1] Baseline model**:**

FullModel( (CRNN_model): CRNN( (sepconv): Sequential( (0): Conv1d(40, 40, kernel_size=(5,), stride=(2,), groups=40) (1): Conv1d(40, 128, kernel_size=(1,), stride=(8,), groups=2) ) (gru): GRU(128, 128, num_layers=2, dropout=0.1, bidirectional=True) ) (attn_layer): AttnMech( (Wx_b): Linear(in_features=256, out_features=256, bias=True) (Vt): Linear(in_features=256, out_features=1, bias=False) ) (U): Linear(in_features=256, out_features=2, bias=False) )

Train/test results: [with 2direct gru] wandb link

Train results [with 1direct gru] wandb link

Baseline test results on the val subset

2] Post-training quantization experiments:

1] Dynamic quantization by PyTorch API link, (Note: dynamic quantization available only on CPU) available blocks: GRU, Linear

Test df: val_df * 5 - for more stable test

torch.jit.script was applied after dynamic quantization.

Test dynamic quantized models [CPU] bidirectional gru

Test dynamic quantized models [CPU] unidirectional gru

2] Static quantization by PyTorch API link, (Note: static ****quantization available only on CPU) available blocks: Conv, Linear

Test df: val_df * 5 - for more stable test

• [1] config: static quantization of attention energy computation
• [2] config: static quantization of Conv1D and attention energy computation
  • Problem: GRU layer is not supported by static quantization proof, it is possible that quantizing and dequantizing SepConv in front of the GRU layer caused problems in this configuration

Test static quantized models [CPU] unidirectional gru

Here base model was updated...

Train results: wandb_link

Path: saved/models/kws_sheila_crnn/1115_144515/model_acc_0.99_epoch_30.pth