search

xiaomi1024 / code_SAMS

9 stars 5 forks source link

Could you please explain these shorts used in your train.py file? #3

Open AnuragIISC opened 1 year ago

AnuragIISC commented 1 year ago

I am not getting what reason you are using these arguments; I read full article, but there is no mention.

parser.add_argument('-num_steps', type=int, default=150)
parser.add_argument('-m', default='st')
parser.add_argument('-nn', type=int, default=1)

Why Number of steps separately?? Epochs are not enough? Then the training and testing of the model is not clear sir, why test file separately when you are doing model.eval() in same file. Help would be appreciated.

Coding511 commented 1 year ago

@AnuragIISC this nn is never used in training. I think you should leave it. m is something I am also unaware of.

xiaomi1024 commented 1 year ago

I am not getting what reason you are using these arguments; I read full article, but there is no mention.

parser.add_argument('-num_steps', type=int, default=150)
parser.add_argument('-m', default='st')
parser.add_argument('-nn', type=int, default=1)

Why Number of steps separately?? Epochs are not enough? Then the training and testing of the model is not clear sir, why test file separately when you are doing model.eval() in same file. Help would be appreciated.

-num_steps is the number of batch. This setting needs to consider the number of training samples and the size of the batch. The relationship between them is num_steps*batch_size/num_training_samples >=1 epoch -m is modal input marker. 's' is speech and 't' is text. Default is two modalities 'st'. In the uploaded model.py, this parameter is not used and can be ignored. It can be added when modifying the model for ablation experiments. -nn is a training version identification. This parameter can be ignored when reproducing the code. It can be considered for use when adjusting the model.