Max pool layer strides param is different to paper

WeidiXie / VGG-Speaker-Recognition

Utterance-level Aggregation For Speaker Recognition In The Wild

362 stars 98 forks source link

Closed liangyanfeng closed 5 years ago

liangyanfeng commented 5 years ago

y = MaxPooling2D((3, 1), strides=(2, 1), name='mpool2')(x5)

seungwonpark commented 5 years ago

There are other inconsistencies between paper and code worth noting:

Angular margin m: 0.35 (code) / 0.4 (paper)
Learning rate schedule: Warmup, then decay with restart (code) / Exponential decay (paper)

Still, I would like to thank the author for publishing this code with pretrained weights.

WeidiXie commented 5 years ago

Thanks. This was my very first work on this project, so I think there are a lot of room for improving this work.