neuraxio / neuraxle-tensorflow

TensorFlow steps, savers, and utilities for Neuraxle.

Neuraxio

neuraxle-tensorflow's Issues

Ability to use a TensorBoard class that accepts a hyperparameter repository for logging all experiment results.

More info here perhaps:

• https://tensorboard.dev/ (it would be cool to integrate with this managed board hosting website)
• https://www.tensorflow.org/tensorboard/get_started

Note: this might be a new project in itself. Like Neuraxle-TensorBoard.

The goal would be to make it compatible with other models (e.g.: also sklearn models and pytorch) so that's why coding it in a new project might be better.

Not urgent: to review in 2 weeks or more.

Create entries for Neuraxle-TensorFlow in the documentation website

I have those ideas in mind:

• an example subgallery folder for just the tf examples

• neuraxle-tensorflow is another page of neuraxle complete API documentation below the existing API Documentation page, like the pytorch full api documentation which is below too.

• for the rest of the documentation site, I'd be okay to mix everything in the pages that aren't examples nor api documentation. For instance, there could be tf and pytorch pages at the same level than the model rest api serving example.

Type annotations for methods in README and make those methods importable as a demo.

Make the Adam Optimiser use the hyperparameter dict for instance.

code issue/bug?

Hi Guillaume,

Thank you for sharing these codes! When i am trying to run your code, i always met following issue( i did not change anything, but jut re-run your code). Any ideas what should I do? i am using neuraxle==0.5.2. Thank you !

AttributeError Traceback (most recent call last)
~/.local/lib/python3.6/site-packages/neuraxle/base.py in apply(self, method, ra, *args, **kwargs)
718 try:
--> 719 results = _method(*kargs, **ra.kwargs)
720 if not isinstance(results, RecursiveDict):

~/.local/lib/python3.6/site-packages/neuraxle/base.py in _set_hyperparams(self, hyperparams)
1434 hyperparams = HyperparameterSamples(hyperparams).to_flat()
-> 1435 self.hyperparams = hyperparams if len(hyperparams) > 0 else self.hyperparams
1436 return self.hyperparams

AttributeError: 'Tensorflow2ModelStep' object has no attribute 'hyperparams'
~/.local/lib/python3.6/site-packages/neuraxle/base.py in set_hyperparams(self, hyperparams)
1427 :func:_HasChildrenMixin._set_train
1428 """
-> 1429 self.apply(method='_set_hyperparams', hyperparams=HyperparameterSamples(hyperparams).to_flat())
1430 return self
1431

~/.local/lib/python3.6/site-packages/neuraxle/base.py in apply(self, method, ra, *args, **kwargs)
723 return results
724 except Exception as err:
--> 725 print('{}: Failed to apply method {}.'.format(self.name, method))
726 print(traceback.format_stack())
727 raise err

AttributeError: 'Tensorflow2ModelStep' object has no attribute 'name'

Bug : Tensorflow step should reuse "context" variable to load its checkpoint.

Neuraxle-TensorFlow/neuraxle_tensorflow/tensorflow_v2.py

Line 203 in e25f2d1

step.checkpoint.restore(step.checkpoint_manager.latest_checkpoint)

Pass context in create_mode, create_optimizer, create_loss...

Some models could need to change their dimensions based on the ExecutionContext.
Neuraxio/Neuraxle#365

Perhaps rename this library to NeuraxFlow and have a better README

TODO: have a README like that:
https://github.com/skorch-dev/skorch

Their readme is really good.

Tensorflow Training Progress Bar

Is there a way to integrate this while still printing out the progress bar that usually happens in Tensorflow when verbose is set to 1? I've tried several combinations after turning off the many low level outputs, but there is no documentation regarding regarding the progress bar.

Invalid: in the TFv2 step, it seems odd that the _create_inputs calls tf.convert_to_tensor

I would think this is unnecessary as NumPy arrays should work as well.

tf.convert_to_tensor is also called on the expected outputs in the computing of the test loss and I think it may not be needed there too.

Refactor: Simplify the TensorFlow step class

After simplifying the saver.

Feature: make saver simple

Pseudocode with probably the wrong function names, but you get it:

class TF1Saver: 
    def save(self, step): 
        stripped_graphs = dict()
        stripped_sessions = dict()

        for k, v in step.__dict__:

            if isinstance(v, graph): 
                self.stripped_graphs[k] = v
                step = self._strip_graph(step, k)

            elif isinstance(v, session): 
                self.stripped_sessions[k] = v
                step = self._strip_session(step, k)

        self._save_stripped_things(stripped_graphs, stripped_sessions)
        return step