num_train_examples

[gowthamvenkatsairam]

In hparam_utils.py why is num_train_examples=200000 * 128. I have a train.tsv containing 100 questions,so,what should assign for num_train_examples

[ghost]

100 questions is small you wouldn't want to do more than a few epochs. I would start with 1000 examples (10 epochs) and see how the test performance develops over time.

For such a small set you might want to reduce the steps between checkpoints. You can change these options in run_task_main.py:

      save_checkpoints_steps=1000,
      keep_checkpoint_max=5,
      keep_checkpoint_every_n_hours=4.0,

For example this would give you one checkpoint per epoch:

      save_checkpoints_steps=100,
      keep_checkpoint_max=5,
      keep_checkpoint_every_n_hours=4.0,

[gowthamvenkatsairam]

what about num_train_examples I think its too large

[gowthamvenkatsairam]

how can I log the loss for every epoch,instead of getting loss after final step

[ghost]

Yes, as I said I would set num_train_examples=1000.

To get the loss and other metrics after ~ an epochs you can set save_checkpoints_steps=100 Then you can run a separate eval job to compute the loss and metrics for every checkpoint.

[eisenjulian]

Will close this issue for the time being, let us know and feel free to reopen if there are any further questions.