name='_pred_binary'))
pred_lane_fn = os.path.join(
save_dir, to_png_fn(img_names[i], name='_pred_lane'))
pred_instance_fn = os.path.join(
save_dir, to_png_fn(img_names[i], name='_pred_instance'))
dirname = os.path.dirname(pred_binary_fn)
makedirs(dirname)
mask_image = postprocess_result['mask_image']
for i in range(4):
instance_seg_image[:, :, i] = minmax_scale(instance_seg_image[:, :, i])
instance_seg_image[:, :, i] = minmax_scale(
instance_seg_image[:, :, i])
embedding_image = np.array(instance_seg_image).astype(np.uint8)
plt.figure('mask_image')
......@@ -189,13 +197,13 @@ def visualize(cfg,
plt.imshow(binary_seg_image * 255, cmap='gray')
plt.show()
cv2.imwrite(pred_binary_fn, np.array(binary_seg_image * 255).astype(np.uint8))
cv2.imwrite(pred_binary_fn,
np.array(binary_seg_image * 255).astype(np.uint8))
cv2.imwrite(pred_lane_fn, postprocess_result['source_image'])
cv2.imwrite(pred_instance_fn, mask_image)
print(pred_lane_fn, 'saved!')
if __name__ == '__main__':
args = parse_args()
if args.cfg_file is not None:
......
contrib/MechanicalIndustryMeter/download_mini_mechanical_industry_meter.py
浏览文件 @ 7835c042
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/MechanicalIndustryMeter/download_unet_mechanical_industry_meter.py
浏览文件 @ 7835c042
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......@@ -23,7 +24,8 @@ from test_utils import download_file_and_uncompress
if __name__ == "__main__":
download_file_and_uncompress(
url='https://paddleseg.bj.bcebos.com/models/unet_mechanical_industry_meter.tar',
url=
'https://paddleseg.bj.bcebos.com/models/unet_mechanical_industry_meter.tar',
savepath=LOCAL_PATH,
extrapath=LOCAL_PATH)
......
contrib/README.md
浏览文件 @ 7835c042
......@@ -3,6 +3,7 @@
提供基于PaddlSeg最新的分割特色模型:
- [人像分割](./HumanSeg)
- [遥感分割](./RemoteSensing)
- [人体解析](./ACE2P)
- [车道线分割](./LaneNet)
- [工业表盘分割](#工业表盘分割)
......@@ -12,6 +13,14 @@
HumanSeg系列全新升级，提供三个适用于不同场景，包含适用于移动端实时分割场景的模型`HumanSeg-lite`，提供了包含光流的后处理的优化，使人像分割在视频场景中更加顺畅，更多详情请参考[HumanSeg](./HumanSeg)
## 遥感分割 Remote Sensing Segmentation
PaddleSeg遥感影像分割涵盖图像预处理、数据增强、模型训练、预测流程。
针对遥感数据多通道、分布范围大、分布不均的特点，我们支持多通道训练预测，内置10+多通道预处理和数据增强的策略，可结合实际业务场景进行定制组合，提升模型泛化能力和鲁棒性。
内置U-Net, HRNet两种主流分割网络，可选择不同的损失函数如Dice Loss, BCE Loss等方式强化小目标和不均衡样本场景下的分割精度。更多详情请参考[RemoteSensing](./RemoteSensing)
以下是遥感云检测的示例效果：
![](./RemoteSensing/docs/imgs/rs.png)
## 人体解析 Human Parsing
......
contrib/RemoteSensing/README.md
浏览文件 @ 7835c042
# 遥感分割（RemoteSensing）
# PaddleSeg遥感影像分割
遥感影像分割是图像分割领域中的重要应用场景，广泛应用于土地测绘、环境监测、城市建设等领域。遥感影像分割的目标多种多样，有诸如积雪、农作物、道路、建筑、水源等地物目标，也有例如云层的空中目标。
PaddleSeg提供了针对遥感专题的语义分割库RemoteSensing，涵盖图像预处理、数据增强、模型训练、预测流程，帮助用户利用深度学习技术解决遥感影像分割问题。
PaddleSeg遥感影像分割涵盖图像预处理、数据增强、模型训练、预测流程，帮助用户利用深度学习技术解决遥感影像分割问题。
## 特点
针对遥感数据多通道、分布范围大、分布不均的特点，我们支持多通道训练预测，内置一系列多通道预处理和数据增强的策略，可结合实际业务场景进行定制组合，提升模型泛化能力和鲁棒性。
- 针对遥感数据多通道、分布范围大、分布不均的特点，我们支持多通道训练预测，内置10+多通道预处理和数据增强的策略，可结合实际业务场景进行定制组合，提升模型泛化能力和鲁棒性。
**Note:** 所有命令需要在`PaddleSeg/contrib/RemoteSensing/`目录下执行。
- 内置U-Net, HRNet两种主流分割网络，可选择不同的损失函数如Dice Loss, BCE Loss等方式强化小目标和不均衡样本场景下的分割精度。
以下是遥感云检测的示例效果：
![](./docs/imgs/rs.png)
## 前置依赖
**Note:** 若没有特殊说明，以下所有命令需要在`PaddleSeg/contrib/RemoteSensing/`目录下执行。
- Paddle 1.7.1+
由于图像分割模型计算开销大，推荐在GPU版本的PaddlePaddle下使用。
PaddlePaddle的安装, 请按照[官网指引](https://paddlepaddle.org.cn/install/quick)安装合适自己的版本。
......@@ -18,7 +24,6 @@ PaddlePaddle的安装, 请按照[官网指引](https://paddlepaddle.org.cn/insta
- 其他依赖安装
通过以下命令安装python包依赖，请确保至少执行过一次以下命令：
```
cd RemoteSensing
pip install -r requirements.txt
```
......@@ -63,9 +68,9 @@ RemoteSensing # 根目录
```
其中，相应的文件名可根据需要自行定义。
遥感领域图像格式多种多样，不同传感器产生的数据格式可能不同。为方便数据加载，本分割库统一采用numpy存储格式`npy`作为原图格式，采用`png`无损压缩格式作为标注图片格式。
原图的前两维是图像的尺寸，第3维是图像的通道数。
标注图像为单通道图像，像素值即为对应的类别,像素标注类别需要从0开始递增，
遥感影像的格式多种多样，不同传感器产生的数据格式也可能不同。PaddleSeg以numpy.ndarray数据类型进行图像预处理。为统一接口并方便数据加载，我们采用numpy存储格式`npy`作为原图格式，采用`png`无损压缩格式作为标注图片格式。
原图的尺寸应为(h, w, channel)，其中h, w为图像的高和宽，channel为图像的通道数。
标注图像为单通道图像，像素值即为对应的类别,像素标注类别需要从0开始递增。
例如0，1，2，3表示有4种类别，标注类别最多为256类。其中可以指定特定的像素值用于表示该值的像素不参与训练和评估（默认为255）。
`train_list.txt`和`val_list.txt`文本以空格为分割符分为两列，第一列为图像文件相对于dataset的相对路径，第二列为标注图像文件相对于dataset的相对路径。如下所示：
......@@ -93,154 +98,38 @@ labelB
### 1. 准备数据集
为了快速体验，我们准备了一个小型demo数据集，已位于`RemoteSensing/dataset/demo/`目录下.
对于您自己的数据集，您需要按照上述的数据协议进行格式转换，可分别使用numpy和pil库保存遥感数据和标注图片。其中numpy api示例如下：
对于您自己的数据集，您需要按照上述的数据协议进行格式转换，可分别使用numpy和Pillow库保存遥感数据和标注图片。其中numpy API示例如下：
```python
import numpy as np
# 保存遥感数据
# 将遥感数据保存到以 .npy 为扩展名的文件中
# img类型：numpy.ndarray
np.save(save_path, img)
```
### 2. 训练代码开发
通过如下`train_demo.py`代码进行训练。
> 导入RemoteSensing api
```python
import transforms.transforms as T
from readers.reader import Reader
from models import UNet
```
> 定义训练和验证时的数据处理和增强流程, 在`train_transforms`中加入了`RandomVerticalFlip`,`RandomHorizontalFlip`等数据增强方式。
```python
train_transforms = T.Compose([
T.RandomVerticalFlip(0.5),
T.RandomHorizontalFlip(0.5),
T.ResizeStepScaling(0.5, 2.0, 0.25),
T.RandomPaddingCrop(256),
T.Normalize(mean=[0.5] * channel, std=[0.5] * channel),
])
eval_transforms = T.Compose([
T.Normalize(mean=[0.5] * channel, std=[0.5] * channel),
])
```
> 定义数据读取器
```python
import os
import os.path as osp
train_list = osp.join(data_dir, 'train.txt')
val_list = osp.join(data_dir, 'val.txt')
label_list = osp.join(data_dir, 'labels.txt')
train_reader = Reader(
data_dir=data_dir,
file_list=train_list,
label_list=label_list,
transforms=train_transforms,
num_workers=8,
buffer_size=16,
shuffle=True,
parallel_method='thread')
eval_reader = Reader(
data_dir=data_dir,
file_list=val_list,
label_list=label_list,
transforms=eval_transforms,
num_workers=8,
buffer_size=16,
shuffle=False,
parallel_method='thread')
```
> 模型构建
```python
model = UNet(
num_classes=2, input_channel=channel, use_bce_loss=True, use_dice_loss=True)
```
> 模型训练，并开启边训边评估
```python
model.train(
num_epochs=num_epochs,
train_reader=train_reader,
train_batch_size=train_batch_size,
eval_reader=eval_reader,
save_interval_epochs=5,
log_interval_steps=10,
save_dir=save_dir,
pretrain_weights=None,
optimizer=None,
learning_rate=lr,
use_vdl=True
)
```
### 3. 模型训练
> 设置GPU卡号
### 2. 模型训练
#### （1） 设置GPU卡号
```shell script
export CUDA_VISIBLE_DEVICES=0
```
> 在RemoteSensing目录下运行`train_demo.py`即可开始训练。
#### （2） 以U-Net为例，在RemoteSensing目录下运行`train_demo.py`即可开始训练。
```shell script
python train_demo.py --data_dir dataset/demo/ --save_dir saved_model/unet/ --channel 3 --num_epochs 20
```
### 4. 模型预测代码开发
通过如下`predict_demo.py`代码进行预测。
> 导入RemoteSensing api
```python
from models import load_model
```
> 加载训练过程中最好的模型，设置预测结果保存路径。
```python
import os
import os.path as osp
model = load_model(osp.join(save_dir, 'best_model'))
pred_dir = osp.join(save_dir, 'pred')
if not osp.exists(pred_dir):
os.mkdir(pred_dir)
```
> 使用模型对验证集进行测试，并保存预测结果。
```python
import numpy as np
from PIL import Image as Image
val_list = osp.join(data_dir, 'val.txt')
color_map = [0, 0, 0, 255, 255, 255]
with open(val_list) as f:
lines = f.readlines()
for line in lines:
img_path = line.split(' ')[0]
print('Predicting {}'.format(img_path))
img_path_ = osp.join(data_dir, img_path)
pred = model.predict(img_path_)
# 以伪彩色png图片保存预测结果
pred_name = osp.basename(img_path).rstrip('npy') + 'png'
pred_path = osp.join(pred_dir, pred_name)
pred_mask = Image.fromarray(pred.astype(np.uint8), mode='P')
pred_mask.putpalette(color_map)
pred_mask.save(pred_path)
python train_demo.py --model_type unet --data_dir dataset/demo/ --save_dir saved_model/unet/ --channel 3 --num_epochs 20
```
### 5. 模型预测
> 设置GPU卡号
### 3. 模型预测
#### （1） 设置GPU卡号
```shell script
export CUDA_VISIBLE_DEVICES=0
```
> 在RemoteSensing目录下运行`predict_demo.py`即可开始训练。
#### （2） 以刚训练好的U-Net最优模型为例，在RemoteSensing目录下运行`predict_demo.py`即可开始训练。
```shell script
python predict_demo.py --data_dir dataset/demo/ --load_model_dir saved_model/unet/best_model/
python predict_demo.py --data_dir dataset/demo/ --file_list val.txt --load_model_dir saved_model/unet/best_model
```
## Api说明
## API说明
您可以使用`RemoteSensing`目录下提供的api构建自己的分割代码。
您可以使用`RemoteSensing`目录下提供的API构建自己的分割代码。
- [数据处理-transforms](docs/transforms.md)
contrib/RemoteSensing/__init__.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/RemoteSensing/models/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .load_model import *
from .unet import *
from .hrnet import *
contrib/RemoteSensing/models/hrnet.py 0 → 100644
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
import paddle.fluid as fluid
import os
from os import path as osp
import numpy as np
from collections import OrderedDict
import copy
import math
import time
import tqdm
import cv2
import yaml
import utils
import utils.logging as logging
from utils.utils import seconds_to_hms, get_environ_info
from utils.metrics import ConfusionMatrix
import nets
import transforms.transforms as T
from .base import BaseModel
def dict2str(dict_input):
out = ''
for k, v in dict_input.items():
try:
v = round(float(v), 6)
except:
pass
out = out + '{}={}, '.format(k, v)
return out.strip(', ')
class HRNet(BaseModel):
def __init__(self,
num_classes=2,
input_channel=3,
stage1_num_modules=1,
stage1_num_blocks=[4],
stage1_num_channels=[64],
stage2_num_modules=1,
stage2_num_blocks=[4, 4],
stage2_num_channels=[18, 36],
stage3_num_modules=4,
stage3_num_blocks=[4, 4, 4],
stage3_num_channels=[18, 36, 72],
stage4_num_modules=3,
stage4_num_blocks=[4, 4, 4, 4],
stage4_num_channels=[18, 36, 72, 144],
use_bce_loss=False,
use_dice_loss=False,
class_weight=None,
ignore_index=255,
sync_bn=True):
super().__init__(
num_classes=num_classes,
use_bce_loss=use_bce_loss,
use_dice_loss=use_dice_loss,
class_weight=class_weight,
ignore_index=ignore_index,
sync_bn=sync_bn)
self.init_params = locals()
self.input_channel = input_channel
self.stage1_num_modules = stage1_num_modules
self.stage1_num_blocks = stage1_num_blocks
self.stage1_num_channels = stage1_num_channels
self.stage2_num_modules = stage2_num_modules
self.stage2_num_blocks = stage2_num_blocks
self.stage2_num_channels = stage2_num_channels
self.stage3_num_modules = stage3_num_modules
self.stage3_num_blocks = stage3_num_blocks
self.stage3_num_channels = stage3_num_channels
self.stage4_num_modules = stage4_num_modules
self.stage4_num_blocks = stage4_num_blocks
self.stage4_num_channels = stage4_num_channels
def build_net(self, mode='train'):
"""应根据不同的情况进行构建"""
model = nets.HRNet(
self.num_classes,
self.input_channel,
mode=mode,
stage1_num_modules=self.stage1_num_modules,
stage1_num_blocks=self.stage1_num_blocks,
stage1_num_channels=self.stage1_num_channels,
stage2_num_modules=self.stage2_num_modules,
stage2_num_blocks=self.stage2_num_blocks,
stage2_num_channels=self.stage2_num_channels,
stage3_num_modules=self.stage3_num_modules,
stage3_num_blocks=self.stage3_num_blocks,
stage3_num_channels=self.stage3_num_channels,
stage4_num_modules=self.stage4_num_modules,
stage4_num_blocks=self.stage4_num_blocks,
stage4_num_channels=self.stage4_num_channels,
use_bce_loss=self.use_bce_loss,
use_dice_loss=self.use_dice_loss,
class_weight=self.class_weight,
ignore_index=self.ignore_index)
inputs = model.generate_inputs()
model_out = model.build_net(inputs)
outputs = OrderedDict()
if mode == 'train':
self.optimizer.minimize(model_out)
outputs['loss'] = model_out
else:
outputs['pred'] = model_out[0]
outputs['logit'] = model_out[1]
return inputs, outputs
def train(self,
num_epochs,
train_reader,
train_batch_size=2,
eval_reader=None,
eval_best_metric='kappa',
save_interval_epochs=1,
log_interval_steps=2,
save_dir='output',
pretrain_weights=None,
resume_weights=None,
optimizer=None,
learning_rate=0.01,
lr_decay_power=0.9,
regularization_coeff=5e-4,
use_vdl=False):
super().train(
num_epochs=num_epochs,
train_reader=train_reader,
train_batch_size=train_batch_size,
eval_reader=eval_reader,
eval_best_metric=eval_best_metric,
save_interval_epochs=save_interval_epochs,
log_interval_steps=log_interval_steps,
save_dir=save_dir,
pretrain_weights=pretrain_weights,
resume_weights=resume_weights,
optimizer=optimizer,
learning_rate=learning_rate,
lr_decay_power=lr_decay_power,
regularization_coeff=regularization_coeff,
use_vdl=use_vdl)
contrib/RemoteSensing/models/load_model.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -25,7 +26,7 @@ import models
def load_model(model_dir):
if not osp.exists(osp.join(model_dir, "model.yml")):
raise Exception("There's not model.yml in {}".format(model_dir))
raise Exception("There's no model.yml in {}".format(model_dir))
with open(osp.join(model_dir, "model.yml")) as f:
info = yaml.load(f.read(), Loader=yaml.Loader)
status = info['status']
......@@ -35,8 +36,7 @@ def load_model(model_dir):
info['Model']))
model = getattr(models, info['Model'])(**info['_init_params'])
if status == "Normal" or \
status == "Prune":
if status == "Normal":
startup_prog = fluid.Program()
model.test_prog = fluid.Program()
with fluid.program_guard(model.test_prog, startup_prog):
......@@ -45,17 +45,12 @@ def load_model(model_dir):
mode='test')
model.test_prog = model.test_prog.clone(for_test=True)
model.exe.run(startup_prog)
if status == "Prune":
from .slim.prune import update_program
model.test_prog = update_program(model.test_prog, model_dir,
model.places[0])
import pickle
with open(osp.join(model_dir, 'model.pdparams'), 'rb') as f:
load_dict = pickle.load(f)
fluid.io.set_program_state(model.test_prog, load_dict)
elif status == "Infer" or \
status == "Quant":
elif status == "Infer":
[prog, input_names, outputs] = fluid.io.load_inference_model(
model_dir, model.exe, params_filename='__params__')
model.test_prog = prog
......@@ -67,8 +62,8 @@ def load_model(model_dir):
for i, out in enumerate(outputs):
var_desc = test_outputs_info[i]
model.test_outputs[var_desc[0]] = out
if 'Transforms' in info:
model.test_transforms = build_transforms(info['Transforms'])
if 'test_transforms' in info:
model.test_transforms = build_transforms(info['test_transforms'])
model.eval_transforms = copy.deepcopy(model.test_transforms)
if '_Attributes' in info:
......
contrib/RemoteSensing/models/unet.py
浏览文件 @ 7835c042
#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
import os.path as osp
import numpy as np
import math
import cv2
import paddle.fluid as fluid
import utils.logging as logging
from collections import OrderedDict
from .base import BaseAPI
from .base import BaseModel
from utils.metrics import ConfusionMatrix
import nets
class UNet(BaseAPI):
class UNet(BaseModel):
"""实现UNet网络的构建并进行训练、评估、预测和模型导出。
Args:
......@@ -55,9 +55,16 @@ class UNet(BaseAPI):
use_bce_loss=False,
use_dice_loss=False,
class_weight=None,
ignore_index=255):
ignore_index=255,
sync_bn=True):
super().__init__(
num_classes=num_classes,
use_bce_loss=use_bce_loss,
use_dice_loss=use_dice_loss,
class_weight=class_weight,
ignore_index=ignore_index,
sync_bn=sync_bn)
self.init_params = locals()
super(UNet, self).__init__()
# dice_loss或bce_loss只适用两类分割中
if num_classes > 2 and (use_bce_loss or use_dice_loss):
raise ValueError(
......@@ -115,24 +122,6 @@ class UNet(BaseAPI):
outputs['logit'] = model_out[1]
return inputs, outputs
def default_optimizer(self,
learning_rate,
num_epochs,
num_steps_each_epoch,
lr_decay_power=0.9):
decay_step = num_epochs * num_steps_each_epoch
lr_decay = fluid.layers.polynomial_decay(
learning_rate,
decay_step,
end_learning_rate=0,
power=lr_decay_power)
optimizer = fluid.optimizer.Momentum(
lr_decay,
momentum=0.9,
regularization=fluid.regularizer.L2Decay(
regularization_coeff=4e-05))
return optimizer
def train(self,
num_epochs,
train_reader,
......@@ -142,13 +131,13 @@ class UNet(BaseAPI):
save_interval_epochs=1,
log_interval_steps=2,
save_dir='output',
pretrain_weights='COCO',
pretrain_weights=None,
resume_weights=None,
optimizer=None,
learning_rate=0.01,
lr_decay_power=0.9,
use_vdl=False,
sensitivities_file=None,
eval_metric_loss=0.05):
regularization_coeff=5e-4,
use_vdl=False):
"""训练。
Args:
......@@ -160,46 +149,17 @@ class UNet(BaseAPI):
save_interval_epochs (int): 模型保存间隔（单位：迭代轮数）。默认为1。
log_interval_steps (int): 训练日志输出间隔（单位：迭代次数）。默认为2。
save_dir (str): 模型保存路径。默认'output'。
pretrain_weights (str): 若指定为路径时，则加载路径下预训练模型；若为字符串'COCO'，
则自动下载在COCO图片数据上预训练的模型权重；若为None，则不使用预训练模型。默认为'COCO'。
pretrain_weights (str): 若指定为路径时，则加载路径下预训练模型；若为None，则不使用预训练模型。
optimizer (paddle.fluid.optimizer): 优化器。当改参数为None时，使用默认的优化器：使用
fluid.optimizer.Momentum优化方法，polynomial的学习率衰减策略。
learning_rate (float): 默认优化器的初始学习率。默认0.01。
lr_decay_power (float): 默认优化器学习率多项式衰减系数。默认0.9。
use_vdl (bool): 是否使用VisualDL进行可视化。默认False。
sensitivities_file (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，
则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
Raises:
ValueError: 模型从inference model进行加载。
"""
if not self.trainable:
raise ValueError(
"Model is not trainable since it was loaded from a inference model."
)
self.labels = train_reader.labels
if optimizer is None:
num_steps_each_epoch = train_reader.num_samples // train_batch_size
optimizer = self.default_optimizer(
learning_rate=learning_rate,
num_epochs=num_epochs,
num_steps_each_epoch=num_steps_each_epoch,
lr_decay_power=lr_decay_power)
self.optimizer = optimizer
# 构建训练、验证、预测网络
self.build_program()
# 初始化网络权重
self.net_initialize(
startup_prog=fluid.default_startup_program(),
pretrain_weights=pretrain_weights,
save_dir=save_dir,
sensitivities_file=sensitivities_file,
eval_metric_loss=eval_metric_loss)
# 训练
self.train_loop(
super().train(
num_epochs=num_epochs,
train_reader=train_reader,
train_batch_size=train_batch_size,
......@@ -208,6 +168,12 @@ class UNet(BaseAPI):
save_interval_epochs=save_interval_epochs,
log_interval_steps=log_interval_steps,
save_dir=save_dir,
pretrain_weights=pretrain_weights,
resume_weights=resume_weights,
optimizer=optimizer,
learning_rate=learning_rate,
lr_decay_power=lr_decay_power,
regularization_coeff=regularization_coeff,
use_vdl=use_vdl)
def evaluate(self,
......@@ -231,7 +197,7 @@ class UNet(BaseAPI):
tuple (metrics, eval_details)：当return_details为True时，增加返回dict (eval_details)，
包含关键字：'confusion_matrix'，表示评估的混淆矩阵。
"""
self.arrange_transforms(transforms=eval_reader.transforms, mode='eval')
self.arrange_transform(transforms=eval_reader.transforms, mode='eval')
total_steps = math.ceil(eval_reader.num_samples * 1.0 / batch_size)
conf_mat = ConfusionMatrix(self.num_classes, streaming=True)
data_generator = eval_reader.generator(
......@@ -272,11 +238,16 @@ class UNet(BaseAPI):
category_iou, miou = conf_mat.mean_iou()
category_acc, macc = conf_mat.accuracy()
precision, recall = conf_mat.precision_recall()
metrics = OrderedDict(
zip(['miou', 'category_iou', 'macc', 'category_acc', 'kappa'],
[miou, category_iou, macc, category_acc,
conf_mat.kappa()]))
zip([
'miou', 'category_iou', 'macc', 'category_acc', 'kappa',
'precision', 'recall'
], [
miou, category_iou, macc, category_acc,
conf_mat.kappa(), precision, recall
]))
if return_details:
eval_details = {
'confusion_matrix': conf_mat.confusion_matrix.tolist()
......@@ -296,11 +267,10 @@ class UNet(BaseAPI):
if transforms is None and not hasattr(self, 'test_transforms'):
raise Exception("transforms need to be defined, now is None.")
if transforms is not None:
self.arrange_transforms(transforms=transforms, mode='test')
self.arrange_transform(transforms=transforms, mode='test')
im, im_info = transforms(im_file)
else:
self.arrange_transforms(
transforms=self.test_transforms, mode='test')
self.arrange_transform(transforms=self.test_transforms, mode='test')
im, im_info = self.test_transforms(im_file)
im = im.astype(np.float32)
im = np.expand_dims(im, axis=0)
......@@ -319,4 +289,4 @@ class UNet(BaseAPI):
h, w = im_info[k][0], im_info[k][1]
pred = pred[0:h, 0:w]
return pred
return {'label_map': pred}
contrib/RemoteSensing/nets/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .unet import UNet
from .hrnet import HRNet
contrib/RemoteSensing/nets/hrnet.py 0 → 100644
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from collections import OrderedDict
import paddle.fluid as fluid
from paddle.fluid.initializer import MSRA
from paddle.fluid.param_attr import ParamAttr
from .loss import softmax_with_loss
from .loss import dice_loss
from .loss import bce_loss
from .libs import sigmoid_to_softmax
class HRNet(object):
def __init__(self,
num_classes,
input_channel=3,
mode='train',
stage1_num_modules=1,
stage1_num_blocks=[4],
stage1_num_channels=[64],
stage2_num_modules=1,
stage2_num_blocks=[4, 4],
stage2_num_channels=[18, 36],
stage3_num_modules=4,
stage3_num_blocks=[4, 4, 4],
stage3_num_channels=[18, 36, 72],
stage4_num_modules=3,
stage4_num_blocks=[4, 4, 4, 4],
stage4_num_channels=[18, 36, 72, 144],
use_bce_loss=False,
use_dice_loss=False,
class_weight=None,
ignore_index=255):
# dice_loss或bce_loss只适用两类分割中
if num_classes > 2 and (use_bce_loss or use_dice_loss):
raise ValueError(
"dice loss and bce loss is only applicable to binary classfication"
)
if class_weight is not None:
if isinstance(class_weight, list):
if len(class_weight) != num_classes:
raise ValueError(
"Length of class_weight should be equal to number of classes"
)
elif isinstance(class_weight, str):
if class_weight.lower() != 'dynamic':
raise ValueError(
"if class_weight is string, must be dynamic!")
else:
raise TypeError(
'Expect class_weight is a list or string but receive {}'.
format(type(class_weight)))
self.num_classes = num_classes
self.input_channel = input_channel
self.mode = mode
self.use_bce_loss = use_bce_loss
self.use_dice_loss = use_dice_loss
self.class_weight = class_weight
self.ignore_index = ignore_index
self.stage1_num_modules = stage1_num_modules
self.stage1_num_blocks = stage1_num_blocks
self.stage1_num_channels = stage1_num_channels
self.stage2_num_modules = stage2_num_modules
self.stage2_num_blocks = stage2_num_blocks
self.stage2_num_channels = stage2_num_channels
self.stage3_num_modules = stage3_num_modules
self.stage3_num_blocks = stage3_num_blocks
self.stage3_num_channels = stage3_num_channels
self.stage4_num_modules = stage4_num_modules
self.stage4_num_blocks = stage4_num_blocks
self.stage4_num_channels = stage4_num_channels
def build_net(self, inputs):
if self.use_dice_loss or self.use_bce_loss:
self.num_classes = 1
image = inputs['image']
logit = self._high_resolution_net(image, self.num_classes)
if self.num_classes == 1:
out = sigmoid_to_softmax(logit)
out = fluid.layers.transpose(out, [0, 2, 3, 1])
else:
out = fluid.layers.transpose(logit, [0, 2, 3, 1])
pred = fluid.layers.argmax(out, axis=3)
pred = fluid.layers.unsqueeze(pred, axes=[3])
if self.mode == 'train':
label = inputs['label']
mask = label != self.ignore_index
return self._get_loss(logit, label, mask)
else:
if self.num_classes == 1:
logit = sigmoid_to_softmax(logit)
else:
logit = fluid.layers.softmax(logit, axis=1)
return pred, logit
return logit
def generate_inputs(self):
inputs = OrderedDict()
inputs['image'] = fluid.data(
dtype='float32',
shape=[None, self.input_channel, None, None],
name='image')
if self.mode == 'train':
inputs['label'] = fluid.data(
dtype='int32', shape=[None, 1, None, None], name='label')
elif self.mode == 'eval':
inputs['label'] = fluid.data(
dtype='int32', shape=[None, 1, None, None], name='label')
return inputs
def _get_loss(self, logit, label, mask):
avg_loss = 0
if not (self.use_dice_loss or self.use_bce_loss):
avg_loss += softmax_with_loss(
logit,
label,
mask,
num_classes=self.num_classes,
weight=self.class_weight,
ignore_index=self.ignore_index)
else:
if self.use_dice_loss:
avg_loss += dice_loss(logit, label, mask)
if self.use_bce_loss:
avg_loss += bce_loss(
logit, label, mask, ignore_index=self.ignore_index)
return avg_loss
def _conv_bn_layer(self,
input,
filter_size,
num_filters,
stride=1,
padding=1,
num_groups=1,
if_act=True,
name=None):
conv = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=num_groups,
act=None,
param_attr=ParamAttr(initializer=MSRA(), name=name + '_weights'),
bias_attr=False)
bn_name = name + '_bn'
bn = fluid.layers.batch_norm(
input=conv,
param_attr=ParamAttr(
name=bn_name + "_scale",
initializer=fluid.initializer.Constant(1.0)),
bias_attr=ParamAttr(
name=bn_name + "_offset",
initializer=fluid.initializer.Constant(0.0)),
moving_mean_name=bn_name + '_mean',
moving_variance_name=bn_name + '_variance')
if if_act:
bn = fluid.layers.relu(bn)
return bn
def _basic_block(self,
input,
num_filters,
stride=1,
downsample=False,
name=None):
residual = input
conv = self._conv_bn_layer(
input=input,
filter_size=3,
num_filters=num_filters,
stride=stride,
name=name + '_conv1')
conv = self._conv_bn_layer(
input=conv,
filter_size=3,
num_filters=num_filters,
if_act=False,
name=name + '_conv2')
if downsample:
residual = self._conv_bn_layer(
input=input,
filter_size=1,
num_filters=num_filters,
if_act=False,
name=name + '_downsample')
return fluid.layers.elementwise_add(x=residual, y=conv, act='relu')
def _bottleneck_block(self,
input,
num_filters,
stride=1,
downsample=False,
name=None):
residual = input
conv = self._conv_bn_layer(
input=input,
filter_size=1,
num_filters=num_filters,
name=name + '_conv1')
conv = self._conv_bn_layer(
input=conv,
filter_size=3,
num_filters=num_filters,
stride=stride,
name=name + '_conv2')
conv = self._conv_bn_layer(
input=conv,
filter_size=1,
num_filters=num_filters * 4,
if_act=False,
name=name + '_conv3')
if downsample:
residual = self._conv_bn_layer(
input=input,
filter_size=1,
num_filters=num_filters * 4,
if_act=False,
name=name + '_downsample')
return fluid.layers.elementwise_add(x=residual, y=conv, act='relu')
def _fuse_layers(self, x, channels, multi_scale_output=True, name=None):
out = []
for i in range(len(channels) if multi_scale_output else 1):
residual = x[i]
shape = fluid.layers.shape(residual)[-2:]
for j in range(len(channels)):
if j > i:
y = self._conv_bn_layer(
x[j],
filter_size=1,
num_filters=channels[i],
if_act=False,
name=name + '_layer_' + str(i + 1) + '_' + str(j + 1))
y = fluid.layers.resize_bilinear(input=y, out_shape=shape)
residual = fluid.layers.elementwise_add(
x=residual, y=y, act=None)
elif j < i:
y = x[j]
for k in range(i - j):
if k == i - j - 1:
y = self._conv_bn_layer(
y,
filter_size=3,
num_filters=channels[i],
stride=2,
if_act=False,
name=name + '_layer_' + str(i + 1) + '_' +
str(j + 1) + '_' + str(k + 1))
else:
y = self._conv_bn_layer(
y,
filter_size=3,
num_filters=channels[j],
stride=2,
name=name + '_layer_' + str(i + 1) + '_' +
str(j + 1) + '_' + str(k + 1))
residual = fluid.layers.elementwise_add(
x=residual, y=y, act=None)
residual = fluid.layers.relu(residual)
out.append(residual)
return out
def _branches(self, x, block_num, channels, name=None):
out = []
for i in range(len(channels)):
residual = x[i]
for j in range(block_num[i]):
residual = self._basic_block(
residual,
channels[i],
name=name + '_branch_layer_' + str(i + 1) + '_' +
str(j + 1))
out.append(residual)
return out
def _high_resolution_module(self,
x,
blocks,
channels,
multi_scale_output=True,
name=None):
residual = self._branches(x, blocks, channels, name=name)
out = self._fuse_layers(
residual,
channels,
multi_scale_output=multi_scale_output,
name=name)
return out
def _transition_layer(self, x, in_channels, out_channels, name=None):
num_in = len(in_channels)
num_out = len(out_channels)
out = []
for i in range(num_out):
if i < num_in:
if in_channels[i] != out_channels[i]:
residual = self._conv_bn_layer(
x[i],
filter_size=3,
num_filters=out_channels[i],
name=name + '_layer_' + str(i + 1))
out.append(residual)
else:
out.append(x[i])
else:
residual = self._conv_bn_layer(
x[-1],
filter_size=3,
num_filters=out_channels[i],
stride=2,
name=name + '_layer_' + str(i + 1))
out.append(residual)
return out
def _stage(self,
x,
num_modules,
num_blocks,
num_channels,
multi_scale_output=True,
name=None):
out = x
for i in range(num_modules):
if i == num_modules - 1 and multi_scale_output == False:
out = self._high_resolution_module(
out,
num_blocks,
num_channels,
multi_scale_output=False,
name=name + '_' + str(i + 1))
else:
out = self._high_resolution_module(
out, num_blocks, num_channels, name=name + '_' + str(i + 1))
return out
def _layer1(self, input, num_modules, num_blocks, num_channels, name=None):
# num_modules 默认为1,是否增加处理，官网实现为[1]，是否对齐。
conv = input
for i in range(num_blocks[0]):
conv = self._bottleneck_block(
conv,
num_filters=num_channels[0],
downsample=True if i == 0 else False,
name=name + '_' + str(i + 1))
return conv
def _high_resolution_net(self, input, num_classes):
x = self._conv_bn_layer(
input=input,
filter_size=3,
num_filters=self.stage1_num_channels[0],
stride=2,
if_act=True,
name='layer1_1')
x = self._conv_bn_layer(
input=x,
filter_size=3,
num_filters=self.stage1_num_channels[0],
stride=2,
if_act=True,
name='layer1_2')
la1 = self._layer1(
x,
self.stage1_num_modules,
self.stage1_num_blocks,
self.stage1_num_channels,
name='layer2')
tr1 = self._transition_layer([la1],
self.stage1_num_channels,
self.stage2_num_channels,
name='tr1')
st2 = self._stage(
tr1,
self.stage2_num_modules,
self.stage2_num_blocks,
self.stage2_num_channels,
name='st2')
tr2 = self._transition_layer(
st2, self.stage2_num_channels, self.stage3_num_channels, name='tr2')
st3 = self._stage(
tr2,
self.stage3_num_modules,
self.stage3_num_blocks,
self.stage3_num_channels,
name='st3')
tr3 = self._transition_layer(
st3, self.stage3_num_channels, self.stage4_num_channels, name='tr3')
st4 = self._stage(
tr3,
self.stage4_num_modules,
self.stage4_num_blocks,
self.stage4_num_channels,
name='st4')
# upsample
shape = fluid.layers.shape(st4[0])[-2:]
st4[1] = fluid.layers.resize_bilinear(st4[1], out_shape=shape)
st4[2] = fluid.layers.resize_bilinear(st4[2], out_shape=shape)
st4[3] = fluid.layers.resize_bilinear(st4[3], out_shape=shape)
out = fluid.layers.concat(st4, axis=1)
last_channels = sum(self.stage4_num_channels)
out = self._conv_bn_layer(
input=out,
filter_size=1,
num_filters=last_channels,
stride=1,
if_act=True,
name='conv-2')
out = fluid.layers.conv2d(
input=out,
num_filters=num_classes,
filter_size=1,
stride=1,
padding=0,
act=None,
param_attr=ParamAttr(initializer=MSRA(), name='conv-1_weights'),
bias_attr=False)
input_shape = fluid.layers.shape(input)[-2:]
out = fluid.layers.resize_bilinear(out, input_shape)
return out
contrib/RemoteSensing/nets/libs.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/nets/loss.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/nets/unet.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/predict_demo.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import os.path as osp
import sys
import numpy as np
from PIL import Image as Image
import argparse
......@@ -8,46 +24,81 @@ from models import load_model
def parse_args():
parser = argparse.ArgumentParser(description='RemoteSensing predict')
parser.add_argument(
'--single_img',
dest='single_img',
help='single image path to predict',
default=None,
type=str)
parser.add_argument(
'--data_dir',
dest='data_dir',
help='dataset directory',
default=None,
type=str)
parser.add_argument(
'--file_list',
dest='file_list',
help='file name of predict file list',
default=None,
type=str)
parser.add_argument(
'--load_model_dir',
dest='load_model_dir',
help='model load directory',
default=None,
type=str)
parser.add_argument(
'--save_img_dir',
dest='save_img_dir',
help='save directory name of predict results',
default='predict_results',
type=str)
if len(sys.argv) < 2:
parser.print_help()
sys.exit(1)
return parser.parse_args()
args = parse_args()
data_dir = args.data_dir
file_list = args.file_list
single_img = args.single_img
load_model_dir = args.load_model_dir
save_img_dir = args.save_img_dir
if not osp.exists(save_img_dir):
os.makedirs(save_img_dir)
# predict
model = load_model(load_model_dir)
pred_dir = osp.join(load_model_dir, 'predict')
if not osp.exists(pred_dir):
os.mkdir(pred_dir)
val_list = osp.join(data_dir, 'val.txt')
color_map = [0, 0, 0, 255, 255, 255]
with open(val_list) as f:
lines = f.readlines()
for line in lines:
img_path = line.split(' ')[0]
print('Predicting {}'.format(img_path))
img_path_ = osp.join(data_dir, img_path)
pred = model.predict(img_path_)
# 以伪彩色png图片保存预测结果
pred_name = osp.basename(img_path).rstrip('npy') + 'png'
pred_path = osp.join(pred_dir, pred_name)
pred_mask = Image.fromarray(pred.astype(np.uint8), mode='P')
pred_mask.putpalette(color_map)
pred_mask.save(pred_path)
color_map = [0, 0, 0, 0, 255, 0]
if single_img is not None:
pred = model.predict(single_img)
# 以伪彩色png图片保存预测结果
pred_name = osp.basename(single_img).rstrip('npy') + 'png'
pred_path = osp.join(save_img_dir, pred_name)
pred_mask = Image.fromarray(pred['label_map'].astype(np.uint8), mode='P')
pred_mask.putpalette(color_map)
pred_mask.save(pred_path)
elif (file_list is not None) and (data_dir is not None):
with open(osp.join(data_dir, file_list)) as f:
lines = f.readlines()
for line in lines:
img_path = line.split(' ')[0]
print('Predicting {}'.format(img_path))
img_path_ = osp.join(data_dir, img_path)
pred = model.predict(img_path_)
# 以伪彩色png图片保存预测结果
pred_name = osp.basename(img_path).rstrip('npy') + 'png'
pred_path = osp.join(save_img_dir, pred_name)
pred_mask = Image.fromarray(
pred['label_map'].astype(np.uint8), mode='P')
pred_mask.putpalette(color_map)
pred_mask.save(pred_path)
else:
raise Exception(
'You should either set the parameter single_img, or set the parameters data_dir, file_list.'
)
contrib/RemoteSensing/readers/__init__.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/readers/base.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/readers/reader.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/tools/create_dataset_list.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/tools/split_dataset_list.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/train_demo.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os.path as osp
import argparse
import transforms.transforms as T
from readers.reader import Reader
from models import UNet
from models import UNet, HRNet
def parse_args():
parser = argparse.ArgumentParser(description='RemoteSensing training')
parser.add_argument(
'--model_type',
dest='model_type',
help="Model type for traing, which is one of ('unet', 'hrnet')",
type=str,
default='hrnet')
parser.add_argument(
'--data_dir',
dest='data_dir',
......@@ -43,7 +64,6 @@ def parse_args():
args = parse_args()
data_dir = args.data_dir
save_dir = args.save_dir
channel = args.channel
......@@ -52,17 +72,9 @@ train_batch_size = args.train_batch_size
lr = args.lr
# 定义训练和验证时的transforms
train_transforms = T.Compose([
T.RandomVerticalFlip(0.5),
T.RandomHorizontalFlip(0.5),
T.ResizeStepScaling(0.5, 2.0, 0.25),
T.RandomPaddingCrop(256),
T.Normalize(mean=[0.5] * channel, std=[0.5] * channel),
])
train_transforms = T.Compose([T.RandomHorizontalFlip(0.5), T.Normalize()])
eval_transforms = T.Compose([
T.Normalize(mean=[0.5] * channel, std=[0.5] * channel),
])
eval_transforms = T.Compose([T.Normalize()])
train_list = osp.join(data_dir, 'train.txt')
val_list = osp.join(data_dir, 'val.txt')
......@@ -74,23 +86,30 @@ train_reader = Reader(
file_list=train_list,
label_list=label_list,
transforms=train_transforms,
num_workers=8,
buffer_size=16,
shuffle=True,
parallel_method='thread')
shuffle=True)
eval_reader = Reader(
data_dir=data_dir,
file_list=val_list,
label_list=label_list,
transforms=eval_transforms,
num_workers=8,
buffer_size=16,
shuffle=False,
parallel_method='thread')
transforms=eval_transforms)
model = UNet(
num_classes=2, input_channel=channel, use_bce_loss=True, use_dice_loss=True)
if args.model_type == 'unet':
model = UNet(
num_classes=2,
input_channel=channel,
use_bce_loss=True,
use_dice_loss=True)
elif args.model_type == 'hrnet':
model = HRNet(
num_classes=2,
input_channel=channel,
use_bce_loss=True,
use_dice_loss=True)
else:
raise ValueError(
"--model_type: {} is set wrong, it shold be one of ('unet', "
"'hrnet')".format(args.model_type))
model.train(
num_epochs=num_epochs,
......@@ -100,7 +119,5 @@ model.train(
save_interval_epochs=5,
log_interval_steps=10,
save_dir=save_dir,
pretrain_weights=None,
optimizer=None,
learning_rate=lr,
use_vdl=True)
contrib/RemoteSensing/transforms/__init__.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/RemoteSensing/transforms/ops.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/transforms/transforms.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/utils/__init__.py
浏览文件 @ 7835c042
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/RemoteSensing/utils/logging.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......
contrib/RemoteSensing/utils/metrics.py
浏览文件 @ 7835c042
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -143,3 +143,14 @@ class ConfusionMatrix(object):
kappa = (po - pe) / (1 - pe)
return kappa
def precision_recall(self):
'''
precision, recall of foreground(value=1) for 2 categories
'''
TP = self.confusion_matrix[1, 1]
FN = self.confusion_matrix[1, 0]
FP = self.confusion_matrix[0, 1]
recall = TP / (TP + FN)
precision = TP / (TP + FP)
return precision, recall
contrib/RemoteSensing/utils/pretrain_weights.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os.path as osp
......
contrib/RemoteSensing/utils/utils.py
浏览文件 @ 7835c042
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -12,13 +13,10 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import time
import os
import os.path as osp
import numpy as np
import six
import yaml
import math
from . import logging
......@@ -204,11 +202,9 @@ def load_pretrain_weights(exe, main_prog, weights_dir, fuse_bn=False):
vars_to_load.append(var)
logging.debug("Weight {} will be load".format(var.name))
fluid.io.load_vars(
executor=exe,
dirname=weights_dir,
main_program=main_prog,
vars=vars_to_load)
params_dict = fluid.io.load_program_state(
weights_dir, var_list=vars_to_load)
fluid.io.set_program_state(main_prog, params_dict)
if len(vars_to_load) == 0:
logging.warning(
"There is no pretrain weights loaded, maybe you should check you pretrain model!"
......
contrib/RoadLine/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
contrib/RoadLine/config.py
浏览文件 @ 7835c042
# -*- coding: utf-8 -*-
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from utils.util import AttrDict, merge_cfg_from_args, get_arguments
import os
......@@ -6,20 +20,20 @@ args = get_arguments()
cfg = AttrDict()
# 待预测图像所在路径
cfg.data_dir = os.path.join(args.example , "data", "test_images")
cfg.data_dir = os.path.join(args.example, "data", "test_images")
# 待预测图像名称列表
cfg.data_list_file = os.path.join(args.example , "data", "test.txt")
cfg.data_list_file = os.path.join(args.example, "data", "test.txt")
# 模型加载路径
cfg.model_path = os.path.join(args.example , "model")
cfg.model_path = os.path.join(args.example, "model")
# 预测结果保存路径
cfg.vis_dir = os.path.join(args.example , "result")
cfg.vis_dir = os.path.join(args.example, "result")
# 预测类别数
cfg.class_num = 2
# 均值, 图像预处理减去的均值
cfg.MEAN = 127.5, 127.5, 127.5
# 标准差，图像预处理除以标准差
cfg.STD = 127.5, 127.5, 127.5
cfg.STD = 127.5, 127.5, 127.5
# 待预测图像输入尺寸
cfg.input_size = 1536, 576
......
contrib/RoadLine/download_RoadLine.py
浏览文件 @ 7835c042
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
......
contrib/RoadLine/infer.py
浏览文件 @ 7835c042
# -*- coding: utf-8 -*-
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import cv2
import numpy as np
......@@ -12,18 +26,19 @@ config = importlib.import_module('config')
cfg = getattr(config, 'cfg')
# paddle垃圾回收策略FLAG，ACE2P模型较大，当显存不够时建议开启
os.environ['FLAGS_eager_delete_tensor_gb']='0.0'
os.environ['FLAGS_eager_delete_tensor_gb'] = '0.0'
import paddle.fluid as fluid
# 预测数据集类
class TestDataSet():
def __init__(self):
self.data_dir = cfg.data_dir
self.data_dir = cfg.data_dir
self.data_list_file = cfg.data_list_file
self.data_list = self.get_data_list()
self.data_num = len(self.data_list)
def get_data_list(self):
# 获取预测图像路径列表
data_list = []
......@@ -40,7 +55,7 @@ class TestDataSet():
def preprocess(self, img):
# 图像预处理
if cfg.example == 'ACE2P':
reader = importlib.import_module(args.example+'.reader')
reader = importlib.import_module(args.example + '.reader')
ACE2P_preprocess = getattr(reader, 'preprocess')
img = ACE2P_preprocess(img)
else:
......@@ -56,10 +71,10 @@ class TestDataSet():
img_path = self.data_list[index]
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
if img is None:
return img, img,img_path, None
return img, img, img_path, None
img_name = img_path.split(os.sep)[-1]
name_prefix = img_name.replace('.'+img_name.split('.')[-1],'')
name_prefix = img_name.replace('.' + img_name.split('.')[-1], '')
img_shape = img.shape[:2]
img_process = self.preprocess(img)
......@@ -90,39 +105,44 @@ def infer():
if image is None:
print(im_name, 'is None')
continue
# 预测
if cfg.example == 'ACE2P':
# ACE2P模型使用多尺度预测
reader = importlib.import_module(args.example+'.reader')
reader = importlib.import_module(args.example + '.reader')
multi_scale_test = getattr(reader, 'multi_scale_test')
parsing, logits = multi_scale_test(exe, test_prog, feed_name, fetch_list, image, im_shape)
parsing, logits = multi_scale_test(exe, test_prog, feed_name,
fetch_list, image, im_shape)
else:
# HumanSeg,RoadLine模型单尺度预测
result = exe.run(program=test_prog, feed={feed_name[0]: image}, fetch_list=fetch_list)
result = exe.run(
program=test_prog,
feed={feed_name[0]: image},
fetch_list=fetch_list)
parsing = np.argmax(result[0][0], axis=0)
parsing = cv2.resize(parsing.astype(np.uint8), im_shape[::-1])
# 预测结果保存
result_path = os.path.join(cfg.vis_dir, im_name + '.png')
if cfg.example == 'HumanSeg':
logits = result[0][0][1]*255
logits = result[0][0][1] * 255
logits = cv2.resize(logits, im_shape[::-1])
ret, logits = cv2.threshold(logits, thresh, 0, cv2.THRESH_TOZERO)
logits = 255 *(logits - thresh)/(255 - thresh)
logits = 255 * (logits - thresh) / (255 - thresh)
# 将分割结果添加到alpha通道
rgba = np.concatenate((ori_img, np.expand_dims(logits, axis=2)), axis=2)
rgba = np.concatenate((ori_img, np.expand_dims(logits, axis=2)),
axis=2)
cv2.imwrite(result_path, rgba)
else:
else:
output_im = PILImage.fromarray(np.asarray(parsing, dtype=np.uint8))
output_im.putpalette(palette)
output_im.save(result_path)
if (idx + 1) % 100 == 0:
print('%d processd' % (idx + 1))
print('%d processd done' % (idx + 1))
print('%d processd done' % (idx + 1))
return 0
......
contrib/RoadLine/utils/__init__.py
浏览文件 @ 7835c042
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
contrib/RoadLine/utils/palette.py
浏览文件 @ 7835c042
##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
## Created by: RainbowSecret
## Microsoft Research
## yuyua@microsoft.com
## Copyright (c) 2018
##
## This source code is licensed under the MIT-style license found in the
## LICENSE file in the root directory of this source tree
##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
# coding: utf8
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
......
contrib/RoadLine/utils/util.py
浏览文件 @ 7835c042
此差异已折叠。
dataset/convert_voc2012.py
浏览文件 @ 7835c042
此差异已折叠。
dataset/download_cityscapes.py
浏览文件 @ 7835c042
此差异已折叠。
dataset/download_optic.py
浏览文件 @ 7835c042
此差异已折叠。
dataset/download_pet.py
浏览文件 @ 7835c042
此差异已折叠。
deploy/python/README.md
浏览文件 @ 7835c042
此差异已折叠。
deploy/python/infer.py
浏览文件 @ 7835c042
此差异已折叠。
docs/annotation/jingling2seg.md
浏览文件 @ 7835c042
此差异已折叠。
docs/annotation/labelme2seg.md
浏览文件 @ 7835c042
此差异已折叠。
pdseg/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/check.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/data_aug.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/data_utils.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/eval.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/export_model.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/loss.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/lovasz_losses.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/metrics.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/backbone/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/backbone/resnet.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/backbone/vgg.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/backbone/xception.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/libs/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/libs/model_libs.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/model_builder.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/deeplab.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/fast_scnn.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/hrnet.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/icnet.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/pspnet.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/models/modeling/unet.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/reader.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/solver.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/tools/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/tools/create_dataset_list.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/tools/gray2pseudo_color.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/tools/jingling2seg.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/tools/labelme2seg.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/train.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/__init__.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/collect.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/config.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/dist_utils.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/fp16_utils.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/load_model_utils.py 0 → 100644
浏览文件 @ 7835c042
此差异已折叠。
pdseg/utils/timer.py
浏览文件 @ 7835c042
此差异已折叠。
pdseg/vis.py
浏览文件 @ 7835c042
此差异已折叠。
pretrained_model/download_model.py
浏览文件 @ 7835c042
此差异已折叠。
slim/distillation/model_builder.py
浏览文件 @ 7835c042
此差异已折叠。
slim/distillation/train_distill.py
浏览文件 @ 7835c042
此差异已折叠。
slim/nas/deeplab.py
浏览文件 @ 7835c042
此差异已折叠。
slim/nas/eval_nas.py
浏览文件 @ 7835c042
此差异已折叠。
slim/nas/model_builder.py
浏览文件 @ 7835c042
此差异已折叠。
slim/nas/train_nas.py
浏览文件 @ 7835c042
此差异已折叠。
slim/prune/eval_prune.py
浏览文件 @ 7835c042
此差异已折叠。
slim/prune/train_prune.py
浏览文件 @ 7835c042
此差异已折叠。
slim/quantization/eval_quant.py
浏览文件 @ 7835c042
此差异已折叠。
slim/quantization/export_model.py
浏览文件 @ 7835c042
此差异已折叠。
slim/quantization/train_quant.py
浏览文件 @ 7835c042
此差异已折叠。
test/local_test_cityscapes.py
浏览文件 @ 7835c042
此差异已折叠。
test/local_test_pet.py
浏览文件 @ 7835c042
此差异已折叠。
test/test_utils.py
浏览文件 @ 7835c042
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或