Update figure in README (#1117) (#1125)

Co-authored-by: Nwhs <wanghaoshuang@baidu.com>

demo/auto_compression/detection/README.md

@@ -23,7 +23,7 @@
 | 模型  |  策略  | 输入尺寸 | mAP<sup>val<br>0.5:0.95 | 预测时延<sup><small>FP32</small><sup><br><sup>(ms) |预测时延<sup><small>FP32</small><sup><br><sup>(ms) | 预测时延<sup><small>INT8</small><sup><br><sup>(ms) |  配置文件 | Inference模型  |
 | :-------- |:-------- |:--------: | :---------------------: | :----------------: | :----------------: | :---------------: | :-----------------------------: | :-----------------------------: |
 | PP-YOLOE-l |  Base模型 | 640*640  |  50.9   |   11.2  |   7.7ms   |  -  |  [config](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | [Model](https://bj.bcebos.com/v1/paddle-slim-models/detection/ppyoloe_crn_l_300e_coco.tar) |
-| PP-YOLOE-l |  量化+蒸馏 | 640*640  |  50.6   |   - |   -   |  6.7ms  |  [config](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/auto_compression/detection/configs/ppyoloe_l_qat_dist.yaml) | [Model](https://bj.bcebos.com/v1/paddle-slim-models/act/ppyoloe_crn_l_300e_coco_quant.tar) |
+| PP-YOLOE-l |  量化+蒸馏 | 640*640  |  50.6   |   - |   -   |  6.7ms  |  [config](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/auto_compression/detection/configs/ppyoloe_l_qat_dis.yaml) | [Model](https://bj.bcebos.com/v1/paddle-slim-models/act/ppyoloe_crn_l_300e_coco_quant.tar) |
 
 - mAP的指标均在COCO val2017数据集中评测得到。
 - PP-YOLOE模型在Tesla V100的GPU环境下测试，测试脚本是[benchmark demo](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.4/deploy/python)
@@ -95,7 +95,8 @@ tar -xf ppyoloe_crn_l_300e_coco.tar
 
 使用run.py脚本得到模型的mAP：
 ```
-python run.py --config_path=./configs/ppyoloe_l_qat_dist.yaml --eval=True
+export CUDA_VISIBLE_DEVEICES=0
+python run.py --config_path=./configs/ppyoloe_l_qat_dis.yaml --eval=True
 ```
 
 **注意**：TinyPose模型暂不支持精度测试。
@@ -104,7 +105,8 @@ python run.py --config_path=./configs/ppyoloe_l_qat_dist.yaml --eval=True
 
 蒸馏量化自动压缩示例通过run.py脚本启动，会使用接口```paddleslim.auto_compression.AutoCompression```对模型进行自动压缩。配置config文件中模型路径、蒸馏、量化、和训练等部分的参数，配置完成后便可对模型进行量化和蒸馏。具体运行命令为：
 ```
-python run.py --config_path=./configs/ppyoloe_l_qat_dist.yaml --save_dir='./output/' 
+export CUDA_VISIBLE_DEVEICES=0
+python run.py --config_path=./configs/ppyoloe_l_qat_dis.yaml --save_dir='./output/'
 ```
 
 

demo/auto_compression/detection/run.py

@@ -66,16 +66,16 @@ def reader_wrapper(reader, input_list):
     return gen
 
 
-def eval(compress_config):
+def eval(config):
 
     place = paddle.CUDAPlace(0) if FLAGS.devices == 'gpu' else paddle.CPUPlace()
     exe = paddle.static.Executor(place)
 
     val_program, feed_target_names, fetch_targets = paddle.fluid.io.load_inference_model(
-        compress_config["model_dir"],
+        config["model_dir"],
         exe,
-        model_filename=compress_config["model_filename"],
-        params_filename=compress_config["params_filename"], )
+        model_filename=config["model_filename"],
+        params_filename=config["params_filename"], )
     clsid2catid = {v: k for k, v in dataset.catid2clsid.items()}
 
     anno_file = dataset.get_anno()
@@ -85,7 +85,7 @@ def eval(compress_config):
         data_all = {k: np.array(v) for k, v in data.items()}
         data_input = {}
         for k, v in data.items():
-            if k in compress_config['input_list']:
+            if k in config['input_list']:
                 data_input[k] = np.array(v)
         outs = exe.run(val_program,
                        feed=data_input,
@@ -142,13 +142,14 @@ def eval_function(exe, compiled_test_program, test_feed_names, test_fetch_list):
 
 
 def main():
-    compress_config, train_config = load_slim_config(FLAGS.config_path)
-    reader_cfg = load_config(compress_config['reader_config'])
+    compress_config, train_config, global_config = load_slim_config(
+        FLAGS.config_path)
+    reader_cfg = load_config(global_config['reader_config'])
 
     train_loader = create('EvalReader')(reader_cfg['TrainDataset'],
                                         reader_cfg['worker_num'],
                                         return_list=True)
-    train_loader = reader_wrapper(train_loader, compress_config['input_list'])
+    train_loader = reader_wrapper(train_loader, global_config['input_list'])
 
     global dataset
     dataset = reader_cfg['EvalDataset']
@@ -158,18 +159,18 @@ def main():
                                       return_list=True)
 
     if FLAGS.eval:
-        eval(compress_config)
+        eval(global_config)
         sys.exit(0)
 
-    if 'Evaluation' in compress_config.keys() and compress_config['Evaluation']:
+    if 'Evaluation' in global_config.keys() and global_config['Evaluation']:
         eval_func = eval_function
     else:
         eval_func = None
 
     ac = AutoCompression(
-        model_dir=compress_config["model_dir"],
-        model_filename=compress_config["model_filename"],
-        params_filename=compress_config["params_filename"],
+        model_dir=global_config["model_dir"],
+        model_filename=global_config["model_filename"],
+        params_filename=global_config["params_filename"],
         save_dir=FLAGS.save_dir,
         strategy_config=compress_config,
         train_config=train_config,

demo/auto_compression/image_classification/README.md

@@ -69,6 +69,7 @@ tar -xf MobileNetV1_infer.tar
 
 ```shell
 # 单卡启动
+export CUDA_VISIBLE_DEVEICES=0
 python run.py \
     --model_dir='MobileNetV1_infer' \
     --model_filename='inference.pdmodel' \
@@ -97,3 +98,4 @@ python -m paddle.distributed.launch run.py \
 - [Paddle Lite部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/lite/lite.md)
 
 ## 5.FAQ
+[1.] 如果遇到报错 ```ValueError: var inputs not in this block``` ，则说明模型中的输入变量的名字不是 ```inputs``` ，可以先用netron可视化查看输入变量的名称，然后修改 ```run.py``` 中的第35行中 ``` yield {"inputs": imgs}``` 为 ```yield {${input_tensor_name}: imgs}```。一般PaddleClas产出部署模型的输入名字如果不是 ```inputs```，则是 ```x```。

demo/auto_compression/image_classification/run.py

 import os
 import sys
-sys.path[0] = os.path.join(os.path.dirname("__file__"), os.path.pardir, os.path.pardir)
+sys.path[0] = os.path.join(
+    os.path.dirname("__file__"), os.path.pardir, os.path.pardir)
 import argparse
 import functools
 from functools import partial
@@ -32,7 +33,7 @@ def reader_wrapper(reader):
     def gen():
         for i, data in enumerate(reader()):
             imgs = np.float32([item[0] for item in data])
-            yield {"x": imgs}
+            yield {"inputs": imgs}
 
     return gen
 

demo/auto_compression/image_classification/run.sh

 # 单卡启动
+export CUDA_VISIBLE_DEVEICES=0
 python run.py \
     --model_dir='MobileNetV1_infer' \
     --model_filename='inference.pdmodel' \

demo/auto_compression/nlp/README.md

@@ -62,7 +62,7 @@ pip install paddleslim
 安装paddlenlp：
 ```shell
 pip install paddlenlp
-``` 
+```
 
 注：安装PaddleNLP的目的是为了下载PaddleNLP中的数据集和Tokenizer。
 
@@ -88,6 +88,7 @@ tar -zxvf afqmc.tar
 数据集为CLUE，不同任务名称代表CLUE上不同的任务，可选择的任务名称有：afqmc, tnews, iflytek, ocnli, cmnli, cluewsc2020, csl。具体运行命令为
 ：
 ```shell
+export CUDA_VISIBLE_DEVEICES=0
 python run.py \
     --model_type='ppminilm' \
     --model_dir='./afqmc/' \
@@ -98,7 +99,7 @@ python run.py \
     --batch_size=16 \
     --max_seq_length=128 \
     --task_name='afqmc' \
-    --config_path='./configs/afqmc.yaml' 
+    --config_path='./configs/afqmc.yaml'
 ```
 
 ## 4. 压缩配置介绍
@@ -184,4 +185,3 @@ Quantization:
 - [Paddle Lite部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/lite/lite.md)
 
 ## 6. FAQ
-

demo/auto_compression/nlp/run.py

 import os
 import sys
-sys.path[0] = os.path.join(os.path.dirname("__file__"), os.path.pardir, os.path.pardir)
+sys.path[0] = os.path.join(
+    os.path.dirname("__file__"), os.path.pardir, os.path.pardir)
 import argparse
 import functools
 from functools import partial
@@ -58,7 +59,9 @@ def convert_example(example,
                     label_list,
                     max_seq_length=512,
                     is_test=False):
-    assert args.dataset in ['glue', 'clue'], "This demo only supports for dataset glue or clue"
+    assert args.dataset in [
+        'glue', 'clue'
+    ], "This demo only supports for dataset glue or clue"
     """Convert a glue example into necessary features."""
     if args.dataset == 'glue':
         if not is_test:
@@ -74,13 +77,14 @@ def convert_example(example,
             return example['input_ids'], example['token_type_ids'], label
         else:
             return example['input_ids'], example['token_type_ids']
-        
-    else: #if args.dataset == 'clue':
+
+    else:  #if args.dataset == 'clue':
         if not is_test:
             # `label_list == None` is for regression task
             label_dtype = "int64" if label_list else "float32"
             # Get the label
-            example['label'] = np.array(example["label"], dtype="int64").reshape((-1, 1))
+            example['label'] = np.array(
+                example["label"], dtype="int64").reshape((-1, 1))
             label = example['label']
         # Convert raw text to feature
         if 'keyword' in example:  # CSL
@@ -91,12 +95,13 @@ def convert_example(example,
                 'label': example['label']
             }
         elif 'target' in example:  # wsc
-            text, query, pronoun, query_idx, pronoun_idx = example['text'], example[
-                'target']['span1_text'], example['target']['span2_text'], example[
-                    'target']['span1_index'], example['target']['span2_index']
+            text, query, pronoun, query_idx, pronoun_idx = example[
+                'text'], example['target']['span1_text'], example['target'][
+                    'span2_text'], example['target']['span1_index'], example[
+                        'target']['span2_index']
             text_list = list(text)
-            assert text[pronoun_idx:(pronoun_idx + len(pronoun)
-                                     )] == pronoun, "pronoun: {}".format(pronoun)
+            assert text[pronoun_idx:(pronoun_idx + len(
+                pronoun))] == pronoun, "pronoun: {}".format(pronoun)
             assert text[query_idx:(query_idx + len(query)
                                    )] == query, "query: {}".format(query)
             if pronoun_idx > query_idx:
@@ -125,8 +130,6 @@ def convert_example(example,
         else:
             return example['input_ids'], example['token_type_ids']
 
-        
-
 
 def create_data_holder(task_name):
     """
@@ -148,7 +151,7 @@ def reader():
     # Create the tokenizer and dataset
     if args.model_type == 'bert':
         tokenizer = BertTokenizer.from_pretrained(args.model_dir)
-    else: # ppminilm
+    else:  # ppminilm
         tokenizer = PPMiniLMTokenizer.from_pretrained(args.model_dir)
     train_ds, dev_ds = load_dataset(
         args.dataset, args.task_name, splits=('train', 'dev'))
@@ -239,7 +242,7 @@ if __name__ == '__main__':
     print_arguments(args)
     paddle.enable_static()
 
-    compress_config, train_config = load_config(args.config_path)
+    compress_config, train_config, _ = load_config(args.config_path)
     if train_config is not None and 'optim_args' in train_config:
         train_config['optim_args'][
             'apply_decay_param_fun'] = apply_decay_param_fun
@@ -256,8 +259,8 @@ if __name__ == '__main__':
         strategy_config=compress_config,
         train_config=train_config,
         train_dataloader=train_dataloader,
-        eval_callback=eval_function
-        if compress_config is None or 'HyperParameterOptimization' not in compress_config else
+        eval_callback=eval_function if compress_config is None or
+        'HyperParameterOptimization' not in compress_config else
         eval_dataloader,
         eval_dataloader=eval_dataloader)
 

demo/auto_compression/semantic_segmentation/README.md

@@ -74,7 +74,7 @@ pip install paddleseg
 - 如果想快速体验，可直接下载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
+wget 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
 ```
 
@@ -87,16 +87,18 @@ tar -xzf ppseg_lite_portrait_398x224_with_softmax.tar.gz
 当只设置训练参数，并传入``deploy_hardware``字段时，将自动搜索压缩策略进行压缩。以骁龙710（SD710）为部署硬件，进行自动压缩的运行命令如下：
 
 ```shell
+export CUDA_VISIBLE_DEVICES=0
 python run.py \
     --model_dir='./ppseg_lite_portrait_398x224_with_softmax' \
     --model_filename='model.pdmodel' \
     --params_filename='model.pdiparams' \
     --save_dir='./save_model' \
-    --config_path='configs/pp_humanseg_auto.yaml'
+    --config_path='configs/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：
 ```shell
+export CUDA_VISIBLE_DEVICES=0
 python run.py \
     --model_dir='./ppseg_lite_portrait_398x224_with_softmax' \
     --model_filename='model.pdmodel' \
@@ -107,6 +109,7 @@ python run.py \
 
 - 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：
 ```shell
+export CUDA_VISIBLE_DEVICES=0
 python run.py \
     --model_dir='./ppseg_lite_portrait_398x224_with_softmax' \
     --model_filename='model.pdmodel' \

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_auto.yaml

 Global:
-  reader_config: configs/pp_humanseg_lite.yml
+  reader_config: configs/pp_humanseg_lite.yaml
 
 TrainConfig:
   epochs: 14
@@ -8,4 +8,4 @@ TrainConfig:
   optim_args:
     weight_decay: 0.0005
   optimizer: SGD
-  
+  
\ No newline at end of file

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_lite.yaml

@@ -25,7 +25,7 @@ train_dataset:
 val_dataset:
   type: Dataset
   dataset_root: data/humanseg
-  train_path: data/humanseg/val.txt
+  val_path: data/humanseg/val.txt
   num_classes: 2
   transforms:
   - type: PaddingByAspectRatio

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_quant_dis.yaml

 Global:
-  reader_config: configs/pp_humanseg_lite.yml
+  reader_config: configs/pp_humanseg_lite.yaml
 
 Distillation:
   distill_lambda: 1.0
@@ -26,4 +26,4 @@ TrainConfig:
   learning_rate: 0.0005
   optimizer: SGD
   optim_args:
-    weight_decay: 4.0e-05
+    weight_decay: 4.0e-05
\ No newline at end of file

demo/auto_compression/semantic_segmentation/configs/pp_humanseg_sparse_dis.yaml

 Global:
-  reader_config: configs/pp_humanseg_lite.yml
+  reader_config: configs/pp_humanseg_lite.yaml
 
 Distillation:
   distill_lambda: 1.0
@@ -31,4 +31,4 @@ TrainConfig:
   optim_args:
     weight_decay: 0.0005
   optimizer: SGD
-  
+  
\ No newline at end of file

demo/auto_compression/semantic_segmentation/run.py

@@ -148,8 +148,8 @@ if __name__ == '__main__':
 
     args = parse_args()
 
-    compress_config, train_config = load_config(args.config_path)
-    cfg = Config(compress_config['reader_config'])
+    compress_config, train_config, global_config = load_config(args.config_path)
+    cfg = Config(global_config['reader_config'])
 
     train_dataset = cfg.train_dataset
     eval_dataset = cfg.val_dataset
@@ -167,7 +167,7 @@ if __name__ == '__main__':
     ac = AutoCompression(
         model_dir=args.model_dir,
         model_filename=args.model_filename,
-        params_filename=args.param_filename,
+        params_filename=args.params_filename,
         save_dir=args.save_dir,
         strategy_config=compress_config,
         train_config=train_config,

paddleslim/auto_compression/auto_strategy.py

@@ -27,7 +27,7 @@ __all__ = [
 
 # config tester to test the loss of quant_post
 hpo_config_tester = {
-    "ptq_algo": ["avg"],
+    "ptq_algo": ["avg", "mse", "KL"],
     "weight_quantize_type": ['channel_wise_abs_max', 'abs_max'],
     "bias_correct": [False],
     "batch_num": [5],
@@ -72,7 +72,8 @@ EXPERIENCE_STRATEGY_WITHOUT_LOSS = [
 ]
 MAGIC_SPARSE_RATIO = 0.75
 ### TODO: 0.03 threshold maybe not suitable, need to check
-MAGIC_EMD_DISTANCE = 0.03
+MAGIC_MAX_EMD_DISTANCE = 0.03
+MAGIC_MIN_EMD_DISTANCE = 0.01
 
 DEFAULT_TRANSFORMER_STRATEGY = 'prune_0.25_int8'
 DEFAULT_STRATEGY = 'origin_int8'
@@ -224,9 +225,13 @@ def prepare_strategy(model_dir,
     return strategy_config
 
 
-def get_final_quant_config(ptq_loss, mode='DistilQuant'):
+def get_final_quant_config(ptq_loss):
     """ transform quantization tester config to real quantization config """
-    if mode == 'HPO':
+    ### if emd loss less than MAGIC_MIN_EMD_DISTANCE, final compress.
+    if ptq_loss < MAGIC_MIN_EMD_DISTANCE:
+        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)
         hpo_config = HyperParameterOptimization(**default_hpo_config)
         configs = [{
@@ -234,7 +239,8 @@ def get_final_quant_config(ptq_loss, mode='DistilQuant'):
             'HyperParameterOptimization': hpo_config
         }]
 
-    if mode == 'DistilQuant':
+    ### if emd loss greater than MAGIC_MAX_EMD_DISTANCE, select qat & dist.
+    else:
         quant_config = Quantization(**default_quant_config)
         dis_config = Distillation()
         configs = [{'Quantization': quant_config, 'Distillation': dis_config}]

paddleslim/auto_compression/compressor.py

@@ -357,26 +357,27 @@ class AutoCompression:
         ).lower() == 'linux':
             ptq_loss = quant_post_hpo.g_min_emd_loss
 
-            final_quant_config = get_final_quant_config(
-                ptq_loss, mode='DistilQuant')
-            quant_strategy, quant_config = self._prepare_strategy(
-                final_quant_config)
-            self.single_strategy_compress(quant_strategy[0], quant_config[0],
-                                          strategy_idx)
+            final_quant_config = get_final_quant_config(ptq_loss)
+            if final_quant_config is not None:
+                quant_strategy, quant_config = self._prepare_strategy(
+                    final_quant_config)
+                self.single_strategy_compress(quant_strategy[0],
+                                              quant_config[0], strategy_idx)
         tmp_model_path = os.path.join(
             self.save_dir, 'strategy_{}'.format(str(strategy_idx + 1)))
         final_model_path = os.path.join(self.final_dir)
         if not os.path.exists(final_model_path):
             os.makedirs(final_model_path)
-        tmp_model_file = os.path.join(tmp_model_path, 'model.pdmodel')
-        tmp_params_file = os.path.join(tmp_model_path, 'model.pdiparams')
-        final_model_file = os.path.join(final_model_path, 'model.pdmodel')
-        final_params_file = os.path.join(final_model_path, 'model.pdiparams')
-        shutil.move(tmp_model_file, final_model_file)
-        shutil.move(tmp_params_file, final_params_file)
-        _logger.info(
-            "==> Finished the ACT process and the final model is saved in:{}".
-            format(final_model_path))
+        tmp_model_file = os.path.join(tmp_model_path, self.model_filename)
+        tmp_params_file = os.path.join(tmp_model_path, self.params_filename)
+        final_model_file = os.path.join(final_model_path, self.model_filename)
+        final_params_file = os.path.join(final_model_path, self.params_filename)
+        if paddle.distributed.get_rank() == 0:
+            shutil.move(tmp_model_file, final_model_file)
+            shutil.move(tmp_params_file, final_params_file)
+            _logger.info(
+                "==> Finished the ACT process and the final model is saved in:{}".
+                format(final_model_path))
         os._exit(0)
 
     def single_strategy_compress(self, strategy, config, strategy_idx):
@@ -569,5 +570,5 @@ class AutoCompression:
             target_vars=test_program_info.fetch_targets,
             executor=self._exe,
             main_program=test_program,
-            model_filename='model.pdmodel',
-            params_filename='model.pdiparams')
+            model_filename=self.model_filename,
+            params_filename=self.params_filename)

paddleslim/auto_compression/config_helpers.py

@@ -25,14 +25,16 @@ def load_config(config_path):
     cfg = yaml.load(f, Loader=yaml.FullLoader)
     f.close()
 
+    global_config = {}
+    if 'Global' in cfg:
+        for g_key, g_value in cfg["Global"].items():
+            global_config[g_key] = g_value
+        cfg.pop('Global')
+
     compress_config = {}
     for key, value in cfg.items():
-        if key == "Global":
-            for g_key, g_value in cfg["Global"].items():
-                compress_config[g_key] = g_value
-        else:
-            default_key = eval(key)(**value)
-            compress_config[key] = default_key
+        default_key = eval(key)(**value)
+        compress_config[key] = default_key
 
     if compress_config.get('TrainConfig') != None:
         train_config = compress_config.pop('TrainConfig')
@@ -42,7 +44,7 @@ def load_config(config_path):
     if len(compress_config) == 0:
         compress_config = None
 
-    return compress_config, train_config
+    return compress_config, train_config, global_config
 
 
 def save_config(config, config_path):

paddleslim/version.py

@@ -14,4 +14,4 @@
 # limitations under the License.
 """ PaddleSlim version string """
 __all__ = ["slim_version"]
-slim_version = "1.0.0"
+slim_version = "2.3.0"