Alexnet

[qbilius]

Changes and observations:

• I moved sess.close outside train and test. An opened session is passed into these functions, so it doesn't make sense to close it here. If should be closed by whoever opened it. Alternatively, we start the session inside these functions – but you didn't seem to be in favor of this.
• Switched back in train to ı̇̀f step == 0: ...`. Not sure why you had it removed. I think unpacking the run/save logic there was useful.
• More feedback during validation is necessary since it may take a while for validation to run. So I added a tqdm progress bar.
• Your tabs are still not converting properly to spaces
• Can you please use a linter? Not only does it help to keep the styling consistent but it also helps to reveal important issues like typos in variable names.

Not solved issues (let me know if these should be posted as issues):

• Total batches in HDF5Provider is incorrect for the purpose of determining num_steps. It returns the number of batches for the batch size that we use to enqueue. However, the number of batches should be computed using the dequeue batch size. In the current implementation, get_validation_target does not have queue passed to it, so fixing it there is not possible. Not sure about the best way to go about it though.
• cache_dir is not passed anywhere. Haven't fixed that as it will depend on the exact load/save division.
• Maybe need an option to not run validation on step 0 as it takes too long during development if you just want to run like a thousand steps to see if the loss is going down.

[yamins81]

The "step==0" logic was removed because the new logic did the whole thing more cleanly. I'm going to switch back to the thing I wrote. Other than that I'll merge all the changes.

Scratch that -- I like what you did better. I'll merge it.

[yamins81]

@qbilius do you recommend a linter?

[qbilius]

It will probably depend on what your editor can do, but I use flake8 in VS Code, seems to do a good job.

On Mon, Jan 2, 2017, 20:58 Dan Yamins notifications@github.com wrote:

@qbilius https://github.com/qbilius do you recommend a linter?

— You are receiving this because you were mentioned.

Reply to this email directly, view it on GitHub https://github.com/neuroailab/tfutils/pull/31#issuecomment-270041463, or mute the thread https://github.com/notifications/unsubscribe-auth/ABMaMSWsASJFeqrro6AdzjsJ2ecXGFAFks5rOaswgaJpZM4LZMvj .