TF-CRNN : A TensorFlow implementation of Convolutional Recurrent Neural Network

Quickstart

Installation

tf_crnn uses tensorflow-gpu package, which needs CUDA and CuDNN libraries for GPU support. Tensorflow GPU support page lists the requirements.

Using Anaconda

When using Anaconda (or Miniconda), conda will install automatically the compatible versions of CUDA and CuDNN

conda env create -f environment.yml

From this page:

When the GPU accelerated version of TensorFlow is installed using conda, by the command “conda install tensorflow-gpu”, these libraries are installed automatically, with versions known to be compatible with the tensorflow-gpu package. Furthermore, conda installs these libraries into a location where they will not interfere with other instances of these libraries that may have been installed via another method. Regardless of using pip or conda-installed tensorflow-gpu, the NVIDIA driver must be installed separately.

How to train a model

sacred package is used to deal with experiments. If you are not yet familiar with it, have a quick look at the documentation.

Input data

In order to train a model, you should input a csv file with each row containing the filename of the image (full path) and its label (plain text) separated by a delimiting character (let’s say ;). Also, each character should be separated by a splitting character (let’s say |), this in order to deal with arbitrary alphabets (especially characters that cannot be encoded with utf-8 format).

An example of such csv file would look like :

/full/path/to/image1.{jpg,png};|s|t|r|i|n|g|_|l|a|b|e|l|1|
/full/path/to/image2.{jpg,png};|s|t|r|i|n|g|_|l|a|b|e|l|2| |w|i|t|h| |special_char|
...

Input lookup alphabet file

You also need to provide a lookup table for the alphabet that will be used. The term alphabet refers to all the symbols you want the network to learn, whether they are characters, digits, symbols, abbreviations, or any other graphical element.

The lookup table is a dictionary mapping alphabet units to integer codes (i.e {‘char’ : <int_code>}). Some lookup tables are already provided as examples in data/alphabet/.

For example to transcribe words that contain only the characters ‘abcdefg’, one possible lookup table would be :

{'a': 1, 'b': 2, 'c': 3, 'd': 4. 'e': 5, 'f': 6, 'g': 7}

The lookup table / dictionary needs to be saved in a json file.

Config file (with sacred)

Set the parameters of the experiment in config.json. The file looks like this :

{
  "lookup_alphabet_file" : "./data/alphabet/lookup.json",
  "csv_files_train" : "./data/csv_experiments/train_data.csv",
  "csv_files_eval" : "./data/csv_experiments/validation_data.csv",
  "output_model_dir" : "./output_model",
  "num_beam_paths" : 1,
  "max_chars_per_string" : 80,
  "n_epochs" : 50,
  "train_batch_size" : 64,
  "eval_batch_size" : 64,
  "learning_rate": 1e-4,
  "input_shape" : [128, 1400],
  "restore_model" : false
}

In order to use your data, you should change the parameters csv_files_train, csv_files_eval and lookup_alphabet_file.

All the configurable parameters can be found in class tf_crnn.config.Params, which can be added to the config file if needed.

Training

Once you have your input csv and alphabet file completed, and the parameters set in config.json, we will use sacred syntax to launch the training :

python training.py with config.json

The saved model and logs will then be exported to the folder specified in the config file (output_model_dir).

Example of training

We will use the IAM Database [marti2002iam] as an example to generate the data in the correct input data and train a model.

Go to the official page to download the dataset and create an account in order to access the data. You don’t need to download the data yourself, the prepare_iam.py script will take care of that for you.

Generating data

First create the IAM_USER and IAM_PWD environment variable to store your credentials, they will be used by the download script

export IAM_USER=<your-username>
export IAM_PWD=<your-password>

Run the script hlp/prepare_iam.py in order to download the data, extract it and format it correctly to train a model.

cd hlp
python prepare_iam.py --download_dir ../data/iam --generated_data_dir ../data/iam/generated
cd ..

The images of the lines are extracted in data/iam/lines/ and the folder data/generated/ contains all the additional files necessary to run the experiment. The csv files are saved in data/generated/generated_csv and the alphabet is placed in data/generated/generated_alphabet.

Training the model

Make sure the config.json file has the correct paths for training and validation data, as well as for the alphabet lookup file and run:

python training.py with config.json

Reference guide

Data handling for input function

dataset_generator(csv_filename, params[, …]) Generates the dataset for the experiment.
padding_inputs_width(image, target_shape, …) Given an input image, will pad it to return a target_shape size padded image.
augment_data(image[, max_rotation, …]) Data augmentation on an image (padding, brightness, contrast, rotation)
random_rotation(img[, max_rotation, crop, …]) Rotates an image with a random angle.

Model definitions

ConvBlock(features, kernel_size, stride, …) Convolutional block class.
get_model_train(parameters) Constructs the full model for training.
get_model_inference(parameters[, weights_path]) Constructs the full model for prediction.
get_crnn_output(input_images[, parameters]) Creates the CRNN network and returns it’s output.

Config for training

Alphabet([lookup_alphabet_file, blank_symbol]) Class for alphabet / symbols units.
Params(**kwargs) Class for parameters of the model and the experiment
import_params_from_json([model_directory, …]) Read the exported json file with parameters of the experiment.

Custom Callbacks

CustomSavingCallback(output_dir, saving_freq) Callback to save weights, architecture, and optimizer at the end of training.
LRTensorBoard(log_dir, **kwargs) Adds learning rate to TensorBoard scalars.
CustomLoaderCallback(loading_dir) Callback to load necessary weight and parameters for training, evaluation and prediction.
CustomPredictionSaverCallback(output_dir, …) Callback to save prediction decoded outputs.

Preprocessing data

data_preprocessing(params) Preporcesses the data for the experiment (training and evaluation data).
preprocess_csv(csv_filename, parameters, …) Converts the original csv data to the format required by the experiment.

tf_crnn.data_handler.augment_data(image, max_rotation=0.1, minimum_width=0)[source]

Data augmentation on an image (padding, brightness, contrast, rotation)

Parameters:
  • image (MockObject) – Tensor
  • max_rotation (float) – float, maximum permitted rotation (in radians)
  • minimum_width (int) – minimum width of image after data augmentation
Return type:

MockObject

Returns:

Tensor

tf_crnn.data_handler.dataset_generator(csv_filename, params, use_labels=True, batch_size=64, data_augmentation=False, num_epochs=None, shuffle=True)[source]

Generates the dataset for the experiment.

Parameters:
  • csv_filename (Union[List[str], str]) – Path to csv file containing the data
  • params (Params) – parameters df the experiment (Params)
  • use_labels (bool) – boolean to indicate dataset generation during training / evaluation (true) or prediction (false)
  • batch_size (int) – size of the generated batches
  • data_augmentation (bool) – whether to use data augmentation strategies or not
  • num_epochs (Optional[int]) – number of epochs to repeat the dataset generation
  • shuffle (bool) – whether to suffle the data
Returns:

tf.data.Dataset

tf_crnn.data_handler.get_resized_width(image, target_height, increment)[source]

Resizes the image according to target_height.

Parameters:
  • image (MockObject) – image to resize
  • target_height (int) – height of the resized image
  • increment (int) – reduction factor due to pooling between input width and output width, this makes sure that the final width will be a multiple of increment
Returns:

resized image

tf_crnn.data_handler.padding_inputs_width(image, target_shape, increment)[source]

Given an input image, will pad it to return a target_shape size padded image. There are 3 cases:

  • image width > target width : simple resizing to shrink the image
  • image width >= 0.5*target width : pad the image
  • image width < 0.5*target width : replicates the image segment and appends it
Parameters:
  • image (MockObject) – Tensor of shape [H,W,C]
  • target_shape (Tuple[int, int]) – final shape after padding [H, W]
  • increment (int) – reduction factor due to pooling between input width and output width, this makes sure that the final width will be a multiple of increment
Return type:

Tuple[MockObject, MockObject]

Returns:

(image padded, output width)

tf_crnn.data_handler.random_rotation(img, max_rotation=0.1, crop=True, minimum_width=0)[source]

Rotates an image with a random angle. See https://stackoverflow.com/questions/16702966/rotate-image-and-crop-out-black-borders for formulae

Parameters:
  • img (MockObject) – Tensor
  • max_rotation (float) – maximum angle to rotate (radians)
  • crop (bool) – boolean to crop or not the image after rotation
  • minimum_width (int) – minimum width of image after data augmentation
Return type:

MockObject

Returns:

class tf_crnn.config.Alphabet(lookup_alphabet_file=None, blank_symbol='$')[source]

Class for alphabet / symbols units.

Variables:
  • _blank_symbol (str) – Blank symbol used for CTC
  • _alphabet_units (List[str]) – list of elements composing the alphabet. The units may be a single character or multiple characters.
  • _codes (List[int]) – Each alphabet unit has a unique corresponding code.
  • _nclasses (int) – number of alphabet units.
alphabet_units
blank_symbol
check_input_file_alphabet(csv_filenames, discarded_chars=';|\t\n\r\x0b\x0c', csv_delimiter=';')[source]

Checks if labels of input files contains only characters that are in the Alphabet.

Parameters:
  • csv_filenames (List[str]) – list of the csv filename
  • discarded_chars (str) – discarded characters
  • csv_delimiter (str) – character delimiting field in the csv file
Return type:

None

Returns:

codes
classmethod create_lookup_from_labels(csv_files, export_lookup_filename, original_lookup_filename=None)[source]

Create a lookup dictionary for csv files containing labels. Exports a json file with the Alphabet.

Parameters:
  • csv_files (List[str]) – list of files to get the labels from (should be of format path;label)
  • export_lookup_filename (str) – filename to export alphabet lookup dictionary
  • original_lookup_filename (Optional[str]) – original lookup filename to update (optional)
Returns:

classmethod load_lookup_from_json(json_filenames)[source]

Load a lookup table from a json file to a dictionnary :type json_filenames: Union[List[str], str] :param json_filenames: either a filename or a list of filenames :rtype: dict :return:

classmethod make_json_lookup_alphabet(string_chars=None)[source]
Parameters:string_chars (Optional[str]) – for example string.ascii_letters, string.digits
Return type:dict
Returns:
classmethod map_lookup(lookup_table, unique_entry=True)[source]

Converts an existing lookup table with minimal range code ([1, len(lookup_table)-1]) and avoids multiple instances of the same code label (bijectivity)

Parameters:
  • lookup_table (dict) – dictionary to be mapped {alphabet_unit : code label}
  • unique_entry (bool) – If each alphabet unit has a unique code and each code a unique alphabet unique (‘bijective’), only True is implemented for now
Return type:

dict

Returns:

a mapped dictionary

n_classes
class tf_crnn.config.Params(**kwargs)[source]

Class for parameters of the model and the experiment

Variables:
  • input_shape (Tuple[int, int]) – input shape of the image to batch (this is the shape after data augmentation). The original will either be resized or pad depending on its original size
  • input_channels (int) – number of color channels for input image (default: 1)
  • cnn_features_list (List(int)) – a list of length n_layers containing the number of features for each convolutionl layer (default: [16, 32, 64, 96, 128])
  • cnn_kernel_size (List(int)) – a list of length n_layers containing the size of the kernel for each convolutionl layer (default: [3, 3, 3, 3, 3])
  • cnn_stride_size (List((int, int))) – a list of length n_layers containing the stride size each convolutionl layer (default: [(1, 1), (1, 1), (1, 1), (1, 1), (1, 1)])
  • cnn_pool_size (List((int, int))) – a list of length n_layers containing the pool size each MaxPool layer default: ([(2, 2), (2, 2), (2, 2), (2, 2), (1, 1)])
  • cnn_batch_norm (List(bool)) – a list of length n_layers containing a bool that indicated wether or not to use batch normalization (default: [False, False, False, False, False])
  • rnn_units (List(int)) – a list containing the number of units per rnn layer (default: 256)
  • num_beam_paths (int) – number of paths (transcriptions) to return for ctc beam search (only used when predicting)
  • csv_delimiter (str) – character to delimit csv input files (default: ‘;’)
  • string_split_delimiter (str) – character that delimits each alphabet unit in the labels (default: ‘|’)
  • csv_files_train (str) – csv filename which contains the (path;label) of each training sample
  • csv_files_eval (str) – csv filename which contains the (path;label) of each eval sample
  • lookup_alphabet_file (str) – json file that contains the mapping alphabet units <-> codes
  • blank_symbol (str) – symbol for to be considered as blank by the CTC decoder (default: ‘$’)
  • max_chars_per_string (int) – maximum characters per sample (to avoid CTC decoder errors) (default: 75)
  • data_augmentation (bool) – if True augments data on the fly (default: true)
  • data_augmentation_max_rotation (float) – max permitted roation to apply to image during training in radians (default: 0.005)
  • data_augmentation_max_slant (float) – maximum angle for slant augmentation (default: 0.7)
  • n_epochs (int) – numbers of epochs to run the training (default: 50)
  • train_batch_size (int) – batch size during training (default: 64)
  • eval_batch_size (int) – batch size during evaluation (default: 128)
  • learning_rate (float) – initial learning rate (default: 1e-4)
  • evaluate_every_epoch (int) – evaluate every ‘evaluate_every_epoch’ epoch (default: 5)
  • save_interval (int) – save the model every ‘save_interval’ epoch (default: 20)
  • optimizer (str) – which optimizer to use (‘adam’, ‘rms’, ‘ada’) (default: ‘adam’)
  • output_model_dir (str) – output directory where the model will be saved and exported
  • restore_model (bool) – boolean to continue training with saved weights (default: False)
classmethod from_json_file(json_file)[source]

Given a json file, creates a Params object.

Parameters:json_file (str) – path to the json file
Returns:Params object
show_experiment_params()[source]

Returns a dictionary with the variables of the class.

Return type:dict
Returns:
to_dict()[source]

Returns the parameters as a dictionary

Return type:dict
Returns:
tf_crnn.config.import_params_from_json(model_directory=None, json_filename=None)[source]

Read the exported json file with parameters of the experiment.

Parameters:
  • model_directory (Optional[str]) – Direcoty where the odel was exported
  • json_filename (Optional[str]) – filename of the file
Return type:

dict

Returns:

a dictionary containing the parameters of the experiment

class tf_crnn.model.ConvBlock(features, kernel_size, stride, cnn_padding, pool_size, batchnorm, **kwargs)[source]

Convolutional block class. It is composed of a Conv2D layer, a BatchNormaization layer (optional), a MaxPool2D layer (optional) and a ReLu activation.

Variables:
  • features (int) – number of features of the convolutional layer
  • kernel_size (int) – size of the convolutional kernel
  • stride (int, int) – stride of the convolutional layer
  • cnn_padding – padding of the convolution (‘same’ or ‘valid’)
  • pool_size (int, int) – size of the maxpooling
  • batchnorm (bool) – use batch norm or not
call(inputs, training=False)[source]
get_config()[source]

Get a dictionary with all the necessary properties to recreate the same layer.

Return type:dict
Returns:dictionary containing the properties of the layer
tf_crnn.model.get_crnn_output(input_images, parameters=None)[source]

Creates the CRNN network and returns it’s output. Passes the input_images through the network and returns its output

Parameters:
  • input_images – images to process (B, H, W, C)
  • parameters (Optional[Params]) – parameters of the model (Params)
Return type:

MockObject

Returns:

the output of the CRNN model

tf_crnn.model.get_model_inference(parameters, weights_path=None)[source]

Constructs the full model for prediction. Defines inputs and outputs, and loads the weights.

Parameters:
  • parameters (Params) – parameters of the model (Params)
  • weights_path (Optional[str]) – path to the weights (.h5 file)
Returns:

the model (tf.Keras.Model)

tf_crnn.model.get_model_train(parameters)[source]

Constructs the full model for training. Defines inputs and outputs, loss function and metric (CER).

Parameters:parameters (Params) – parameters of the model (Params)
Returns:the model (tf.Keras.Model)
class tf_crnn.callbacks.CustomLoaderCallback(loading_dir)[source]

Callback to load necessary weight and parameters for training, evaluation and prediction.

Variables:loading_dir (str) – path to directory to save logs
set_model(model)[source]
class tf_crnn.callbacks.CustomPredictionSaverCallback(output_dir, parameters)[source]

Callback to save prediction decoded outputs. This will save the decoded outputs into a file.

Variables:
  • output_dir (str) – path to directory to save logs
  • parameters (Params) – parameters of the experiment (Params)
on_predict_batch_end(batch, logs)[source]
on_predict_begin(logs=None)[source]
class tf_crnn.callbacks.CustomSavingCallback(output_dir, saving_freq, save_best_only=False, keep_max_models=5)[source]

Callback to save weights, architecture, and optimizer at the end of training. Inspired by ModelCheckpoint.

Variables:
  • output_dir (str) – path to the folder where files will be saved
  • saving_freq (int) – save every n epochs
  • save_best_only (bool) – wether to save a model if it is best thant the last saving
  • keep_max_models (int) – number of models to keep, the older ones will be deleted
on_epoch_begin(epoch, logs=None)[source]
on_epoch_end(epoch, logs=None)[source]
on_train_end(logs=None)[source]
class tf_crnn.callbacks.LRTensorBoard(log_dir, **kwargs)[source]

Adds learning rate to TensorBoard scalars.

Variables:logdir (str) – path to directory to save logs
on_epoch_end(epoch, logs=None)[source]
tf_crnn.preprocessing.data_preprocessing(params)[source]

Preporcesses the data for the experiment (training and evaluation data). Exports the updated csv files into <output_model_dir>/preprocessed/updated_{eval,train}.csv

Parameters:params (Params) – parameters of the experiment (Params)
Return type:(<class ‘str’>, <class ‘str’>, <class ‘int’>, <class ‘int’>)
Returns:output path files, number of samples (for train and evaluation data)
tf_crnn.preprocessing.preprocess_csv(csv_filename, parameters, output_csv_filename)[source]

Converts the original csv data to the format required by the experiment. Removes the samples which labels have too many characters. Computes the widths of input images and removes the samples which have more characters per label than image width. Converts the string labels to dense codes. The output csv file contains the path to the image, the dense list of codes corresponding to the alphabets units (which are padded with 0 if len(label) < max_len) and the length of the label sequence.

Parameters:
  • csv_filename (str) – path to csv file
  • parameters (Params) – parameters of the experiment (Params)
  • output_csv_filename (str) – path to the output csv file
Return type:

int

Returns:

number of samples in the output csv file

References

[MB02]U-V Marti and Horst Bunke. The iam-database: an english sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39–46, 2002.

A TensorFlow implementation of the Convolutional Recurrent Neural Network (CRNN) for image-based sequence recognition tasks, such as scene text recognition and OCR.

This implementation uses tf.keras to build the model and tf.data modules to handle input data.

Indices and tables