Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Transition model creation to tf.estimator.Estimator #3

Open
yweweler opened this issue Nov 16, 2018 · 1 comment
Open

Transition model creation to tf.estimator.Estimator #3

yweweler opened this issue Nov 16, 2018 · 1 comment
Labels
enhancement New feature or request
Milestone
Transition away f...

Comments

@yweweler
Copy link
Owner

Instead of relying on the manual creation and implementation of training, evaluation and prediction we should rewrite the code to build the model using an Estimator.
This should substantially reduce the effort necessary to export, serve, optimize, train, evaluate and predict using the architecture. As a side effect the codebase required can also be minimized.
@yweweler yweweler added the enhancement New feature or request label Nov 16, 2018
@yweweler yweweler changed the title Transition model creating to tf.estimator.Estimator Transition model creation to tf.estimator.Estimator Nov 16, 2018
yweweler added a commit that referenced this issue Nov 17, 2018
@yweweler
Rework the model to be usable with an estimator
e44fbf2 
WIP (#3)

* The model_fn is not finished yet.
* The model_fn is still stateful, as it saves the input placeholders and all intermediate tensors and summaries in the Model object. This state should be removed to enable reuse of a created model. Otherwise multiple calls to the model_fn will corrupt the internal state.
* Both the inference and the evaluate code are broken in the current state.
* The model_fn does not return EstimatorSpecs right now. Before this is possible, I have to move the gradient clipping and the batch normalization code from `train.py` into the model_fn.
yweweler added a commit that referenced this issue Nov 17, 2018
@yweweler
Refactor model_fn until the model trains
eaf83cf 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Add hooks to the training code
9e9fb60 
WIP (#3)

* The NanTensorHook is still missing.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Add todos and remove unused code
1d57dae 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Rework evaluation to use the estimator API
93dc485 
WIP (#3)
* Add todos.
* Remove unused code.
* Writing of the summaries does not work currently.
* Evaluation is extremely slow for some reason.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Fix missing summaries during training
128a2a2 
WIP (#3)

* Training summaries are now written during training.
* Removed training_summary from the Model constructor.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Implement the NanTensorHook and remove train summary flag
70ea397 
WIP (#3)

* Remove the unnecessary write summary flag from the model constructor.
* NanTensorHook is now attached in the model fn via the EstimatorSpec object (only for training)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Remove unused code
ea2e15d 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Fix checkpoint loading and add missing summaries during evaluation
e98769b 
WIP (#3)

* The correct checkpoints are now loaded for evaluation.
* Evaluation summaries are now written.
* For some reason the mean loss is wrong however.
* The evaluation metrics still pollute tensorboard.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Fix invalid evaluation summary
0d60746 
WIP (#3)

* The wrong summary was used in the eval metrics.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Remove NanTensorHook from model_fn
df9e88b 
WIP (#3)

* The estimator adds a NanTensorHook itself.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Fix the estimator creating a duplicate 'loss_1' summary
ce8aa70 
WIP (#3)

* The estimator has an hardcoded 'loss' summary. In case the is no loss summary with exactly this name, or there is a scope with this name, the estimator will create a new duplicate summary or will alter the scope name.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Move evaluation summaries into a named scope
90fc394 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Remove unused evaluation code
478eefa 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Refactor inference to use the tf.estimator API
84709c0 
WIP (#3)

* Implement dataset loading for the inference/predict code.
* Rewrite the model for prediction using the estimator.
* Rewrite the inference code to use the estimator.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Rewrite the model and the inference code
3623569 
WIP (#3)

* Inference does now wotk with the estimator API.
* Pre-processing and multi-threading are currently a bit hacky.
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Estimator inference now actually synthezises the requested sentences
8e75cf5 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Rewrite the experimental model exporter to use the tf.estimator API
07ff0f0 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Rewritten model export and poor mans serving to the estimator API
6707c04 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Add todos to the exporting hyper-params
b58613a 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Update model.py docstrings
8e9b23e 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Cleanup of train.py
e89b65b 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Cleanup evaluate.py and add comments
f007586 
WIP (#3)
yweweler added a commit that referenced this issue Nov 30, 2018
@yweweler
Cleanup inference.py and add comments
dda63ae 
WIP(#3)
@yweweler
Copy link
Owner Author

yweweler commented Dec 8, 2018

Fixed by #6

@yweweler yweweler mentioned this issue Dec 8, 2018
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
enhancement New feature or request
Projects
None yet
Milestone
Transition away from deprecated funct...
Development

No branches or pull requests
1 participant
@yweweler