[WIP] Testing VGG19, VGG16, and VGG-like (from scratch versions)

[sdv4]

Will include testing various parameters for the architecture, and labelling more images.

[sagarvijaygupta]

You must change VGG19 model to

def create_vgg19_network(input_shape):
    base_model = VGG19(input_shape=input_shape)
    return Model(input=base_model.input, output=base_model.get_layer('fc2').output)

There was some unseen error in saving of models. I will create a PR for this shortly.

[codecov-io]

Codecov Report

Merging #207 into master will increase coverage by 1.3%. The diff coverage is n/a.

@@            Coverage Diff            @@
##           master     #207     +/-   ##
=========================================
+ Coverage    20.8%   22.11%   +1.3%     
=========================================
  Files          11       11             
  Lines        1216     1126     -90     
  Branches      160      148     -12     
=========================================
- Hits          253      249      -4     
+ Misses        961      875     -86     
  Partials        2        2

Impacted Files	Coverage Δ
autowebcompat/network.py	27.85% <0%> (-1.4%)	:arrow_down:
collect.py	0% <0%> (ø)	:arrow_up:
label.py	0% <0%> (ø)	:arrow_up:
train.py	0% <0%> (ø)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 0aded3e...94e6155. Read the comment docs.

[sdv4]

@sagarvijaygupta I made that change now, thanks!

[marco-c]

@sdv4 are you labeling with bounding boxes too? It is optional, but it would be better as it would give us more information about the incompatibilities.

[sdv4]

@marco-c Yes, I will do that. Just fyi: the first run resulted in validation accuracy of 91.19% after the first epoch, with no improvement by the fourth epoch when I killed the process (was just running locally on my machine).

[marco-c]

@sdv4 could you split the labeling into a separate PR, so we can merge it?

[marco-c]

And we should look at the confusion matrix, that percentage looks too good to be true!

[sdv4]

@marco-c Will do!

As for the accuracy percentages, were you expecting them to be much lower because the networks aren't using the ImageNet weights? How should I go about investigating this further?

[marco-c]

As for the accuracy percentages, were you expecting them to be much lower because the networks aren't using the ImageNet weights?

I was pessimistic, the problem is not so easy so I was expecting some tuning to be needed.

How should I go about investigating this further?

@sagarvijaygupta added support for outputting the confusion matrix, the first step would be to retrain with one of the networks and see what the confusion matrix looks like.

[sdv4]

@marco-c I did include the output confusion matrices:

VGG19: [[136 39] [ 26 215]]

VGG16: [[142 61] [ 22 191]]

Is this what you were referring to?

[marco-c]

Is this what you were referring to?

Yes, sorry, I had missed them, that issue was getting too long :P They don't look too imbalanced!