Daniel Povey 0d94783e76 Instead of a pooling operation, use the first bottleneck_dim dimensions of the preceding self_attn.forward2 as the input to the squeeze-excite module.

2022-11-04 15:16:59 +08:00

..

ASR

Instead of a pooling operation, use the first bottleneck_dim dimensions of the preceding self_attn.forward2 as the input to the squeeze-excite module.

2022-11-04 15:16:59 +08:00