search

calico / basenji

Sequential regulatory activity predictions with deep convolutional neural networks.
Apache License 2.0
395 stars 121 forks source link

tutorial training with --sample 0.05 #182

Closed zyj1729 closed 11 months ago

zyj1729 commented 11 months ago

I followed the tutorial.ipynb and use --sample 0.05 to train a test model, the training stops in 38 epochs. The results seem far off. Is it because I only used 5% training data? Screenshot 2023-10-23 at 10 58 12 PM Screenshot 2023-10-23 at 11 00 03 PM

zyj1729 commented 11 months ago

I realize this is the sample size thing. Once I remove --sample 0.05, the results make sense now. I would still suggest to maybe explicitly mention at the end of tutorial.ipynb that we shouldn't expect meaningful results using --sample 0.05. Because I thought the model loaded to explore_model.ipynb was the one trained from tutorial.ipynb.

zyj1729 commented 11 months ago

close this

amingc3 commented 4 months ago

Hello, I've also encountered this issue. I would like to ask, after you removed '--sample 0.05', how many training iterations did you perform on the author-provided data to achieve a Pearson correlation coefficient of 0.6?