TF implementation of Inception Time - optimized for each task using Grid Search

[Bsingstad]

In this pull request, we propose a new implementation of Inception Time (Tensorflow) which parameters are optimized using grid search. This work will be further explained in a paper published in ReScience C journal : http://rescience.github.io/

[helme]

Hi @Bsingstad , first of all I have to excuse myself for the very (very) late reply on this PR! big sorry for that!

Overall, your PR looks great, however I need to request some changes. In particular:

• copying and renaming the template files (your_config.pyand your_model.py) to a new files with better names, i.e. tf_inception_time_config.pyand inception_time.py
• we want to keep code/reproduce_results.py as is in order to allow others to reproduce our results. Instead we suggest to create a new file called code/reproduce_inception_time_results.py
• in addition we observed, that your current version of code/reproduce_results.py would train all your models for each of the tasks, i.e. a model designed for form (conf_tf_inception_form) is also trained on other tasks. Is this intended, or do you prefer specific models for specific tasks?
• an lastly, we prefer to keep the case elif modeltype == "YOUR_MODEL_TYPE" as a template, instead we would suggest to just add an additional case for your models here

other than that, your PR looks great, I'm curious to do pull request once the issues are resolved. What do you think about our suggestions?

Best, Patrick

[Bsingstad]

Hi @helme

Thank you for your suggestions. I will address them as soon as I can.

Best regards Bjørn-Jostein

[Bsingstad]

Hi @helme

Sorry for taking so long time.

I have now updated the pull request according to your requirements:

1. Good point! I have renamed the files containing the code for our proposed model to inception_time.py and configs tf_inception_time_config.py. We also added the original template files your_model.py and your_configs.py to our PR.
2. I have renamed code/reproduce_results.py to code/reproduce_inception_time_results.py in my PR
3. Actually this was an unintended behavior. In our replication paper we only report the specific model for each specific task, like this:

Model	All	Diagnostic	Subdiagnostic	Superdiagnostic	Form	Rhythm
Inception Time (all)	0.926(08)
Inception Time (diagnostic)		0.929(09)
Inception Time (subdiagnostic)			0.927(08)
Inception Time (superdiagnostic)				0.922(06)
Inception Time (form)					0.840(11)
Inception Time (rhythm)						0.923(32)

but we didn't find a way to only train the specific models without doing major changes in your base code.

4. Of course! I have added elif modeltype == "YOUR_MODEL_TYPE" to scp_experiment.py and added an additional case for my model: elif modeltype == "inception_time_model":

[helme]

Hi @Bsingstad, I hope my change requests for this pull request are somehow understandable?

[Bsingstad]

I have now committed the changes based on your last review