Refine arguments of ocr models. (#3261)

* Refine arguments of ocr models.

* Add comments for issue of attention model in paddle1.5

PaddleCV/ocr_recognition/README.md

+>注意：在paddle1.5版本上训练attention model有收敛问题，建议您暂时使用paddle1.4版本，后续我们会修复该问题。
+
 ## 代码结构
 ```
 ├── data_reader.py  # 下载、读取、处理数据。
@@ -6,7 +8,9 @@
 ├── train.py   # 用于模型的训练。
 ├── infer.py   # 加载训练好的模型文件，对新数据进行预测。
 ├── eval.py     # 评估模型在指定数据集上的效果。
-└── utils.py    # 定义通用的函数。
+├─ utils.py    # 定义通用的函数。
+├── run_crnn_ctc.sh     # 执行crnn_ctc模型训练任务
+└── run_attention.sh    # 执行attention模型训练任务
 ```
 
 
@@ -136,14 +140,14 @@ env CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --parallel=True
 
 执行`python train.py --help`可查看更多使用方式和参数详细说明。
 
-图2为使用默认参数在默认数据集上训练`CTC model`的收敛曲线，其中横坐标轴为训练迭代次数，纵轴为样本级错误率。其中，蓝线为训练集上的样本错误率，红线为测试集上的样本错误率。测试集上最低错误率为22.0%.
+图2为执行脚本`run_crnn_ctc.sh`在默认数据集上训练`CTC model`的收敛曲线，其中横坐标轴为训练迭代次数，纵轴为样本级错误率。其中，蓝线为训练集上的样本错误率，红线为测试集上的样本错误率。测试集上最低错误率为22.0%.
 
 <p align="center">
 <img src="images/train.jpg" width="400" hspace='10'/> <br/>
 <strong>图 2</strong>
 </p>
 
-图3为使用默认参数在默认数据集上训练`attention model`的收敛曲线，其中横坐标轴为训练迭代次数，纵轴为样本级错误率。其中，蓝线为训练集上的样本错误率，红线为测试集上的样本错误率。测试集上最低错误率为16.25%.
+图3为执行脚本`run_attention.sh`在默认数据集上训练`attention model`的收敛曲线，其中横坐标轴为训练迭代次数，纵轴为样本级错误率。其中，蓝线为训练集上的样本错误率，红线为测试集上的样本错误率。测试集上最低错误率为16.25%.
 
 <p align="center">
 <img src="images/train_attention.jpg" width="400" hspace='10'/> <br/>

PaddleCV/ocr_recognition/attention_model.py

@@ -22,11 +22,7 @@ word_vector_dim = 128
 max_length = 100
 sos = 0
 eos = 1
-gradient_clip = 10
-LR = 1.0
 beam_size = 1
-learning_rate_decay = None
-
 
 def conv_bn_pool(input,
                  group,
@@ -192,7 +188,7 @@ def attention_train_net(args, data_shape, num_classes):
     prediction = gru_decoder_with_attention(trg_embedding, encoded_vector,
                                             encoded_proj, decoder_boot,
                                             decoder_size, num_classes)
-    fluid.clip.set_gradient_clip(fluid.clip.GradientClipByValue(gradient_clip))
+    fluid.clip.set_gradient_clip(fluid.clip.GradientClipByValue(args.gradient_clip))
     label_out = fluid.layers.cast(x=label_out, dtype='int64')
 
     _, maxid = fluid.layers.topk(input=prediction, k=1)
@@ -203,8 +199,8 @@ def attention_train_net(args, data_shape, num_classes):
 
     cost = fluid.layers.cross_entropy(input=prediction, label=label_out)
     sum_cost = fluid.layers.reduce_sum(cost)
-
-    if learning_rate_decay == "piecewise_decay":
+    LR = args.lr
+    if args.lr_decay_strategy == "piecewise_decay":
         learning_rate = fluid.layers.piecewise_decay([50000], [LR, LR * 0.01])
     else:
         learning_rate = LR

PaddleCV/ocr_recognition/crnn_ctc_model.py

@@ -184,9 +184,9 @@ def encoder_net(images,
 
 
 def ctc_train_net(args, data_shape, num_classes):
-    L2_RATE = 0.0004
-    LR = 1.0e-3
-    MOMENTUM = 0.9
+    L2_RATE = args.l2decay
+    LR = args.lr
+    MOMENTUM = args.momentum
     learning_rate_decay = None
     regularizer = fluid.regularizer.L2Decay(L2_RATE)
 

PaddleCV/ocr_recognition/run_attention.sh

+export CUDA_VISIBLE_DEVICES=0
+nohup python train.py \
+--lr=1.0 \
+--gradient_clip=10 \
+--model="attention" \
+--log_period=10 \
+> attention.log 2>&1 &
+
+tailf attention.log

PaddleCV/ocr_recognition/run_crnn_ctc.sh

+export CUDA_VISIBLE_DEVICES=0
+nohup python train.py \
+--lr=1e-3 \
+--l2decay=4e-4 \
+--momentum=0.9 \
+--model="crnn_ctc" \
+--log_period=10 \
+> crnn_ctc.log 2>&1 &
+
+tailf crnn_ctc.log

PaddleCV/ocr_recognition/train.py

@@ -32,26 +32,33 @@ import numpy as np
 parser = argparse.ArgumentParser(description=__doc__)
 add_arg = functools.partial(add_arguments, argparser=parser)
 # yapf: disable
-add_arg('batch_size',        int,   32,         "Minibatch size.")
-add_arg('total_step',        int,   720000,    "The number of iterations. Zero or less means whole training set. More than 0 means the training set might be looped until # of iterations is reached.")
-add_arg('log_period',        int,   1000,       "Log period.")
-add_arg('save_model_period', int,   15000,      "Save model period. '-1' means never saving the model.")
-add_arg('eval_period',       int,   15000,      "Evaluate period. '-1' means never evaluating the model.")
-add_arg('save_model_dir',    str,   "./models", "The directory the model to be saved to.")
-add_arg('train_images',      str,   None,       "The directory of images to be used for training.")
-add_arg('train_list',        str,   None,       "The list file of images to be used for training.")
-add_arg('test_images',       str,   None,       "The directory of images to be used for test.")
-add_arg('test_list',         str,   None,       "The list file of images to be used for training.")
-add_arg('model',    str,   "crnn_ctc",           "Which type of network to be used. 'crnn_ctc' or 'attention'")
-add_arg('init_model',        str,   None,       "The init model file of directory.")
-add_arg('use_gpu',           bool,  True,      "Whether use GPU to train.")
-add_arg('min_average_window',int,   10000,     "Min average window.")
-add_arg('max_average_window',int,   12500,     "Max average window. It is proposed to be set as the number of minibatch in a pass.")
-add_arg('average_window',    float, 0.15,      "Average window.")
-add_arg('parallel',          bool,  False,     "Whether use parallel training.")
-add_arg('profile',           bool,  False,      "Whether to use profiling.")
-add_arg('skip_batch_num',    int,   0,          "The number of first minibatches to skip as warm-up for better performance test.")
-add_arg('skip_test',         bool,  False,      "Whether to skip test phase.")
+add_arg('batch_size',        int,     32,         "Minibatch size.")
+
+add_arg('lr',                float,   1e-3,         "Learning rate.")
+add_arg('lr_decay_strategy', str,     None,         "Learning rate decay strategy. 'piecewise_decay' or None is valid.")
+add_arg('l2decay',           float,   4e-4,         "L2 decay rate.")
+add_arg('momentum',          float,   0.9,         "Momentum rate.")
+add_arg('gradient_clip',     float,   10.0,         "The threshold of gradient clipping.")
+
+add_arg('total_step',        int,     720000,    "The number of iterations. Zero or less means whole training set. More than 0 means the training set might be looped until # of iterations is reached.")
+add_arg('log_period',        int,     1000,       "Log period.")
+add_arg('save_model_period', int,     15000,      "Save model period. '-1' means never saving the model.")
+add_arg('eval_period',       int,     15000,      "Evaluate period. '-1' means never evaluating the model.")
+add_arg('save_model_dir',    str,     "./models", "The directory the model to be saved to.")
+add_arg('train_images',      str,     None,       "The directory of images to be used for training.")
+add_arg('train_list',        str,     None,       "The list file of images to be used for training.")
+add_arg('test_images',       str,     None,       "The directory of images to be used for test.")
+add_arg('test_list',         str,     None,       "The list file of images to be used for training.")
+add_arg('model',             str,     "crnn_ctc",           "Which type of network to be used. 'crnn_ctc' or 'attention'")
+add_arg('init_model',        str,     None,       "The init model file of directory.")
+add_arg('use_gpu',           bool,    True,      "Whether use GPU to train.")
+add_arg('min_average_window',int,     10000,     "Min average window.")
+add_arg('max_average_window',int,     12500,     "Max average window. It is proposed to be set as the number of minibatch in a pass.")
+add_arg('average_window',    float,   0.15,      "Average window.")
+add_arg('parallel',          bool,    False,     "Whether use parallel training.")
+add_arg('profile',           bool,    False,      "Whether to use profiling.")
+add_arg('skip_batch_num',    int,     0,          "The number of first minibatches to skip as warm-up for better performance test.")
+add_arg('skip_test',         bool,    False,      "Whether to skip test phase.")
 # yapf: enable
 
 
@@ -131,12 +138,13 @@ def train(args):
         for data in test_reader():
             exe.run(inference_program, feed=get_feeder_data(data, place))
         _, test_seq_error = error_evaluator.eval(exe)
-        print("\nTime: %s; Iter[%d]; Test seq error: %s.\n" %
-              (time.time(), iter_num, str(test_seq_error[0])))
+        print("\n[%s] - Iter[%d]; Test seq error: %s.\n" %
+              (time.asctime( time.localtime(time.time())), iter_num, str(test_seq_error[0])))
 
         #Note: The following logs are special for CE monitoring.
         #Other situations do not need to care about these logs.
-        print("kpis	test_acc	%f" % (1 - test_seq_error[0]))
+        if 'ce_mode' in os.environ:
+            print("kpis	test_acc	%f" % (1 - test_seq_error[0]))
 
     def save_model(args, exe, iter_num):
         filename = "model_%05d" % iter_num
@@ -171,14 +179,15 @@ def train(args):
             iter_num += 1
             # training log
             if iter_num % args.log_period == 0:
-                print("\nTime: %s; Iter[%d]; Avg loss: %.3f; Avg seq err: %.3f"
-                      % (time.time(), iter_num,
+                print("\n[%s] - Iter[%d]; Avg loss: %.3f; Avg seq err: %.3f"
+                      % (time.asctime( time.localtime(time.time())), iter_num,
                          total_loss / (args.log_period * args.batch_size),
                          total_seq_error / (args.log_period * args.batch_size)))
-                print("kpis	train_cost	%f" % (total_loss / (args.log_period *
+                if 'ce_mode' in os.environ:
+                    print("kpis	train_cost	%f" % (total_loss / (args.log_period *
                                                             args.batch_size)))
-                print("kpis	train_acc	%f" % (
-                    1 - total_seq_error / (args.log_period * args.batch_size)))
+                    print("kpis	train_acc	%f" % (
+                        1 - total_seq_error / (args.log_period * args.batch_size)))
                 total_loss = 0.0
                 total_seq_error = 0.0
 
@@ -198,7 +207,8 @@ def train(args):
                 else:
                     save_model(args, exe, iter_num)
         end_time = time.time()
-        print("kpis	train_duration	%f" % (end_time - start_time))
+        if 'ce_mode' in os.environ:
+            print("kpis	train_duration	%f" % (end_time - start_time))
         # Postprocess benchmark data
         latencies = batch_times[args.skip_batch_num:]
         latency_avg = np.average(latencies)