From dcaa0465773f357a2698a0274e08d81857c849f4 Mon Sep 17 00:00:00 2001 From: zhouzj <41366441+zzjjay@users.noreply.github.com> Date: Thu, 12 May 2022 11:36:26 +0800 Subject: [PATCH] add demo for auto-compress (#1078) --- demo/auto_compression/pp-humanseg/README.md | 164 +++++++++++++++ .../configs/humanseg_quant_dis.yaml | 56 ++++++ .../configs/humanseg_sparse_dis.yaml | 61 ++++++ demo/auto_compression/pp-humanseg/run.py | 190 ++++++++++++++++++ 4 files changed, 471 insertions(+) create mode 100644 demo/auto_compression/pp-humanseg/README.md create mode 100644 demo/auto_compression/pp-humanseg/configs/humanseg_quant_dis.yaml create mode 100644 demo/auto_compression/pp-humanseg/configs/humanseg_sparse_dis.yaml create mode 100644 demo/auto_compression/pp-humanseg/run.py diff --git a/demo/auto_compression/pp-humanseg/README.md b/demo/auto_compression/pp-humanseg/README.md new file mode 100644 index 00000000..38de8450 --- /dev/null +++ b/demo/auto_compression/pp-humanseg/README.md @@ -0,0 +1,164 @@ +# 使用预测模型进行自动压缩示例 + +本示例将介绍如何使用PaddleSeg中预测模型进行自动压缩训练。 + +以[PP-HumanSeg-Lite](https://github.com/PaddlePaddle/PaddleSeg/tree/develop/contrib/PP-HumanSeg#portrait-segmentation)模型为例，使用自动压缩接口分别进行了蒸馏稀疏训练和蒸馏量化训练实验，并在SD710上使用单线程测试加速效果，其压缩结果和测速结果如下所示： +| 压缩方式 | Total IoU | 耗时(ms)
thread=1 | 加速比 | +|:-----:|:----------:|:---------:| :------:| +| Baseline | 0.9287 | 56.363 | - | +| 非结构化稀疏 | 0.9235 | 37.712 | 49.456% | +| 量化 | 0.9284 | 49.656 | 13.506% | + +## 自动压缩训练流程 + +### 1. 准备数据集 + +参考[PaddleSeg数据准备文档](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/data/marker/marker_cn.md) + +### 2. 准备待压缩模型 + +PaddleSeg 是基于飞桨 PaddlePaddle 开发的端到端图像分割开发套件，涵盖了高精度和轻量级等不同方向的大量高质量分割模型。 +安装 PaddleSeg 指令如下： +``` +pip install paddleseg +``` +PaddleSeg 环境依赖详见[安装文档](https://github.com/PaddlePaddle/PaddleSeg/blob/develop/docs/install_cn.md)。 + +#### 2.1 下载代码 +``` +git clone https://github.com/PaddlePaddle/PaddleSeg.git +``` +#### 2.2 准备预训练模型 + +在 PaddleSeg 目录下执行如下指令，下载预训练模型。 +``` shell +wget https://paddleseg.bj.bcebos.com/dygraph/ppseg/ppseg_lite_portrait_398x224.tar.gz +tar -xzf ppseg_lite_portrait_398x224.tar.gz +``` +#### 2.3 导出预测模型 + +在 PaddleSeg 目录下执行如下命令，则预测模型会保存在 inference_model 文件夹。 +```shell +# 设置1张可用的卡 +export CUDA_VISIBLE_DEVICES=0 +# windows下请执行以下命令 +# set CUDA_VISIBLE_DEVICES=0 +python export.py \ + --config configs/pp_humanseg_lite/pp_humanseg_lite_export_398x224.yml \ + --model_path ppseg_lite_portrait_398x224/model.pdparams \ + --save_dir inference_model + --with_softmax +``` +或直接下载 PP-HumanSeg-Lite 的预测模型： +```shell +wegt https://paddleseg.bj.bcebos.com/dygraph/ppseg/ppseg_lite_portrait_398x224_with_softmax.tar.gz +tar -xzf ppseg_lite_portrait_398x224_with_softmax.tar.gz +``` + +### 3. 多策略融合压缩 + +每一个小章节代表一种多策略融合压缩方式。 + +### 3.1 进行蒸馏稀疏压缩 +自动压缩训练需要准备 config 文件、数据集 dataloader 以及测试函数（``eval_function``）。 +#### 3.1.1 配置config + +使用自动压缩进行蒸馏和非结构化稀疏的联合训练，首先要配置 config 文件，包含蒸馏、稀疏和训练三部分参数。 + +- 蒸馏参数 + +蒸馏参数主要设置蒸馏节点（``distill_node_pair``）和教师网络测预测模型路径。蒸馏节点需包含教师网络节点和对应的学生网络节点，其中教师网络节点名称将在程序中自动添加 “teacher_” 前缀，如下所示。 +```yaml +Distillation: + distill_lambda: 1.0 + distill_loss: l2_loss + distill_node_pair: + - teacher_relu_30.tmp_0 + - relu_30.tmp_0 + merge_feed: true + teacher_model_dir: ./inference_model + teacher_model_filename: model.pdmodel + teacher_params_filename: model.pdiparams +``` +- 稀疏参数 + +稀疏参数设置如下所示，其中参数含义详见[非结构化稀疏API文档](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/dygraph/pruners/unstructured_pruner.rst)。 +```yaml +UnstructurePrune: + prune_strategy: gmp + prune_mode: ratio + pruned_ratio: 0.75 + gmp_config: + stable_iterations: 0 + pruning_iterations: 4500 + tunning_iterations: 4500 + resume_iteration: -1 + pruning_steps: 100 + initial_ratio: 0.15 + prune_params_type: conv1x1_only + local_sparsity: True +``` + +- 训练参数 + +训练参数主要设置学习率、训练次数（epochs）和优化器等。 +```yaml +TrainConfig: + epochs: 14 + eval_iter: 400 + learning_rate: 5.0e-03 + optimizer: SGD + optim_args: + weight_decay: 0.0005 +``` +#### 3.1.2 准备 dataloader 和测试函数 +准备好数据集后，需将训练数据封装成 dict 类型传入自动压缩接口，可参考以下函数进行封装。测试函数用于测试模型精度，需在静态图模式下实现。 +```python +def reader_wrapper(reader): + def gen(): + for i, data in enumerate(reader()): + imgs = np.array(data[0]) + yield {"x": imgs} + return gen +``` +> 注：该dict类型的key值要和保存预测模型时的输入名称保持一致。 + +#### 3.1.3 开启训练 + +将训练数据集 dataloader 和测试函数传入接口 ``paddleslim.auto_compression.AutoCompression``，对模型进行非结构化稀疏训练。运行指令如下： +```shell +python run.py \ + --model_dir='inference_model' \ + --model_filename='inference.pdmodel' \ + --params_filename='./inference.pdiparams' \ + --save_dir='./save_model' \ + --config_path='configs/humanseg_sparse_dis.yaml' +``` + +### 3.2 进行蒸馏量化压缩 +#### 3.2.1 配置config +使用自动压缩进行量化训练，首先要配置config文件，包含蒸馏、量化和训练三部分参数。其中蒸馏和训练参数与稀疏训练类似，下面主要介绍量化参数的设置。 +- 量化参数 + +量化参数主要设置量化比特数和量化op类型，其中量化op包含卷积层（conv2d, depthwise_conv2d）和全连接层（matmul）。以下为只量化卷积层的示例： +```yaml +Quantization: + activation_bits: 8 + weight_bits: 8 + is_full_quantize: false + not_quant_pattern: + - skip_quant + quantize_op_types: + - conv2d + - depthwise_conv2d +``` +#### 3.2.2 开启训练 +将数据集 dataloader 和测试函数（``eval_function``）传入接口``paddleslim.auto_compression.AutoCompression``，对模型进行量化训练。运行指令如下： +``` +python run.py \ + --model_dir='inference_model' \ + --model_filename='inference.pdmodel' \ + --params_filename='./inference.pdiparams' \ + --save_dir='./save_model' \ + --config_path='configs/humanseg_quant_dis.yaml' +``` diff --git a/demo/auto_compression/pp-humanseg/configs/humanseg_quant_dis.yaml b/demo/auto_compression/pp-humanseg/configs/humanseg_quant_dis.yaml new file mode 100644 index 00000000..ba05eff4 --- /dev/null +++ b/demo/auto_compression/pp-humanseg/configs/humanseg_quant_dis.yaml @@ -0,0 +1,56 @@ +Distillation: + distill_lambda: 1.0 + distill_loss: l2_loss + distill_node_pair: + - teacher_reshape2_1.tmp_0 # + - reshape2_1.tmp_0 + - teacher_reshape2_3.tmp_0 # + - reshape2_3.tmp_0 + - teacher_reshape2_5.tmp_0 # + - reshape2_5.tmp_0 + - teacher_reshape2_7.tmp_0 #block1 + - reshape2_7.tmp_0 + - teacher_reshape2_9.tmp_0 # + - reshape2_9.tmp_0 + - teacher_reshape2_11.tmp_0 # + - reshape2_11.tmp_0 + - teacher_reshape2_13.tmp_0 # + - reshape2_13.tmp_0 + - teacher_reshape2_15.tmp_0 # + - reshape2_15.tmp_0 + - teacher_reshape2_17.tmp_0 # + - reshape2_17.tmp_0 + - teacher_reshape2_19.tmp_0 # + - reshape2_19.tmp_0 + - teacher_reshape2_21.tmp_0 # + - reshape2_21.tmp_0 + - teacher_depthwise_conv2d_14.tmp_0 # block2 + - depthwise_conv2d_14.tmp_0 + - teacher_depthwise_conv2d_15.tmp_0 + - depthwise_conv2d_15.tmp_0 + - teacher_reshape2_23.tmp_0 #block1 + - reshape2_23.tmp_0 + - teacher_relu_30.tmp_0 # final_conv + - relu_30.tmp_0 + - teacher_bilinear_interp_v2_1.tmp_0 + - bilinear_interp_v2_1.tmp_0 + merge_feed: true + teacher_model_dir: ./inference_model + teacher_model_filename: inference.pdmodel + teacher_params_filename: inference.pdiparams +Quantization: + activation_bits: 8 + is_full_quantize: false + not_quant_pattern: + - skip_quant + quantize_op_types: + - conv2d + - depthwise_conv2d + weight_bits: 8 +TrainConfig: + epochs: 1 + eval_iter: 400 + learning_rate: 0.0005 + optimizer: SGD + optim_args: + weight_decay: 4.0e-05 \ No newline at end of file diff --git a/demo/auto_compression/pp-humanseg/configs/humanseg_sparse_dis.yaml b/demo/auto_compression/pp-humanseg/configs/humanseg_sparse_dis.yaml new file mode 100644 index 00000000..47b1b647 --- /dev/null +++ b/demo/auto_compression/pp-humanseg/configs/humanseg_sparse_dis.yaml @@ -0,0 +1,61 @@ +Distillation: + distill_lambda: 1.0 + distill_loss: l2_loss + distill_node_pair: + - teacher_reshape2_1.tmp_0 + - reshape2_1.tmp_0 + - teacher_reshape2_3.tmp_0 + - reshape2_3.tmp_0 + - teacher_reshape2_5.tmp_0 + - reshape2_5.tmp_0 + - teacher_reshape2_7.tmp_0 #block1 + - reshape2_7.tmp_0 + - teacher_reshape2_9.tmp_0 + - reshape2_9.tmp_0 + - teacher_reshape2_11.tmp_0 + - reshape2_11.tmp_0 + - teacher_reshape2_13.tmp_0 + - reshape2_13.tmp_0 + - teacher_reshape2_15.tmp_0 + - reshape2_15.tmp_0 + - teacher_reshape2_17.tmp_0 + - reshape2_17.tmp_0 + - teacher_reshape2_19.tmp_0 + - reshape2_19.tmp_0 + - teacher_reshape2_21.tmp_0 + - reshape2_21.tmp_0 + - teacher_depthwise_conv2d_14.tmp_0 # block2 + - depthwise_conv2d_14.tmp_0 + - teacher_depthwise_conv2d_15.tmp_0 + - depthwise_conv2d_15.tmp_0 + - teacher_reshape2_23.tmp_0 #block1 + - reshape2_23.tmp_0 + - teacher_relu_30.tmp_0 # final_conv + - relu_30.tmp_0 + - teacher_bilinear_interp_v2_1.tmp_0 + - bilinear_interp_v2_1.tmp_0 + merge_feed: true + teacher_model_dir: ./inference_model + teacher_model_filename: inference.pdmodel + teacher_params_filename: inference.pdiparams +UnstructurePrune: + prune_strategy: gmp + prune_mode: ratio + pruned_ratio: 0.75 + gmp_config: + stable_iterations: 0 + pruning_iterations: 4500 + tunning_iterations: 4500 + resume_iteration: -1 + pruning_steps: 100 + initial_ratio: 0.15 + prune_params_type: conv1x1_only + local_sparsity: True +TrainConfig: + epochs: 14 + eval_iter: 400 + learning_rate: 5.0e-03 + optim_args: + weight_decay: 0.0005 + optimizer: SGD + \ No newline at end of file diff --git a/demo/auto_compression/pp-humanseg/run.py b/demo/auto_compression/pp-humanseg/run.py new file mode 100644 index 00000000..ef81ae72 --- /dev/null +++ b/demo/auto_compression/pp-humanseg/run.py @@ -0,0 +1,190 @@ +import os +import argparse +import random +import paddle +import numpy as np +import paddleseg.transforms as T +from paddleseg.datasets import Dataset +from paddleseg.utils import worker_init_fn +from paddleslim.auto_compression.config_helpers import load_config +from paddleslim.auto_compression import AutoCompression + +from paddleseg.core.infer import reverse_transform +from paddleseg.utils import metrics +import paddle.nn.functional as F +import cv2 +import paddle.fluid as fluid + + +def parse_args(): + parser = argparse.ArgumentParser(description='Model training') + parser.add_argument( + '--model_dir', + type=str, + default=None, + help="inference model directory.") + parser.add_argument( + '--model_filename', + type=str, + default=None, + help="inference model filename.") + parser.add_argument( + '--params_filename', + type=str, + default=None, + help="inference params filename.") + parser.add_argument( + '--save_dir', + type=str, + default=None, + help="directory to save compressed model.") + parser.add_argument( + '--config_path', + type=str, + default=None, + help="path of compression strategy config.") + return parser.parse_args() + + +def eval_function(exe, compiled_test_program, test_feed_names, test_fetch_list): + + nranks = paddle.distributed.ParallelEnv().local_rank + + batch_sampler = paddle.io.DistributedBatchSampler( + eval_dataset, batch_size=1, shuffle=False, drop_last=False) + loader = paddle.io.DataLoader( + eval_dataset, + batch_sampler=batch_sampler, + num_workers=1, + return_list=True, ) + + total_iters = len(loader) + intersect_area_all = 0 + pred_area_all = 0 + label_area_all = 0 + + print("Start evaluating (total_samples: {}, total_iters: {})...".format( + len(eval_dataset), total_iters)) + print("nranks:", nranks) + + for iter, (image, label) in enumerate(loader): + paddle.enable_static() + + label = np.array(label).astype('int64') + ori_shape = np.array(label).shape[-2:] + + image = np.array(image) + logits = exe.run(compiled_test_program, + feed={test_feed_names[0]: image}, + fetch_list=test_fetch_list, + return_numpy=True) + + paddle.disable_static() + logit = logits[0] + + logit = reverse_transform( + paddle.to_tensor(logit), + ori_shape, + eval_dataset.transforms.transforms, + mode='bilinear') + + pred = paddle.argmax( + paddle.to_tensor(logit), axis=1, keepdim=True, dtype='int32') + + intersect_area, pred_area, label_area = metrics.calculate_area( + pred, + paddle.to_tensor(label), + eval_dataset.num_classes, + ignore_index=eval_dataset.ignore_index) + + if nranks > 1: + intersect_area_list = [] + pred_area_list = [] + label_area_list = [] + paddle.distributed.all_gather(intersect_area_list, intersect_area) + paddle.distributed.all_gather(pred_area_list, pred_area) + paddle.distributed.all_gather(label_area_list, label_area) + + # Some image has been evaluated and should be eliminated in last iter + if (iter + 1) * nranks > len(eval_dataset): + valid = len(eval_dataset) - iter * nranks + intersect_area_list = intersect_area_list[:valid] + pred_area_list = pred_area_list[:valid] + label_area_list = label_area_list[:valid] + + for i in range(len(intersect_area_list)): + intersect_area_all = intersect_area_all + intersect_area_list[i] + pred_area_all = pred_area_all + pred_area_list[i] + label_area_all = label_area_all + label_area_list[i] + else: + intersect_area_all = intersect_area_all + intersect_area + pred_area_all = pred_area_all + pred_area + label_area_all = label_area_all + label_area + + class_iou, miou = metrics.mean_iou(intersect_area_all, pred_area_all, + label_area_all) + class_acc, acc = metrics.accuracy(intersect_area_all, pred_area_all) + kappa = metrics.kappa(intersect_area_all, pred_area_all, label_area_all) + class_dice, mdice = metrics.dice(intersect_area_all, pred_area_all, + label_area_all) + + infor = "[EVAL] #Images: {} mIoU: {:.4f} Acc: {:.4f} Kappa: {:.4f} Dice: {:.4f}".format( + len(eval_dataset), miou, acc, kappa, mdice) + print(infor) + + paddle.enable_static() + return miou + + +def reader_wrapper(reader): + def gen(): + for i, data in enumerate(reader()): + imgs = np.array(data[0]) + yield {"x": imgs} + + return gen + + +if __name__ == '__main__': + + args = parse_args() + + transforms = [T.RandomPaddingCrop(crop_size=(512, 512)), T.Normalize()] + train_dataset = Dataset( + transforms=transforms, + dataset_root='dataset_root', # Need to fill in + num_classes=2, + train_path='train_path', # Need to fill in + mode='train') + eval_dataset = Dataset( + transforms=transforms, + dataset_root='dataset_root', # Need to fill in + num_classes=2, + train_path='val_path', # Need to fill in + mode='val') + + batch_sampler = paddle.io.DistributedBatchSampler( + train_dataset, batch_size=128, shuffle=True, drop_last=True) + + train_loader = paddle.io.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + num_workers=2, + return_list=True, + worker_init_fn=worker_init_fn, ) + train_dataloader = reader_wrapper(train_loader) + + # set auto_compression + compress_config, train_config = load_config(args.config_path) + + ac = AutoCompression( + model_dir=args.model_dir, + model_filename=args.model_filename, + params_filename=args.param_filename, + save_dir=args.save_dir, + strategy_config=compress_config, + train_config=train_config, + train_dataloader=train_dataloader, + eval_callback=eval_function) + + ac.compress() -- GitLab