Add automatic region division and Fix pre-tensor quant (#1517)

* Fixed naming conflicts and fc layer quantification

* Update fine_tune.py

* Update readme

* Add automatic region division and Fix pre-tensor quant

* Update unit test file

* Fix bugs.

Co-authored-by: zhouzj <[email protected]>

Author: gushiqiao

docs/zh_cn/tutorials/quant/post_training_quantization.md

@@ -74,7 +74,7 @@ $$
 说明：
 - 如果想使用bias_correction，可以在PaddleSlim的[离线量化接口](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/static/quant/quantization_api.rst#quant_post_static)修改`bias_correction`参数为True即可，默认为False。
 - 如果想使用Adaround方法，可以在PaddleSlim的[离线量化接口](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/static/quant/quantization_api.rst#quant_post_static)修改`round_type`参数为`adaround`即可，默认为`round`。
-- 如果想使用BRECQ方法，可以在PaddleSlim的[量化重构接口](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/static/quant/quantization_api.rst#quant_post_static)修改`recon_level`参数为`regionn-wise`即可，默认为`layer-wise`。
+- 如果想使用BRECQ方法，可以在PaddleSlim的[量化重构接口](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/static/quant/quantization_api.rst#quant_post_static)修改`recon_level`参数为`region-wise`即可，默认为`layer-wise`。
 - 如果想使用QDrop方法，可以在PaddleSlim的[量化重构接口](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/static/quant/quantization_api.rst#quant_post_static)修改`simulate_activation_quant`参数为`True`即可，默认为`False`。
 
 ### 效果对比

example/post_training_quantization/pytorch_yolo_series/configs/yolov6s_fine_tune.yaml

@@ -1,5 +1,5 @@
 arch: YOLOv6
-model_dir: ./yolov6s.onnx
+model_dir:  ./yolov6s.onnx
 dataset_dir: /dataset/coco/
 model_filename: model.pdmodel
 params_filename: model.pdiparams
@@ -8,25 +8,3 @@ val_image_dir: val2017
 train_anno_path: annotations/instances_train2017.json
 val_anno_path: annotations/instances_val2017.json
 skip_tensor_list: None
-regions: [['x2paddle_image_arrays','relu_8.tmp_0'],
-          ['relu_8.tmp_0','relu_15.tmp_0'],
-          ['relu_15.tmp_0','relu_21.tmp_0'],
-          ['concat_1.tmp_0','relu_26.tmp_0'],
-          ['concat_2.tmp_0', 'relu_30.tmp_0'],
-          ['relu_30.tmp_0', 'concat_4.tmp_0'],
-          ['relu_30.tmp_0', 'relu_31.tmp_0'],
-          ['concat_3.tmp_0', 'relu_35.tmp_0'],
-          ['relu_35.tmp_0', 'relu_36.tmp_0'],
-          ['concat_5.tmp_0', 'concat_10.tmp_0'],
-          ['relu_35.tmp_0', 'concat_8.tmp_0']]
-region_weights_names: [['conv2d_0.w_0','conv2d_1.w_0','conv2d_2.w_0','conv2d_3.w_0','conv2d_4.w_0','conv2d_5.w_0','conv2d_6.w_0','conv2d_7.w_0','conv2d_8.w_0'],
-                    ['conv2d_9.w_0','conv2d_10.w_0','conv2d_11.w_0','conv2d_12.w_0','conv2d_13.w_0','conv2d_14.w_0','conv2d_15.w_0'],
-                    ['conv2d_16.w_0','conv2d_17.w_0','conv2d_18.w_0','conv2d_19.w_0','conv2d_20.w_0','conv2d_21.w_0'],
-                    ['conv2d_22.w_0','conv2d_23.w_0','conv2d_24.w_0','conv2d_25.w_0','conv2d_26.w_0'],
-                    ['conv2d_27.w_0','conv2d_28.w_0','conv2d_29.w_0','conv2d_30.w_0'],
-                    ['conv2d_32.w_0','conv2d_34.w_0','conv2d_35.w_0','conv2d_37.w_0','conv2d_38.w_0','conv2d_39.w_0'],
-                    ['conv2d_31.w_0'],
-                    ['conv2d_33.w_0','conv2d_36.w_0','conv2d_40.w_0','conv2d_41.w_0'],
-                    ['conv2d_42.w_0'],
-                    ['conv2d_44.w_0','conv2d_47.w_0','conv2d_51.w_0','conv2d_52.w_0','conv2d_53.w_0','conv2d_54.w_0','conv2d_55.w_0','conv2d_56.w_0','conv2d_57.w_0','conv2d_58.w_0'],
-                    ['conv2d_43.w_0','conv2d_45.w_0','conv2d_46.w_0','conv2d_49.w_0','conv2d_48.w_0','conv2d_50.w_0'],] 

example/post_training_quantization/pytorch_yolo_series/fine_tune.py

@@ -43,8 +43,6 @@ def argsparser():
         help="which device used to compress.")
     parser.add_argument(
         '--algo', type=str, default='avg', help="post quant algo.")
-    parser.add_argument(
-        '--round_type', type=str, default='adaround', help="round type.")
     parser.add_argument('--gpu', type=int, default=0, help='gpu index')
 
     parser.add_argument(
@@ -59,6 +57,10 @@ def argsparser():
         help='simulate activation quant')
     parser.add_argument(
         '--epochs', type=int, default=20, help='steps to reconstruct')
+    parser.add_argument(
+        '--lr', type=float, default=0.1, help='learning rate of reconstruct')
+    parser.add_argument(
+        '--limit', type=int, default=5, help='size of each region')
 
     return parser
 
@@ -104,10 +106,11 @@ def main():
         weight_quantize_type='channel_wise_abs_max',
         recon_level=FLAGS.recon_level,
         simulate_activation_quant=FLAGS.simulate_activation_quant,
-        regions=config['regions'],
-        region_weights_names=config['region_weights_names'],
+        regions=None,
+        region_weights_names=None,
         epochs=FLAGS.epochs,
-        lr=0.1)
+        lr=FLAGS.lr,
+        limit=FLAGS.limit)
 
 
 if __name__ == '__main__':