Set default models to training mode in the train_step.

[drewoldag]

It turns out that we have to explicitly set the model to training mode to actually train it. Oddly this isn't mentioned in the pytorch example code that was used when adding the ExampleCNN model.

For now, I've added model.train() to the create_trainer and model.eval() to the create_evaluator functions.

Confusion matrix before adding model.train()

Confusion matrix after adding model.train() 10 epochs

Confusion matrix after adding model.train() 50 epochs

Overall accuracy for this is about 64%, which is a little better than the pytorch example code that got to about 54%.

[github-actions]

Before [dad7062] <v0.1.1>	After [096ecfe]	Ratio	Benchmark (Parameter)
2.50±0.3s	2.36±0.6s	0.94	benchmarks.time_computation
2.8k	440	0.16	benchmarks.mem_list

Click here to view all benchmarks.

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.

Project coverage is 34.75%. Comparing base (dad7062) to head (411de9d).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/fibad/pytorch_ignite.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #103      +/-   ##
==========================================
- Coverage   34.83%   34.75%   -0.08%     
==========================================
  Files          18       18              
  Lines         887      889       +2     
==========================================
  Hits          309      309              
- Misses        578      580       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mtauraso]

It really seems like we should add a test for this.

Maybe we do a few dozen epochs in CI on CIFAR10 and ExampleAutoencoder, and threshold either loss or confusion matrix values?

That being said I have not objection to merging this as-is.

[drewoldag]

Yeah, I'm not sure what the best way to test this would be. It seems like there is an attribute model.training that is a boolean that we could use as a signal. I'm surprised by this, it feels like a gotcha that isn't obviously documented in the pytorch getting started guides (and I would have assumed it would be)

I'll create a follow up issue to add some kind of testing around this.

[drewoldag]

new issue here: #104