• Support GPU accelaration. Please pay attention that only when the module is complicated enough the GPU effectiveness can be seen
• Support LogCTC, which prevents from overflow issue
• Support batch mode, which means that different size of image are allowed in a single batch

• x is a list, each of which is H x W image. Here H should be identical, while W varies.
• x_mask is a matrix, each row of which means the valid region of corresponding image, thus is a mask. For example, [[1, 1, 1, 0, 0], [1, 1, 0, 0, 0]] means the width of first image is 3 and the second is 2
• y is a matrix, each row of which is the labels
• y_clip is a vector, each element of which means the length of corresponding label

Please refer to prepare_testing_data and prepare_traing_data functions in src/utee.py for more details

Once the data is prepared, simplily run python solver.py to start the training procedure

Deamo contains online version of feedforward stage, which means you can simply invoke the API to recognize a given images, with given model path.

My own experiments shows that 92.8% accuracy can be obtained if trained on andequate training samples, say, 20W. In my own case, images are simply handwrite english words.