WhirlFirst
commented
2 weeks ago Hi, for DeepCDR dependencies, you could also contact the author of the original repository (https://github.com/kimmo1019/DeepCDR) for help. I didn't encounter your problem. But I did notice that their environment didn't work with the A100 GPU, so I switched to V100 for training. For the DeepCDR task, we used the max-pooling embeddings for best performance, so the dimension is 768. We used the model with the same parameter size for downstream tasks.
Hello, I have been trying to run the DeepCDR scripts to generate the predictions for IC50. However I have been running into issues with the dependencies. I installed Keras, TensorFlow, and Hickle as specified but it seems that Hickle is not compatible with python3, but when I used python 2.7 I could not install the associated versions of Keras/Tensorflow. I was also looking into the embedding file in DeepCDR directory and it seems the scFoundation embeddings here are of dimension 768. When I ran the script in the model directory to generate embeddings for some other bulk data, it gave me embeddings with dimension 3072. I was wondering why there is a difference, is it because the model version used for DeepCDR was 50M params vs the model directory maybe using the 100M params version?