Number of trainable parameters (when froozen)


So am using ResNet34 and have 7 output classes, the end looks like this:

Dropout [512] 0 False

Linear [7] 3,591 True

Total params: 21,816,135
Total trainable params: 548,487
Total non-trainable params: 21,267,648

I assume its a fully connection between the second last and the last layer, so 512*7 weights to train, which sums up tp 3584. But the learner.summary() displays 7 more values (3591). What are the last 7 trainable parameters? Is it the biases?