Please run python hw3.py --train --bi. The rest hyperparameters are set to
default in my script.
The best performance is achieved from the following architecture:
-
A wide variant of ResNet: the first convolutional layer is with stride 1 and kernel size 3x3, the output dimension is 512; I used 3 sets of layers, expanding the channel dimension from 512 to 2048.
-
A 1D bidirectional rnn operating on the time dimension. The hidden size is 1024, with a dropout probability of 0.5.
-
A final dense layer.
I used CTCLoss and the Adam Optimizer.
I used ctcdecoder (beam search).
The learning rate is 2e-3, the weight decay is 5e-5; batch size is 32, I also used a scheduler to cut the lr in half every 5 epochs.
I tested different model architectures, optimizers, and various hyperparameters. The models I tested can be found in
models.py