Experiments

[mkassm]

Model training (with only 100 samples per epoch)

Using K-Tesla GPU and 3564 classes

CNN-GRU Model

loss: 5.1932 - accuracy: 0.2562 - val_loss: 6.5540 - val_accuracy: 0.139

DenseNet

loss: 3.2234 - accuracy: 0.4100 - val_loss: 3.6075 - val_accuracy: 0.368

Custom DenseNet with GRU

loss: 4.8542 - accuracy: 0.2878 - val_loss: 4.5494 - val_accuracy: 0.336

Using CPU and 65 classes

CNN-GRU Model with image size (64*64)

loss: 0.3073 - accuracy: 0.9337 - val_loss: 1.2696 - val_accuracy: 0.7050

DenseNet

loss: 0.2475 - accuracy: 0.9420 - val_loss: 0.1999 - val_accuracy: 0.9538

Custom DenseNet with GRU

loss: 0.2517 - accuracy: 0.9428 - val_loss: 0.2979 - val_accuracy: 0.9275

Using K-Tesla GPU and 6181 classes (full dataset)

DenseNet

loss: 4.1546 - accuracy: 0.3313 - val_loss: 4.4678 - val_accuracy: 0.315
test loss: 3009.866455078125 - test acc: 0.0
precision recall f1-score
accuracy 0.00 32236
macro avg 0.00 0.00 0.00 32236
weighted avg 0.00 0.00 0.00 32236

DenseNet with GRU (using pre-trained ImageNet weights)

loss: 5.8829 - accuracy: 0.2244 - val_loss: 6.5092 - val_accuracy: 0.200
test loss: 8.13744831085205
test acc: 0.019999999552965164
precision recall f1-score
accuracy 0.01 32236
macro avg 0.00 0.00 0.00 32236
weighted avg 0.00 0.01 0.00 32236