[cherry-pick] cherry pick some fix on ocr model, gan and PaddleSlim into model-1.4 (#2108)

* Fix eval.py in ocr ctc demo. (#1970)

* Refine readme in cycle gan. (#1972)

* 1. Refine readme
2. Fix some issues

* Refine README.md.

* Fix atention model. (#2059)

* 1. Enable GC strategy in run.sh (#2063)

2. Fix loading checkpoint in filter pruning strategy by adding init_model option into config file.

PaddleCV/gan/cycle_gan/README.md

-
 
 运行本目录下的程序示例需要使用PaddlePaddle develop最新版本。如果您的PaddlePaddle安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新PaddlePaddle安装版本。
 
@@ -73,8 +72,8 @@ env CUDA_VISIBLE_DEVICES=0 python train.py
 执行以下命令读取多张图片进行预测：
 
 ```
-env CUDA_VISIBLE_DEVICE=0 python infer.py \
-    --init_model="checkpoints/1" --input="./data/inputA/*" \
+env CUDA_VISIBLE_DEVICES=0 python infer.py \
+    --init_model="output/checkpoints/1" --input="./data/horse2zebra/trainA/*" \
     --input_style A --output="./output"
 ```
 
@@ -89,3 +88,5 @@ env CUDA_VISIBLE_DEVICE=0 python infer.py \
 <img src="images/B2A.jpg" width="620" hspace='10'/> <br/>
 <strong>图 3</strong>
 </p>
+
+>在本文示例中，均可通过修改`CUDA_VISIBLE_DEVICES`改变使用的显卡号。

PaddleCV/gan/cycle_gan/infer.py

@@ -44,7 +44,6 @@ def infer(args):
     if not os.path.exists(args.output):
         os.makedirs(args.output)
     for file in glob.glob(args.input):
-        print "read %s" % file
         image_name = os.path.basename(file)
         image = Image.open(file)
         image = image.resize((256, 256))

PaddleCV/gan/cycle_gan/train.py

@@ -22,7 +22,7 @@ add_arg = functools.partial(add_arguments, argparser=parser)
 # yapf: disable
 add_arg('batch_size',        int,   1,          "Minibatch size.")
 add_arg('epoch',             int,   2,        "The number of epoched to be trained.")
-add_arg('output',            str,   "./output_0", "The directory the model and the test result to be saved to.")
+add_arg('output',            str,   "./output", "The directory the model and the test result to be saved to.")
 add_arg('init_model',        str,   None,       "The init model file of directory.")
 add_arg('save_checkpoints',  bool,  True,       "Whether to save checkpoints.")
 add_arg('run_test',          bool,  True,       "Whether to run test.")

PaddleCV/ocr_recognition/README.md

-
 
 运行本目录下的程序示例需要使用PaddlePaddle develop最新版本。如果您的PaddlePaddle安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新PaddlePaddle安装版本。
 
@@ -156,12 +155,13 @@ env CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --parallel=True
 通过以下命令调用评估脚本用指定数据集对模型进行评估：
 
 ```
-env CUDA_VISIBLE_DEVICE=0 python eval.py \
+env CUDA_VISIBLE_DEVICES=0 python eval.py \
     --model_path="./models/model_0" \
     --input_images_dir="./eval_data/images/" \
-    --input_images_list="./eval_data/eval_list\" \
+    --input_images_list="./eval_data/eval_list"
 ```
 
+
 执行`python train.py --help`可查看参数详细说明。
 
 
@@ -170,7 +170,7 @@ env CUDA_VISIBLE_DEVICE=0 python eval.py \
 从标准输入读取一张图片的路径，并对齐进行预测：
 
 ```
-env CUDA_VISIBLE_DEVICE=0 python infer.py \
+env CUDA_VISIBLE_DEVICES=0 python infer.py \
     --model_path="models/model_00044_15000"
 ```
 
@@ -193,7 +193,7 @@ result: [2067 2067 8187 8477 5027 7191 2431 1462]
 从文件中批量读取图片路径，并对其进行预测：
 
 ```
-env CUDA_VISIBLE_DEVICE=0 python infer.py \
+env CUDA_VISIBLE_DEVICES=0 python infer.py \
     --model_path="models/model_00044_15000" \
     --input_images_list="data/test.list"
 ```
@@ -204,3 +204,5 @@ env CUDA_VISIBLE_DEVICE=0 python infer.py \
 |- |:-: |
 |[ocr_ctc_params](https://paddle-ocr-models.bj.bcebos.com/ocr_ctc.zip) | 22.3% |
 |[ocr_attention_params](https://paddle-ocr-models.bj.bcebos.com/ocr_attention.zip) | 15.8%|
+
+>在本文示例中，均可通过修改`CUDA_VISIBLE_DEVICES`改变当前任务使用的显卡号。

PaddleCV/ocr_recognition/attention_model.py

@@ -339,7 +339,7 @@ def attention_infer(images, num_classes, use_cudnn=True):
     return ids
 
 
-def attention_eval(data_shape, num_classes):
+def attention_eval(data_shape, num_classes, use_cudnn=True):
     images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
     label_in = fluid.layers.data(
         name='label_in', shape=[1], dtype='int32', lod_level=1)
@@ -349,7 +349,7 @@ def attention_eval(data_shape, num_classes):
     label_in = fluid.layers.cast(x=label_in, dtype='int64')
 
     gru_backward, encoded_vector, encoded_proj = encoder_net(
-        images, is_test=True)
+        images, is_test=True, use_cudnn=use_cudnn)
 
     backward_first = fluid.layers.sequence_pool(
         input=gru_backward, pool_type='first')

PaddleCV/ocr_recognition/crnn_ctc_model.py

@@ -213,12 +213,12 @@ def ctc_train_net(args, data_shape, num_classes):
     return sum_cost, error_evaluator, inference_program, model_average
 
 
-def ctc_infer(images, num_classes, use_cudnn):
+def ctc_infer(images, num_classes, use_cudnn=True):
     fc_out = encoder_net(images, num_classes, is_test=True, use_cudnn=use_cudnn)
     return fluid.layers.ctc_greedy_decoder(input=fc_out, blank=num_classes)
 
 
-def ctc_eval(data_shape, num_classes, use_cudnn):
+def ctc_eval(data_shape, num_classes, use_cudnn=True):
     images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
     label = fluid.layers.data(
         name='label', shape=[1], dtype='int32', lod_level=1)

PaddleCV/ocr_recognition/data_reader.py

@@ -10,6 +10,11 @@ from os import path
 from paddle.dataset.image import load_image
 import paddle
 
+try:
+    input = raw_input
+except NameError:
+    pass
+
 SOS = 0
 EOS = 1
 NUM_CLASSES = 95
@@ -175,7 +180,7 @@ class DataGenerator(object):
                         yield img, label
             else:
                 while True:
-                    img_path = raw_input("Please input the path of image: ")
+                    img_path = input("Please input the path of image: ")
                     img = Image.open(img_path).convert('L')
                     img = np.array(img) - 127.5
                     img = img[np.newaxis, ...]

PaddleCV/ocr_recognition/eval.py

@@ -31,7 +31,8 @@ def evaluate(args):
     num_classes = data_reader.num_classes()
     data_shape = data_reader.data_shape()
     # define network
-    evaluator, cost = eval(data_shape, num_classes)
+    evaluator, cost = eval(
+        data_shape, num_classes, use_cudnn=True if args.use_gpu else False)
 
     # data reader
     test_reader = data_reader.test(
@@ -62,8 +63,8 @@ def evaluate(args):
         count += 1
         exe.run(fluid.default_main_program(), feed=get_feeder_data(data, place))
     avg_distance, avg_seq_error = evaluator.eval(exe)
-    print("Read %d samples; avg_distance: %s; avg_seq_error: %s" % (
-        count, avg_distance, avg_seq_error))
+    print("Read %d samples; avg_distance: %s; avg_seq_error: %s" %
+          (count, avg_distance, avg_seq_error))
 
 
 def main():

PaddleSlim/configs/filter_pruning_sen.yaml

@@ -20,6 +20,7 @@ strategies:
         metric_name: 'acc_top1'
 compressor:
     epoch: 200
+    #init_model: './checkpoints/0' # Please enable this option for loading checkpoint.
     checkpoint_path: './checkpoints/'
     strategies:
         - sensitive_pruning_strategy

PaddleSlim/configs/filter_pruning_uniform.yaml

@@ -16,6 +16,7 @@ strategies:
         metric_name: 'acc_top1'
 compressor:
     epoch: 200
+    #init_model: './checkpoints/0' # Please enable this option for loading checkpoint.
     checkpoint_path: './checkpoints/'
     strategies:
         - uniform_pruning_strategy

PaddleSlim/configs/quantization_pruning.yaml

@@ -30,6 +30,7 @@ strategies:
         activation_quantize_type: 'abs_max'
 compressor:
     epoch: 142
+    #init_model: './checkpoints/0' # Please enable this option for loading checkpoint.
     checkpoint_path: './checkpoints/'
     strategies:
         - uniform_pruning_strategy

PaddleSlim/run.sh

@@ -24,6 +24,10 @@ fi
 
 cd -
 
+# enable GC strategy
+export FLAGS_fast_eager_deletion_mode=1
+export FLAGS_eager_delete_tensor_gb=0.0
+
 # for distillation
 #-----------------
 export CUDA_VISIBLE_DEVICES=0,1,2,3