Anyone attempted a heterogeneous sized colored image Dataset like FGVCAircraft

[ankitgooner]

HI all,

Please let me know if anyone tried creating a model for such datasets.I am facing some issues in reducing loss of my CNN which is way off.

The issue i am facing in this case is since the output_features(the classes of aircrafts are close to 100).
Model takes a lot of time to run(15 mins so far with 3 epochs) and the results are still not accurate.

I tried everything from tweaking the data to change other model parameters.

[LuluW8071]

For such large dataset you may need to limit dataset to 25 labels or introduce random crops image augmentation
And maybe increase conv2d layers and add dropout with ReLU layers for VGG model architecture

[ankitgooner]

i tried with reducing number of output features as well. Getting some bizzare results

[LuluW8071]

Looks like your model isn't learning anything since test loss and accuracy is same