Add demo of ACT for some segmentation models (#1186)

demo/auto_compression/semantic_segmentation/README.md

@@ -19,16 +19,37 @@
 
 - [PP-HumanSeg-Lite](https://github.com/PaddlePaddle/PaddleSeg/tree/develop/contrib/PP-HumanSeg#portrait-segmentation)
 
-| 模型 | 策略  | Total IoU | 耗时(ms)<br>thread=1 | 配置文件 | Inference模型  |
-|:-----:|:-----:|:----------:|:---------:| :------:| :------:|
-| PP-HumanSeg-Lite | Baseline |  0.9287 | 56.363 | - | [model](https://paddleseg.bj.bcebos.com/dygraph/ppseg/ppseg_lite_portrait_398x224_with_softmax.tar.gz) |
-| PP-HumanSeg-Lite | 非结构化稀疏+蒸馏 |  0.9235 | 37.712 | [config](./configs/pp_human_sparse_dis.yaml)| - |
-| PP-HumanSeg-Lite | 量化+蒸馏 |  0.9284 | 49.656 | [config](./configs/pp_human_sparse_dis.yaml) | - |
+| 模型 | 策略  | Total IoU | ARM CPU耗时(ms)<br>thread=1 |Nvidia GPU耗时(ms)| 配置文件 | Inference模型  |
+|:-----:|:-----:|:----------:|:---------:| :------:|:------:|:------:|
+| PP-HumanSeg-Lite | Baseline |  92.87 | 56.363 |-| - | [model](https://paddleseg.bj.bcebos.com/dygraph/ppseg/ppseg_lite_portrait_398x224_with_softmax.tar.gz) |
+| PP-HumanSeg-Lite | 非结构化稀疏+蒸馏 |  92.35 | 37.712 |-| [config](./configs/pp_human/pp_human_sparse.yaml)| - |
+| PP-HumanSeg-Lite | 量化+蒸馏 |  92.84 | 49.656 |-| [config](./configs/pp_human/pp_human_qat.yaml) | - |
+| PP-Liteseg | Baseline |  77.04| - | 1.425| - |[model](https://paddleseg.bj.bcebos.com/tipc/easyedge/RES-paddle2-PPLIteSegSTDC1.zip)|
+| PP-Liteseg | 量化训练 |  76.93 | - | 1.158|[config](./configs/pp_liteseg/pp_liteseg_qat.yaml) | - |
+| HRNet | Baseline |  78.97 | - |8.188|-| [model](https://paddleseg.bj.bcebos.com/tipc/easyedge/RES-paddle2-HRNetW18-Seg.zip)|
+| HRNet | 量化训练 |  78.90 | - |5.812| [config](./configs/hrnet/hrnet_qat.yaml) | - |
+| UNet | Baseline | 65.00  | - |15.291|-| [model](https://paddleseg.bj.bcebos.com/tipc/easyedge/RES-paddle2-UNet.zip) |
+| UNet | 量化训练 |  64.93 | - |10.228| [config](./configs/unet/unet_qat.yaml) | - |
+| Deeplabv3-ResNet50 | Baseline |  79.90 | -|12.766| -| [model](https://paddleseg.bj.bcebos.com/tipc/easyedge/RES-paddle2-Deeplabv3-ResNet50.zip)|
+| Deeplabv3-ResNet50 | 量化训练 |  78.89 | - |8.839|[config](./configs/deeplabv3/deeplabv3_qat.yaml) | - |
 
-- 测试环境：`SDM710 2*A75(2.2GHz) 6*A55(1.7GHz)`；
-- 数据集：AISegment + PP-HumanSeg14K + 内部自建数据集。其中 AISegment 是开源数据集，可从[链接](https://github.com/aisegmentcn/matting_human_datasets)处获取；PP-HumanSeg14K 是 PaddleSeg 自建数据集，可从[官方渠道](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/contrib/PP-HumanSeg/paper.md#pp-humanseg14k-a-large-scale-teleconferencing-video-dataset)获取；内部数据集不对外公开。
+- ARM CPU测试环境：`SDM710 2*A75(2.2GHz) 6*A55(1.7GHz)`；
 
-下面将以开源数据集为例介绍如何进行自动压缩。
+- Nvidia GPU测试环境：
+
+  - 硬件：NVIDIA Tesla T4 单卡
+  - 软件：CUDA 11.0, cuDNN 8.0, TensorRT 8.0
+  - 测试配置：batch_size: 40, max_seq_len: 128
+
+- PP-HumanSeg-Lite数据集
+
+  - 数据集：AISegment + PP-HumanSeg14K + 内部自建数据集。其中 AISegment 是开源数据集，可从[链接](https://github.com/aisegmentcn/matting_human_datasets)处获取；PP-HumanSeg14K 是 PaddleSeg 自建数据集，可从[官方渠道](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/contrib/PP-HumanSeg/paper.md#pp-humanseg14k-a-large-scale-teleconferencing-video-dataset)获取；内部数据集不对外公开。
+
+- PP-Liteseg，HRNet，UNet，Deeplabv3-ResNet50数据集
+
+  - cityscapes: 请从[cityscapes官网](https://www.cityscapes-dataset.com/login/)下载完整数据
+
+下面将以开源数据集为例介绍如何对PP-HumanSeg-Lite进行自动压缩。
 
 ## 3. 自动压缩流程
 
@@ -93,7 +114,7 @@ python run.py \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_auto.yaml' \
+    --config_path='configs/pp_humanseg/pp_humanseg_auto.yaml' \
     --deploy_hardware='SD710'
 
 # 多卡启动
@@ -103,7 +124,7 @@ python -m paddle.distributed.launch run.py \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_auto.yaml' \
+    --config_path='configs/pp_humanseg/pp_humanseg_auto.yaml' \
     --deploy_hardware='SD710'
 ```
 - 自行配置稀疏参数进行非结构化稀疏和蒸馏训练，配置参数含义详见[自动压缩超参文档](https://github.com/PaddlePaddle/PaddleSlim/blob/27dafe1c722476f1b16879f7045e9215b6f37559/demo/auto_compression/hyperparameter_tutorial.md)。具体命令如下所示：
@@ -115,7 +136,7 @@ python run.py \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_sparse_dis.yaml'
+    --config_path='configs/pp_humanseg/pp_humanseg_sparse.yaml'
 
 # 多卡启动
 export CUDA_VISIBLE_DEVICES=0,1
@@ -124,7 +145,7 @@ python -m paddle.distributed.launch run.py \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_sparse_dis.yaml'
+    --config_path='configs/pp_humanseg/pp_humanseg_sparse.yaml'
 ```
 
 - 自行配置量化参数进行量化和蒸馏训练，配置参数含义详见[自动压缩超参文档](https://github.com/PaddlePaddle/PaddleSlim/blob/27dafe1c722476f1b16879f7045e9215b6f37559/demo/auto_compression/hyperparameter_tutorial.md)。具体命令如下所示：
@@ -136,7 +157,7 @@ python run.py \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_quant_dis.yaml'
+    --config_path='configs/pp_humanseg/pp_humanseg_qat.yaml'
 
 # 多卡启动
 export CUDA_VISIBLE_DEVICES=0,1
@@ -145,7 +166,7 @@ python -m paddle.distributed.launch run.py \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_quant_dis.yaml'
+    --config_path='configs/pp_humanseg/pp_humanseg_qat.yaml'
 ```
 
 压缩完成后会在`save_dir`中产出压缩好的预测模型，可直接预测部署。

demo/auto_compression/semantic_segmentation/configs/dataset/cityscapes_1024x512_scale1.0.yml

+batch_size: 16
+train_dataset:
+  type: Cityscapes
+  dataset_root: data/cityscapes
+  transforms:
+    - type: ResizeStepScaling
+      min_scale_factor: 0.5
+      max_scale_factor: 2.0
+      scale_step_size: 0.25
+    - type: RandomPaddingCrop
+      crop_size: [1024, 512]
+    - type: RandomHorizontalFlip
+    - type: RandomDistort
+      brightness_range: 0.5
+      contrast_range: 0.5
+      saturation_range: 0.5
+    - type: Normalize
+  mode: train
+val_dataset:
+  type: Cityscapes
+  dataset_root: data/cityscapes
+  transforms:
+    - type: Normalize
+  mode: val

demo/auto_compression/semantic_segmentation/configs/deeplabv3/deeplabv3_qat.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - conv2d_123.tmp_1
+
+Quantization:
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+TrainConfig:
+  epochs: 10
+  eval_iter: 360
+  learning_rate: 0.0001
+  optimizer_builder:
+    optimizer: 
+      type: SGD
+    weight_decay: 0.0005

demo/auto_compression/semantic_segmentation/configs/hrnet/hrnet_qat.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - conv2d_613.tmp_1
+
+Quantization:
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+TrainConfig:
+  epochs: 20
+  eval_iter: 360
+  learning_rate: 0.0001
+  optimizer_builder:
+    optimizer:
+      type: SGD
+    weight_decay: 4.0e-05

demo/auto_compression/semantic_segmentation/configs/hrnet/hrnet_sparse.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node: 
+  - conv2d_613.tmp_1
+
+UnstructurePrune:
+  prune_strategy: gmp
+  prune_mode: ratio
+  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: 90
+  learning_rate: 
+    type: PiecewiseDecay
+    boundaries: [4500]
+    values: [0.0005, 0.00025]
+  optimizer_builder:
+    optimizer: 
+      type: SGD
+    weight_decay: 0.0005

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_auto.yaml → demo/auto_compression/semantic_segmentation/configs/pp_humanseg/pp_humanseg_auto.yaml

 Global:
-  reader_config: configs/pp_humanseg_lite.yaml
+  reader_config: configs/dataset/pp_humanseg_dataset.yaml
 
 TrainConfig:
   epochs: 14

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_quant_dis.yaml → demo/auto_compression/semantic_segmentation/configs/pp_humanseg/pp_humanseg_qat.yaml

 Global:
-  reader_config: configs/pp_humanseg_lite.yaml
+  reader_config: configs/dataset/pp_humanseg_lite.yaml
 
 Distillation:
   alpha: 1.0

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_sparse_dis.yaml → demo/auto_compression/semantic_segmentation/configs/pp_humanseg/pp_humanseg_sparse.yaml

 Global:
-  reader_config: configs/pp_humanseg_lite.yaml
+  reader_config: configs/dataset/pp_humanseg_lite.yaml
 
 Distillation:
   alpha: 1.0

demo/auto_compression/semantic_segmentation/configs/pp_liteseg/pp_liteseg_auto.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+TrainConfig:
+ epochs: 14
+ eval_iter: 90
+ learning_rate: 
+   type: PiecewiseDecay
+   boundaries: [900]
+   values: [0.001, 0.0005]
+ optimizer_builder:
+   optimizer: 
+     type: SGD
+   weight_decay: 0.0005  

demo/auto_compression/semantic_segmentation/configs/pp_liteseg/pp_liteseg_qat.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - conv2d_95.tmp_0
+
+Quantization:
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+TrainConfig:
+  epochs: 20
+  eval_iter: 180
+  learning_rate: 0.0005
+  optimizer_builder:
+    optimizer: 
+      type: SGD
+    weight_decay: 4.0e-05  
+

demo/auto_compression/semantic_segmentation/configs/pp_liteseg/pp_liteseg_sparse.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node: 
+  - conv2d_95.tmp_0
+
+UnstructurePrune:
+  prune_strategy: gmp
+  prune_mode: ratio
+  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: 50
+  eval_iter: 180
+  learning_rate: 
+    type: PiecewiseDecay
+    boundaries: [4500]
+    values: [0.0025, 0.00025]
+  optimizer_builder:
+    optimizer: 
+      type: SGD
+    weight_decay: 0.0005

demo/auto_compression/semantic_segmentation/configs/unet/unet_channel_prune.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - conv2d_37.tmp_1
+
+ChannelPrune:
+  # pruned_ratio: 裁剪比例
+  pruned_ratio: 0.25
+  # prune_params_name: 需要裁剪的参数名字
+  prune_params_name:
+  - conv2d_17.w_0
+  - conv2d_16.w_0
+  - conv2d_15.w_0
+  - conv2d_14.w_0
+  - conv2d_13.w_0
+  - conv2d_12.w_0
+  - conv2d_11.w_0
+  - conv2d_10.w_0
+  - conv2d_9.w_0
+  - conv2d_8.w_0
+  - conv2d_7.w_0
+  - conv2d_6.w_0
+  - conv2d_5.w_0
+  - conv2d_4.w_0
+  - conv2d_3.w_0
+  - conv2d_2.w_0
+  - conv2d_1.w_0
+  - conv2d_0.w_0
+  # criterion: 评估一个卷积层内通道重要性所参考的指标
+  criterion: l1_norm
+TrainConfig:
+  epochs: 861
+  eval_iter: 180
+  learning_rate: 
+    type: PiecewiseDecay
+    boundaries: [1000, 2000, 3000, 40000, 80000, 120000]
+    values: [0.00001, 0.0001, 0.001, 0.01, 0.001, 0.0001, 0.00001]
+  optimizer_builder:
+    optimizer: 
+      type: Momentum
+      momentum: 0.9
+    weight_decay: 4.0e-05

demo/auto_compression/semantic_segmentation/configs/unet/unet_qat.yaml

+Global:
+  reader_config: configs/dataset/cityscapes_1024x512_scale1.0.yml
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - conv2d_37.tmp_1
+
+Quantization:
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+
+TrainConfig:
+  epochs: 20
+  eval_iter: 360
+  learning_rate: 0.0001
+  optimizer: SGD
+  optim_args:
+    weight_decay: 4.0e-05