Skip to content

M
models

PaddlePaddle / models
大约 1 年 前同步成功

通知 222
Star 6828
Fork 2962
M
models
未验证 提交 6b882d42 编写于 作者: C ceci3 提交者： GitHub
浏览文件
操作

delete cycle_gan (#4476)

上级 90fcad46
显示空白变更内容
内联 并排
Showing with 4 addition and 1145 deletion
PaddleCV/gan/cycle_gan/.run_ce.sh 已删除 100755 → 0
浏览文件 @ 90fcad46
#!/bin/bash
# This file is only used for continuous evaluation.
export FLAGS_cudnn_deterministic=True
export ce_mode=1
CUDA_VISIBLE_DEVICES=0 python train.py --batch_size=1 --epoch=10 --run_ce=True --use_gpu=True | python _ce.py
PaddleCV/gan/cycle_gan/README.md 已删除 100644 → 0
浏览文件 @ 90fcad46
运行本目录下的程序示例需要使用PaddlePaddle develop最新版本。如果您的PaddlePaddle安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新PaddlePaddle安装版本。
## 代码结构
```
├── data_reader.py # 读取、处理数据。
├── layers.py # 封装定义基础的layers。
├── model.py # 定义基础生成网络和判别网络。
├── trainer.py # 构造loss和训练网络。
├── train.py # 训练脚本。
└── infer.py # 预测脚本。
```
## 简介
TODO
## 数据准备
本教程使用 horse2zebra 数据集 来进行模型的训练测试工作，该数据集是用关键字'wild horse'和'zebra'过滤[ImageNet](http://www.image-net.org/)数据集并下载得到的。
horse2zebra训练集包含1069张野马图片，1336张斑马图片。测试集包含121张野马图片和141张斑马图片。
数据下载处理完毕后，并组织为以下路径结构：
```
data
|-- horse2zebra
| |-- testA
| |-- testA.txt
| |-- testB
| |-- testB.txt
| |-- trainA
| |-- trainA.txt
| |-- trainB
| `-- trainB.txt
```
以上数据文件中，`data`文件夹需要放在训练脚本`train.py`同级目录下。`testA`为存放野马测试图片的文件夹，`testB`为存放斑马测试图片的文件夹，`testA.txt`和`testB.txt`分别为野马和斑马测试图片路径列表文件，格式如下：
```
testA/n02381460_9243.jpg
testA/n02381460_9244.jpg
testA/n02381460_9245.jpg
```
训练数据组织方式与测试数据相同。
## 模型训练与预测
### 训练
在GPU单卡上训练:
```
env CUDA_VISIBLE_DEVICES=0 python train.py
```
执行`python train.py --help`可查看更多使用方式和参数详细说明。
图1为训练152轮的训练损失示意图，其中横坐标轴为训练轮数，纵轴为在训练集上的损失。其中，'g_A_loss','g_B_loss','d_A_loss'和'd_B_loss'分别为生成器A、生成器B、判别器A和判别器B的训练损失。
<p align="center">
<img src="images/cycleGAN_loss.png" width="620" hspace='10'/> <br/>
<strong>图 1</strong>
</p>
### 预测
执行以下命令读取多张图片进行预测：
```
env CUDA_VISIBLE_DEVICES=0 python infer.py \
--init_model="output/checkpoints/1" --input="./data/horse2zebra/trainA/*" \
--input_style A --output="./output"
```
训练150轮的模型预测效果如图2和图3所示：
<p align="center">
<img src="images/A2B.jpg" width="620" hspace='10'/> <br/>
<strong>图 2</strong>
</p>
<p align="center">
<img src="images/B2A.jpg" width="620" hspace='10'/> <br/>
<strong>图 3</strong>
</p>
>在本文示例中，均可通过修改`CUDA_VISIBLE_DEVICES`改变使用的显卡号。
PaddleCV/gan/cycle_gan/_ce.py 已删除 100644 → 0
浏览文件 @ 90fcad46
####this file is only used for continuous evaluation test!
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
sys.path.append(os.environ['ceroot'])
from kpi import CostKpi, DurationKpi, AccKpi
#### NOTE kpi.py should shared in models in some way!!!!
d_train_cost_kpi = CostKpi('d_train_cost', 0.05, 0, actived=True, desc='train cost of discriminator')
g_train_cost_kpi = CostKpi('g_train_cost', 0.05, 0, actived=True, desc='train cost of generator')
train_speed_kpi = DurationKpi(
'duration',
0.05,
0,
actived=True,
unit_repr='second',
desc='train time used in one GPU card')
tracking_kpis = [d_train_cost_kpi, g_train_cost_kpi, train_speed_kpi]
def parse_log(log):
'''
This method should be implemented by model developers.
The suggestion:
each line in the log should be key, value, for example:
"
train_cost\t1.0
test_cost\t1.0
train_cost\t1.0
train_cost\t1.0
train_acc\t1.2
"
'''
for line in log.split('\n'):
fs = line.strip().split(',')
print(fs)
if len(fs) == 3 and fs[0] == 'kpis':
kpi_name = fs[1]
kpi_value = float(fs[2])
print("kpi {}={}".format(kpi_name, kpi_value))
yield kpi_name, kpi_value
def log_to_ce(log):
kpi_tracker = {}
for kpi in tracking_kpis:
kpi_tracker[kpi.name] = kpi
for (kpi_name, kpi_value) in parse_log(log):
print(kpi_name, kpi_value)
kpi_tracker[kpi_name].add_record(kpi_value)
kpi_tracker[kpi_name].persist()
if __name__ == '__main__':
log = sys.stdin.read()
# print("*****")
# print(log)
# print("****")
log_to_ce(log)
PaddleCV/gan/cycle_gan/data_reader.py 已删除 100644 → 0
浏览文件 @ 90fcad46
# Copyright (c) 2019 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
from __future__ import division
from __future__ import print_function
import os
from PIL import Image
import numpy as np
A_LIST_FILE = "./data/horse2zebra/trainA.txt"
B_LIST_FILE = "./data/horse2zebra/trainB.txt"
A_TEST_LIST_FILE = "./data/horse2zebra/testA.txt"
B_TEST_LIST_FILE = "./data/horse2zebra/testB.txt"
IMAGES_ROOT = "./data/horse2zebra/"
def image_shape():
return [3, 256, 256]
def max_images_num():
return 1335
def reader_creater(list_file, cycle=True, shuffle=True, return_name=False):
images = [IMAGES_ROOT + line for line in open(list_file, 'r').readlines()]
def reader():
while True:
if shuffle:
np.random.shuffle(images)
for file in images:
file = file.strip("\n\r\t ")
image = Image.open(file)
image = image.resize((256, 256))
image = np.array(image) / 127.5 - 1
if len(image.shape) != 3:
continue
image = image[:, :, 0:3].astype("float32")
image = image.transpose([2, 0, 1])
if return_name:
yield image[np.newaxis, :], os.path.basename(file)
else:
yield image
if not cycle:
break
return reader
def a_reader(shuffle=True):
"""
Reader of images with A style for training.
"""
return reader_creater(A_LIST_FILE, shuffle=shuffle)
def b_reader(shuffle=True):
"""
Reader of images with B style for training.
"""
return reader_creater(B_LIST_FILE, shuffle=shuffle)
def a_test_reader():
"""
Reader of images with A style for test.
"""
return reader_creater(A_TEST_LIST_FILE, cycle=False, return_name=True)
def b_test_reader():
"""
Reader of images with B style for test.
"""
return reader_creater(B_TEST_LIST_FILE, cycle=False, return_name=True)
PaddleCV/gan/cycle_gan/infer.py 已删除 100644 → 0
浏览文件 @ 90fcad46
# Copyright (c) 2019 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.
import argparse
import functools
import os
from PIL import Image
import paddle.fluid as fluid
import paddle
import numpy as np
from scipy.misc import imsave
from model import build_generator_resnet_9blocks, build_gen_discriminator
import glob
from utility import add_arguments, print_arguments
parser = argparse.ArgumentParser(description=__doc__)
add_arg = functools.partial(add_arguments, argparser=parser)
# yapf: disable
add_arg('input', str, None, "The images to be infered.")
add_arg('output', str, "./infer_result", "The directory the infer result to be saved to.")
add_arg('init_model', str, None, "The init model file of directory.")
add_arg('input_style', str, "A", "The style of the input, A or B")
add_arg('use_gpu', bool, True, "Whether to use GPU to train.")
# yapf: enable
def infer(args):
data_shape = [-1, 3, 256, 256]
input = fluid.layers.data(name='input', shape=data_shape, dtype='float32')
if args.input_style == "A":
model_name = 'g_a'
fake = build_generator_resnet_9blocks(input, name="g_A")
elif args.input_style == "B":
model_name = 'g_b'
fake = build_generator_resnet_9blocks(input, name="g_B")
else:
raise "Input with style [%s] is not supported." % args.input_style
# prepare environment
place = fluid.CPUPlace()
if args.use_gpu:
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fluid.io.load_persistables(exe, args.init_model + "/" + model_name)
if not os.path.exists(args.output):
os.makedirs(args.output)
for file in glob.glob(args.input):
image_name = os.path.basename(file)
image = Image.open(file)
image = image.resize((256, 256))
image = np.array(image) / 127.5 - 1
if len(image.shape) != 3:
continue
data = image.transpose([2, 0, 1])[np.newaxis, :].astype("float32")
tensor = fluid.LoDTensor()
tensor.set(data, place)
fake_temp = exe.run(fetch_list=[fake.name], feed={"input": tensor})
fake_temp = np.squeeze(fake_temp[0]).transpose([1, 2, 0])
input_temp = np.squeeze(data).transpose([1, 2, 0])
imsave(args.output + "/fake_" + image_name, (
(fake_temp + 1) * 127.5).astype(np.uint8))
if __name__ == "__main__":
args = parser.parse_args()
print_arguments(args)
infer(args)
PaddleCV/gan/cycle_gan/layers.py 已删除 100644 → 0
浏览文件 @ 90fcad46
# Copyright (c) 2019 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 division
import paddle.fluid as fluid
import numpy as np
import os
# cudnn is not better when batch size is 1.
use_cudnn_conv2d_transpose = False
use_cudnn_conv2d = True
use_layer_norm = True
def cal_padding(img_size, stride, filter_size, dilation=1):
"""Calculate padding size."""
valid_filter_size = dilation * (filter_size - 1) + 1
if img_size % stride == 0:
out_size = max(filter_size - stride, 0)
else:
out_size = max(filter_size - (img_size % stride), 0)
return out_size // 2, out_size - out_size // 2
def instance_norm(input, name=None):
# TODO(lvmengsi@baidu.com): Check the accuracy when using fluid.layers.layer_norm.
if use_layer_norm:
return fluid.layers.layer_norm(input, begin_norm_axis=2)
helper = fluid.layer_helper.LayerHelper("instance_norm", **locals())
dtype = helper.input_dtype()
epsilon = 1e-5
mean = fluid.layers.reduce_mean(input, dim=[2, 3], keep_dim=True)
var = fluid.layers.reduce_mean(
fluid.layers.square(input - mean), dim=[2, 3], keep_dim=True)
if name is not None:
scale_name = name + "_scale"
offset_name = name + "_offset"
scale_param = fluid.ParamAttr(
name=scale_name,
initializer=fluid.initializer.TruncatedNormal(1.0, 0.02),
trainable=True)
offset_param = fluid.ParamAttr(
name=offset_name,
initializer=fluid.initializer.Constant(0.0),
trainable=True)
scale = helper.create_parameter(
attr=scale_param, shape=input.shape[1:2], dtype=dtype)
offset = helper.create_parameter(
attr=offset_param, shape=input.shape[1:2], dtype=dtype)
tmp = fluid.layers.elementwise_mul(x=(input - mean), y=scale, axis=1)
tmp = tmp / fluid.layers.sqrt(var + epsilon)
tmp = fluid.layers.elementwise_add(tmp, offset, axis=1)
return tmp
def conv2d(input,
num_filters=64,
filter_size=7,
stride=1,
stddev=0.02,
padding="VALID",
name="conv2d",
norm=True,
relu=True,
relufactor=0.0):
"""Wrapper for conv2d op to support VALID and SAME padding mode."""
need_crop = False
if padding == "SAME":
top_padding, bottom_padding = cal_padding(input.shape[2], stride,
filter_size)
left_padding, right_padding = cal_padding(input.shape[2], stride,
filter_size)
height_padding = bottom_padding
width_padding = right_padding
if top_padding != bottom_padding or left_padding != right_padding:
height_padding = top_padding + stride
width_padding = left_padding + stride
need_crop = True
else:
height_padding = 0
width_padding = 0
padding = [height_padding, width_padding]
param_attr = fluid.ParamAttr(
name=name + "_w",
initializer=fluid.initializer.TruncatedNormal(scale=stddev))
bias_attr = fluid.ParamAttr(
name=name + "_b", initializer=fluid.initializer.Constant(0.0))
conv = fluid.layers.conv2d(
input,
num_filters,
filter_size,
name=name,
stride=stride,
padding=padding,
use_cudnn=use_cudnn_conv2d,
param_attr=param_attr,
bias_attr=bias_attr)
if need_crop:
conv = fluid.layers.crop(
conv,
shape=(-1, conv.shape[1], conv.shape[2] - 1, conv.shape[3] - 1),
offsets=(0, 0, 1, 1))
if norm:
conv = instance_norm(input=conv, name=name + "_norm")
if relu:
conv = fluid.layers.leaky_relu(conv, alpha=relufactor)
return conv
def deconv2d(input,
out_shape,
num_filters=64,
filter_size=7,
stride=1,
stddev=0.02,
padding="VALID",
name="conv2d",
norm=True,
relu=True,
relufactor=0.0):
"""Wrapper for deconv2d op to support VALID and SAME padding mode."""
need_crop = False
if padding == "SAME":
top_padding, bottom_padding = cal_padding(out_shape[0], stride,
filter_size)
left_padding, right_padding = cal_padding(out_shape[1], stride,
filter_size)
height_padding = top_padding
width_padding = left_padding
if top_padding != bottom_padding or left_padding != right_padding:
need_crop = True
else:
height_padding = 0
width_padding = 0
padding = [height_padding, width_padding]
param_attr = fluid.ParamAttr(
name=name + "_w",
initializer=fluid.initializer.TruncatedNormal(scale=stddev))
bias_attr = fluid.ParamAttr(
name=name + "_b", initializer=fluid.initializer.Constant(0.0))
conv = fluid.layers.conv2d_transpose(
input,
num_filters,
name=name,
filter_size=filter_size,
stride=stride,
padding=padding,
use_cudnn=use_cudnn_conv2d_transpose,
param_attr=param_attr,
bias_attr=bias_attr)
if need_crop:
conv = fluid.layers.crop(
conv,
shape=(-1, conv.shape[1], conv.shape[2] - 1, conv.shape[3] - 1),
offsets=(0, 0, 0, 0))
if norm:
conv = instance_norm(input=conv, name=name + "_norm")
if relu:
conv = fluid.layers.leaky_relu(conv, alpha=relufactor)
return conv
PaddleCV/gan/cycle_gan/model.py 已删除 100644 → 0
浏览文件 @ 90fcad46
# Copyright (c) 2019 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 layers import conv2d, deconv2d
import paddle.fluid as fluid
def build_resnet_block(inputres, dim, name="resnet"):
out_res = fluid.layers.pad2d(inputres, [1, 1, 1, 1], mode="reflect")
out_res = conv2d(out_res, dim, 3, 1, 0.02, "VALID", name + "_c1")
out_res = fluid.layers.pad2d(out_res, [1, 1, 1, 1], mode="reflect")
out_res = conv2d(
out_res, dim, 3, 1, 0.02, "VALID", name + "_c2", relu=False)
return fluid.layers.relu(out_res + inputres)
def build_generator_resnet_9blocks(inputgen, name="generator"):
'''The shape of input should be equal to the shape of output.'''
pad_input = fluid.layers.pad2d(inputgen, [3, 3, 3, 3], mode="reflect")
o_c1 = conv2d(pad_input, 32, 7, 1, 0.02, name=name + "_c1")
o_c2 = conv2d(o_c1, 64, 3, 2, 0.02, "SAME", name + "_c2")
o_c3 = conv2d(o_c2, 128, 3, 2, 0.02, "SAME", name + "_c3")
o_r1 = build_resnet_block(o_c3, 128, name + "_r1")
o_r2 = build_resnet_block(o_r1, 128, name + "_r2")
o_r3 = build_resnet_block(o_r2, 128, name + "_r3")
o_r4 = build_resnet_block(o_r3, 128, name + "_r4")
o_r5 = build_resnet_block(o_r4, 128, name + "_r5")
o_r6 = build_resnet_block(o_r5, 128, name + "_r6")
o_r7 = build_resnet_block(o_r6, 128, name + "_r7")
o_r8 = build_resnet_block(o_r7, 128, name + "_r8")
o_r9 = build_resnet_block(o_r8, 128, name + "_r9")
o_c4 = deconv2d(o_r9, [128, 128], 64, 3, 2, 0.02, "SAME", name + "_c4")
o_c5 = deconv2d(o_c4, [256, 256], 32, 3, 2, 0.02, "SAME", name + "_c5")
o_c6 = conv2d(o_c5, 3, 7, 1, 0.02, "SAME", name + "_c6", relu=False)
out_gen = fluid.layers.tanh(o_c6, name + "_t1")
return out_gen
def build_gen_discriminator(inputdisc, name="discriminator"):
o_c1 = conv2d(
inputdisc,
64,
4,
2,
0.02,
"SAME",
name + "_c1",
norm=False,
relufactor=0.2)
o_c2 = conv2d(o_c1, 128, 4, 2, 0.02, "SAME", name + "_c2", relufactor=0.2)
o_c3 = conv2d(o_c2, 256, 4, 2, 0.02, "SAME", name + "_c3", relufactor=0.2)
o_c4 = conv2d(o_c3, 512, 4, 1, 0.02, "SAME", name + "_c4", relufactor=0.2)
o_c5 = conv2d(
o_c4, 1, 4, 1, 0.02, "SAME", name + "_c5", norm=False, relu=False)
return o_c5
PaddleCV/gan/cycle_gan/train.py 已删除 100644 → 0
浏览文件 @ 90fcad46
# Copyright (c) 2019 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
from __future__ import division
from __future__ import print_function
import os
def set_paddle_flags(flags):
for key, value in flags.items():
if os.environ.get(key, None) is None:
os.environ[key] = str(value)
use_cudnn_deterministic = os.environ.get('FLAGS_cudnn_deterministic', None)
if use_cudnn_deterministic:
use_cudnn_exhaustive_search = 0
else:
use_cudnn_exhaustive_search = 1
# NOTE(paddle-dev): All of these flags should be
# set before `import paddle`. Otherwise, it would
# not take any effect.
set_paddle_flags({
'FLAGS_cudnn_exhaustive_search': use_cudnn_exhaustive_search,
'FLAGS_conv_workspace_size_limit': 256,
'FLAGS_eager_delete_tensor_gb': 0, # enable gc
# You can omit the following settings, because the default
# value of FLAGS_memory_fraction_of_eager_deletion is 1,
# and default value of FLAGS_fast_eager_deletion_mode is 1
'FLAGS_memory_fraction_of_eager_deletion': 1,
'FLAGS_fast_eager_deletion_mode': 1
})
import random
import sys
import paddle
import argparse
import functools
import time
import numpy as np
from scipy.misc import imsave
import paddle.fluid as fluid
import paddle.fluid.profiler as profiler
import data_reader
from utility import add_arguments, print_arguments, ImagePool
from trainer import GATrainer, GBTrainer, DATrainer, DBTrainer
parser = argparse.ArgumentParser(description=__doc__)
add_arg = functools.partial(add_arguments, argparser=parser)
# yapf: disable
add_arg('batch_size', int, 1, "Minibatch size.")
add_arg('epoch', int, 2, "The number of epoched to be trained.")
add_arg('output', str, "./output", "The directory the model and the test result to be saved to.")
add_arg('init_model', str, None, "The init model file of directory.")
add_arg('save_checkpoints', bool, True, "Whether to save checkpoints.")
add_arg('run_test', bool, True, "Whether to run test.")
add_arg('use_gpu', bool, True, "Whether to use GPU to train.")
add_arg('profile', bool, False, "Whether to profile.")
# NOTE: args for profiler, used for benchmark
add_arg('profiler_path', str, './profiler_cyclegan', "the path of profiler output files. used for benchmark")
add_arg('max_iter', int, 0, "the max batch nums to train. used for benchmark")
add_arg('run_ce', bool, False, "Whether to run for model ce.")
# yapf: enable
def train(args):
max_images_num = data_reader.max_images_num()
shuffle = True
if args.run_ce:
np.random.seed(10)
fluid.default_startup_program().random_seed = 90
max_images_num = 1
shuffle = False
data_shape = [-1] + data_reader.image_shape()
input_A = fluid.layers.data(
name='input_A', shape=data_shape, dtype='float32')
input_B = fluid.layers.data(
name='input_B', shape=data_shape, dtype='float32')
fake_pool_A = fluid.layers.data(
name='fake_pool_A', shape=data_shape, dtype='float32')
fake_pool_B = fluid.layers.data(
name='fake_pool_B', shape=data_shape, dtype='float32')
g_A_trainer = GATrainer(input_A, input_B)
g_B_trainer = GBTrainer(input_A, input_B)
d_A_trainer = DATrainer(input_A, fake_pool_A)
d_B_trainer = DBTrainer(input_B, fake_pool_B)
# prepare environment
place = fluid.CPUPlace()
if args.use_gpu:
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
A_pool = ImagePool()
B_pool = ImagePool()
A_reader = paddle.batch(
data_reader.a_reader(shuffle=shuffle), args.batch_size)()
B_reader = paddle.batch(
data_reader.b_reader(shuffle=shuffle), args.batch_size)()
if not args.run_ce:
A_test_reader = data_reader.a_test_reader()
B_test_reader = data_reader.b_test_reader()
def test(epoch):
out_path = args.output + "/test"
if not os.path.exists(out_path):
os.makedirs(out_path)
i = 0
for data_A, data_B in zip(A_test_reader(), B_test_reader()):
A_name = data_A[1]
B_name = data_B[1]
tensor_A = fluid.LoDTensor()
tensor_B = fluid.LoDTensor()
tensor_A.set(data_A[0], place)
tensor_B.set(data_B[0], place)
fake_A_temp, fake_B_temp, cyc_A_temp, cyc_B_temp = exe.run(
g_A_trainer.infer_program,
fetch_list=[
g_A_trainer.fake_A, g_A_trainer.fake_B, g_A_trainer.cyc_A,
g_A_trainer.cyc_B
],
feed={"input_A": tensor_A,
"input_B": tensor_B})
fake_A_temp = np.squeeze(fake_A_temp[0]).transpose([1, 2, 0])
fake_B_temp = np.squeeze(fake_B_temp[0]).transpose([1, 2, 0])
cyc_A_temp = np.squeeze(cyc_A_temp[0]).transpose([1, 2, 0])
cyc_B_temp = np.squeeze(cyc_B_temp[0]).transpose([1, 2, 0])
input_A_temp = np.squeeze(data_A[0]).transpose([1, 2, 0])
input_B_temp = np.squeeze(data_B[0]).transpose([1, 2, 0])
imsave(out_path + "/fakeB_" + str(epoch) + "_" + A_name, (
(fake_B_temp + 1) * 127.5).astype(np.uint8))
imsave(out_path + "/fakeA_" + str(epoch) + "_" + B_name, (
(fake_A_temp + 1) * 127.5).astype(np.uint8))
imsave(out_path + "/cycA_" + str(epoch) + "_" + A_name, (
(cyc_A_temp + 1) * 127.5).astype(np.uint8))
imsave(out_path + "/cycB_" + str(epoch) + "_" + B_name, (
(cyc_B_temp + 1) * 127.5).astype(np.uint8))
imsave(out_path + "/inputA_" + str(epoch) + "_" + A_name, (
(input_A_temp + 1) * 127.5).astype(np.uint8))
imsave(out_path + "/inputB_" + str(epoch) + "_" + B_name, (
(input_B_temp + 1) * 127.5).astype(np.uint8))
i += 1
def checkpoints(epoch):
out_path = args.output + "/checkpoints/" + str(epoch)
if not os.path.exists(out_path):
os.makedirs(out_path)
fluid.io.save_persistables(
exe, out_path + "/g_a", main_program=g_A_trainer.program)
fluid.io.save_persistables(
exe, out_path + "/g_b", main_program=g_B_trainer.program)
fluid.io.save_persistables(
exe, out_path + "/d_a", main_program=d_A_trainer.program)
fluid.io.save_persistables(
exe, out_path + "/d_b", main_program=d_B_trainer.program)
print("saved checkpoint to {}".format(out_path))
sys.stdout.flush()
def init_model():
assert os.path.exists(
args.init_model), "[%s] cann't be found." % args.init_mode
fluid.io.load_persistables(
exe, args.init_model + "/g_a", main_program=g_A_trainer.program)
fluid.io.load_persistables(
exe, args.init_model + "/g_b", main_program=g_B_trainer.program)
fluid.io.load_persistables(
exe, args.init_model + "/d_a", main_program=d_A_trainer.program)
fluid.io.load_persistables(
exe, args.init_model + "/d_b", main_program=d_B_trainer.program)
print("Load model from {}".format(args.init_model))
if args.init_model:
init_model()
losses = [[], []]
t_time = 0
build_strategy = fluid.BuildStrategy()
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_threads = 1
exec_strategy.use_experimental_executor = True
g_A_trainer_program = fluid.CompiledProgram(
g_A_trainer.program).with_data_parallel(
loss_name=g_A_trainer.g_loss_A.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
g_B_trainer_program = fluid.CompiledProgram(
g_B_trainer.program).with_data_parallel(
loss_name=g_B_trainer.g_loss_B.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
d_B_trainer_program = fluid.CompiledProgram(
d_B_trainer.program).with_data_parallel(
loss_name=d_B_trainer.d_loss_B.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
d_A_trainer_program = fluid.CompiledProgram(
d_A_trainer.program).with_data_parallel(
loss_name=d_A_trainer.d_loss_A.name,
build_strategy=build_strategy,
exec_strategy=exec_strategy)
total_batch_num = 0 # this is for benchmark
for epoch in range(args.epoch):
batch_id = 0
for i in range(max_images_num):
if args.max_iter and total_batch_num == args.max_iter: # this for benchmark
return
data_A = next(A_reader)
data_B = next(B_reader)
tensor_A = fluid.LoDTensor()
tensor_B = fluid.LoDTensor()
tensor_A.set(data_A, place)
tensor_B.set(data_B, place)
s_time = time.time()
# optimize the g_A network
g_A_loss, fake_B_tmp = exe.run(
g_A_trainer_program,
fetch_list=[g_A_trainer.g_loss_A, g_A_trainer.fake_B],
feed={"input_A": tensor_A,
"input_B": tensor_B})
fake_pool_B = B_pool.pool_image(fake_B_tmp)
# optimize the d_B network
d_B_loss = exe.run(
d_B_trainer_program,
fetch_list=[d_B_trainer.d_loss_B],
feed={"input_B": tensor_B,
"fake_pool_B": fake_pool_B})[0]
# optimize the g_B network
g_B_loss, fake_A_tmp = exe.run(
g_B_trainer_program,
fetch_list=[g_B_trainer.g_loss_B, g_B_trainer.fake_A],
feed={"input_A": tensor_A,
"input_B": tensor_B})
fake_pool_A = A_pool.pool_image(fake_A_tmp)
# optimize the d_A network
d_A_loss = exe.run(
d_A_trainer_program,
fetch_list=[d_A_trainer.d_loss_A],
feed={"input_A": tensor_A,
"fake_pool_A": fake_pool_A})[0]
batch_time = time.time() - s_time
t_time += batch_time
print(
"epoch{}; batch{}; g_A_loss: {}; d_B_loss: {}; g_B_loss: {}; d_A_loss: {}; "
"Batch_time_cost: {}".format(epoch, batch_id, g_A_loss[
0], d_B_loss[0], g_B_loss[0], d_A_loss[0], batch_time))
losses[0].append(g_A_loss[0])
losses[1].append(d_A_loss[0])
sys.stdout.flush()
batch_id += 1
total_batch_num = total_batch_num + 1 # this is for benchmark
# profiler tools for benchmark
if args.profile and epoch == 0 and batch_id == 10:
profiler.reset_profiler()
elif args.profile and epoch == 0 and batch_id == 15:
return
if args.run_test and not args.run_ce:
test(epoch)
if args.save_checkpoints and not args.run_ce:
checkpoints(epoch)
if args.run_ce:
print("kpis,g_train_cost,{}".format(np.mean(losses[0])))
print("kpis,d_train_cost,{}".format(np.mean(losses[1])))
print("kpis,duration,{}".format(t_time / args.epoch))
if __name__ == "__main__":
args = parser.parse_args()
print_arguments(args)
if args.profile:
if args.use_gpu:
with profiler.profiler('All', 'total', args.profiler_path) as prof:
train(args)
else:
with profiler.profiler("CPU", sorted_key='total') as cpuprof:
train(args)
else:
train(args)
PaddleCV/gan/cycle_gan/trainer.py 已删除 100644 → 0
浏览文件 @ 90fcad46
# Copyright (c) 2019 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
from __future__ import division
from __future__ import print_function
from model import build_generator_resnet_9blocks, build_gen_discriminator
import paddle.fluid as fluid
step_per_epoch = 1335
cycle_loss_factor = 10.0
class GATrainer():
def __init__(self, input_A, input_B):
self.program = fluid.default_main_program().clone()
with fluid.program_guard(self.program):
self.fake_B = build_generator_resnet_9blocks(input_A, name="g_A")
self.fake_A = build_generator_resnet_9blocks(input_B, name="g_B")
self.cyc_A = build_generator_resnet_9blocks(self.fake_B, "g_B")
self.cyc_B = build_generator_resnet_9blocks(self.fake_A, "g_A")
self.infer_program = self.program.clone()
diff_A = fluid.layers.abs(
fluid.layers.elementwise_sub(
x=input_A, y=self.cyc_A))
diff_B = fluid.layers.abs(
fluid.layers.elementwise_sub(
x=input_B, y=self.cyc_B))
self.cyc_loss = (
fluid.layers.reduce_mean(diff_A) +
fluid.layers.reduce_mean(diff_B)) * cycle_loss_factor
self.fake_rec_B = build_gen_discriminator(self.fake_B, "d_B")
self.disc_loss_B = fluid.layers.reduce_mean(
fluid.layers.square(self.fake_rec_B - 1))
self.g_loss_A = fluid.layers.elementwise_add(self.cyc_loss,
self.disc_loss_B)
vars = []
for var in self.program.list_vars():
if fluid.io.is_parameter(var) and var.name.startswith("g_A"):
vars.append(var.name)
self.param = vars
lr = 0.0002
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.piecewise_decay(
boundaries=[
100 * step_per_epoch, 120 * step_per_epoch,
140 * step_per_epoch, 160 * step_per_epoch,
180 * step_per_epoch
],
values=[
lr, lr * 0.8, lr * 0.6, lr * 0.4, lr * 0.2, lr * 0.1
]),
beta1=0.5,
name="g_A")
optimizer.minimize(self.g_loss_A, parameter_list=vars)
class GBTrainer():
def __init__(self, input_A, input_B):
self.program = fluid.default_main_program().clone()
with fluid.program_guard(self.program):
self.fake_B = build_generator_resnet_9blocks(input_A, name="g_A")
self.fake_A = build_generator_resnet_9blocks(input_B, name="g_B")
self.cyc_A = build_generator_resnet_9blocks(self.fake_B, "g_B")
self.cyc_B = build_generator_resnet_9blocks(self.fake_A, "g_A")
self.infer_program = self.program.clone()
diff_A = fluid.layers.abs(
fluid.layers.elementwise_sub(
x=input_A, y=self.cyc_A))
diff_B = fluid.layers.abs(
fluid.layers.elementwise_sub(
x=input_B, y=self.cyc_B))
self.cyc_loss = (
fluid.layers.reduce_mean(diff_A) +
fluid.layers.reduce_mean(diff_B)) * cycle_loss_factor
self.fake_rec_A = build_gen_discriminator(self.fake_A, "d_A")
disc_loss_A = fluid.layers.reduce_mean(
fluid.layers.square(self.fake_rec_A - 1))
self.g_loss_B = fluid.layers.elementwise_add(self.cyc_loss,
disc_loss_A)
vars = []
for var in self.program.list_vars():
if fluid.io.is_parameter(var) and var.name.startswith("g_B"):
vars.append(var.name)
self.param = vars
lr = 0.0002
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.piecewise_decay(
boundaries=[
100 * step_per_epoch, 120 * step_per_epoch,
140 * step_per_epoch, 160 * step_per_epoch,
180 * step_per_epoch
],
values=[
lr, lr * 0.8, lr * 0.6, lr * 0.4, lr * 0.2, lr * 0.1
]),
beta1=0.5,
name="g_B")
optimizer.minimize(self.g_loss_B, parameter_list=vars)
class DATrainer():
def __init__(self, input_A, fake_pool_A):
self.program = fluid.default_main_program().clone()
with fluid.program_guard(self.program):
self.rec_A = build_gen_discriminator(input_A, "d_A")
self.fake_pool_rec_A = build_gen_discriminator(fake_pool_A, "d_A")
self.d_loss_A = (fluid.layers.square(self.fake_pool_rec_A) +
fluid.layers.square(self.rec_A - 1)) / 2.0
self.d_loss_A = fluid.layers.reduce_mean(self.d_loss_A)
optimizer = fluid.optimizer.Adam(learning_rate=0.0002, beta1=0.5)
optimizer._name = "d_A"
vars = []
for var in self.program.list_vars():
if fluid.io.is_parameter(var) and var.name.startswith("d_A"):
vars.append(var.name)
self.param = vars
lr = 0.0002
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.piecewise_decay(
boundaries=[
100 * step_per_epoch, 120 * step_per_epoch,
140 * step_per_epoch, 160 * step_per_epoch,
180 * step_per_epoch
],
values=[
lr, lr * 0.8, lr * 0.6, lr * 0.4, lr * 0.2, lr * 0.1
]),
beta1=0.5,
name="d_A")
optimizer.minimize(self.d_loss_A, parameter_list=vars)
class DBTrainer():
def __init__(self, input_B, fake_pool_B):
self.program = fluid.default_main_program().clone()
with fluid.program_guard(self.program):
self.rec_B = build_gen_discriminator(input_B, "d_B")
self.fake_pool_rec_B = build_gen_discriminator(fake_pool_B, "d_B")
self.d_loss_B = (fluid.layers.square(self.fake_pool_rec_B) +
fluid.layers.square(self.rec_B - 1)) / 2.0
self.d_loss_B = fluid.layers.reduce_mean(self.d_loss_B)
optimizer = fluid.optimizer.Adam(learning_rate=0.0002, beta1=0.5)
vars = []
for var in self.program.list_vars():
if fluid.io.is_parameter(var) and var.name.startswith("d_B"):
vars.append(var.name)
self.param = vars
lr = 0.0002
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.piecewise_decay(
boundaries=[
100 * step_per_epoch, 120 * step_per_epoch,
140 * step_per_epoch, 160 * step_per_epoch,
180 * step_per_epoch
],
values=[
lr, lr * 0.8, lr * 0.6, lr * 0.4, lr * 0.2, lr * 0.1
]),
beta1=0.5,
name="d_B")
optimizer.minimize(self.d_loss_B, parameter_list=vars)
PaddleCV/gan/cycle_gan/utility.py 已删除 100644 → 0
浏览文件 @ 90fcad46
"""Contains common utility functions."""
# 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
import distutils.util
import six
import random
import glob
import numpy as np
def print_arguments(args):
"""Print argparse's arguments.
Usage:
.. code-block:: python
parser = argparse.ArgumentParser()
parser.add_argument("name", default="Jonh", type=str, help="User name.")
args = parser.parse_args()
print_arguments(args)
:param args: Input argparse.Namespace for printing.
:type args: argparse.Namespace
"""
print("----------- Configuration Arguments -----------")
for arg, value in sorted(six.iteritems(vars(args))):
print("%s: %s" % (arg, value))
print("------------------------------------------------")
def add_arguments(argname, type, default, help, argparser, **kwargs):
"""Add argparse's argument.
Usage:
.. code-block:: python
parser = argparse.ArgumentParser()
add_argument("name", str, "Jonh", "User name.", parser)
args = parser.parse_args()
"""
type = distutils.util.strtobool if type == bool else type
argparser.add_argument(
"--" + argname,
default=default,
type=type,
help=help + ' Default: %(default)s.',
**kwargs)
class ImagePool(object):
def __init__(self, pool_size=50):
self.pool = []
self.count = 0
self.pool_size = pool_size
def pool_image(self, image):
if self.count < self.pool_size:
self.pool.append(image)
self.count += 1
return image
else:
p = random.random()
if p > 0.5:
random_id = random.randint(0, self.pool_size - 1)
temp = self.pool[random_id]
self.pool[random_id] = image
return temp
else:
return image
PaddleCV/gan/infer.py
浏览文件 @ 6b882d42
...@@ -204,6 +204,7 @@ def infer(args): ...@@ -204,6 +204,7 @@ def infer(args):
if not os.path.exists(args.output): if not os.path.exists(args.output):
os.makedirs(args.output) os.makedirs(args.output)
attr_names = args.selected_attrs.split(',') attr_names = args.selected_attrs.split(',')
if args.model_net == 'AttGAN' or args.model_net == 'STGAN': if args.model_net == 'AttGAN' or args.model_net == 'STGAN':
...@@ -311,6 +312,9 @@ def infer(args): ...@@ -311,6 +312,9 @@ def infer(args):
fake_temp = save_batch_image(fake_temp[0]) fake_temp = save_batch_image(fake_temp[0])
input_temp = save_batch_image(np.array(real_img)) input_temp = save_batch_image(np.array(real_img))
if len(image_names) == 1:
fake_temp = np.expand_dims(fake_temp, axis=0)
input_temp = np.expand_dims(input_temp, axis=0)
for i, name in enumerate(image_names): for i, name in enumerate(image_names):
fake_image = Image.fromarray(((fake_temp[i] + 1) * 127.5).astype(np.uint8)) fake_image = Image.fromarray(((fake_temp[i] + 1) * 127.5).astype(np.uint8))
fake_image.save(os.path.join(args.output, "fake_" + name)) fake_image.save(os.path.join(args.output, "fake_" + name))
......
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