Suggest parameters for training word classifier CTC with IAM dataset

[yasersakkaf]

Following are the default parameters used in the word_classifier_CTC notebook: char_size = 67 PAD = 0 # Padding

num_new_images = 2 # Number of new images per image fac_alpha = 2.0 # Factors for image preprocessing fac_sigma = 0.08

num_buckets = 10 slider_size = (60, 2) step_size = 2 N_INPUT = slider_size[0]*slider_size[1] vocab_size = char_size + 2 # Number of different chars + <PAD> and <EOS>

layers = 2 residual_layers = 1 # HAVE TO be smaller than layers units = 256 num_hidden = 2*units

learning_rate = 1e-4 # 1e-4 max_gradient_norm = 5.0 # For gradient clipping dropout = 0.4 train_per = 0.8 # Percentage of training data

TRAIN_STEPS = 100000 # Number of training steps! TEST_ITER = 150 LOSS_ITER = 50 SAVE_ITER = 2000 BATCH_SIZE = 20 EPOCH = 2000 # Number of batches in epoch - not accurate

We used the same parameters for training the word classifier CTC mdel with IAM dataset and it took 7 hours for 1 Epoch on GPU. Please suggest what should be the parameters for training the word classifier CTC mdel with IAM dataset so that the training gets completed in less time.

[Breta01]

Hi, I am sorry the CTC notebook right now is quite slow. I already have quicker version (it's not fully tested), but I will update it here during today and let you know. What do you prefer python script or jupyter notebook (currently I prefer python scripts)?

[Breta01]

I just updated the code. Please try it out. If you encounter any errors let me know.

[yasersakkaf]

Thanks. I will try it out.