refine quantization demo (#52)

demo/quant/quant_aware/README.md

@@ -6,9 +6,29 @@
 
 请参考 <a href='../../../paddleslim/quant/quantization_api_doc.md'>量化API文档</a>。
 
-## 分类模型的离线量化流程
+## 分类模型的量化训练流程
 
-### 1. 配置量化参数
+### 准备数据
+
+在``demo``文件夹下创建``data``文件夹，将``ImageNet``数据集解压在``data``文件夹下，解压后``data``文件夹下应包含以下文件：
+- ``'train'``文件夹，训练图片
+- ``'train_list.txt'``文件
+- ``'val'``文件夹，验证图片
+- ``'val_list.txt'``文件
+
+### 准备需要量化的模型
+
+使用以下命令下载训练好的模型并解压。
+
+```
+mkdir pretrain
+cd pretrain
+wget http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar
+tar xf MobileNetV1_pretrained.tar
+cd ..
+```
+
+### 配置量化参数
 
 ```
 quant_config = {
@@ -24,7 +44,7 @@ quant_config = {
 }
 ```
 
-### 2. 对训练和测试program插入可训练量化op
+### 对训练和测试program插入可训练量化op
 
 ```
 val_program = quant_aware(val_program, place, quant_config, scope=None, for_test=True)
@@ -32,7 +52,7 @@ val_program = quant_aware(val_program, place, quant_config, scope=None, for_test
 compiled_train_prog = quant_aware(train_prog, place, quant_config, scope=None, for_test=False)
 ```
 
-### 3.关掉指定build策略
+### 关掉指定build策略
 
 ```
 build_strategy = fluid.BuildStrategy()
@@ -45,32 +65,10 @@ compiled_train_prog = compiled_train_prog.with_data_parallel(
         exec_strategy=exec_strategy)
 ```
 
-### 4. freeze program
-
-```
-float_program, int8_program = convert(val_program,
-                                      place,
-                                      quant_config,
-                                      scope=None,
-                                      save_int8=True)
-```
 
-### 5.保存预测模型
+### 训练命令
 
 ```
-fluid.io.save_inference_model(
-    dirname=float_path,
-    feeded_var_names=[image.name],
-    target_vars=[out], executor=exe,
-    main_program=float_program,
-    model_filename=float_path + '/model',
-    params_filename=float_path + '/params')
-
-fluid.io.save_inference_model(
-    dirname=int8_path,
-    feeded_var_names=[image.name],
-    target_vars=[out], executor=exe,
-    main_program=int8_program,
-    model_filename=int8_path + '/model',
-    params_filename=int8_path + '/params')
+python train.py --model MobileNet --pretrained_model ./pretrain/MobileNetV1_pretrained --checkpoint_dir ./output/mobilenetv1 --num_epochs 30
 ```
+运行之后，可看到``best_model``的最后测试结果，和MobileNet量化前的精度top1=70.99%, top5=89.68%非常相近。

demo/quant/quant_aware/train.py

@@ -8,8 +8,8 @@ import math
 import time
 import numpy as np
 import paddle.fluid as fluid
-sys.path.append(sys.path[0] + "../../../")
-sys.path.append(sys.path[0] + "../../")
+sys.path.append(sys.path[0] + "/../../../")
+sys.path.append(sys.path[0] + "/../../")
 from paddleslim.common import get_logger
 from paddleslim.analysis import flops
 from paddleslim.quant import quant_aware, quant_post, convert
@@ -37,7 +37,7 @@ parser.add_argument('--step_epochs', nargs='+', type=int, default=[30, 60, 90],
 add_arg('config_file',      str, None,                 "The config file for compression with yaml format.")
 add_arg('data',             str, "imagenet",             "Which data to use. 'mnist' or 'imagenet'")
 add_arg('log_period',       int, 10,                 "Log period in batches.")
-add_arg('test_period',      int, 10,                 "Test period in epoches.")
+add_arg('checkpoint_dir',         str, "output",           "checkpoint save dir")
 # yapf: enable
 
 model_list = [m for m in dir(models) if "__" not in m]
@@ -192,16 +192,6 @@ def compress(args):
         return np.mean(np.array(acc_top1_ns))
 
     def train(epoch, compiled_train_prog):
-        build_strategy = fluid.BuildStrategy()
-        build_strategy.memory_optimize = False
-        build_strategy.enable_inplace = False
-        build_strategy.fuse_all_reduce_ops = False
-        build_strategy.sync_batch_norm = False
-        exec_strategy = fluid.ExecutionStrategy()
-        compiled_train_prog = compiled_train_prog.with_data_parallel(
-            loss_name=avg_cost.name,
-            build_strategy=build_strategy,
-            exec_strategy=exec_strategy)
 
         batch_id = 0
         for data in train_reader():
@@ -221,14 +211,41 @@ def compress(args):
                            end_time - start_time))
             batch_id += 1
 
+    build_strategy = fluid.BuildStrategy()
+    build_strategy.memory_optimize = False
+    build_strategy.enable_inplace = False
+    build_strategy.fuse_all_reduce_ops = False
+    build_strategy.sync_batch_norm = False
+    exec_strategy = fluid.ExecutionStrategy()
+    compiled_train_prog = compiled_train_prog.with_data_parallel(
+        loss_name=avg_cost.name,
+        build_strategy=build_strategy,
+        exec_strategy=exec_strategy)
+
     ############################################################################################################
     # train loop
     ############################################################################################################
+    best_acc1 = 0.0
+    best_epoch = 0
     for i in range(args.num_epochs):
         train(i, compiled_train_prog)
-        if i % args.test_period == 0:
-            test(i, val_program)
-
+        acc1 = test(i, val_program)
+        fluid.io.save_persistables(
+            exe,
+            dirname=os.path.join(args.checkpoint_dir, str(i)),
+            main_program=val_program)
+        if acc1 > best_acc1:
+            best_acc1 = acc1
+            best_epoch = i
+            fluid.io.save_persistables(
+                exe,
+                dirname=os.path.join(args.checkpoint_dir, 'best_model'),
+                main_program=val_program)
+
+    fluid.io.load_persistables(
+        exe,
+        dirname=os.path.join(args.checkpoint_dir, 'best_model'),
+        main_program=val_program)
     ############################################################################################################
     # 3. Freeze the graph after training by adjusting the quantize
     #    operators' order for the inference.
@@ -237,7 +254,8 @@ def compress(args):
     float_program, int8_program = convert(val_program, place, quant_config, \
                                                         scope=None, \
                                                         save_int8=True)
-
+    print("eval best_model after convert")
+    final_acc1 = test(best_epoch, float_program)
     ############################################################################################################
     # 4. Save inference model
     ############################################################################################################

demo/quant/quant_post/README.md

@@ -10,7 +10,7 @@
 
 ### 准备数据
 
-在当前文件夹下创建``data``文件夹，将``imagenet``数据集解压在``data``文件夹下，解压后``data``文件夹下应包含以下文件：
+在``demo``文件夹下创建``data``文件夹，将``ImageNet``数据集解压在``data``文件夹下，解压后``data``文件夹下应包含以下文件：
 - ``'train'``文件夹，训练图片
 - ``'train_list.txt'``文件
 - ``'val'``文件夹，验证图片
@@ -69,4 +69,4 @@ python eval.py --model_path ./quant_model_train/MobileNet
 ```
 top1_acc/top5_acc= [0.70141864 0.89086477]
 ```
-从以上精度对比可以看出，对``mobilenet``在``imagenet``上的分类模型进行离线量化后 ``top1``精度损失为``0.77%``， ``top5``精度损失为``0.46%``. 
+从以上精度对比可以看出，对``mobilenet``在``imagenet``上的分类模型进行离线量化后 ``top1``精度损失为``0.77%``， ``top5``精度损失为``0.46%``.

demo/quant/quant_post/quant_post.py

@@ -9,13 +9,12 @@ import time
 import numpy as np
 import paddle.fluid as fluid
 
-import reader
 sys.path.append(sys.path[0] + "/../../../")
 from paddleslim.common import get_logger
 from paddleslim.quant import quant_post
 sys.path.append(sys.path[0] + "/../../")
 from utility import add_arguments, print_arguments
-
+import imagenet_reader as reader
 _logger = get_logger(__name__, level=logging.INFO)
 
 parser = argparse.ArgumentParser(description=__doc__)