update ACT quick start demo (#1116)

demo/auto_compression/README.md

-# 自动压缩工具ACT（Auto Compression Tookit）
+# 自动化压缩工具ACT（Auto Compression Tookit）
 
 ## 简介
-PaddleSlim推出全新自动压缩工具（ACT），旨在通过Source-Free的方式，自动对预测模型进行压缩，压缩后模型可直接部署应用。ACT自动压缩工具主要特性如下：
+PaddleSlim推出全新自动化压缩工具（ACT），旨在通过Source-Free的方式，自动对预测模型进行压缩，压缩后模型可直接部署应用。ACT自动化压缩工具主要特性如下：
 - **『更便捷』**：开发者无需了解或修改模型源码，直接使用导出的预测模型进行压缩；
 - **『更智能』**：开发者简单配置即可启动压缩，ACT工具会自动优化得到最好预测模型；
 - **『更丰富』**：ACT中提供了量化训练、蒸馏、结构化剪枝、非结构化剪枝、多种离线量化方法及超参搜索等等，可任意搭配使用。
@@ -14,32 +14,68 @@ PaddleSlim推出全新自动压缩工具（ACT），旨在通过Source-Free的
 
 ## 快速上手
 
+- 1.准备模型及数据集
+
+```shell
+# 下载MobileNet预测模型
+wget https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileNetV1_infer.tar
+tar -xf MobileNetV1_infer.tar
+# 下载ImageNet小型数据集
+wget https://sys-p0.bj.bcebos.com/slim_ci/ILSVRC2012_data_demo.tar.gz
+tar xf ILSVRC2012_data_demo.tar.gz
+```
+
+- 2.运行
+
 ```python
 # 导入依赖包
-from paddleslim.auto_compression.config_helpers import load_config
+import paddle
+from PIL import Image
+from paddle.vision.datasets import DatasetFolder
+from paddle.vision.transforms import transforms
 from paddleslim.auto_compression import AutoCompression
-from paddleslim.common.imagenet_reader import reader
-# 加载配置文件
-compress_config, train_config = load_slim_config("./image_classification/mobilenetv1_qat_dis.yaml")
+paddle.enable_static()
+# 定义DataSet
+class ImageNetDataset(DatasetFolder):
+    def __init__(self, path, image_size=224):
+        super(ImageNetDataset, self).__init__(path)
+        normalize = transforms.Normalize(
+            mean=[123.675, 116.28, 103.53], std=[58.395, 57.120, 57.375])
+        self.transform = transforms.Compose([
+            transforms.Resize(256),
+            transforms.CenterCrop(image_size), transforms.Transpose(),
+            normalize
+        ])
+
+    def __getitem__(self, idx):
+        img_path, _ = self.samples[idx]
+        return self.transform(Image.open(img_path).convert('RGB'))
+
+    def __len__(self):
+        return len(self.samples)
+
 # 定义DataLoader
-train_loader = reader(mode='test') # DataLoader
+train_dataset = ImageNetDataset("./ILSVRC2012_data_demo/ILSVRC2012/train/")
+image = paddle.static.data(
+    name='inputs', shape=[None] + [3, 224, 224], dtype='float32')
+train_loader = paddle.io.DataLoader(train_dataset, feed_list=[image], batch_size=32)
 # 开始自动压缩
 ac = AutoCompression(
-    model_dir="./mobilenetv1_infer",
-    model_filename="model.pdmodel",
-    params_filename="model.pdiparams",
+    model_dir="./MobileNetV1_infer/",
+    model_filename="inference.pdmodel",
+    params_filename="inference.pdiparams",
     save_dir="output",
-    strategy_config=compress_config,
-    train_config=train_config,
+    strategy_config=None,
+    train_config=None,
     train_dataloader=train_loader,
-    eval_callback=None)  # eval_function to verify accuracy
+    eval_dataloader=train_loader)  # eval_function to verify accuracy
 ac.compress()
 ```
 
 **提示：**
 - DataLoader传入的数据集是待压缩模型所用的数据集，DataLoader继承自`paddle.io.DataLoader`。
-- 如无需验证自动压缩过程中模型的精度，`eval_callback`可不传入function，程序会自动根据损失来选择最优模型。
-- 自动压缩Config中定义量化、蒸馏、剪枝等压缩算法会合并执行，压缩策略有：量化+蒸馏，剪枝+蒸馏等等。
+- 如无需验证自动化压缩过程中模型的精度，`eval_callback`可不传入function，程序会自动根据损失来选择最优模型。
+- 自动化压缩Config中定义量化、蒸馏、剪枝等压缩算法会合并执行，压缩策略有：量化+蒸馏，剪枝+蒸馏等等。
 
 ## 应用示例
 
@@ -52,11 +88,11 @@ ac.compress()
 #### [NLP](./nlp)
 
 #### 即将发布
-- [ ] 更多自动压缩应用示例
-- [ ] X2Paddle模型自动压缩示例
+- [ ] 更多自动化压缩应用示例
+- [ ] X2Paddle模型自动化压缩示例
 
 ## 其他
 
 - ACT可以自动处理常见的预测模型，如果有更特殊的改造需求，可以参考[ACT超参配置教程](./hyperparameter_tutorial.md)来进行单独配置压缩策略。
 
-- 如果你发现任何关于ACT自动压缩工具的问题或者是建议, 欢迎通过[GitHub Issues](https://github.com/PaddlePaddle/PaddleSlim/issues)给我们提issues。同时欢迎贡献更多优秀模型，共建开源生态。
+- 如果你发现任何关于ACT自动化压缩工具的问题或者是建议, 欢迎通过[GitHub Issues](https://github.com/PaddlePaddle/PaddleSlim/issues)给我们提issues。同时欢迎贡献更多优秀模型，共建开源生态。

demo/auto_compression/image_classification/README.md

@@ -18,9 +18,9 @@
 ## 2. Benchmark
 - MobileNetV1模型
 
-| 模型 | 策略 | Top-1 Acc | 耗时(ms) threads=4 | 
+| 模型 | 策略 | Top-1 Acc | 耗时(ms) threads=4 |
 |:------:|:------:|:------:|:------:|
-| MobileNetV1 | Base模型 | 70.90 | 39.041 | 
+| MobileNetV1 | Base模型 | 70.90 | 39.041 |
 | MobileNetV1 | 量化+蒸馏 | 70.49 | 29.238|
 
 - 测试环境：`SDM710 2*A75(2.2GHz) 6*A55(1.7GHz)`
@@ -59,7 +59,7 @@ pip install paddleslim
 
 ```shell
 wget https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileNetV1_infer.tar
-tar -zxvf MobileNetV1_infer.tar
+tar -xf MobileNetV1_infer.tar
 ```
 也可根据[PaddleClas文档](https://github.com/PaddlePaddle/PaddleClas/blob/release/2.3/docs/zh_CN/inference_deployment/export_model.md)导出Inference模型。
 
@@ -73,20 +73,20 @@ python run.py \
     --model_dir='MobileNetV1_infer' \
     --model_filename='inference.pdmodel' \
     --params_filename='inference.pdiparams' \
-    --save_dir='./save_quant_mobilev1/' \
+    --save_dir='./output' \
     --batch_size=128 \
-    --config_path='./configs/mobilev1.yaml'\
+    --config_path='./configs/mobilenetv1_qat_dis.yaml'\
     --data_dir='ILSVRC2012'
-    
+
 # 多卡启动
 python -m paddle.distributed.launch run.py \
     --model_dir='MobileNetV1_infer' \
     --model_filename='inference.pdmodel' \
     --params_filename='inference.pdiparams' \
-    --save_dir='./save_quant_mobilev1/' \
+    --save_dir='./output' \
     --batch_size=128 \
-    --config_path='./configs/mobilev1.yaml'\
-    --data_dir='ILSVRC2012' 
+    --config_path='./configs/mobilenetv1_qat_dis.yaml'\
+    --data_dir='ILSVRC2012'
 ```
 
 
@@ -96,4 +96,4 @@ python -m paddle.distributed.launch run.py \
 - [Paddle Inference C++部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/inference/cpp_inference.md)
 - [Paddle Lite部署](https://github.com/PaddlePaddle/PaddleSeg/blob/release/2.5/docs/deployment/lite/lite.md)
 
-## 5.FAQ
\ No newline at end of file
+## 5.FAQ

paddleslim/auto_compression/auto_strategy.py

@@ -25,16 +25,16 @@ __all__ = [
     "prepare_strategy", "create_strategy_config", "get_final_quant_config"
 ]
 
-### config tester to test the loss of quant_post
+# config tester to test the loss of quant_post
 hpo_config_tester = {
-    "ptq_algo": ["avg", "mse", "KL"],
+    "ptq_algo": ["avg"],
     "weight_quantize_type": ['channel_wise_abs_max', 'abs_max'],
     "bias_correct": [False],
-    "batch_num": [2, 3],
+    "batch_num": [5],
     "max_quant_count": 1,
 }
 
-### default hpo config
+# default hpo config
 default_hpo_config = {
     "ptq_algo": ["KL", "hist", "avg", "mse"],
     "weight_quantize_type": ['channel_wise_abs_max', 'abs_max'],
@@ -44,11 +44,26 @@ default_hpo_config = {
     "max_quant_count": 20,
 }
 
-### default quant config, can be used by ptq&hpo and qat&distillation
+# default quant config, can be used by ptq&hpo and qat&distillation
 default_quant_config = {
     'quantize_op_types': ['conv2d', 'depthwise_conv2d', 'mul', 'matmul'],
     'weight_bits': 8,
-    'activation_bits': 8
+    'activation_bits': 8,
+    "is_full_quantize": False,
+    "activation_quantize_type": 'range_abs_max',
+    "weight_quantize_type": 'abs_max',
+    "not_quant_pattern": ["skip_quant"],
+}
+
+# default train config
+DefaultTrainConfig = {
+    "epochs": 1,
+    "eval_iter": 500,
+    "learning_rate": 0.0001,
+    "optimizer": "Momentum",
+    "optim_args": {
+        "weight_decay": 4.0e-05
+    },
 }
 
 EXPERIENCE_STRATEGY_WITHOUT_LOSS = [
@@ -118,6 +133,12 @@ def create_strategy_config(strategy_str, model_type):
     return configs
 
 
+def create_train_config(strategy_str, model_type):
+    # TDOD: support more strategy and model_type
+    train_config = TrainConfig(**DefaultTrainConfig)
+    return train_config
+
+
 def prepare_strategy(model_dir,
                      model_filename,
                      params_filename,
@@ -203,9 +224,9 @@ def prepare_strategy(model_dir,
     return strategy_config
 
 
-def get_final_quant_config(ptq_loss):
+def get_final_quant_config(ptq_loss, mode='DistilQuant'):
     """ transform quantization tester config to real quantization config """
-    if ptq_loss <= MAGIC_EMD_DISTANCE:
+    if mode == 'HPO':
         quant_config = Quantization(**default_quant_config)
         hpo_config = HyperParameterOptimization(**default_hpo_config)
         configs = [{
@@ -213,10 +234,11 @@ def get_final_quant_config(ptq_loss):
             'HyperParameterOptimization': hpo_config
         }]
 
-    else:
+    if mode == 'DistilQuant':
         quant_config = Quantization(**default_quant_config)
         dis_config = Distillation()
         configs = [{'Quantization': quant_config, 'Distillation': dis_config}]
+        _logger.info("Start Quantization and Distillation Training.")
 
     return configs
 

paddleslim/auto_compression/compressor.py

@@ -24,14 +24,14 @@ import paddle
 import paddle.distributed.fleet as fleet
 if platform.system().lower() == 'linux':
     from ..quant import quant_post_hpo
-from ..quant.quanter import convert
+from ..quant.quanter import convert, quant_post
 from ..common.recover_program import recover_inference_program
 from ..common import get_logger
 from ..common.patterns import get_patterns
 from ..analysis import TableLatencyPredictor
 from .create_compressed_program import build_distill_program, build_quant_program, build_prune_program, remove_unused_var_nodes
 from .strategy_config import ProgramInfo, merge_config
-from .auto_strategy import prepare_strategy, get_final_quant_config, create_strategy_config
+from .auto_strategy import prepare_strategy, get_final_quant_config, create_strategy_config, create_train_config
 
 _logger = get_logger(__name__, level=logging.INFO)
 
@@ -143,6 +143,11 @@ class AutoCompression:
         self._strategy, self._config = self._prepare_strategy(
             self.strategy_config)
 
+        # If train_config is None, set default train_config
+        if self.train_config is None:
+            self.train_config = create_train_config(self.strategy_config,
+                                                    self.model_type)
+
     def _prepare_envs(self):
         devices = paddle.device.get_device().split(':')[0]
         places = paddle.device._convert_to_place(devices)
@@ -248,8 +253,9 @@ class AutoCompression:
                                          feed_target_names, fetch_targets)
 
         config_dict = dict(config._asdict())
-        if config_dict["prune_strategy"] == "gmp" and config_dict[
-                'gmp_config'] is None:
+        if "prune_strategy" in config_dict and config_dict[
+                "prune_strategy"] == "gmp" and config_dict[
+                    'gmp_config'] is None:
             _logger.info(
                 "Calculating the iterations per epoch……(It will take some time)")
             # NOTE:XXX: This way of calculating the iters needs to be improved.
@@ -351,20 +357,57 @@ class AutoCompression:
         ).lower() == 'linux':
             ptq_loss = quant_post_hpo.g_min_emd_loss
 
-            final_quant_config = get_final_quant_config(ptq_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)
-        old_model_path = os.path.join(
+        tmp_model_path = os.path.join(
             self.save_dir, 'strategy_{}'.format(str(strategy_idx + 1)))
         final_model_path = os.path.join(self.final_dir)
-        shutil.move(old_model_path, final_model_path)
+        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))
         os._exit(0)
 
     def single_strategy_compress(self, strategy, config, strategy_idx):
-        ### start compress, including train/eval model
-        if strategy == 'ptq_hpo':
+        # start compress, including train/eval model
+        # TODO: add the emd loss of evaluation model.
+        if strategy == 'quant_post':
+            quant_post(
+                self._exe,
+                model_dir=self.model_dir,
+                quantize_model_path=os.path.join(
+                    self.save_dir, 'strategy_{}'.format(str(strategy_idx + 1))),
+                data_loader=self.train_dataloader,
+                model_filename=self.model_filename,
+                params_filename=self.params_filename,
+                save_model_filename=self.model_filename,
+                save_params_filename=self.params_filename,
+                batch_size=1,
+                batch_nums=config.batch_num,
+                algo=config.ptq_algo,
+                round_type='round',
+                bias_correct=config.bias_correct,
+                hist_percent=config.hist_percent,
+                quantizable_op_type=config.quantize_op_types,
+                is_full_quantize=config.is_full_quantize,
+                weight_bits=config.weight_bits,
+                activation_bits=config.activation_bits,
+                activation_quantize_type='range_abs_max',
+                weight_quantize_type=config.weight_quantize_type,
+                onnx_format=False)
+
+        elif strategy == 'ptq_hpo':
             if platform.system().lower() != 'linux':
                 raise NotImplementedError(
                     "post-quant-hpo is not support in system other than linux")
@@ -503,11 +546,12 @@ class AutoCompression:
             test_program_info.program,
             paddle.static.CompiledProgram) else test_program_info.program
 
-        paddle.static.load(test_program,
-                           os.path.join(self.save_dir, 'best_model'))
-        os.remove(os.path.join(self.save_dir, 'best_model.pdmodel'))
-        os.remove(os.path.join(self.save_dir, 'best_model.pdopt'))
-        os.remove(os.path.join(self.save_dir, 'best_model.pdparams'))
+        if os.path.exists(os.path.join(self.save_dir, 'best_model.pdparams')):
+            paddle.static.load(test_program,
+                               os.path.join(self.save_dir, 'best_model'))
+            os.remove(os.path.join(self.save_dir, 'best_model.pdmodel'))
+            os.remove(os.path.join(self.save_dir, 'best_model.pdopt'))
+            os.remove(os.path.join(self.save_dir, 'best_model.pdparams'))
 
         if 'qat' in strategy:
             float_program, int8_program = convert(test_program_info.program._program, self._places, self._quant_config, \