From 0fa990bbc9de9d2afa94b87cd38987389712d285 Mon Sep 17 00:00:00 2001 From: peterzhang2029 Date: Wed, 25 Oct 2017 19:09:31 +0800 Subject: [PATCH] add config and refine doc --- scene_text_recognition/README.md | 116 ++++++++++--------- scene_text_recognition/config.py | 75 ++++++++++++ scene_text_recognition/data_provider.py | 100 ---------------- scene_text_recognition/index.html | 116 ++++++++++--------- scene_text_recognition/infer.py | 51 +++++--- scene_text_recognition/model.py | 74 +++++++----- scene_text_recognition/reader.py | 62 ++++++++++ scene_text_recognition/requirements.txt | 2 + scene_text_recognition/train.py | 148 +++++++++++------------- scene_text_recognition/utils.py | 59 ++++++++++ 10 files changed, 461 insertions(+), 342 deletions(-) create mode 100644 scene_text_recognition/config.py delete mode 100644 scene_text_recognition/data_provider.py create mode 100644 scene_text_recognition/reader.py create mode 100644 scene_text_recognition/requirements.txt create mode 100644 scene_text_recognition/utils.py diff --git a/scene_text_recognition/README.md b/scene_text_recognition/README.md index de1418dd..5e83a68e 100644 --- a/scene_text_recognition/README.md +++ b/scene_text_recognition/README.md @@ -4,7 +4,7 @@ 在现实生活中，包括路牌、菜单、大厦标语在内的很多场景均会有文字出现，这些场景的照片中的文字为图片场景的理解提供了更多信息，\[[1](#参考文献)\]使用深度学习模型自动识别路牌中的文字，帮助街景应用获取更加准确的地址信息。 -本例将演示如何用 PaddlePaddle 完成 **场景文字识别 (STR, Scene Text Recognition)** 任务。以下图为例，给定一个场景图片，STR需要从图片中识别出对应的文字"keep": +本例将演示如何用 PaddlePaddle 完成 **场景文字识别 (STR, Scene Text Recognition)** 任务。以下图为例，给定一个场景图片，STR需要从图片中识别出对应的文字"keep"。


@@ -14,70 +14,66 @@ ## 使用 PaddlePaddle 训练与预测 +### 安装依赖包 +```bash +pip install -r requirements.txt +``` + +### 指定训练配置参数 + +通过 `config.py` 脚本修改训练和模型配置参数，脚本中有对可配置参数的详细解释，示例如下： +```python +class TrainerConfig(object): + + # Whether to use GPU in training or not. + use_gpu = True + # The number of computing threads. + trainer_count = 1 + + # The training batch size. + batch_size = 10 + + ... + + +class ModelConfig(object): + + # Number of the filters for convolution group. + filter_num = 8 + + ... +``` +修改 `config.py` 对参数进行调整。例如，通过修改 `use_gpu` 参数来指定是否使用 GPU 进行训练。 + ### 模型训练 训练脚本 [./train.py](./train.py) 中设置了如下命令行参数： ``` -usage: train.py [-h] --image_shape IMAGE_SHAPE --train_file_list - TRAIN_FILE_LIST --test_file_list TEST_FILE_LIST - [--batch_size BATCH_SIZE] - [--model_output_prefix MODEL_OUTPUT_PREFIX] - [--trainer_count TRAINER_COUNT] - [--save_period_by_batch SAVE_PERIOD_BY_BATCH] - [--num_passes NUM_PASSES] - -PaddlePaddle CTC example - -optional arguments: - -h, --help show this help message and exit - --image_shape IMAGE_SHAPE - image's shape, format is like '173,46' - --train_file_list TRAIN_FILE_LIST - path of the file which contains path list of train - image files - --test_file_list TEST_FILE_LIST - path of the file which contains path list of test - image files - --batch_size BATCH_SIZE - size of a mini-batch - --model_output_prefix MODEL_OUTPUT_PREFIX - prefix of path for model to store (default: - ./model.ctc) - --trainer_count TRAINER_COUNT - number of training threads - --save_period_by_batch SAVE_PERIOD_BY_BATCH - save model to disk every N batches - --num_passes NUM_PASSES - number of passes to train (default: 1) -``` +Options: + --train_file_list_path TEXT The path of the file which contains path list + of train image files. [required] + --test_file_list_path TEXT The path of the file which contains path list + of test image files. [required] + --model_save_dir TEXT The path to save the trained models (default: + 'models'). + --help Show this message and exit. -重要的几个参数包括： +``` -- `image_shape` 图片的尺寸 - `train_file_list` 训练数据的列表文件，每行一个路径加对应的text，具体格式为： ``` word_1.png, "PROPER" word_2.png, "FOOD" ``` -- `test_file_list` 测试数据的列表文件，格式同上 - -### 预测 -预测部分由infer.py完成，使用的是最优路径解码算法，即：在每个时间步选择一个概率最大的字符。在使用过程中，需要在infer.py中指定具体的模型目录、图片固定尺寸、batch_size和图片文件的列表文件。例如： -```python -model_path = "model.ctc-pass-9-batch-150-test.tar.gz" -image_shape = "173,46" -batch_size = 50 -infer_file_list = 'data/test_data/Challenge2_Test_Task3_GT.txt' -``` -然后运行```python infer.py``` - +- `test_file_list` 测试数据的列表文件，格式同上。 +- `model_save_dir` 模型参数会的保存目录目录， 默认为当前目录下的`models`目录。 ### 具体执行的过程： 1.从官方网站下载数据\[[2](#参考文献)\]（Task 2.3: Word Recognition (2013 edition)），会有三个文件: Challenge2_Training_Task3_Images_GT.zip、Challenge2_Test_Task3_Images.zip和 Challenge2_Test_Task3_GT.txt。 分别对应训练集的图片和图片对应的单词，测试集的图片，测试数据对应的单词，然后执行以下命令，对数据解压并移动至目标文件夹： -``` +```bash mkdir -p data/train_data mkdir -p data/test_data unzip Challenge2_Training_Task3_Images_GT.zip -d data/train_data @@ -85,16 +81,26 @@ unzip Challenge2_Test_Task3_Images.zip -d data/test_data mv Challenge2_Test_Task3_GT.txt data/test_data ``` -2.获取训练数据文件夹中 `gt.txt` 的路径 (data/train_data）和测试数据文件夹中`Challenge2_Test_Task3_GT.txt`的路径(data/test_data) +2.获取训练数据文件夹中 `gt.txt` 的路径 (data/train_data）和测试数据文件夹中`Challenge2_Test_Task3_GT.txt`的路径(data/test_data)。 -3.执行命令 +3.执行如下命令进行训练： +```bash +python train.py \ +--train_file_list_path 'data/train_data/gt.txt' \ +--test_file_list_path 'data/test_data/Challenge2_Test_Task3_GT.txt' ``` -python train.py --train_file_list data/train_data/gt.txt --test_file_list data/test_data/Challenge2_Test_Task3_GT.txt --image_shape '173,46' -``` -4.训练过程中，模型参数会自动备份到指定目录，默认为 ./model.ctc +4.训练过程中，模型参数会自动备份到指定目录，默认会保存在 `./models` 目录下。 -5.设置infer.py中的相关参数(模型所在路径)，运行```python infer.py``` 进行预测 +### 预测 +预测部分由 `infer.py` 完成，使用的是最优路径解码算法，即：在每个时间步选择一个概率最大的字符。在使用过程中，需要在 `infer.py` 中指定具体的模型目录、图片固定尺寸、batch_size（默认设置为10）和图片文件的列表文件。执行如下代码： +```bash +python infer.py \ +--model_path 'models/params_pass_00000.tar.gz' \ +--image_shape '173,46' \ +--infer_file_list_path 'data/test_data/Challenge2_Test_Task3_GT.txt' +``` +即可进行预测。 ### 其他数据集 @@ -104,7 +110,7 @@ python train.py --train_file_list data/train_data/gt.txt --test_file_list data/t ### 注意事项 - 由于模型依赖的 `warp CTC` 只有CUDA的实现，本模型只支持 GPU 运行 -- 本模型参数较多，占用显存比较大，实际执行时可以调节batch_size 控制显存占用 +- 本模型参数较多，占用显存比较大，实际执行时可以调节`batch_size`控制显存占用 - 本模型使用的数据集较小，可以选用其他更大的数据集\[[3](#参考文献)\]来训练需要的模型 ## 参考文献 diff --git a/scene_text_recognition/config.py b/scene_text_recognition/config.py new file mode 100644 index 00000000..9cc56354 --- /dev/null +++ b/scene_text_recognition/config.py @@ -0,0 +1,75 @@ +__all__ = ["TrainerConfig", "ModelConfig"] + + +class TrainerConfig(object): + + # Whether to use GPU in training or not. + use_gpu = True + + # The number of computing threads. + trainer_count = 1 + + # The training batch size. + batch_size = 10 + + # The epoch number. + num_passes = 10 + + # Parameter updates momentum. + momentum = 0 + + # The shape of images. + image_shape = (173, 46) + + # The buffer size of the data reader. + # The number of buffer size samples will be shuffled in training. + buf_size = 1000 + + # The parameter is used to control logging period. + # Training log will be printed every log_period. + log_period = 50 + + +class ModelConfig(object): + + # Number of the filters for convolution group. + filter_num = 8 + + # Use batch normalization or not in image convolution group. + with_bn = True + + # The number of channels for block expand layer. + num_channels = 128 + + # The parameter stride_x in block expand layer. + stride_x = 1 + + # The parameter stride_y in block expand layer. + stride_y = 1 + + # The parameter block_x in block expand layer. + block_x = 1 + + # The parameter block_y in block expand layer. + block_y = 11 + + # The hidden size for gru. + hidden_size = num_channels + + # Use norm_by_times or not in warp ctc layer. + norm_by_times = True + + # The list for number of filter in image convolution group layer. + filter_num_list = [16, 32, 64, 128] + + # The parameter conv_padding in image convolution group layer. + conv_padding = 1 + + # The parameter conv_filter_size in image convolution group layer. + conv_filter_size = 3 + + # The parameter pool_size in image convolution group layer. + pool_size = 2 + + # The parameter pool_stride in image convolution group layer. + pool_stride = 2 diff --git a/scene_text_recognition/data_provider.py b/scene_text_recognition/data_provider.py deleted file mode 100644 index f33a102e..00000000 --- a/scene_text_recognition/data_provider.py +++ /dev/null @@ -1,100 +0,0 @@ -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -import os -import cv2 - -from paddle.v2.image import load_image - - -class AsciiDic(object): - UNK = 0 - - def __init__(self): - self.dic = { - '': self.UNK, - } - self.chars = [chr(i) for i in range(40, 171)] - for id, c in enumerate(self.chars): - self.dic[c] = id + 1 - - def lookup(self, w): - return self.dic.get(w, self.UNK) - - def id2word(self): - self.id2word = {} - for key, value in self.dic.items(): - self.id2word[value] = key - - return self.id2word - - def word2ids(self, sent): - ''' - transform a word to a list of ids. - ''' - return [self.lookup(c) for c in list(sent)] - - def size(self): - return len(self.dic) - - -class ImageDataset(object): - def __init__(self, - train_image_paths_generator, - test_image_paths_generator, - infer_image_paths_generator, - fixed_shape=None, - is_infer=False): - ''' - :param train_image_paths_generator: - return list of train images' paths. - :type train_image_paths_generator: function - :param fixed_shape: fixed shape of images. - :type fixed_shape: tuple - ''' - if is_infer == False: - self.train_filelist = [p for p in train_image_paths_generator] - self.test_filelist = [p for p in test_image_paths_generator] - else: - self.infer_filelist = [p for p in infer_image_paths_generator] - - self.fixed_shape = fixed_shape - self.ascii_dic = AsciiDic() - - def train(self): - for i, (image, label) in enumerate(self.train_filelist): - yield self.load_image(image), self.ascii_dic.word2ids(label) - - def test(self): - for i, (image, label) in enumerate(self.test_filelist): - yield self.load_image(image), self.ascii_dic.word2ids(label) - - def infer(self): - for i, (image, label) in enumerate(self.infer_filelist): - yield self.load_image(image), label - - def load_image(self, path): - ''' - load image and transform to 1-dimention vector - ''' - image = load_image(path) - image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) - # resize all images to a fixed shape - - if self.fixed_shape: - image = cv2.resize( - image, self.fixed_shape, interpolation=cv2.INTER_CUBIC) - - image = image.flatten() / 255. - return image - - -def get_file_list(image_file_list): - pwd = os.path.dirname(image_file_list) - with open(image_file_list) as f: - for line in f: - fs = line.strip().split(',', 1) - file = fs[0].strip() - path = os.path.join(pwd, file) - yield path, fs[1][2:-1] diff --git a/scene_text_recognition/index.html b/scene_text_recognition/index.html index 46996528..64a1160a 100644 --- a/scene_text_recognition/index.html +++ b/scene_text_recognition/index.html @@ -46,7 +46,7 @@ 在现实生活中，包括路牌、菜单、大厦标语在内的很多场景均会有文字出现，这些场景的照片中的文字为图片场景的理解提供了更多信息，\[[1](#参考文献)\]使用深度学习模型自动识别路牌中的文字，帮助街景应用获取更加准确的地址信息。 -本例将演示如何用 PaddlePaddle 完成 **场景文字识别 (STR, Scene Text Recognition)** 任务。以下图为例，给定一个场景图片，STR需要从图片中识别出对应的文字"keep": +本例将演示如何用 PaddlePaddle 完成 **场景文字识别 (STR, Scene Text Recognition)** 任务。以下图为例，给定一个场景图片，STR需要从图片中识别出对应的文字"keep"。


@@ -56,70 +56,66 @@ ## 使用 PaddlePaddle 训练与预测 +### 安装依赖包 +```bash +pip install -r requirements.txt +``` + +### 指定训练配置参数 + +通过 `config.py` 脚本修改训练和模型配置参数，脚本中有对可配置参数的详细解释，示例如下： +```python +class TrainerConfig(object): + + # Whether to use GPU in training or not. + use_gpu = True + # The number of computing threads. + trainer_count = 1 + + # The training batch size. + batch_size = 10 + + ... + + +class ModelConfig(object): + + # Number of the filters for convolution group. + filter_num = 8 + + ... +``` +修改 `config.py` 对参数进行调整。例如，通过修改 `use_gpu` 参数来指定是否使用 GPU 进行训练。 + ### 模型训练 训练脚本 [./train.py](./train.py) 中设置了如下命令行参数： ``` -usage: train.py [-h] --image_shape IMAGE_SHAPE --train_file_list - TRAIN_FILE_LIST --test_file_list TEST_FILE_LIST - [--batch_size BATCH_SIZE] - [--model_output_prefix MODEL_OUTPUT_PREFIX] - [--trainer_count TRAINER_COUNT] - [--save_period_by_batch SAVE_PERIOD_BY_BATCH] - [--num_passes NUM_PASSES] - -PaddlePaddle CTC example - -optional arguments: - -h, --help show this help message and exit - --image_shape IMAGE_SHAPE - image's shape, format is like '173,46' - --train_file_list TRAIN_FILE_LIST - path of the file which contains path list of train - image files - --test_file_list TEST_FILE_LIST - path of the file which contains path list of test - image files - --batch_size BATCH_SIZE - size of a mini-batch - --model_output_prefix MODEL_OUTPUT_PREFIX - prefix of path for model to store (default: - ./model.ctc) - --trainer_count TRAINER_COUNT - number of training threads - --save_period_by_batch SAVE_PERIOD_BY_BATCH - save model to disk every N batches - --num_passes NUM_PASSES - number of passes to train (default: 1) -``` +Options: + --train_file_list_path TEXT The path of the file which contains path list + of train image files. [required] + --test_file_list_path TEXT The path of the file which contains path list + of test image files. [required] + --model_save_dir TEXT The path to save the trained models (default: + 'models'). + --help Show this message and exit. -重要的几个参数包括： +``` -- `image_shape` 图片的尺寸 - `train_file_list` 训练数据的列表文件，每行一个路径加对应的text，具体格式为： ``` word_1.png, "PROPER" word_2.png, "FOOD" ``` -- `test_file_list` 测试数据的列表文件，格式同上 - -### 预测 -预测部分由infer.py完成，使用的是最优路径解码算法，即：在每个时间步选择一个概率最大的字符。在使用过程中，需要在infer.py中指定具体的模型目录、图片固定尺寸、batch_size和图片文件的列表文件。例如： -```python -model_path = "model.ctc-pass-9-batch-150-test.tar.gz" -image_shape = "173,46" -batch_size = 50 -infer_file_list = 'data/test_data/Challenge2_Test_Task3_GT.txt' -``` -然后运行```python infer.py``` - +- `test_file_list` 测试数据的列表文件，格式同上。 +- `model_save_dir` 模型参数会的保存目录目录， 默认为当前目录下的`models`目录。 ### 具体执行的过程： 1.从官方网站下载数据\[[2](#参考文献)\]（Task 2.3: Word Recognition (2013 edition)），会有三个文件: Challenge2_Training_Task3_Images_GT.zip、Challenge2_Test_Task3_Images.zip和 Challenge2_Test_Task3_GT.txt。 分别对应训练集的图片和图片对应的单词，测试集的图片，测试数据对应的单词，然后执行以下命令，对数据解压并移动至目标文件夹： -``` +```bash mkdir -p data/train_data mkdir -p data/test_data unzip Challenge2_Training_Task3_Images_GT.zip -d data/train_data @@ -127,16 +123,26 @@ unzip Challenge2_Test_Task3_Images.zip -d data/test_data mv Challenge2_Test_Task3_GT.txt data/test_data ``` -2.获取训练数据文件夹中 `gt.txt` 的路径 (data/train_data）和测试数据文件夹中`Challenge2_Test_Task3_GT.txt`的路径(data/test_data) +2.获取训练数据文件夹中 `gt.txt` 的路径 (data/train_data）和测试数据文件夹中`Challenge2_Test_Task3_GT.txt`的路径(data/test_data)。 -3.执行命令 +3.执行如下命令进行训练： +```bash +python train.py \ +--train_file_list_path 'data/train_data/gt.txt' \ +--test_file_list_path 'data/test_data/Challenge2_Test_Task3_GT.txt' ``` -python train.py --train_file_list data/train_data/gt.txt --test_file_list data/test_data/Challenge2_Test_Task3_GT.txt --image_shape '173,46' -``` -4.训练过程中，模型参数会自动备份到指定目录，默认为 ./model.ctc +4.训练过程中，模型参数会自动备份到指定目录，默认会保存在 `./models` 目录下。 -5.设置infer.py中的相关参数(模型所在路径)，运行```python infer.py``` 进行预测 +### 预测 +预测部分由 `infer.py` 完成，使用的是最优路径解码算法，即：在每个时间步选择一个概率最大的字符。在使用过程中，需要在 `infer.py` 中指定具体的模型目录、图片固定尺寸、batch_size（默认设置为10）和图片文件的列表文件。执行如下代码： +```bash +python infer.py \ +--model_path 'models/params_pass_00000.tar.gz' \ +--image_shape '173,46' \ +--infer_file_list_path 'data/test_data/Challenge2_Test_Task3_GT.txt' +``` +即可进行预测。 ### 其他数据集 @@ -146,7 +152,7 @@ python train.py --train_file_list data/train_data/gt.txt --test_file_list data/t ### 注意事项 - 由于模型依赖的 `warp CTC` 只有CUDA的实现，本模型只支持 GPU 运行 -- 本模型参数较多，占用显存比较大，实际执行时可以调节batch_size 控制显存占用 +- 本模型参数较多，占用显存比较大，实际执行时可以调节`batch_size`控制显存占用 - 本模型使用的数据集较小，可以选用其他更大的数据集\[[3](#参考文献)\]来训练需要的模型 ## 参考文献 diff --git a/scene_text_recognition/infer.py b/scene_text_recognition/infer.py index ff1f43be..b53c600b 100644 --- a/scene_text_recognition/infer.py +++ b/scene_text_recognition/infer.py @@ -1,11 +1,11 @@ -import logging -import argparse +import click import gzip import paddle.v2 as paddle from model import Model -from data_provider import get_file_list, AsciiDic, ImageDataset +from reader import DataGenerator from decoder import ctc_greedy_decoder +from utils import AsciiDic, get_file_list def infer_batch(inferer, test_batch, labels): @@ -15,9 +15,8 @@ def infer_batch(inferer, test_batch, labels): infer_results[i * num_steps:(i + 1) * num_steps] for i in xrange(0, len(test_batch)) ] - results = [] - # best path decode + # Best path decode. for i, probs in enumerate(probs_split): output_transcription = ctc_greedy_decoder( probs_seq=probs, vocabulary=AsciiDic().id2word()) @@ -28,21 +27,42 @@ def infer_batch(inferer, test_batch, labels): (result, label)) -def infer(model_path, image_shape, batch_size, infer_file_list): +@click.command('infer') +@click.option( + "--model_path", type=str, required=True, help=("The path of saved model.")) +@click.option( + "--image_shape", + type=str, + required=True, + help=("The fixed size for image dataset (format is like: '173,46').")) +@click.option( + "--batch_size", + type=int, + default=10, + help=("The number of examples in one batch (default: 10).")) +@click.option( + "--infer_file_list_path", + type=str, + required=True, + help=("The path of the file which contains " + "path list of image files for inference.")) +def infer(model_path, image_shape, batch_size, infer_file_list_path): image_shape = tuple(map(int, image_shape.split(','))) - infer_generator = get_file_list(infer_file_list) - - dataset = ImageDataset(None, None, infer_generator, image_shape, True) + infer_file_list = get_file_list(infer_file_list_path) + char_dict = AsciiDic() + dict_size = char_dict.size() + data_generator = DataGenerator(char_dict=char_dict, image_shape=image_shape) - paddle.init(use_gpu=True, trainer_count=4) + paddle.init(use_gpu=True, trainer_count=1) parameters = paddle.parameters.Parameters.from_tar(gzip.open(model_path)) - model = Model(AsciiDic().size(), image_shape, is_infer=True) + model = Model(dict_size, image_shape, is_infer=True) inferer = paddle.inference.Inference( output_layer=model.log_probs, parameters=parameters) test_batch = [] labels = [] - for i, (image, label) in enumerate(dataset.infer()): + for i, (image, + label) in enumerate(data_generator.infer_reader(infer_file_list)()): test_batch.append([image]) labels.append(label) if len(test_batch) == batch_size: @@ -54,9 +74,4 @@ def infer(model_path, image_shape, batch_size, infer_file_list): if __name__ == "__main__": - model_path = "model.ctc-pass-9-batch-150-test.tar.gz" - image_shape = "173,46" - batch_size = 50 - infer_file_list = 'data/test_data/Challenge2_Test_Task3_GT.txt' - - infer(model_path, image_shape, batch_size, infer_file_list) + infer() diff --git a/scene_text_recognition/model.py b/scene_text_recognition/model.py index 2ea1240d..86dd852c 100644 --- a/scene_text_recognition/model.py +++ b/scene_text_recognition/model.py @@ -3,16 +3,17 @@ from paddle.v2 import layer from paddle.v2 import evaluator from paddle.v2.activation import Relu, Linear from paddle.v2.networks import img_conv_group, simple_gru +from config import ModelConfig as conf class Model(object): def __init__(self, num_classes, shape, is_infer=False): ''' - :param num_classes: size of the character dict. + :param num_classes: The size of the character dict. :type num_classes: int - :param shape: size of the input images. + :param shape: The size of the input images. :type shape: tuple of 2 int - :param is_infer: infer mode or not + :param is_infer: For inference or not :type shape: bool ''' self.num_classes = num_classes @@ -24,39 +25,50 @@ class Model(object): self.__build_nn__() def __declare_input_layers__(self): - # image input as a float vector + ''' + Define the input layer. + ''' + # Image input as a float vector. self.image = layer.data( name='image', type=paddle.data_type.dense_vector(self.image_vector_size), height=self.shape[0], width=self.shape[1]) - # label input as a ID list - if self.is_infer == False: + # Label input as an ID list + if not self.is_infer: self.label = layer.data( name='label', type=paddle.data_type.integer_value_sequence(self.num_classes)) def __build_nn__(self): - # CNN output image features, 128 float matrixes - conv_features = self.conv_groups(self.image, 8, True) + ''' + Build the network topology. + ''' + # CNN output image features. + conv_features = self.conv_groups(self.image, conf.filter_num, + conf.with_bn) - # cutting CNN output into a sequence of feature vectors, which are + # Cut CNN output into a sequence of feature vectors, which are # 1 pixel wide and 11 pixel high. sliced_feature = layer.block_expand( input=conv_features, - num_channels=128, - stride_x=1, - stride_y=1, - block_x=1, - block_y=11) + num_channels=conf.num_channels, + stride_x=conf.stride_x, + stride_y=conf.stride_y, + block_x=conf.block_x, + block_y=conf.block_y) # RNNs to capture sequence information forwards and backwards. - gru_forward = simple_gru(input=sliced_feature, size=128, act=Relu()) + gru_forward = simple_gru( + input=sliced_feature, size=conf.hidden_size, act=Relu()) gru_backward = simple_gru( - input=sliced_feature, size=128, act=Relu(), reverse=True) + input=sliced_feature, + size=conf.hidden_size, + act=Relu(), + reverse=True) - # map each step of RNN to character distribution. + # Map each step of RNN to character distribution. self.output = layer.fc( input=[gru_forward, gru_backward], size=self.num_classes + 1, @@ -66,31 +78,31 @@ class Model(object): input=paddle.layer.identity_projection(input=self.output), act=paddle.activation.Softmax()) - # warp CTC to calculate cost for a CTC task. - if self.is_infer == False: + # Use warp CTC to calculate cost for a CTC task. + if not self.is_infer: self.cost = layer.warp_ctc( input=self.output, label=self.label, size=self.num_classes + 1, - norm_by_times=True, + norm_by_times=conf.norm_by_times, blank=self.num_classes) self.eval = evaluator.ctc_error(input=self.output, label=self.label) - def conv_groups(self, input_image, num, with_bn): + def conv_groups(self, input, num, with_bn): ''' - :param input_image: input image. - :type input_image: LayerOutput - :param num: number of CONV filters. + :param input: Input layer. + :type input: LayerOutput + :param num: Number of the filters. :type num: int - :param with_bn: whether with batch normal. + :param with_bn: Whether with batch normalization. :type with_bn: bool ''' assert num % 4 == 0 - filter_num_list = [16, 32, 64, 128] + filter_num_list = conf.filter_num_list is_input_image = True - tmp = input_image + tmp = input for num_filter in filter_num_list: @@ -103,12 +115,12 @@ class Model(object): tmp = img_conv_group( input=tmp, num_channels=num_channels, - conv_padding=1, + conv_padding=conf.conv_padding, conv_num_filter=[num_filter] * (num / 4), - conv_filter_size=3, + conv_filter_size=conf.conv_filter_size, conv_act=Relu(), conv_with_batchnorm=with_bn, - pool_size=2, - pool_stride=2, ) + pool_size=conf.pool_size, + pool_stride=conf.pool_stride, ) return tmp diff --git a/scene_text_recognition/reader.py b/scene_text_recognition/reader.py new file mode 100644 index 00000000..013477ad --- /dev/null +++ b/scene_text_recognition/reader.py @@ -0,0 +1,62 @@ +import os +import cv2 + +from paddle.v2.image import load_image + + +class DataGenerator(object): + def __init__(self, char_dict, image_shape): + ''' + :param char_dict: The dictionary class for labels. + :type char_dict: class + :param image_shape: The fixed shape of images. + :type image_shape: tuple + ''' + self.image_shape = image_shape + self.char_dict = char_dict + + def train_reader(self, file_list): + ''' + Reader interface for training. + + :param file_list: The path list of the image file for training. + :type file_list: list + ''' + + def reader(): + for i, (image, label) in enumerate(file_list): + yield self.load_image(image), self.char_dict.word2ids(label) + + return reader + + def infer_reader(self, file_list): + ''' + Reader interface for inference. + + :param file_list: The path list of the image file for inference. + :type file_list: list + ''' + + def reader(): + for i, (image, label) in enumerate(file_list): + yield self.load_image(image), label + + return reader + + def load_image(self, path): + ''' + Load image and transform to 1-dimention vector. + + :param path: The path of the image data. + :type path: str + ''' + image = load_image(path) + image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) + + # Resize all images to a fixed shape. + if self.image_shape: + image = cv2.resize( + image, self.image_shape, interpolation=cv2.INTER_CUBIC) + + image = image.flatten() / 255. + return image diff --git a/scene_text_recognition/requirements.txt b/scene_text_recognition/requirements.txt new file mode 100644 index 00000000..eb8ed79b --- /dev/null +++ b/scene_text_recognition/requirements.txt @@ -0,0 +1,2 @@ +click +opencv-python \ No newline at end of file diff --git a/scene_text_recognition/train.py b/scene_text_recognition/train.py index 212102c5..557f1ba5 100644 --- a/scene_text_recognition/train.py +++ b/scene_text_recognition/train.py @@ -1,109 +1,91 @@ -import logging -import argparse import gzip +import os +import click import paddle.v2 as paddle +from config import TrainerConfig as conf from model import Model -from data_provider import get_file_list, AsciiDic, ImageDataset +from reader import DataGenerator +from utils import get_file_list, AsciiDic -parser = argparse.ArgumentParser(description="PaddlePaddle CTC example") -parser.add_argument( - '--image_shape', - type=str, - required=True, - help="image's shape, format is like '173,46'") -parser.add_argument( - '--train_file_list', + +@click.command('train') +@click.option( + "--train_file_list_path", type=str, required=True, - help='path of the file which contains path list of train image files') -parser.add_argument( - '--test_file_list', + help=("The path of the file which contains " + "path list of train image files.")) +@click.option( + "--test_file_list_path", type=str, required=True, - help='path of the file which contains path list of test image files') -parser.add_argument( - '--batch_size', type=int, default=5, help='size of a mini-batch') -parser.add_argument( - '--model_output_prefix', + help=("The path of the file which contains " + "path list of test image files.")) +@click.option( + "--model_save_dir", type=str, - default='model.ctc', - help='prefix of path for model to store (default: ./model.ctc)') -parser.add_argument( - '--trainer_count', type=int, default=4, help='number of training threads') -parser.add_argument( - '--save_period_by_batch', - type=int, - default=150, - help='save model to disk every N batches') -parser.add_argument( - '--num_passes', - type=int, - default=10, - help='number of passes to train (default: 1)') - -args = parser.parse_args() - - -def main(): - image_shape = tuple(map(int, args.image_shape.split(','))) - - print 'image_shape', image_shape - print 'batch_size', args.batch_size - print 'train_file_list', args.train_file_list - print 'test_file_list', args.test_file_list - - train_generator = get_file_list(args.train_file_list) - test_generator = get_file_list(args.test_file_list) - infer_generator = None - - dataset = ImageDataset( - train_generator, - test_generator, - infer_generator, - fixed_shape=image_shape, - is_infer=False) - - paddle.init(use_gpu=True, trainer_count=args.trainer_count) - - model = Model(AsciiDic().size(), image_shape, is_infer=False) + default="models", + help="The path to save the trained models (default: 'models').") +def train(train_file_list_path, test_file_list_path, model_save_dir): + + if not os.path.exists(model_save_dir): + os.mkdir(model_save_dir) + train_file_list = get_file_list(train_file_list_path) + test_file_list = get_file_list(test_file_list_path) + char_dict = AsciiDic() + dict_size = char_dict.size() + data_generator = DataGenerator( + char_dict=char_dict, image_shape=conf.image_shape) + + paddle.init(use_gpu=conf.use_gpu, trainer_count=conf.trainer_count) + # Create optimizer. + optimizer = paddle.optimizer.Momentum(momentum=conf.momentum) + # Define network topology. + model = Model(dict_size, conf.image_shape, is_infer=False) + # Create all the trainable parameters. params = paddle.parameters.create(model.cost) - optimizer = paddle.optimizer.Momentum(momentum=0) + trainer = paddle.trainer.SGD( cost=model.cost, parameters=params, update_equation=optimizer, extra_layers=model.eval) + # Feeding dictionary. + feeding = {'image': 0, 'label': 1} def event_handler(event): if isinstance(event, paddle.event.EndIteration): - if event.batch_id % 100 == 0: - print "Pass %d, batch %d, Samples %d, Cost %f, Eval %s" % ( - event.pass_id, event.batch_id, - event.batch_id * args.batch_size, event.cost, event.metrics) - - if event.batch_id > 0 and event.batch_id % args.save_period_by_batch == 0: - result = trainer.test( - reader=paddle.batch(dataset.test, batch_size=10), - feeding={'image': 0, - 'label': 1}) - print "Test %d-%d, Cost %f, Eval %s" % ( - event.pass_id, event.batch_id, result.cost, result.metrics) - - path = "{}-pass-{}-batch-{}-test.tar.gz".format( - args.model_output_prefix, event.pass_id, event.batch_id) - with gzip.open(path, 'w') as f: - params.to_tar(f) + if event.batch_id % conf.log_period == 0: + print("Pass %d, batch %d, Samples %d, Cost %f, Eval %s" % + (event.pass_id, event.batch_id, event.batch_id * + conf.batch_size, event.cost, event.metrics)) + + if isinstance(event, paddle.event.EndPass): + # Here, because training and testing data share a same format, + # we still use the reader.train_reader to read the testing data. + result = trainer.test( + reader=paddle.batch( + data_generator.train_reader(test_file_list), + batch_size=conf.batch_size), + feeding=feeding) + print("Test %d, Cost %f, Eval %s" % + (event.pass_id, result.cost, result.metrics)) + with gzip.open( + os.path.join(model_save_dir, "params_pass_%05d.tar.gz" % + event.pass_id), "w") as f: + trainer.save_parameter_to_tar(f) trainer.train( reader=paddle.batch( - paddle.reader.shuffle(dataset.train, buf_size=500), - batch_size=args.batch_size), - feeding={'image': 0, - 'label': 1}, + paddle.reader.shuffle( + data_generator.train_reader(train_file_list), + buf_size=conf.buf_size), + batch_size=conf.batch_size), + feeding=feeding, event_handler=event_handler, - num_passes=args.num_passes) + num_passes=conf.num_passes) if __name__ == "__main__": - main() + train() diff --git a/scene_text_recognition/utils.py b/scene_text_recognition/utils.py new file mode 100644 index 00000000..dd43113a --- /dev/null +++ b/scene_text_recognition/utils.py @@ -0,0 +1,59 @@ +import os + + +class AsciiDic(object): + UNK_ID = 0 + + def __init__(self): + self.dic = { + '': self.UNK_ID, + } + self.chars = [chr(i) for i in range(40, 171)] + for id, c in enumerate(self.chars): + self.dic[c] = id + 1 + + def lookup(self, w): + return self.dic.get(w, self.UNK_ID) + + def id2word(self): + ''' + Return a reversed char dict. + ''' + self.id2word = {} + for key, value in self.dic.items(): + self.id2word[value] = key + + return self.id2word + + def word2ids(self, word): + ''' + Transform a word to a list of ids. + + :param word: The word appears in image data. + :type word: str + ''' + return [self.lookup(c) for c in list(word)] + + def size(self): + return len(self.dic) + + +def get_file_list(image_file_list): + ''' + Generate the file list for training and testing data. + + :param image_file_list: The path of the file which contains + path list of image files. + :type image_file_list: str + ''' + dirname = os.path.dirname(image_file_list) + path_list = [] + with open(image_file_list) as f: + for line in f: + line_split = line.strip().split(',', 1) + filename = line_split[0].strip() + path = os.path.join(dirname, filename) + label = line_split[1][2:-1] + path_list.append((path, label)) + + return path_list -- GitLab