Lab: Feature Extraction with Traffic Signs

I'd like you to now do the same thing but with the German Traffic Sign dataset. The ImageNet dataset with 1000 classes had no traffic sign images. Will the high-level features learned still be transferable to such a different dataset?

Staying with the VGG example:

python feature_extraction.py --training_file bottlenecks/vgg_traffic_100_bottleneck_features_train.p --validation_file bottlenecks/vgg_traffic_bottleneck_features_validation.p

The only difference in the filename is we're swapping "cifar10" with "traffic".

Depending on how you wrote your solution you may have to manually change the number of classes back to 43 as well.

How do the feature extraction results compare with your results from the Traffic Sign project?