Open
Conversation
…ardless eval loss, because of the correctness of the eval loss calculation is questionable.
…hs without improvement.
nshmyrev
reviewed
Apr 16, 2017
| # for the model with size 512: | ||
| --max_steps 150000 | ||
| ``` | ||
|
|
Contributor
There was a problem hiding this comment.
Should work by default, not with options.
| num_iter_cover_train = int(sum(train_bucket_sizes) / | ||
| self.params.batch_size / | ||
| self.params.steps_per_checkpoint) | ||
| current_step, iter_inx, num_epochs_last_impr, max_num_epochs,\ |
Contributor
There was a problem hiding this comment.
"inx" is unclear abbreviation.
| self.params.batch_size / | ||
| self.params.steps_per_checkpoint) | ||
| current_step, iter_inx, num_epochs_last_impr, max_num_epochs,\ | ||
| num_up_trends, num_down_trends = 0, 0, 0, 2, 0, 0 |
Contributor
There was a problem hiding this comment.
"trend" is not a proper word here.
| current_step, iter_inx, num_epochs_last_impr, max_num_epochs,\ | ||
| num_up_trends, num_down_trends = 0, 0, 0, 2, 0, 0 | ||
| prev_train_losses, prev_valid_losses, prev_epoch_valid_losses = [], [], [] | ||
| num_iter_cover_train = max(1, int(sum(train_bucket_sizes) / |
Contributor
There was a problem hiding this comment.
This is called "epoch", no? "cover" is not a good word in this context.
| if len(prev_epoch_valid_losses) > 0: | ||
| print('Previous min epoch eval loss: %f, current epoch eval loss: %f' % | ||
| (min(prev_epoch_valid_losses), epoch_eval_loss)) | ||
| # Check if there was improvement during last epoch |
| # Check if there was improvement during last epoch | ||
| if (epoch_eval_loss < min(prev_epoch_valid_losses)): | ||
| if num_epochs_last_impr > max_num_epochs/1.5: | ||
| max_num_epochs = int(1.5 * num_epochs_last_impr) |
Contributor
There was a problem hiding this comment.
1.5 must be separated into something standalone and properly named, not used in multiple places in the code without name.
|
|
||
|
|
||
| def __calc_epoch_loss(self, epoch_losses): | ||
| """Calculate average loss during the epoch. |
Contributor
There was a problem hiding this comment.
Comment is wrong, this is not really an average.
nshmyrev
requested changes
Apr 18, 2017
| prev_train_losses.append(train_loss) | ||
| prev_valid_losses.append(eval_loss) | ||
| step_time, train_loss = 0.0, 0.0 | ||
| iter_idx += 1 |
Contributor
There was a problem hiding this comment.
Use current step instead of iter_idx
| num_epochs_last_impr = 0 | ||
| else: | ||
| print('No improvement during last epoch.') | ||
| num_epochs_last_impr += 1 |
Contributor
There was a problem hiding this comment.
epochs_without_improvement
| num_iter_total += num_iter_cover_valid | ||
| for batch_id in xrange(num_iter_cover_valid): | ||
| iter_total += iter_per_valid | ||
| for batch_id in xrange(iter_per_valid): |
Contributor
There was a problem hiding this comment.
Use xrange with batch_size step.
| self.params.batch_size)) | ||
| num_iter_total += num_iter_cover_valid | ||
| for batch_id in xrange(num_iter_cover_valid): | ||
| iter_total += iter_per_valid |
Contributor
There was a problem hiding this comment.
Count iter_total in inner loop
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add new stop criteria. Previously number of the allowable number of epochs without improvement was fixed. Training stopped if no improvement was seen during this window. New stop criteria include possibility of increasing of that window depending on the number of epochs needed for previous improvements during training.