Mayuchen

why the lr_mult of "conv4_3_CPM" is so low?

1

Dear Zhe Cao, I've try to train this model. But in "pose_train_test.prototxt", this layer you have not fine_tuning, but the lr_mult of "conv4_3_CPM" is only 1, why the lr_mult of...

the loss is nan

6

It is a very nice work. But there are some problem in my experiments. Training is easy to gradient explosion, the loss is nan, even if my learning rate is...

An issue is found in recurrence.

1

An issue is found in recurrence. Location tokens, {,... , , ... , }. It is used when tokenizer decodes, where the LLM comes out with some offset coordinates relative...