Upgrade ImageNet Demo (PaddlePaddle#1233)

* Upgrade ImageNet Demo

* Upgrade ImageNet Demo

* Upgrade ImageNet Demo

Author: RachelXu7

demo/imagenet_reader.py

@@ -11,6 +11,7 @@
 np.random.seed(0)
 
 DATA_DIM = 224
+RESIZE_DIM = 256
 
 THREAD = 16
 BUF_SIZE = 10240
@@ -34,8 +35,8 @@ def crop_image(img, target_size, center):
     width, height = img.size
     size = target_size
     if center == True:
-        w_start = (width - size) / 2
-        h_start = (height - size) / 2
+        w_start = (width - size) // 2
+        h_start = (height - size) // 2
     else:
         w_start = np.random.randint(0, width - size + 1)
         h_start = np.random.randint(0, height - size + 1)
@@ -98,7 +99,7 @@ def random_color(img, lower=0.5, upper=1.5):
     return img
 
 
-def process_image(sample, mode, color_jitter, rotate):
+def process_image(sample, mode, color_jitter, rotate, crop_size, resize_size):
     img_path = sample[0]
 
     try:
@@ -108,10 +109,10 @@ def process_image(sample, mode, color_jitter, rotate):
         return None
     if mode == 'train':
         if rotate: img = rotate_image(img)
-        img = random_crop(img, DATA_DIM)
+        img = random_crop(img, crop_size)
     else:
-        img = resize_short(img, target_size=256)
-        img = crop_image(img, target_size=DATA_DIM, center=True)
+        img = resize_short(img, target_size=resize_size)
+        img = crop_image(img, target_size=crop_size, center=True)
     if mode == 'train':
         if color_jitter:
             img = distort_color(img)
@@ -185,9 +186,15 @@ def test(data_dir=DATA_DIR):
 
 
 class ImageNetDataset(Dataset):
-    def __init__(self, data_dir=DATA_DIR, mode='train'):
+    def __init__(self,
+                 data_dir=DATA_DIR,
+                 mode='train',
+                 crop_size=DATA_DIM,
+                 resize_size=RESIZE_DIM):
         super(ImageNetDataset, self).__init__()
         self.data_dir = data_dir
+        self.crop_size = crop_size
+        self.resize_size = resize_size
         train_file_list = os.path.join(data_dir, 'train_list.txt')
         val_file_list = os.path.join(data_dir, 'val_list.txt')
         test_file_list = os.path.join(data_dir, 'test_list.txt')
@@ -211,21 +218,27 @@ def __getitem__(self, index):
                 [data_path, sample[1]],
                 mode='train',
                 color_jitter=False,
-                rotate=False)
+                rotate=False,
+                crop_size=self.crop_size,
+                resize_size=self.resize_size)
             return data, np.array([label]).astype('int64')
         elif self.mode == 'val':
             data, label = process_image(
                 [data_path, sample[1]],
                 mode='val',
                 color_jitter=False,
-                rotate=False)
+                rotate=False,
+                crop_size=self.crop_size,
+                resize_size=self.resize_size)
             return data, np.array([label]).astype('int64')
         elif self.mode == 'test':
             data = process_image(
                 [data_path, sample[1]],
                 mode='test',
                 color_jitter=False,
-                rotate=False)
+                rotate=False,
+                crop_size=self.crop_size,
+                resize_size=self.resize_size)
             return data
 
     def __len__(self):

example/auto_compression/image_classification/README.md

@@ -24,25 +24,25 @@
 | 模型 | 策略 | Top-1 Acc | GPU 耗时(ms) | ARM CPU 耗时(ms) |
 |:------:|:------:|:------:|:------:|:------:|
 | MobileNetV1 | Baseline | 70.90 | - | 33.15 |
-| MobileNetV1 | 量化+蒸馏 | 70.49 | - | 13.64 |
+| MobileNetV1 | 量化+蒸馏 | 70.57 | - | 13.64 |
 | ResNet50_vd | Baseline | 79.12 | 3.19 | - |
-| ResNet50_vd | 量化+蒸馏 | 78.55 | 0.92 | - |
+| ResNet50_vd | 量化+蒸馏 | 78.74 | 0.92 | - |
 | ShuffleNetV2_x1_0 | Baseline | 68.65 | - | 10.43 |
-| ShuffleNetV2_x1_0 | 量化+蒸馏 | 67.78 | - | 5.51 |
+| ShuffleNetV2_x1_0 | 量化+蒸馏 | 68.32 | - | 5.51 |
 | SqueezeNet1_0_infer | Baseline | 59.60 | - | 35.98 |
-| SqueezeNet1_0_infer | 量化+蒸馏 | 59.13 | - | 16.96 |
+| SqueezeNet1_0_infer | 量化+蒸馏 | 59.45 | - | 16.96 |
 | PPLCNetV2_base | Baseline | 76.86 | - | 36.50 |
 | PPLCNetV2_base | 量化+蒸馏 | 76.43 | - | 15.79 |
 | PPHGNet_tiny | Baseline | 79.59 | 2.82 | - |
-| PPHGNet_tiny | 量化+蒸馏 | 79.19 | 0.98 | - |
+| PPHGNet_tiny | 量化+蒸馏 | 79.20 | 0.98 | - |
+| InceptionV3 | Baseline | 79.14 | 4.79 | - |
+| InceptionV3 | 量化+蒸馏 | 78.32 | 1.47 | - |
 | EfficientNetB0 | Baseline | 77.02 | 1.95 | - |
-| EfficientNetB0 | 量化+蒸馏 | 73.61 | 1.44 | - |
+| EfficientNetB0 | 量化+蒸馏 | 75.39 | 1.44 | - |
 | GhostNet_x1_0 | Baseline | 74.02 | 2.93 | - |
-| GhostNet_x1_0 | 量化+蒸馏 | 71.11 | 1.03 | - |
-| InceptionV3 | Baseline | 79.14 | 4.79 | - |
-| InceptionV3 | 量化+蒸馏 | 73.16 | 1.47 | - |
+| GhostNet_x1_0 | 量化+蒸馏 | 72.62 | 1.03 | - |
 | MobileNetV3_large_x1_0 | Baseline | 75.32 | - | 16.62 |
-| MobileNetV3_large_x1_0 | 量化+蒸馏 | 68.84 | - | 9.85 |
+| MobileNetV3_large_x1_0 | 量化+蒸馏 | 70.93 | - | 9.85 |
 
 - ARM CPU 测试环境：`SDM865(4xA77+4xA55)`
 - Nvidia GPU 测试环境：
@@ -119,7 +119,7 @@ python -m paddle.distributed.launch run.py --save_dir='./save_quant_mobilev1/' -
 
 准备好inference模型后，使用以下命令进行预测：
 ```shell
-python infer.py -c configs/infer.yaml
+python infer.py --config_path="configs/infer.yaml"
 ```
 
 在配置文件```configs/infer.yaml```中有以下字段用于配置预测参数：
@@ -134,7 +134,7 @@ python infer.py -c configs/infer.yaml
 - ```PostProcess.Topk.class_id_map_file```：数据集 label 的映射文件，默认为```./images/imagenet1k_label_list.txt```，该文件为 PaddleClas 所使用的 ImageNet 数据集 label 映射文件
 
 注意：
-- 请注意模型的输入数据尺寸，部分模型需要修改参数：```PreProcess.resize_short```, ```PreProcess.resize```
+- 请注意模型的输入数据尺寸，如InceptionV3输入尺寸为299，所以部分模型需要修改参数：```PreProcess.resize_short```, ```PreProcess.resize```
 - 如果希望提升评测模型速度，使用 ```GPU``` 评测时，建议开启 ```TensorRT``` 加速预测，使用 ```CPU``` 评测时，建议开启 ```MKL-DNN``` 加速预测
 - 若使用 TesorRT 预测引擎，需安装 ```WITH_TRT=ON``` 的Paddle，下载地址：[Python预测库](https://paddleinference.paddlepaddle.org.cn/master/user_guides/download_lib.html#python)
 

example/auto_compression/image_classification/configs/EfficientNetB0/qat_dis.yaml

@@ -15,7 +15,7 @@ Quantization:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
-  activation_quantize_type: range_abs_max
+  activation_quantize_type: moving_average_abs_max
   weight_quantize_type: channel_wise_abs_max
   not_quant_pattern:
   - skip_quant

example/auto_compression/image_classification/configs/GhostNet_x1_0/qat_dis.yaml

@@ -15,7 +15,7 @@ Quantization:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
-  activation_quantize_type: range_abs_max
+  activation_quantize_type: moving_average_abs_max
   weight_quantize_type: channel_wise_abs_max
   not_quant_pattern:
   - skip_quant
@@ -33,3 +33,4 @@ TrainConfig:
     optimizer:
       type: Momentum
     weight_decay: 0.00002
+  origin_metric: 0.7402

example/auto_compression/image_classification/configs/InceptionV3/prune_dis.yaml

@@ -4,6 +4,8 @@ Global:
   model_filename: inference.pdmodel
   params_filename: inference.pdiparams
   batch_size: 32
+  resize_size: 320
+  crop_size: 299
   data_dir: /ILSVRC2012
 
 Distillation:

example/auto_compression/image_classification/configs/InceptionV3/qat_dis.yaml

@@ -1,26 +1,29 @@
 Global:
   input_name: x
-  model_dir: InceptionV3_infer
+  model_dir: save_quant_inception
   model_filename: inference.pdmodel
   params_filename: inference.pdiparams
   batch_size: 32
-  data_dir: /ILSVRC2012
+  resize_size: 320
+  img_size: 299
+  data_dir: /workspace/dataset/ILSVRC2012
 
 Distillation:
-  alpha: 10.0
+  alpha: 1.0
   loss: l2
   node:
   - softmax_1.tmp_0
 Quantization:
   is_full_quantize: false
-  activation_quantize_type: range_abs_max
+  activation_quantize_type: moving_average_abs_max
   weight_quantize_type: channel_wise_abs_max
   not_quant_pattern:
   - skip_quant
   quantize_op_types:
   - conv2d
   - depthwise_conv2d
   weight_bits: 8
+  
 TrainConfig:
   epochs: 1
   eval_iter: 500

example/auto_compression/image_classification/configs/MobileNetV1/qat_dis.yaml

@@ -16,7 +16,7 @@ Quantization:
   activation_bits: 8
   is_full_quantize: false
   activation_quantize_type: moving_average_abs_max
-  weight_quantize_type: abs_max
+  weight_quantize_type: channel_wise_abs_max
   not_quant_pattern:
   - skip_quant
   quantize_op_types:

example/auto_compression/image_classification/configs/MobileNetV3_large_x1_0/qat_dis.yaml

@@ -12,10 +12,10 @@ Distillation:
   node:
   - softmax_0.tmp_0
 Quantization:
+  use_pact: true
   activation_bits: 8
   is_full_quantize: false
-  use_pact: true
-  activation_quantize_type: range_abs_max
+  activation_quantize_type: moving_average_abs_max
   weight_quantize_type: channel_wise_abs_max
   not_quant_pattern:
   - skip_quant
@@ -25,10 +25,10 @@ Quantization:
   weight_bits: 8
 TrainConfig:
   epochs: 1
-  eval_iter: 2000
+  eval_iter: 500
   learning_rate: 
     type: CosineAnnealingDecay 
-    learning_rate: 0.0001
+    learning_rate: 0.015
   optimizer_builder:
     optimizer:
       type: Momentum

example/auto_compression/image_classification/configs/PPHGNet_tiny/prune_dis.yaml

@@ -0,0 +1,37 @@
+Global:
+  input_name: x
+  model_dir: PPHGNet_tiny_infer
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  batch_size: 32
+  data_dir: /ILSVRC2012
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_1.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: 1
+  eval_iter: 500
+  learning_rate: 
+    type: CosineAnnealingDecay 
+    learning_rate: 0.015
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7959

example/auto_compression/image_classification/configs/PPHGNet_tiny/qat_dis.yaml

@@ -0,0 +1,36 @@
+Global:
+  input_name: x
+  model_dir: PPHGNet_tiny_infer
+  model_filename: inference.pdmodel
+  params_filename: inference.pdiparams
+  batch_size: 32
+  data_dir: /ILSVRC2012
+
+Distillation:
+  alpha: 1.0
+  loss: l2
+  node:
+  - softmax_1.tmp_0
+Quantization:
+  use_pact: true
+  activation_bits: 8
+  is_full_quantize: false
+  activation_quantize_type: moving_average_abs_max
+  weight_quantize_type: channel_wise_abs_max
+  not_quant_pattern:
+  - skip_quant
+  quantize_op_types:
+  - conv2d
+  - depthwise_conv2d
+  weight_bits: 8
+TrainConfig:
+  epochs: 1
+  eval_iter: 500
+  learning_rate: 
+    type: CosineAnnealingDecay 
+    learning_rate: 0.015
+  optimizer_builder:
+    optimizer:
+      type: Momentum
+    weight_decay: 0.00002
+  origin_metric: 0.7959