diff --git a/slim/MODEL_ZOO.md b/slim/MODEL_ZOO.md index 146cf8d0fd651ead84e30fbd8e461f7cd67f5e87..87f20fecad4cabc933873b0dff3e8b5ff8594344 100644 --- a/slim/MODEL_ZOO.md +++ b/slim/MODEL_ZOO.md @@ -8,25 +8,25 @@ - cuDNN >=7.4 - NCCL 2.1.2 -## 裁剪模型库 +## 剪裁模型库 ### 训练策略 -- 裁剪模型训练时使用[PaddleDetection模型库](../../docs/MODEL_ZOO_cn.md)发布的模型权重作为预训练权重。 -- 裁剪训练使用模型默认配置，即除`pretrained_weights`外配置不变。 -- 裁剪模型全部为基于敏感度的卷积通道裁剪。 -- YOLOv3模型主要裁剪`yolo_head`部分，即裁剪参数如下。 +- 剪裁模型训练时使用[PaddleDetection模型库](../../docs/MODEL_ZOO_cn.md)发布的模型权重作为预训练权重。 +- 剪裁训练使用模型默认配置，即除`pretrained_weights`外配置不变。 +- 剪裁模型全部为基于敏感度的卷积通道剪裁。 +- YOLOv3模型主要剪裁`yolo_head`部分，即剪裁参数如下。 ``` --pruned_params="yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights,yolo_block.0.1.1.conv.weights,yolo_block.0.2.conv.weights,yolo_block.0.tip.conv.weights,yolo_block.1.0.0.conv.weights,yolo_block.1.0.1.conv.weights,yolo_block.1.1.0.conv.weights,yolo_block.1.1.1.conv.weights,yolo_block.1.2.conv.weights,yolo_block.1.tip.conv.weights,yolo_block.2.0.0.conv.weights,yolo_block.2.0.1.conv.weights,yolo_block.2.1.0.conv.weights,yolo_block.2.1.1.conv.weights,yolo_block.2.2.conv.weights,yolo_block.2.tip.conv.weights" ``` -- YOLOv3模型裁剪中裁剪策略`r578`表示`yolo_head`中三个输出分支一次使用`0.5, 0.7, 0.8`的裁剪率裁剪，即裁剪率如下。 +- YOLOv3模型剪裁中剪裁策略`r578`表示`yolo_head`中三个输出分支一次使用`0.5, 0.7, 0.8`的剪裁率剪裁，即剪裁率如下。 ``` --pruned_ratios="0.5,0.5,0.5,0.5,0.5,0.5,0.7,0.7,0.7,0.7,0.7,0.7,0.8,0.8,0.8,0.8,0.8,0.8" ``` -- YOLOv3模型裁剪中裁剪策略`sensity`表示`yolo_head`中各参数裁剪率如下，该裁剪率为使用`yolov3_mobilnet_v1`模型在COCO数据集上敏感度实验分析得出。 +- YOLOv3模型剪裁中剪裁策略`sensity`表示`yolo_head`中各参数剪裁率如下，该剪裁率为使用`yolov3_mobilnet_v1`模型在COCO数据集上敏感度实验分析得出。 ``` --pruned_ratios="0.1,0.2,0.2,0.2,0.2,0.1,0.2,0.3,0.3,0.3,0.2,0.1,0.3,0.4,0.4,0.4,0.4,0.3" @@ -34,10 +34,10 @@ ### YOLOv3 on COCO -| 骨架网络 | 裁剪策略 | 输入尺寸 | Box AP | 下载 | -| :----------------| :-------: | :------: |:------: | :-----------------------------------------------------: | -| ResNet50-vd-dcn | sensity | 320 | 39.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50_dcn_prune1x.tar) | -| ResNet50-vd-dcn | sensity | 320 | 38.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50_dcn_prune578.tar) | +| 骨架网络 | 剪裁策略 | 输入尺寸 | Box AP | 下载 | +| :----------------| :-------: | :------: | :-----: | :-----------------------------------------------------: | +| ResNet50-vd-dcn | sensity | 608 | 39.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50_dcn_prune1x.tar) | +| ResNet50-vd-dcn | r578 | 608 | 38.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50_dcn_prune578.tar) | | MobileNetV1 | sensity | 608 | 30.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | | MobileNetV1 | sensity | 416 | 29.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | | MobileNetV1 | sensity | 320 | 27.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune1x.tar) | @@ -47,12 +47,36 @@ ### YOLOv3 on Pascal VOC -| 骨架网络 | 裁剪策略 | 输入尺寸 | Box AP | 下载 | -| :----------------| :-------: | :------: |:------: | :-----------------------------------------------------: | +| 骨架网络 | 剪裁策略 | 输入尺寸 | Box AP | 下载 | +| :----------------| :-------: | :------: | :-----: | :-----------------------------------------------------: | +| MobileNetV1 | sensity | 608 | 78.4 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | +| MobileNetV1 | sensity | 416 | 78.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | +| MobileNetV1 | sensity | 320 | 76.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune1x.tar) | | MobileNetV1 | r578 | 608 | 77.6 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | | MobileNetV1 | r578 | 416 | 77.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | | MobileNetV1 | r578 | 320 | 75.5 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578.tar) | +### 蒸馏通道剪裁模型 + +可通过高精度模型蒸馏通道剪裁后模型的方式，训练方法及相关示例见[蒸馏通道剪裁模型](./extensions/distill_pruned_model/distill_pruned_model.ipynb)。 + +COCO数据集上蒸馏通道剪裁模型库如下。 + +| 骨架网络 | 剪裁策略 | 输入尺寸 | teacher模型 | Box AP | 下载 | +| :----------------| :-------: | :------: | :--------------------- | :-----: | :-----------------------------------------------------: | +| ResNet50-vd-dcn | r578 | 608 | YOLOv3-ResNet50-vd-dcn | 39.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_r50_dcn_prune578_distill.tar) | +| MobileNetV1 | r578 | 608 | YOLOv3-ResNet34 | 29.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 416 | YOLOv3-ResNet34 | 28.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 320 | YOLOv3-ResNet34 | 25.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_prune578_distillby_r34.tar) | + +Pascal VOC数据集上蒸馏通道剪裁模型库如下。 + +| 骨架网络 | 剪裁策略 | 输入尺寸 | teacher模型 | Box AP | 下载 | +| :----------------| :-------: | :------: | :--------------------- | :-----: | :-----------------------------------------------------: | +| MobileNetV1 | r578 | 608 | YOLOv3-ResNet34 | 78.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 416 | YOLOv3-ResNet34 | 78.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | +| MobileNetV1 | r578 | 320 | YOLOv3-ResNet34 | 76.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/prune/yolov3_mobilenet_v1_voc_prune578_distillby_r34.tar) | + ## 蒸馏模型库 diff --git a/slim/extensions/distill_pruned_model/README.md b/slim/extensions/distill_pruned_model/README.md new file mode 100644 index 0000000000000000000000000000000000000000..6b7573ad6c49c23bbc6eb8cd9de883fccc9292c9 --- /dev/null +++ b/slim/extensions/distill_pruned_model/README.md @@ -0,0 +1,67 @@ +# 蒸馏通道剪裁模型教程 + +该文档介绍如何使用[PaddleSlim](https://paddlepaddle.github.io/PaddleSlim)的蒸馏接口和卷积通道剪裁接口对检测库中的模型进行卷积层的通道剪裁并使用较高精度模型对其蒸馏。 + +在阅读该示例前，建议您先了解以下内容： + +- [检测库的使用方法](https://github.com/PaddlePaddle/PaddleDetection) +- [PaddleSlim通道剪裁API文档](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/) +- [PaddleSlim蒸馏API文档](https://paddlepaddle.github.io/PaddleSlim/api/single_distiller_api/) +- [检测库模型通道剪裁文档](../../prune/README.md) +- [检测库模型蒸馏文档](../../distillation/README.md) + +请确保已正确[安装PaddleDetection](../../docs/tutorials/INSTALL_cn.md)及其依赖。 + +已发布蒸馏通道剪裁模型见[压缩模型库](../MODEL_ZOO.md) + +蒸馏通道剪裁模型示例见[Ipython notebook示例](./distill_pruned_model_demo.ipynb) + +## 1. 数据准备 + +请参考检测库[数据下载](../../../docs/tutorials/INSTALL_cn.md)文档准备数据。 + +## 2. 模型选择 + +通过`-c`选项指定待剪裁模型的配置文件的相对路径，更多可选配置文件请参考: [检测库配置文件](../../../configs)。 + +蒸馏通道剪裁模型中，我们使用原模型全量权重来初始化待剪裁模型，已发布模型的权重可在[模型库](../../../docs/MODEL_ZOO.md)中获取。 + +通过`-o pretrain_weights`指定待剪裁模型的预训练权重，可以指定url或本地文件系统的路径。如下所示： + +``` +-o pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar +``` + +或 + +``` +-o pretrain_weights=output/yolov3_mobilenet_v1_voc/model_final +``` + +## 4. 启动蒸馏剪裁任务 + +使用`distill_pruned_model.py`启动蒸馏剪裁任务时，通过`--pruned_params`选项指定待剪裁的参数名称列表，参数名之间用空格分隔，通过`--pruned_ratios`选项指定各个参数被裁掉的比例。 获取待裁剪模型参数名称方法可参考[通道剪裁模教程](../../prune/README.md)。 +通过`-t`参数指定teacher模型配置文件，`--teacher_pretrained`指定teacher模型权重，更多关于蒸馏模型设置可参考[模型蒸馏文档](../../distillation/README.md)。 +蒸馏通道检测模型脚本目前只支持使用YOLOv3细粒度损失训练，即训练过程中须指定`-o use_fine_grained_loss=true`。 + +``` +python distill_pruned_model.py \ +-c ../../../configs/yolov3_mobilenet_v1_voc.yml \ +-t ../../../configs/yolov3_r34_voc.yml \ +--teacher_pretrained=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar \ +--pruned_params "yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights" \ +--pruned_ratios="0.2,0.3,0.4" \ +-o use_fine_grained_loss=true pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar +``` + +## 5. 评估模型 + +由于产出模型为通道剪裁模型，训练完成后，可通过通道剪裁中提供的评估脚本`../../prune/eval.py`评估模型精度，通过`--pruned_params`和`--pruned_ratios`指定剪裁的参数名称列表和各参数剪裁比例。 + +``` +python ../../prune/eval.py \ +-c ../../../configs/yolov3_mobilenet_v1_voc.yml \ +--pruned_params "yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights" \ +--pruned_ratios="0.2,0.3,0.4" \ +-o weights=output/yolov3_mobilenet_v1_voc/model_final +``` diff --git a/slim/extensions/distill_pruned_model/distill_pruned_model.py b/slim/extensions/distill_pruned_model/distill_pruned_model.py new file mode 100644 index 0000000000000000000000000000000000000000..6ab641480604e29e0b32b000b49d15d3bb311f46 --- /dev/null +++ b/slim/extensions/distill_pruned_model/distill_pruned_model.py @@ -0,0 +1,365 @@ +# 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 numpy as np +from collections import OrderedDict +from paddleslim.dist.single_distiller import merge, l2_loss +from paddleslim.prune import Pruner +from paddleslim.analysis import flops + +from paddle import fluid +from ppdet.core.workspace import load_config, merge_config, create +from ppdet.data.reader import create_reader +from ppdet.utils.eval_utils import parse_fetches, eval_results, eval_run +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 + +import logging +FORMAT = '%(asctime)s-%(levelname)s: %(message)s' +logging.basicConfig(level=logging.INFO, format=FORMAT) +logger = logging.getLogger(__name__) + + +def split_distill(split_output_names, weight): + """ + Add fine grained distillation losses. + Each loss is composed by distill_reg_loss, distill_cls_loss and + distill_obj_loss + """ + student_var = [] + for name in split_output_names: + student_var.append(fluid.default_main_program().global_block().var( + name)) + s_x0, s_y0, s_w0, s_h0, s_obj0, s_cls0 = student_var[0:6] + s_x1, s_y1, s_w1, s_h1, s_obj1, s_cls1 = student_var[6:12] + s_x2, s_y2, s_w2, s_h2, s_obj2, s_cls2 = student_var[12:18] + teacher_var = [] + for name in split_output_names: + teacher_var.append(fluid.default_main_program().global_block().var( + 'teacher_' + name)) + t_x0, t_y0, t_w0, t_h0, t_obj0, t_cls0 = teacher_var[0:6] + t_x1, t_y1, t_w1, t_h1, t_obj1, t_cls1 = teacher_var[6:12] + t_x2, t_y2, t_w2, t_h2, t_obj2, t_cls2 = teacher_var[12:18] + + def obj_weighted_reg(sx, sy, sw, sh, tx, ty, tw, th, tobj): + loss_x = fluid.layers.sigmoid_cross_entropy_with_logits( + sx, fluid.layers.sigmoid(tx)) + loss_y = fluid.layers.sigmoid_cross_entropy_with_logits( + sy, fluid.layers.sigmoid(ty)) + loss_w = fluid.layers.abs(sw - tw) + loss_h = fluid.layers.abs(sh - th) + loss = fluid.layers.sum([loss_x, loss_y, loss_w, loss_h]) + weighted_loss = fluid.layers.reduce_mean(loss * + fluid.layers.sigmoid(tobj)) + return weighted_loss + + def obj_weighted_cls(scls, tcls, tobj): + loss = fluid.layers.sigmoid_cross_entropy_with_logits( + scls, fluid.layers.sigmoid(tcls)) + weighted_loss = fluid.layers.reduce_mean( + fluid.layers.elementwise_mul( + loss, fluid.layers.sigmoid(tobj), axis=0)) + return weighted_loss + + def obj_loss(sobj, tobj): + obj_mask = fluid.layers.cast(tobj > 0., dtype="float32") + obj_mask.stop_gradient = True + loss = fluid.layers.reduce_mean( + fluid.layers.sigmoid_cross_entropy_with_logits(sobj, obj_mask)) + return loss + + distill_reg_loss0 = obj_weighted_reg(s_x0, s_y0, s_w0, s_h0, t_x0, t_y0, + t_w0, t_h0, t_obj0) + distill_reg_loss1 = obj_weighted_reg(s_x1, s_y1, s_w1, s_h1, t_x1, t_y1, + t_w1, t_h1, t_obj1) + distill_reg_loss2 = obj_weighted_reg(s_x2, s_y2, s_w2, s_h2, t_x2, t_y2, + t_w2, t_h2, t_obj2) + distill_reg_loss = fluid.layers.sum( + [distill_reg_loss0, distill_reg_loss1, distill_reg_loss2]) + + distill_cls_loss0 = obj_weighted_cls(s_cls0, t_cls0, t_obj0) + distill_cls_loss1 = obj_weighted_cls(s_cls1, t_cls1, t_obj1) + distill_cls_loss2 = obj_weighted_cls(s_cls2, t_cls2, t_obj2) + distill_cls_loss = fluid.layers.sum( + [distill_cls_loss0, distill_cls_loss1, distill_cls_loss2]) + + distill_obj_loss0 = obj_loss(s_obj0, t_obj0) + distill_obj_loss1 = obj_loss(s_obj1, t_obj1) + distill_obj_loss2 = obj_loss(s_obj2, t_obj2) + distill_obj_loss = fluid.layers.sum( + [distill_obj_loss0, distill_obj_loss1, distill_obj_loss2]) + loss = (distill_reg_loss + distill_cls_loss + distill_obj_loss) * weight + return loss + + +def main(): + env = os.environ + 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', 1)) + + if 'FLAGS_selected_gpus' in env: + device_id = int(env['FLAGS_selected_gpus']) + else: + device_id = 0 + place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() + exe = fluid.Executor(place) + + # build program + model = create(main_arch) + inputs_def = cfg['TrainReader']['inputs_def'] + train_feed_vars, train_loader = model.build_inputs(**inputs_def) + train_fetches = model.train(train_feed_vars) + loss = train_fetches['loss'] + + start_iter = 0 + train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * + devices_num, cfg) + train_loader.set_sample_list_generator(train_reader, place) + + eval_prog = fluid.Program() + with fluid.program_guard(eval_prog, fluid.default_startup_program()): + with fluid.unique_name.guard(): + model = create(main_arch) + inputs_def = cfg['EvalReader']['inputs_def'] + test_feed_vars, eval_loader = model.build_inputs(**inputs_def) + fetches = model.eval(test_feed_vars) + eval_prog = eval_prog.clone(True) + + eval_reader = create_reader(cfg.EvalReader) + eval_loader.set_sample_list_generator(eval_reader, place) + + teacher_cfg = load_config(FLAGS.teacher_config) + merge_config(FLAGS.opt) + teacher_arch = teacher_cfg.architecture + teacher_program = fluid.Program() + teacher_startup_program = fluid.Program() + + with fluid.program_guard(teacher_program, teacher_startup_program): + with fluid.unique_name.guard(): + 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) + teacher_loss = train_fetches['loss'] + + exe.run(teacher_startup_program) + assert FLAGS.teacher_pretrained, "teacher_pretrained should be set" + checkpoint.load_params(exe, teacher_program, FLAGS.teacher_pretrained) + teacher_program = teacher_program.clone(for_test=True) + + data_name_map = { + 'target0': 'target0', + 'target1': 'target1', + 'target2': 'target2', + 'image': 'image', + 'gt_bbox': 'gt_bbox', + 'gt_class': 'gt_class', + 'gt_score': 'gt_score' + } + merge(teacher_program, fluid.default_main_program(), data_name_map, place) + + yolo_output_names = [ + 'strided_slice_0.tmp_0', 'strided_slice_1.tmp_0', + 'strided_slice_2.tmp_0', 'strided_slice_3.tmp_0', + 'strided_slice_4.tmp_0', 'transpose_0.tmp_0', 'strided_slice_5.tmp_0', + 'strided_slice_6.tmp_0', 'strided_slice_7.tmp_0', + 'strided_slice_8.tmp_0', 'strided_slice_9.tmp_0', 'transpose_2.tmp_0', + 'strided_slice_10.tmp_0', 'strided_slice_11.tmp_0', + 'strided_slice_12.tmp_0', 'strided_slice_13.tmp_0', + 'strided_slice_14.tmp_0', 'transpose_4.tmp_0' + ] + + assert cfg.use_fine_grained_loss, \ + "Only support use_fine_grained_loss=True, Please set it in config file or '-o use_fine_grained_loss=true'" + distill_loss = split_distill(yolo_output_names, 1000) + loss = distill_loss + loss + lr_builder = create('LearningRate') + optim_builder = create('OptimizerBuilder') + lr = lr_builder() + opt = optim_builder(lr) + opt.minimize(loss) + + exe.run(fluid.default_startup_program()) + checkpoint.load_params(exe, + fluid.default_main_program(), cfg.pretrain_weights) + + + assert FLAGS.pruned_params is not None, \ + "FLAGS.pruned_params is empty!!! Please set it by '--pruned_params' option." + pruned_params = FLAGS.pruned_params.strip().split(",") + logger.info("pruned params: {}".format(pruned_params)) + pruned_ratios = [float(n) for n in FLAGS.pruned_ratios.strip().split(",")] + logger.info("pruned ratios: {}".format(pruned_ratios)) + assert len(pruned_params) == len(pruned_ratios), \ + "The length of pruned params and pruned ratios should be equal." + assert pruned_ratios > [0] * len(pruned_ratios) and pruned_ratios < [1] * len(pruned_ratios), \ + "The elements of pruned ratios should be in range (0, 1)." + + pruner = Pruner() + distill_prog = pruner.prune( + fluid.default_main_program(), + fluid.global_scope(), + params=pruned_params, + ratios=pruned_ratios, + place=place, + only_graph=False)[0] + + base_flops = flops(eval_prog) + eval_prog = pruner.prune( + eval_prog, + fluid.global_scope(), + params=pruned_params, + ratios=pruned_ratios, + place=place, + only_graph=True)[0] + pruned_flops = flops(eval_prog) + logger.info("FLOPs -{}; total FLOPs: {}; pruned FLOPs: {}".format( + float(base_flops - pruned_flops) / base_flops, base_flops, + pruned_flops)) + + build_strategy = fluid.BuildStrategy() + build_strategy.fuse_all_reduce_ops = False + build_strategy.fuse_all_optimizer_ops = False + build_strategy.fuse_elewise_add_act_ops = True + # only enable sync_bn in multi GPU devices + sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn' + build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \ + and cfg.use_gpu + + exec_strategy = fluid.ExecutionStrategy() + # iteration number when CompiledProgram tries to drop local execution scopes. + # Set it to be 1 to save memory usages, so that unused variables in + # local execution scopes can be deleted after each iteration. + exec_strategy.num_iteration_per_drop_scope = 1 + + parallel_main = fluid.CompiledProgram(distill_prog).with_data_parallel( + loss_name=loss.name, + build_strategy=build_strategy, + exec_strategy=exec_strategy) + compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog) + + # parse eval fetches + extra_keys = [] + if cfg.metric == 'COCO': + extra_keys = ['im_info', 'im_id', 'im_shape'] + if cfg.metric == 'VOC': + extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] + eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog, + extra_keys) + + # whether output bbox is normalized in model output layer + is_bbox_normalized = False + if hasattr(model, 'is_bbox_normalized') and \ + callable(model.is_bbox_normalized): + is_bbox_normalized = model.is_bbox_normalized() + map_type = cfg.map_type if 'map_type' in cfg else '11point' + best_box_ap_list = [0.0, 0] #[map, iter] + cfg_name = os.path.basename(FLAGS.config).split('.')[0] + save_dir = os.path.join(cfg.save_dir, cfg_name) + + train_loader.start() + for step_id in range(start_iter, cfg.max_iters): + teacher_loss_np, distill_loss_np, loss_np, lr_np = exe.run( + parallel_main, + fetch_list=[ + 'teacher_' + teacher_loss.name, distill_loss.name, loss.name, + lr.name + ]) + if step_id % cfg.log_iter == 0: + logger.info( + "step {} lr {:.6f}, loss {:.6f}, distill_loss {:.6f}, teacher_loss {:.6f}". + format(step_id, lr_np[0], loss_np[0], distill_loss_np[0], + teacher_loss_np[0])) + if step_id % cfg.snapshot_iter == 0 and step_id != 0 or step_id == cfg.max_iters - 1: + save_name = str( + step_id) if step_id != cfg.max_iters - 1 else "model_final" + checkpoint.save(exe, distill_prog, + os.path.join(save_dir, save_name)) + # eval + results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, + eval_values, eval_cls) + resolution = None + box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes, + resolution, is_bbox_normalized, + FLAGS.output_eval, map_type, + cfg['EvalReader']['dataset']) + + if box_ap_stats[0] > best_box_ap_list[0]: + best_box_ap_list[0] = box_ap_stats[0] + best_box_ap_list[1] = step_id + checkpoint.save(exe, distill_prog, + os.path.join("./", "best_model")) + logger.info("Best test box ap: {}, in step: {}".format( + best_box_ap_list[0], best_box_ap_list[1])) + train_loader.reset() + + +if __name__ == '__main__': + parser = ArgsParser() + parser.add_argument( + "-t", + "--teacher_config", + default=None, + type=str, + help="Config file of teacher architecture.") + 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( + "-p", + "--pruned_params", + default=None, + type=str, + help="The parameters to be pruned when calculating sensitivities.") + parser.add_argument( + "--pruned_ratios", + default=None, + type=str, + help="The ratios pruned iteratively for each parameter when calculating sensitivities." + ) + FLAGS = parser.parse_args() + main() diff --git a/slim/extensions/distill_pruned_model/distill_pruned_model_demo.ipynb b/slim/extensions/distill_pruned_model/distill_pruned_model_demo.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..7705a0397f5f332101ff1861c8cbd9eab374c774 --- /dev/null +++ b/slim/extensions/distill_pruned_model/distill_pruned_model_demo.ipynb @@ -0,0 +1,641 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "### 蒸馏通道裁剪模型示例\n", + "本示例介绍使用更高精度的[YOLOv3-ResNet34](../../configs/yolov3_r34.yml)模型蒸馏经通道裁剪的[YOLOv3-MobileNet](../../configs/yolov3_mobilenet_v1.yml)模型。脚本可参照蒸馏脚本[distill.py](../distillation/distill.py)和通道裁剪脚本[prune.py](../prune/prune.py)简单修改得到，蒸馏过程采用细粒度损失来蒸馏YOLOv3输出层特征图。" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "切换到PaddleDetection根目录，设置环境变量" + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "metadata": { + "scrolled": true + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "/workplace/PaddleDetection\n" + ] + } + ], + "source": [ + "% cd ../.." + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "导入依赖包，注意须同时导入蒸馏和通道裁剪的相关接口" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "metadata": {}, + "outputs": [], + "source": [ + "from __future__ import absolute_import\n", + "from __future__ import division\n", + "from __future__ import print_function\n", + "\n", + "import os\n", + "import numpy as np\n", + "from collections import OrderedDict\n", + "from paddleslim.dist.single_distiller import merge, l2_loss\n", + "from paddleslim.prune import Pruner\n", + "from paddleslim.analysis import flops\n", + "\n", + "from paddle import fluid\n", + "from ppdet.core.workspace import load_config, merge_config, create\n", + "from ppdet.data.reader import create_reader\n", + "from ppdet.utils.eval_utils import parse_fetches, eval_results, eval_run\n", + "from ppdet.utils.stats import TrainingStats\n", + "from ppdet.utils.cli import ArgsParser\n", + "from ppdet.utils.check import check_gpu\n", + "import ppdet.utils.checkpoint as checkpoint" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "定义细粒度的蒸馏损失函数" + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "metadata": {}, + "outputs": [], + "source": [ + "def split_distill(split_output_names, weight):\n", + " \"\"\"\n", + " Add fine grained distillation losses.\n", + " Each loss is composed by distill_reg_loss, distill_cls_loss and\n", + " distill_obj_loss\n", + " \"\"\"\n", + " student_var = []\n", + " for name in split_output_names:\n", + " student_var.append(fluid.default_main_program().global_block().var(\n", + " name))\n", + " s_x0, s_y0, s_w0, s_h0, s_obj0, s_cls0 = student_var[0:6]\n", + " s_x1, s_y1, s_w1, s_h1, s_obj1, s_cls1 = student_var[6:12]\n", + " s_x2, s_y2, s_w2, s_h2, s_obj2, s_cls2 = student_var[12:18]\n", + " teacher_var = []\n", + " for name in split_output_names:\n", + " teacher_var.append(fluid.default_main_program().global_block().var(\n", + " 'teacher_' + name))\n", + " t_x0, t_y0, t_w0, t_h0, t_obj0, t_cls0 = teacher_var[0:6]\n", + " t_x1, t_y1, t_w1, t_h1, t_obj1, t_cls1 = teacher_var[6:12]\n", + " t_x2, t_y2, t_w2, t_h2, t_obj2, t_cls2 = teacher_var[12:18]\n", + "\n", + " def obj_weighted_reg(sx, sy, sw, sh, tx, ty, tw, th, tobj):\n", + " loss_x = fluid.layers.sigmoid_cross_entropy_with_logits(\n", + " sx, fluid.layers.sigmoid(tx))\n", + " loss_y = fluid.layers.sigmoid_cross_entropy_with_logits(\n", + " sy, fluid.layers.sigmoid(ty))\n", + " loss_w = fluid.layers.abs(sw - tw)\n", + " loss_h = fluid.layers.abs(sh - th)\n", + " loss = fluid.layers.sum([loss_x, loss_y, loss_w, loss_h])\n", + " weighted_loss = fluid.layers.reduce_mean(loss *\n", + " fluid.layers.sigmoid(tobj))\n", + " return weighted_loss\n", + "\n", + " def obj_weighted_cls(scls, tcls, tobj):\n", + " loss = fluid.layers.sigmoid_cross_entropy_with_logits(\n", + " scls, fluid.layers.sigmoid(tcls))\n", + " weighted_loss = fluid.layers.reduce_mean(\n", + " fluid.layers.elementwise_mul(\n", + " loss, fluid.layers.sigmoid(tobj), axis=0))\n", + " return weighted_loss\n", + "\n", + " def obj_loss(sobj, tobj):\n", + " obj_mask = fluid.layers.cast(tobj > 0., dtype=\"float32\")\n", + " obj_mask.stop_gradient = True\n", + " loss = fluid.layers.reduce_mean(\n", + " fluid.layers.sigmoid_cross_entropy_with_logits(sobj, obj_mask))\n", + " return loss\n", + "\n", + " distill_reg_loss0 = obj_weighted_reg(s_x0, s_y0, s_w0, s_h0, t_x0, t_y0,\n", + " t_w0, t_h0, t_obj0)\n", + " distill_reg_loss1 = obj_weighted_reg(s_x1, s_y1, s_w1, s_h1, t_x1, t_y1,\n", + " t_w1, t_h1, t_obj1)\n", + " distill_reg_loss2 = obj_weighted_reg(s_x2, s_y2, s_w2, s_h2, t_x2, t_y2,\n", + " t_w2, t_h2, t_obj2)\n", + " distill_reg_loss = fluid.layers.sum(\n", + " [distill_reg_loss0, distill_reg_loss1, distill_reg_loss2])\n", + "\n", + " distill_cls_loss0 = obj_weighted_cls(s_cls0, t_cls0, t_obj0)\n", + " distill_cls_loss1 = obj_weighted_cls(s_cls1, t_cls1, t_obj1)\n", + " distill_cls_loss2 = obj_weighted_cls(s_cls2, t_cls2, t_obj2)\n", + " distill_cls_loss = fluid.layers.sum(\n", + " [distill_cls_loss0, distill_cls_loss1, distill_cls_loss2])\n", + "\n", + " distill_obj_loss0 = obj_loss(s_obj0, t_obj0)\n", + " distill_obj_loss1 = obj_loss(s_obj1, t_obj1)\n", + " distill_obj_loss2 = obj_loss(s_obj2, t_obj2)\n", + " distill_obj_loss = fluid.layers.sum(\n", + " [distill_obj_loss0, distill_obj_loss1, distill_obj_loss2])\n", + " loss = (distill_reg_loss + distill_cls_loss + distill_obj_loss) * weight\n", + " return loss" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "读取配置文件，设置use_fined_grained_loss=True" + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "metadata": {}, + "outputs": [], + "source": [ + "cfg = load_config(\"./configs/yolov3_mobilenet_v1.yml\")\n", + "merge_config({'use_fine_grained_loss': True})" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "创建执行器" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "metadata": {}, + "outputs": [], + "source": [ + "devices_num = fluid.core.get_cuda_device_count()\n", + "place = fluid.CUDAPlace(0)\n", + "# devices_num = int(os.environ.get('CPU_NUM', 1))\n", + "# place = fluid.CPUPlace()\n", + "exe = fluid.Executor(place)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "构造训练模型和reader" + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "loading annotations into memory...\n", + "Done (t=24.56s)\n", + "creating index...\n", + "index created!\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "2020-02-07 08:02:05,272-WARNING: Found an invalid bbox in annotations: im_id: 550395, area: 0.0 x1: 9.98, y1: 188.56, x2: 14.52, y2: 188.56.\n", + "2020-02-07 08:02:17,043-WARNING: Found an invalid bbox in annotations: im_id: 200365, area: 0.0 x1: 296.65, y1: 388.33, x2: 296.68, y2: 388.33.\n", + "2020-02-07 08:02:19,653-INFO: 118287 samples in file dataset/coco/annotations/instances_train2017.json\n", + "2020-02-07 08:02:25,768-INFO: places would be ommited when DataLoader is not iterable\n" + ] + }, + { + "data": { + "text/plain": [ + "" + ] + }, + "execution_count": 6, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "main_arch = cfg.architecture\n", + "# build program\n", + "model = create(main_arch)\n", + "inputs_def = cfg['TrainReader']['inputs_def']\n", + "train_feed_vars, train_loader = model.build_inputs(**inputs_def)\n", + "train_fetches = model.train(train_feed_vars)\n", + "loss = train_fetches['loss']\n", + "\n", + "start_iter = 0\n", + "train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * devices_num, cfg)\n", + "train_loader.set_sample_list_generator(train_reader, place)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "构造评估模型和reader" + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "loading annotations into memory...\n", + "Done (t=0.84s)\n", + "creating index...\n", + "index created!\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "2020-02-07 08:02:32,708-INFO: 5000 samples in file dataset/coco/annotations/instances_val2017.json\n", + "2020-02-07 08:02:32,805-INFO: places would be ommited when DataLoader is not iterable\n" + ] + }, + { + "data": { + "text/plain": [ + "" + ] + }, + "execution_count": 7, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "eval_prog = fluid.Program()\n", + "with fluid.program_guard(eval_prog, fluid.default_startup_program()):\n", + " with fluid.unique_name.guard():\n", + " model = create(main_arch)\n", + " inputs_def = cfg['EvalReader']['inputs_def']\n", + " test_feed_vars, eval_loader = model.build_inputs(**inputs_def)\n", + " fetches = model.eval(test_feed_vars)\n", + "eval_prog = eval_prog.clone(True)\n", + "\n", + "eval_reader = create_reader(cfg.EvalReader)\n", + "eval_loader.set_sample_list_generator(eval_reader, place)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "构造teacher模型并导入权重" + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "2020-02-07 08:02:38,933-INFO: Found /root/.cache/paddle/weights/yolov3_r34\n", + "2020-02-07 08:02:38,935-INFO: Loading parameters from /root/.cache/paddle/weights/yolov3_r34...\n" + ] + } + ], + "source": [ + "teacher_cfg = load_config(\"./configs/yolov3_r34.yml\")\n", + "merge_config({'use_fine_grained_loss': True})\n", + "teacher_arch = teacher_cfg.architecture\n", + "teacher_program = fluid.Program()\n", + "teacher_startup_program = fluid.Program()\n", + "\n", + "with fluid.program_guard(teacher_program, teacher_startup_program):\n", + " with fluid.unique_name.guard():\n", + " teacher_feed_vars = OrderedDict()\n", + " for name, var in train_feed_vars.items():\n", + " teacher_feed_vars[name] = teacher_program.global_block(\n", + " )._clone_variable(\n", + " var, force_persistable=False)\n", + " model = create(teacher_arch)\n", + " train_fetches = model.train(teacher_feed_vars)\n", + " teacher_loss = train_fetches['loss']\n", + "\n", + "exe.run(teacher_startup_program)\n", + "checkpoint.load_params(exe, teacher_program, \"https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar\")\n", + "teacher_program = teacher_program.clone(for_test=True)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "合并program" + ] + }, + { + "cell_type": "code", + "execution_count": 9, + "metadata": {}, + "outputs": [], + "source": [ + "data_name_map = {\n", + " 'target0': 'target0',\n", + " 'target1': 'target1',\n", + " 'target2': 'target2',\n", + " 'image': 'image',\n", + " 'gt_bbox': 'gt_bbox',\n", + " 'gt_class': 'gt_class',\n", + " 'gt_score': 'gt_score'\n", + "}\n", + "merge(teacher_program, fluid.default_main_program(), data_name_map, place)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "构造蒸馏损失和优化器" + ] + }, + { + "cell_type": "code", + "execution_count": 10, + "metadata": {}, + "outputs": [], + "source": [ + "yolo_output_names = [\n", + " 'strided_slice_0.tmp_0', 'strided_slice_1.tmp_0',\n", + " 'strided_slice_2.tmp_0', 'strided_slice_3.tmp_0',\n", + " 'strided_slice_4.tmp_0', 'transpose_0.tmp_0', 'strided_slice_5.tmp_0',\n", + " 'strided_slice_6.tmp_0', 'strided_slice_7.tmp_0',\n", + " 'strided_slice_8.tmp_0', 'strided_slice_9.tmp_0', 'transpose_2.tmp_0',\n", + " 'strided_slice_10.tmp_0', 'strided_slice_11.tmp_0',\n", + " 'strided_slice_12.tmp_0', 'strided_slice_13.tmp_0',\n", + " 'strided_slice_14.tmp_0', 'transpose_4.tmp_0'\n", + "]\n", + " \n", + "distill_loss = split_distill(yolo_output_names, 1000)\n", + "loss = distill_loss + loss\n", + "lr_builder = create('LearningRate')\n", + "optim_builder = create('OptimizerBuilder')\n", + "lr = lr_builder()\n", + "opt = optim_builder(lr)\n", + "opt.minimize(loss)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "导入待裁剪模型裁剪前全部权重" + ] + }, + { + "cell_type": "code", + "execution_count": 12, + "metadata": { + "scrolled": true + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "2020-02-07 08:03:39,997-INFO: Found /root/.cache/paddle/weights/yolov3_mobilenet_v1\n", + "2020-02-07 08:03:39,999-INFO: Loading parameters from /root/.cache/paddle/weights/yolov3_mobilenet_v1...\n" + ] + } + ], + "source": [ + "exe.run(fluid.default_startup_program())\n", + "checkpoint.load_params(exe, fluid.default_main_program(), \"https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar\")" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "裁剪训练和评估program" + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "pruned params: ['yolo_block.0.0.0.conv.weights', 'yolo_block.0.0.1.conv.weights', 'yolo_block.0.1.0.conv.weights', 'yolo_block.0.1.1.conv.weights', 'yolo_block.0.2.conv.weights', 'yolo_block.0.tip.conv.weights', 'yolo_block.1.0.0.conv.weights', 'yolo_block.1.0.1.conv.weights', 'yolo_block.1.1.0.conv.weights', 'yolo_block.1.1.1.conv.weights', 'yolo_block.1.2.conv.weights', 'yolo_block.1.tip.conv.weights', 'yolo_block.2.0.0.conv.weights', 'yolo_block.2.0.1.conv.weights', 'yolo_block.2.1.0.conv.weights', 'yolo_block.2.1.1.conv.weights', 'yolo_block.2.2.conv.weights', 'yolo_block.2.tip.conv.weights']\n", + "pruned ratios: [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8]\n", + "FLOPs -0.675602593026; total FLOPs: 24531648.0; pruned FLOPs: 7958003.0\n" + ] + } + ], + "source": [ + "pruned_params = [\"yolo_block.0.0.0.conv.weights\", \n", + " \"yolo_block.0.0.1.conv.weights\", \n", + " \"yolo_block.0.1.0.conv.weights\", \n", + " \"yolo_block.0.1.1.conv.weights\", \n", + " \"yolo_block.0.2.conv.weights\", \n", + " \"yolo_block.0.tip.conv.weights\", \n", + " \"yolo_block.1.0.0.conv.weights\", \n", + " \"yolo_block.1.0.1.conv.weights\", \n", + " \"yolo_block.1.1.0.conv.weights\", \n", + " \"yolo_block.1.1.1.conv.weights\", \n", + " \"yolo_block.1.2.conv.weights\", \n", + " \"yolo_block.1.tip.conv.weights\", \n", + " \"yolo_block.2.0.0.conv.weights\", \n", + " \"yolo_block.2.0.1.conv.weights\", \n", + " \"yolo_block.2.1.0.conv.weights\", \n", + " \"yolo_block.2.1.1.conv.weights\", \n", + " \"yolo_block.2.2.conv.weights\", \n", + " \"yolo_block.2.tip.conv.weights\"]\n", + "pruned_ratios = [0.5] * 6 + [0.7] * 6 + [0.8] * 6\n", + "\n", + "print(\"pruned params: {}\".format(pruned_params))\n", + "print(\"pruned ratios: {}\".format(pruned_ratios))\n", + "\n", + "pruner = Pruner()\n", + "distill_prog = pruner.prune(\n", + " fluid.default_main_program(),\n", + " fluid.global_scope(),\n", + " params=pruned_params,\n", + " ratios=pruned_ratios,\n", + " place=place,\n", + " only_graph=False)[0]\n", + "\n", + "base_flops = flops(eval_prog)\n", + "eval_prog = pruner.prune(\n", + " eval_prog,\n", + " fluid.global_scope(),\n", + " params=pruned_params,\n", + " ratios=pruned_ratios,\n", + " place=place,\n", + " only_graph=True)[0]\n", + "pruned_flops = flops(eval_prog)\n", + "print(\"FLOPs -{}; total FLOPs: {}; pruned FLOPs: {}\".format(float(base_flops - pruned_flops)/base_flops, base_flops, pruned_flops))" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "编译训练和评估program" + ] + }, + { + "cell_type": "code", + "execution_count": 14, + "metadata": {}, + "outputs": [], + "source": [ + "build_strategy = fluid.BuildStrategy()\n", + "build_strategy.fuse_all_reduce_ops = False\n", + "build_strategy.fuse_all_optimizer_ops = False\n", + "build_strategy.fuse_elewise_add_act_ops = True\n", + "# only enable sync_bn in multi GPU devices\n", + "sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn'\n", + "build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \\\n", + " and cfg.use_gpu\n", + "\n", + "exec_strategy = fluid.ExecutionStrategy()\n", + "# iteration number when CompiledProgram tries to drop local execution scopes.\n", + "# Set it to be 1 to save memory usages, so that unused variables in\n", + "# local execution scopes can be deleted after each iteration.\n", + "exec_strategy.num_iteration_per_drop_scope = 1\n", + "\n", + "parallel_main = fluid.CompiledProgram(distill_prog).with_data_parallel(\n", + " loss_name=loss.name,\n", + " build_strategy=build_strategy,\n", + " exec_strategy=exec_strategy)\n", + "compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "开始训练" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "step 0 lr 0.000000, loss 1224.825684, distill_loss 362.955292, teacher_loss 39.179062\n" + ] + } + ], + "source": [ + "# parse eval fetches\n", + "extra_keys = []\n", + "if cfg.metric == 'COCO':\n", + " extra_keys = ['im_info', 'im_id', 'im_shape']\n", + "if cfg.metric == 'VOC':\n", + " extra_keys = ['gt_bbox', 'gt_class', 'is_difficult']\n", + "eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog,\n", + " extra_keys)\n", + "\n", + "# whether output bbox is normalized in model output layer\n", + "is_bbox_normalized = False\n", + "map_type = cfg.map_type if 'map_type' in cfg else '11point'\n", + "best_box_ap_list = [0.0, 0] #[map, iter]\n", + "save_dir = os.path.join(cfg.save_dir, 'yolov3_mobilenet_v1')\n", + "\n", + "train_loader.start()\n", + "for step_id in range(start_iter, cfg.max_iters):\n", + " teacher_loss_np, distill_loss_np, loss_np, lr_np = exe.run(\n", + " parallel_main,\n", + " fetch_list=[\n", + " 'teacher_' + teacher_loss.name, distill_loss.name, loss.name,\n", + " lr.name\n", + " ])\n", + " if step_id % 20 == 0:\n", + " print(\n", + " \"step {} lr {:.6f}, loss {:.6f}, distill_loss {:.6f}, teacher_loss {:.6f}\".\n", + " format(step_id, lr_np[0], loss_np[0], distill_loss_np[0],\n", + " teacher_loss_np[0]))\n", + " if step_id % cfg.snapshot_iter == 0 and step_id != 0 or step_id == cfg.max_iters - 1:\n", + " save_name = str(\n", + " step_id) if step_id != cfg.max_iters - 1 else \"model_final\"\n", + " checkpoint.save(exe,\n", + " distill_prog,\n", + " os.path.join(save_dir, save_name))\n", + " # eval\n", + " results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys,\n", + " eval_values, eval_cls)\n", + " resolution = None\n", + " box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes,\n", + " resolution, is_bbox_normalized,\n", + " FLAGS.output_eval, map_type,\n", + " cfg['EvalReader']['dataset'])\n", + "\n", + " if box_ap_stats[0] > best_box_ap_list[0]:\n", + " best_box_ap_list[0] = box_ap_stats[0]\n", + " best_box_ap_list[1] = step_id\n", + " checkpoint.save(exe,\n", + " distill_prog,\n", + " os.path.join(\"./\", \"best_model\"))\n", + " print(\"Best test box ap: {}, in step: {}\".format(\n", + " best_box_ap_list[0], best_box_ap_list[1]))\n", + " train_loader.reset()" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "我们也提供了一键式启动蒸馏通道裁剪模型训练脚本[distill_pruned_model.py](./distill_pruned_model.py)和蒸馏通道裁剪模型库，请参考[蒸馏通道裁剪模型](./README.md)" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 2", + "language": "python", + "name": "python2" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 2 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython2", + "version": "2.7.15" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/slim/prune/README.md b/slim/prune/README.md index 8ba9b7b43f7706efd2a865dbbbd7ee8c768de15f..f7e733ef5b0630f74f055e8a28bf48797e007373 100644 --- a/slim/prune/README.md +++ b/slim/prune/README.md @@ -20,10 +20,10 @@ 对于剪裁任务，原模型的权重不一定对剪裁后的模型训练的重训练有贡献，所以加载原模型的权重不是必需的步骤。 -通过`-o weights`指定模型的权重，可以指定url或本地文件系统的路径。如下所示： +通过`-o pretrain_weights`指定模型的预训练权重，可以指定url或本地文件系统的路径。如下所示： ``` --o weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar +-o pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar ``` 或 @@ -55,7 +55,7 @@ python prune.py \ python prune.py \ -c ../../configs/yolov3_mobilenet_v1_voc.yml \ --pruned_params "yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights" \ ---pruned_ratios="0.2 0.3 0.4" +--pruned_ratios="0.2,0.3,0.4" ``` ## 5. 评估剪裁模型 @@ -66,7 +66,7 @@ python prune.py \ python eval.py \ -c ../../configs/yolov3_mobilenet_v1_voc.yml \ --pruned_params "yolo_block.0.0.0.conv.weights,yolo_block.0.0.1.conv.weights,yolo_block.0.1.0.conv.weights" \ ---pruned_ratios="0.2 0.3 0.4" \ +--pruned_ratios="0.2,0.3,0.4" \ -o weights=output/yolov3_mobilenet_v1_voc/model_final ```