search

MIV-XJTU / ARTrack

Apache License 2.0
228 stars 33 forks source link

BaselineQuesetion #60

Closed Zsong1106 closed 5 months ago

Zsong1106 commented 5 months ago

When I am using the second stage of training to run the baseline model, when gradient backpropagation calculation loss occurs, it will prompt for an in place replacement error. I looked at the incorrect information and its dimension is [802768]. I first located the dimension in words'embeddings. I haven't made any changes either. Do you know where some specific details went wrong. 1

AlexDotHam commented 5 months ago

I've found that bug before, but when I reproduced my code I didn't find that question. In my opinion, maybe you change the file in /lib/model/layers/head_seq.py 836 line, emb_weight = self.word_embeddings.weight.clone() There is a clone function to avoid that problem, you can try to use another clone operation.

Zsong1106 commented 5 months ago

Thank you very much for your help. I found an environment issue in the code, which is a mismatch between the cuda version and yours. After switching to the cuda version, it can run smoothly