[ACT] Rename 'Quantization' to 'QuantAware'. (#1536)

zzjjay · web-flow · commit c90ab602d7d6 · 2022-11-23T10:17:14.000+08:00
* [ACT] Rename 'Quantization' to 'QuantAware'.

* Fix examples.

* Fix tests.
diff --git a/example/auto_compression/README.md b/example/auto_compression/README.md
@@ -178,7 +178,7 @@ ac = AutoCompression(
     params_filename="inference.pdiparams",
     save_dir="MobileNetV1_quant",
     config={"QuantPost": {}, "HyperParameterOptimization": {'ptq_algo': ['avg'], 'max_quant_count': 3}},
-    ### config={"Quantization": {}, "Distillation": {}}, ### 如果您的系统为Windows系统, 请使用当前这一行配置
+    ### config={"QuantAware": {}, "Distillation": {}}, ### 如果您的系统为Windows系统, 请使用当前这一行配置
     train_dataloader=train_loader,
     eval_dataloader=train_loader)
 ac.compress()
diff --git a/example/auto_compression/detection/configs/picodet_s_qat_dis.yaml b/example/auto_compression/detection/configs/picodet_s_qat_dis.yaml
@@ -18,7 +18,7 @@ Distillation:
   - conv2d_154.tmp_1
   - tmp_8
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
   weight_bits: 8
diff --git a/example/auto_compression/detection/configs/ppyoloe_l_qat_dis.yaml b/example/auto_compression/detection/configs/ppyoloe_l_qat_dis.yaml
@@ -11,7 +11,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
diff --git a/example/auto_compression/detection/configs/ppyoloe_s_qat_dis.yaml b/example/auto_compression/detection/configs/ppyoloe_s_qat_dis.yaml
@@ -11,7 +11,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
diff --git a/example/auto_compression/detection/configs/ssd_mbv1_voc_qat_dis.yaml b/example/auto_compression/detection/configs/ssd_mbv1_voc_qat_dis.yaml
@@ -13,7 +13,7 @@ Distillation:
   - concat_2.tmp_0
   - concat_1.tmp_0
 
-Quantization:
+QuantAware:
   use_pact: True
   weight_quantize_type: 'channel_wise_abs_max'
   activation_quantize_type: 'moving_average_abs_max'
diff --git a/example/auto_compression/detection/configs/tinypose_qat_dis.yaml b/example/auto_compression/detection/configs/tinypose_qat_dis.yaml
@@ -12,7 +12,7 @@ Distillation:
   node:
   -  conv2d_441.tmp_0
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
   weight_quantize_type: 'channel_wise_abs_max'  # 'abs_max' is layer wise quant
diff --git a/example/auto_compression/detection/configs/yolov3_mbv1_qat_dis.yaml b/example/auto_compression/detection/configs/yolov3_mbv1_qat_dis.yaml
@@ -13,7 +13,7 @@ Distillation:
   - conv2d_85.tmp_0
   - conv2d_86.tmp_0
 
-Quantization:
+QuantAware:
   activation_quantize_type: 'range_abs_max'
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/hyperparameter_tutorial.md b/example/auto_compression/hyperparameter_tutorial.md
@@ -3,11 +3,11 @@
 
 ## 1.1 各压缩方法超参解析
 
-### 1.1.1 量化训练（quantization）
+### 1.1.1 量化训练（quant aware）
 
 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（mul, matmul_v2）。以下为只量化卷积层的示例：
 ```yaml
-Quantization:
+QuantAware:
     use_pact: false                               # 量化训练是否使用PACT方法
     weight_quantize_type: 'channel_wise_abs_max'  # 权重量化方式
     quantize_op_types: [conv2d, depthwise_conv2d] # 量化OP列表
diff --git a/example/auto_compression/image_classification/configs/EfficientNetB0/qat_dis.yaml b/example/auto_compression/image_classification/configs/EfficientNetB0/qat_dis.yaml
@@ -11,7 +11,7 @@ Distillation:
   node:
   - softmax_1.tmp_0
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/GhostNet_x1_0/qat_dis.yaml b/example/auto_compression/image_classification/configs/GhostNet_x1_0/qat_dis.yaml
@@ -10,7 +10,7 @@ Distillation:
   loss: l2
   node:
   - softmax_0.tmp_0
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/InceptionV3/qat_dis.yaml b/example/auto_compression/image_classification/configs/InceptionV3/qat_dis.yaml
@@ -12,7 +12,7 @@ Distillation:
   loss: l2
   node:
   - softmax_1.tmp_0
-Quantization:
+QuantAware:
   is_full_quantize: false
   activation_quantize_type: moving_average_abs_max
   weight_quantize_type: channel_wise_abs_max
diff --git a/example/auto_compression/image_classification/configs/MobileNetV1/qat_dis.yaml b/example/auto_compression/image_classification/configs/MobileNetV1/qat_dis.yaml
@@ -10,7 +10,7 @@ Distillation:
   loss: l2
   node:
   - softmax_0.tmp_0
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/MobileNetV3_large_x1_0/qat_dis.yaml b/example/auto_compression/image_classification/configs/MobileNetV3_large_x1_0/qat_dis.yaml
@@ -9,7 +9,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/PPHGNet_tiny/qat_dis.yaml b/example/auto_compression/image_classification/configs/PPHGNet_tiny/qat_dis.yaml
@@ -11,7 +11,7 @@ Distillation:
   node:
   - softmax_1.tmp_0
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/PPLCNetV2_base/qat_dis.yaml b/example/auto_compression/image_classification/configs/PPLCNetV2_base/qat_dis.yaml
@@ -11,7 +11,7 @@ Distillation:
   node:
   - softmax_1.tmp_0
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/PPLCNet_x1_0/qat_dis.yaml b/example/auto_compression/image_classification/configs/PPLCNet_x1_0/qat_dis.yaml
@@ -10,7 +10,7 @@ Distillation:
   loss: l2
   node:
   - softmax_1.tmp_0
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/ResNet50_vd/qat_dis.yaml b/example/auto_compression/image_classification/configs/ResNet50_vd/qat_dis.yaml
@@ -11,7 +11,7 @@ Distillation:
   node:
   - softmax_0.tmp_0
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/ShuffleNetV2_x1_0/qat_dis.yaml b/example/auto_compression/image_classification/configs/ShuffleNetV2_x1_0/qat_dis.yaml
@@ -10,7 +10,7 @@ Distillation:
   loss: l2
   node:
   - softmax_0.tmp_0
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/image_classification/configs/SqueezeNet1_0/qat_dis.yaml b/example/auto_compression/image_classification/configs/SqueezeNet1_0/qat_dis.yaml
@@ -10,7 +10,7 @@ Distillation:
   loss: l2
   node:
   - softmax_0.tmp_0
-Quantization:
+QuantAware:
   activation_bits: 8
   is_full_quantize: false
   activation_quantize_type: moving_average_abs_max
diff --git a/example/auto_compression/image_classification/configs/SwinTransformer_base_patch4_window7_224/qat_dis.yaml b/example/auto_compression/image_classification/configs/SwinTransformer_base_patch4_window7_224/qat_dis.yaml
@@ -10,7 +10,7 @@ Distillation:
   loss: l2
   node:
   - softmax_48.tmp_0
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/ocr/configs/ppocrv3_det_qat_dist.yaml b/example/auto_compression/ocr/configs/ppocrv3_det_qat_dist.yaml
@@ -9,7 +9,7 @@ Distillation:
   alpha: 1.0
   loss: l2
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/auto_compression/pytorch_yolo_series/configs/yolov5s_qat_dis.yaml b/example/auto_compression/pytorch_yolo_series/configs/yolov5s_qat_dis.yaml
@@ -12,7 +12,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
diff --git a/example/auto_compression/pytorch_yolo_series/configs/yolov6s_qat_dis.yaml b/example/auto_compression/pytorch_yolo_series/configs/yolov6s_qat_dis.yaml
@@ -12,7 +12,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   activation_quantize_type: 'moving_average_abs_max'
   quantize_op_types:
diff --git a/example/auto_compression/pytorch_yolo_series/configs/yolov6s_v2_qat_dis.yaml b/example/auto_compression/pytorch_yolo_series/configs/yolov6s_v2_qat_dis.yaml
@@ -13,7 +13,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   activation_quantize_type: 'moving_average_abs_max'
   quantize_op_types:
diff --git a/example/auto_compression/pytorch_yolo_series/configs/yolov7_qat_dis.yaml b/example/auto_compression/pytorch_yolo_series/configs/yolov7_qat_dis.yaml
@@ -12,7 +12,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   activation_quantize_type: 'moving_average_abs_max'
   quantize_op_types:
diff --git a/example/auto_compression/pytorch_yolo_series/configs/yolov7_tiny_qat_dis.yaml b/example/auto_compression/pytorch_yolo_series/configs/yolov7_tiny_qat_dis.yaml
@@ -12,7 +12,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:
+QuantAware:
   onnx_format: true
   activation_quantize_type: 'moving_average_abs_max'
   quantize_op_types:
diff --git a/example/auto_compression/semantic_segmentation/configs/BiSeNetV2/BiSeNetV2_qat.yaml b/example/auto_compression/semantic_segmentation/configs/BiSeNetV2/BiSeNetV2_qat.yaml
@@ -11,7 +11,7 @@ Distillation:
   node:
   - conv2d_103.tmp_1
 
-Quantization:
+QuantAware:
   onnx_format: True
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/semantic_segmentation/configs/deeplabv3/deeplabv3_qat.yaml b/example/auto_compression/semantic_segmentation/configs/deeplabv3/deeplabv3_qat.yaml
@@ -11,7 +11,7 @@ Distillation:
   node:
   - conv2d_123.tmp_1
 
-Quantization:
+QuantAware:
   onnx_format: True
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/semantic_segmentation/configs/hrnet/hrnet_qat.yaml b/example/auto_compression/semantic_segmentation/configs/hrnet/hrnet_qat.yaml
@@ -10,7 +10,7 @@ Distillation:
   node:
   - conv2d_613.tmp_1
 
-Quantization:
+QuantAware:
   onnx_format: True
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/semantic_segmentation/configs/pp_humanseg/pp_humanseg_qat.yaml b/example/auto_compression/semantic_segmentation/configs/pp_humanseg/pp_humanseg_qat.yaml
@@ -10,7 +10,7 @@ Distillation:
   node:
   - batch_norm_47.tmp_2
 
-Quantization:
+QuantAware:
   onnx_format: True
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/semantic_segmentation/configs/pp_liteseg/pp_liteseg_qat.yaml b/example/auto_compression/semantic_segmentation/configs/pp_liteseg/pp_liteseg_qat.yaml
@@ -10,7 +10,7 @@ Distillation:
   node:
   - conv2d_95.tmp_0
 
-Quantization:
+QuantAware:
   onnx_format: True
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/semantic_segmentation/configs/unet/unet_qat.yaml b/example/auto_compression/semantic_segmentation/configs/unet/unet_qat.yaml
@@ -10,7 +10,7 @@ Distillation:
   node:
   - conv2d_37.tmp_1
 
-Quantization:
+QuantAware:
   onnx_format: True
   quantize_op_types:
   - conv2d
diff --git a/example/auto_compression/tensorflow_mobilenet/configs/mbv1_qat_dis.yaml b/example/auto_compression/tensorflow_mobilenet/configs/mbv1_qat_dis.yaml
@@ -38,7 +38,7 @@ Distillation:
   - batch_norm_26.tmp_3
   - conv2d_42.tmp_1
 
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   is_full_quantize: false
diff --git a/example/full_quantization/image_classification/configs/mobilenetv3_large_qat_dis.yaml b/example/full_quantization/image_classification/configs/mobilenetv3_large_qat_dis.yaml
@@ -9,7 +9,7 @@ Global:
 Distillation:
   alpha: 1.0
   loss: soft_label
-Quantization:
+QuantAware:
   use_pact: true
   activation_bits: 8
   activation_quantize_type: moving_average_abs_max
diff --git a/example/full_quantization/picodet/configs/picodet_npu.yaml b/example/full_quantization/picodet/configs/picodet_npu.yaml
@@ -15,7 +15,7 @@ PTQ:   # Post Training Quantization
   batch_size: 32
   batch_nums: 10
 
-Quantization:   # Auto Compression
+QuantAware:   # Auto Compression
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
   weight_bits: 8
diff --git a/example/full_quantization/picodet/configs/picodet_npu_with_postprocess.yaml b/example/full_quantization/picodet/configs/picodet_npu_with_postprocess.yaml
@@ -15,7 +15,7 @@ PTQ:   # Post Training Quantization
   batch_size: 32
   batch_nums: 10
 
-Quantization:   # Auto Compression
+QuantAware:   # Auto Compression
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
   weight_bits: 8
diff --git a/example/full_quantization/ppyoloe/configs/ppyoloe_s_416_qat_dis.yaml b/example/full_quantization/ppyoloe/configs/ppyoloe_s_416_qat_dis.yaml
@@ -20,7 +20,7 @@ Distillation:
   alpha: 1.0
   loss: soft_label
 
-Quantization:   # Auto Compression
+QuantAware:   # Auto Compression
   onnx_format: true
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
diff --git a/example/full_quantization/ppyoloe/configs/ppyoloe_s_qat_dis.yaml b/example/full_quantization/ppyoloe/configs/ppyoloe_s_qat_dis.yaml
@@ -16,7 +16,7 @@ PTQ:   # Post Training Quantization
   batch_size: 10
   batch_nums: 10
 
-Quantization:   # Auto Compression
+QuantAware:   # Auto Compression
   onnx_format: true
   use_pact: true
   activation_quantize_type: 'moving_average_abs_max'
diff --git a/paddleslim/auto_compression/__init__.py b/paddleslim/auto_compression/__init__.py
@@ -20,7 +20,8 @@
 
 __all__ = [
     "AutoCompression",
-    "Quantization",
+    "QuantAware",
+    "QuantPost",
     "Distillation",
     "MultiTeacherDistillation",
     "HyperParameterOptimization",
diff --git a/paddleslim/auto_compression/auto_strategy.py b/paddleslim/auto_compression/auto_strategy.py
@@ -125,17 +125,17 @@ def create_strategy_config(strategy_str, model_type):
         ### only platform is linux can use smac to do hyperparameter optimization
         ### choose quant_aware to do quantization in other platform
         if platform.system().lower() == 'linux':
-            quant_config = Quantization(**default_quant_config)
+            quant_config = QuantAware(**default_quant_config)
             hpo_config = HyperParameterOptimization(**hpo_config_tester)
             configs.append({
                 'QuantPost': quant_config,
                 'HyperParameterOptimization': hpo_config
             })
         else:
-            quant_config = Quantization(**default_quant_config)
+            quant_config = QuantAware(**default_quant_config)
             dis_config = Distillation()
             configs.append({
-                'Quantization': quant_config,
+                'QuantAware': quant_config,
                 'Distillation': dis_config
             })
 
@@ -248,7 +248,7 @@ def get_final_quant_config(ptq_loss, model_type=None):
         return None
     ### if emd loss less than MAGIC_MAX_EMD_DISTANCE, select quant_post & hpo.
     elif ptq_loss < MAGIC_MAX_EMD_DISTANCE:
-        quant_config = Quantization(**default_quant_config)
+        quant_config = QuantAware(**default_quant_config)
         hpo_config = HyperParameterOptimization(**default_hpo_config)
         configs = [{
             'QuantPost': quant_config,
@@ -257,9 +257,9 @@ def get_final_quant_config(ptq_loss, model_type=None):
 
     ### if emd loss greater than MAGIC_MAX_EMD_DISTANCE, select qat & dist.
     else:
-        quant_config = Quantization(**default_quant_config)
+        quant_config = QuantAware(**default_quant_config)
         dis_config = Distillation()
-        configs = [{'Quantization': quant_config, 'Distillation': dis_config}]
+        configs = [{'QuantAware': quant_config, 'Distillation': dis_config}]
         _logger.info("Start Quantization and Distillation Training.")
 
     return configs
diff --git a/paddleslim/auto_compression/compressor.py b/paddleslim/auto_compression/compressor.py
diff --git a/paddleslim/auto_compression/strategy_config.py b/paddleslim/auto_compression/strategy_config.py
diff --git a/tests/act/qat_dist_train.yaml b/tests/act/qat_dist_train.yaml
