diff --git a/PaddleCV/PaddleDetection/slim/distillation/README.md b/PaddleCV/PaddleDetection/slim/distillation/README.md new file mode 100755 index 0000000000000000000000000000000000000000..de6a1eef06405626d2e570c9f964e6334b76dd20 --- /dev/null +++ b/PaddleCV/PaddleDetection/slim/distillation/README.md @@ -0,0 +1,119 @@ +>运行该示例前请安装Paddle1.6或更高版本 + +# 检测模型蒸馏示例 + +## 概述 + +该示例使用PaddleSlim提供的[蒸馏策略](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/tutorial.md#3-蒸馏)对检测库中的模型进行蒸馏训练。 +在阅读该示例前，建议您先了解以下内容： + +- [检测库的常规训练方法](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/PaddleDetection) +- [PaddleSlim使用文档](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md) + + +## 配置文件说明 + +关于配置文件如何编写您可以参考： + +- [PaddleSlim配置文件编写说明](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md#122-%E9%85%8D%E7%BD%AE%E6%96%87%E4%BB%B6%E7%9A%84%E4%BD%BF%E7%94%A8) +- [蒸馏策略配置文件编写说明](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md#23-蒸馏) + +这里以ResNet34-YoloV3蒸馏MobileNetV1-YoloV3模型为例，首先，为了对`student model`和`teacher model`有个总体的认识，从而进一步确认蒸馏的对象，我们通过以下命令分别观察两个网络变量（Variable）的名称和形状： + +```python +# 观察student model的Variable +for v in fluid.default_main_program().list_vars(): + if "py_reader" not in v.name and "double_buffer" not in v.name and "generated_var" not in v.name: + print(v.name, v.shape) +# 观察teacher model的Variable +for v in teacher_program.list_vars(): + print(v.name, v.shape) +``` + +经过对比可以发现，`student model`和`teacher model`的部分中间结果分别为： + +```bash +# student model +conv2d_15.tmp_0 +# teacher model +teacher_teacher_conv2d_1.tmp_0 +``` + + +所以，我们用`l2_distiller`对这两个特征图做蒸馏。在配置文件中进行如下配置： + +```yaml +distillers: + l2_distiller: + class: 'L2Distiller' + teacher_feature_map: 'teacher_teacher_conv2d_1.tmp_0' + student_feature_map: 'conv2d_15.tmp_0' + distillation_loss_weight: 1 +strategies: + distillation_strategy: + class: 'DistillationStrategy' + distillers: ['l2_distiller'] + start_epoch: 0 + end_epoch: 270 +``` + +我们也可以根据上述操作为蒸馏策略选择其他loss，PaddleSlim支持的有`FSP_loss`, `L2_loss`和`softmax_with_cross_entropy_loss` 。 + +## 训练 + +根据[PaddleDetection/tools/train.py](https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/PaddleDetection/tools/train.py)编写压缩脚本compress.py。 +在该脚本中定义了Compressor对象，用于执行压缩任务。 + + + + +您可以通过运行脚本`run.sh`运行该示例。 + + +### 保存断点（checkpoint） + +如果在配置文件中设置了`checkpoint_path`, 则在蒸馏任务执行过程中会自动保存断点，当任务异常中断时， +重启任务会自动从`checkpoint_path`路径下按数字顺序加载最新的checkpoint文件。如果不想让重启的任务从断点恢复， +需要修改配置文件中的`checkpoint_path`，或者将`checkpoint_path`路径下文件清空。 + +>注意：配置文件中的信息不会保存在断点中，重启前对配置文件的修改将会生效。 + + +## 评估 + +如果在配置文件中设置了`checkpoint_path`，则每个epoch会保存一个压缩后的用于评估的模型， +该模型会保存在`${checkpoint_path}/${epoch_id}/eval_model/`路径下，包含`__model__`和`__params__`两个文件。 +其中，`__model__`用于保存模型结构信息，`__params__`用于保存参数（parameters）信息。 + +如果不需要保存评估模型，可以在定义Compressor对象时，将`save_eval_model`选项设置为False（默认为True）。 + +## 预测 + +如果在配置文件中设置了`checkpoint_path`，并且在定义Compressor对象时指定了`prune_infer_model`选项，则每个epoch都会 +保存一个`inference model`。该模型是通过删除eval_program中多余的operators而得到的。 + +该模型会保存在`${checkpoint_path}/${epoch_id}/eval_model/`路径下，包含`__model__.infer`和`__params__`两个文件。 +其中，`__model__.infer`用于保存模型结构信息，`__params__`用于保存参数（parameters）信息。 + +更多关于`prune_infer_model`选项的介绍，请参考：[Compressor介绍](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md#121-%E5%A6%82%E4%BD%95%E6%94%B9%E5%86%99%E6%99%AE%E9%80%9A%E8%AE%AD%E7%BB%83%E8%84%9A%E6%9C%AC) + +### python预测 + +在脚本slim/infer.py中展示了如何使用fluid python API加载使用预测模型进行预测。 + +### PaddleLite + +该示例中产出的预测（inference）模型可以直接用PaddleLite进行加载使用。 +关于PaddleLite如何使用，请参考：[PaddleLite使用文档](https://github.com/PaddlePaddle/Paddle-Lite/wiki#%E4%BD%BF%E7%94%A8) + +## 示例结果 + +### MobileNetV1-YOLO-V3 + +| FLOPS |Box AP| +|---|---| +|baseline|76.2 | +|蒸馏后|- | + + +## FAQ diff --git a/PaddleCV/PaddleDetection/slim/distillation/compress.py b/PaddleCV/PaddleDetection/slim/distillation/compress.py new file mode 100644 index 0000000000000000000000000000000000000000..f1e0189ab4f6285587d4677d0b73e62fbd89619c --- /dev/null +++ b/PaddleCV/PaddleDetection/slim/distillation/compress.py @@ -0,0 +1,325 @@ +# 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 +import time +import multiprocessing +import numpy as np +from collections import deque, OrderedDict +from paddle.fluid.contrib.slim.core import Compressor +from paddle.fluid.framework import IrGraph + + +def set_paddle_flags(**kwargs): + for key, value in kwargs.items(): + if os.environ.get(key, None) is None: + os.environ[key] = str(value) + + +# NOTE(paddle-dev): All of these flags should be set before +# `import paddle`. Otherwise, it would not take any effect. +set_paddle_flags( + FLAGS_eager_delete_tensor_gb=0, # enable GC to save memory +) + +from paddle import fluid + +import sys +sys.path.append("../../") +from ppdet.core.workspace import load_config, merge_config, create +from ppdet.data.data_feed import create_reader +from ppdet.utils.eval_utils import parse_fetches, eval_results +from ppdet.utils.stats import TrainingStats +from ppdet.utils.cli import ArgsParser +from ppdet.utils.check import check_gpu +import ppdet.utils.checkpoint as checkpoint +from ppdet.modeling.model_input import create_feed + +import logging +FORMAT = '%(asctime)s-%(levelname)s: %(message)s' +logging.basicConfig(level=logging.INFO, format=FORMAT) +logger = logging.getLogger(__name__) + + +def eval_run(exe, compile_program, reader, keys, values, cls, test_feed): + """ + Run evaluation program, return program outputs. + """ + iter_id = 0 + results = [] + if len(cls) != 0: + values = [] + for i in range(len(cls)): + _, accum_map = cls[i].get_map_var() + cls[i].reset(exe) + values.append(accum_map) + + images_num = 0 + start_time = time.time() + has_bbox = 'bbox' in keys + for data in reader(): + data = test_feed.feed(data) + feed_data = {'image': data['image'], 'im_size': data['im_size']} + outs = exe.run(compile_program, + feed=feed_data, + fetch_list=[values[0]], + return_numpy=False) + outs.append(data['gt_box']) + outs.append(data['gt_label']) + outs.append(data['is_difficult']) + res = { + k: (np.array(v), v.recursive_sequence_lengths()) + for k, v in zip(keys, outs) + } + results.append(res) + if iter_id % 100 == 0: + logger.info('Test iter {}'.format(iter_id)) + iter_id += 1 + images_num += len(res['bbox'][1][0]) if has_bbox else 1 + logger.info('Test finish iter {}'.format(iter_id)) + + end_time = time.time() + fps = images_num / (end_time - start_time) + if has_bbox: + logger.info('Total number of images: {}, inference time: {} fps.'. + format(images_num, fps)) + else: + logger.info('Total iteration: {}, inference time: {} batch/s.'.format( + images_num, fps)) + + return results + + +def main(): + cfg = load_config(FLAGS.config) + if 'architecture' in cfg: + main_arch = cfg.architecture + else: + raise ValueError("'architecture' not specified in config file.") + + merge_config(FLAGS.opt) + if 'log_iter' not in cfg: + cfg.log_iter = 20 + + # check if set use_gpu=True in paddlepaddle cpu version + check_gpu(cfg.use_gpu) + + if cfg.use_gpu: + devices_num = fluid.core.get_cuda_device_count() + else: + devices_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + + if 'train_feed' not in cfg: + train_feed = create(main_arch + 'TrainFeed') + else: + train_feed = create(cfg.train_feed) + + if 'eval_feed' not in cfg: + eval_feed = create(main_arch + 'EvalFeed') + else: + eval_feed = create(cfg.eval_feed) + + place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() + exe = fluid.Executor(place) + + lr_builder = create('LearningRate') + optim_builder = create('OptimizerBuilder') + + # build program + model = create(main_arch) + train_loader, train_feed_vars = create_feed(train_feed, iterable=True) + train_fetches = model.train(train_feed_vars) + loss = train_fetches['loss'] + lr = lr_builder() + opt = optim_builder(lr) + opt.minimize(loss) + #for v in fluid.default_main_program().list_vars(): + # if "py_reader" not in v.name and "double_buffer" not in v.name and "generated_var" not in v.name: + # print(v.name, v.shape) + + cfg.max_iters = 258 + train_reader = create_reader(train_feed, cfg.max_iters, FLAGS.dataset_dir) + train_loader.set_sample_list_generator(train_reader, place) + + exe.run(fluid.default_startup_program()) + + # parse train fetches + train_keys, train_values, _ = parse_fetches(train_fetches) + train_keys.append('lr') + train_values.append(lr.name) + + train_fetch_list = [] + for k, v in zip(train_keys, train_values): + train_fetch_list.append((k, v)) + print("train_fetch_list: {}".format(train_fetch_list)) + + eval_prog = fluid.Program() + startup_prog = fluid.Program() + with fluid.program_guard(eval_prog, startup_prog): + with fluid.unique_name.guard(): + model = create(main_arch) + _, test_feed_vars = create_feed(eval_feed, iterable=True) + fetches = model.eval(test_feed_vars) + eval_prog = eval_prog.clone(True) + + eval_reader = create_reader(eval_feed, args_path=FLAGS.dataset_dir) + test_data_feed = fluid.DataFeeder(test_feed_vars.values(), place) + + # parse eval fetches + extra_keys = [] + if cfg.metric == 'COCO': + extra_keys = ['im_info', 'im_id', 'im_shape'] + if cfg.metric == 'VOC': + extra_keys = ['gt_box', 'gt_label', 'is_difficult'] + eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog, + extra_keys) + + eval_fetch_list = [] + for k, v in zip(eval_keys, eval_values): + eval_fetch_list.append((k, v)) + print("eval_fetch_list: {}".format(eval_fetch_list)) + + exe.run(startup_prog) + checkpoint.load_params(exe, + fluid.default_main_program(), cfg.pretrain_weights) + + best_box_ap_list = [] + + def eval_func(program, scope): + results = eval_run(exe, program, eval_reader, eval_keys, eval_values, + eval_cls, test_data_feed) + + resolution = None + is_bbox_normalized = False + if 'mask' in results[0]: + resolution = model.mask_head.resolution + box_ap_stats = eval_results(results, eval_feed, cfg.metric, + cfg.num_classes, resolution, + is_bbox_normalized, FLAGS.output_eval) + if len(best_box_ap_list) == 0: + best_box_ap_list.append(box_ap_stats[0]) + elif box_ap_stats[0] > best_box_ap_list[0]: + best_box_ap_list[0] = box_ap_stats[0] + logger.info("Best test box ap: {}".format(best_box_ap_list[0])) + return best_box_ap_list[0] + + test_feed = [('image', test_feed_vars['image'].name), + ('im_size', test_feed_vars['im_size'].name)] + + teacher_cfg = load_config(FLAGS.teacher_config) + teacher_arch = teacher_cfg.architecture + teacher_programs = [] + teacher_program = fluid.Program() + teacher_startup_program = fluid.Program() + with fluid.program_guard(teacher_program, teacher_startup_program): + with fluid.unique_name.guard('teacher_'): + teacher_feed_vars = OrderedDict() + for name, var in train_feed_vars.items(): + teacher_feed_vars[name] = teacher_program.global_block( + )._clone_variable( + var, force_persistable=False) + model = create(teacher_arch) + train_fetches = model.train(teacher_feed_vars) + #print("="*50+"teacher_model_params"+"="*50) + #for v in teacher_program.list_vars(): + # print(v.name, v.shape) + #return + + exe.run(teacher_startup_program) + assert FLAGS.teacher_pretrained and os.path.exists( + FLAGS.teacher_pretrained + ), "teacher_pretrained should be set when teacher_model is not None." + + def if_exist(var): + return os.path.exists(os.path.join(FLAGS.teacher_pretrained, var.name)) + + fluid.io.load_vars( + exe, + FLAGS.teacher_pretrained, + main_program=teacher_program, + predicate=if_exist) + + teacher_programs.append(teacher_program.clone(for_test=True)) + + com = Compressor( + place, + fluid.global_scope(), + fluid.default_main_program(), + train_reader=train_reader, + train_feed_list=[(key, value.name) + for key, value in train_feed_vars.items()], + train_fetch_list=train_fetch_list, + eval_program=eval_prog, + eval_reader=eval_reader, + eval_feed_list=test_feed, + eval_func={'map': eval_func}, + eval_fetch_list=eval_fetch_list[0:1], + save_eval_model=True, + prune_infer_model=[["image", "im_size"], ["multiclass_nms_0.tmp_0"]], + teacher_programs=teacher_programs, + train_optimizer=None, + distiller_optimizer=opt, + log_period=20) + com.config(FLAGS.slim_file) + com.run() + + +if __name__ == '__main__': + parser = ArgsParser() + parser.add_argument( + "-t", + "--teacher_config", + default=None, + type=str, + help="Config file of teacher architecture.") + parser.add_argument( + "-s", + "--slim_file", + default=None, + type=str, + help="Config file of PaddleSlim.") + parser.add_argument( + "-r", + "--resume_checkpoint", + default=None, + type=str, + help="Checkpoint path for resuming training.") + parser.add_argument( + "--eval", + action='store_true', + default=False, + help="Whether to perform evaluation in train") + parser.add_argument( + "--teacher_pretrained", + default=None, + type=str, + help="Whether to use pretrained model.") + parser.add_argument( + "--output_eval", + default=None, + type=str, + help="Evaluation directory, default is current directory.") + parser.add_argument( + "-d", + "--dataset_dir", + default=None, + type=str, + help="Dataset path, same as DataFeed.dataset.dataset_dir") + FLAGS = parser.parse_args() + main() diff --git a/PaddleCV/PaddleDetection/slim/distillation/run.sh b/PaddleCV/PaddleDetection/slim/distillation/run.sh new file mode 100644 index 0000000000000000000000000000000000000000..a5497bdce2464c72e14fa2168b87db60685e83e8 --- /dev/null +++ b/PaddleCV/PaddleDetection/slim/distillation/run.sh @@ -0,0 +1,47 @@ +#!/usr/bin/env bash + +# download pretrain model +root_url="https://paddlemodels.bj.bcebos.com/object_detection" +yolov3_r34_voc="yolov3_r34_voc.tar" +pretrain_dir='./pretrain' + +if [ ! -d ${pretrain_dir} ]; then + mkdir ${pretrain_dir} +fi + +cd ${pretrain_dir} + +if [ ! -f ${yolov3_r34_voc} ]; then + wget ${root_url}/${yolov3_r34_voc} + tar xf ${yolov3_r34_voc} +fi +cd - + +# enable GC strategy +export FLAGS_fast_eager_deletion_mode=1 +export FLAGS_eager_delete_tensor_gb=0.0 + +# for distillation +#----------------- +export CUDA_VISIBLE_DEVICES=0,1,2,3 + + +# Fixing name conflicts in distillation +cd ${pretrain_dir}/yolov3_r34_voc +for files in $(ls teacher_*) + do mv $files ${files#*_} +done +for files in $(ls *) + do mv $files "teacher_"$files +done +cd - + +python -u compress.py \ +-c ../../configs/yolov3_mobilenet_v1_voc.yml \ +-t yolov3_resnet34.yml \ +-s yolov3_mobilenet_v1_yolov3_resnet34_distillation.yml \ +-o YoloTrainFeed.batch_size=64 \ +-d ../../dataset/voc \ +--teacher_pretrained ./pretrain/yolov3_r34_voc \ +> yolov3_distallation.log 2>&1 & +tailf yolov3_distallation.log diff --git a/PaddleCV/PaddleDetection/slim/distillation/yolov3_mobilenet_v1_yolov3_resnet34_distillation.yml b/PaddleCV/PaddleDetection/slim/distillation/yolov3_mobilenet_v1_yolov3_resnet34_distillation.yml new file mode 100644 index 0000000000000000000000000000000000000000..6a2a5a2575b71563c557c528a4dc94f00dce73ca --- /dev/null +++ b/PaddleCV/PaddleDetection/slim/distillation/yolov3_mobilenet_v1_yolov3_resnet34_distillation.yml @@ -0,0 +1,18 @@ +version: 1.0 +distillers: + l2_distiller: + class: 'L2Distiller' + teacher_feature_map: 'teacher_teacher_conv2d_1.tmp_0' + student_feature_map: 'conv2d_15.tmp_0' + distillation_loss_weight: 1 +strategies: + distillation_strategy: + class: 'DistillationStrategy' + distillers: ['l2_distiller'] + start_epoch: 0 + end_epoch: 270 +compressor: + epoch: 271 + checkpoint_path: './checkpoints/' + strategies: + - distillation_strategy diff --git a/PaddleCV/PaddleDetection/slim/distillation/yolov3_resnet34.yml b/PaddleCV/PaddleDetection/slim/distillation/yolov3_resnet34.yml new file mode 100644 index 0000000000000000000000000000000000000000..c04bdde9f6e35a1ce53231e3862a37364ff7dbb8 --- /dev/null +++ b/PaddleCV/PaddleDetection/slim/distillation/yolov3_resnet34.yml @@ -0,0 +1,34 @@ +architecture: YOLOv3 +log_smooth_window: 20 +metric: VOC +map_type: 11point +num_classes: 20 +weight_prefix_name: teacher_ + +YOLOv3: + backbone: ResNet + yolo_head: YOLOv3Head + +ResNet: + norm_type: sync_bn + freeze_at: 0 + freeze_norm: false + norm_decay: 0. + depth: 34 + feature_maps: [3, 4, 5] + +YOLOv3Head: + anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]] + anchors: [[10, 13], [16, 30], [33, 23], + [30, 61], [62, 45], [59, 119], + [116, 90], [156, 198], [373, 326]] + norm_decay: 0. + ignore_thresh: 0.7 + label_smooth: false + nms: + background_label: -1 + keep_top_k: 100 + nms_threshold: 0.45 + nms_top_k: 1000 + normalized: false + score_threshold: 0.01