Training on own data _InvalidArgumentError

[buriburisuri]

Training on own data _InvalidArgumentError

[yannick-couzinie]

Training on the TEDlium 2 Corpus I had the same problem. You can find a quick and dirty work-around using @robbiebarrat's post, by the following steps:

• Make sure you have at least tensorflow >= 1.2. If you don't do this, you will not have the flag to ignore the error messages, as you can see in the source code for tensorflow (just change the branches at the top if you want to check).
• Go to your virtual envs/pythons, navigate down lib then site-packages, find sugartensor open sg_loss.py and change the lines 225-226 from
  out = tf.nn.ctc_loss(opt.target.sg_to_sparse(), tensor, opt.seq_len, ctc_merge_repeated=opt.merge, time_major=False)
  to
  out = tf.nn.ctc_loss(opt.target.sg_to_sparse(), tensor, opt.seq_len, ctc_merge_repeated=opt.merge, ignore_longer_outputs_than_inputs=True, time_major=False)

Training should at least run through. I would have preferred to just add an argument to the function call, but something with sugar-tensor changing how functions work confused me, so I just did this simple thing.

Note what the documentation says about the flag, so you know what happens:

The ignore_longer_outputs_than_inputs option allows to specify the behavior of the CTCLoss when dealing with sequences that have longer outputs than inputs. If true, the CTCLoss will simply return zero gradient for those items, otherwise an InvalidArgument error is returned, stopping training.
So this really is just a workaround until somebody understands what is actually going wrong, or if something is going wrong.

Mind you though, I got another error after that one. So you might get one too. This of course doesn't fix anything, if it's not clear, but at least you can train ( if you don't run into the other error).

[lemmonation]

I got the same error. Have u fix this? @star633669 @buriburisuri

[robbiebarrat]

this may help - i have the same problem. My error shows

`InvalidArgumentError (see above for traceback): Not enough time for target transition sequence (required: 32, available: 31)1You can turn this error into a warning by using the flag ignore_longer_outputs_than_inputs
 [[Node: ctc/CTCLoss = CTCLoss[ctc_merge_repeated=true, ignore_longer_outputs_than_inputs=false, preprocess_collapse_repeated=false, _device="/job:localhost/replica:0/task:0/cpu:0"](ctc/Log, ctc/ToInt64, ctc/ToInt32_2, ctc/ToInt32_1)]]`

The "your can turn this error into a warning by using this flag..." seems like it might help someone - i'll update here if i figure it out.

[dizzyname]

i want to know how can i train my own data.
could you help me please? @star633669