search

vsomnath / chemnet_auxilliaries

Contains the code for preparing the ChEMBL25 dataset to be used for training protocol
4 stars 4 forks source link

Weights for pretrained ChemCeption model #2

Open vsilier opened 3 years ago

vsilier commented 3 years ago

I do not have the time or resources to train the ChemCeption model myself. Could you please provide me with the trained model, via the weights? Could you explain how to load these weights into the ChemCeption model?

I'm looking to evaluate the pretrained model with different metrics and on Molnet datasets.

vsomnath commented 3 years ago

Hi, Thank you for your question. You can find the trained model weights here.

The script finetune_chemception_molnet.py should be used for finetuning the model on the corresponding MolNet dataset. Just set the restore_exp argument to the directory containing the checkpoints.

DeepChem has undergone some degree of rewrite in recent times, and I haven't had the bandwidth to check if the scripts are running as they were 2 years ago.

You can try using this, but just leave an open note here with the errors if you still struggle with it, and I will get back to resolving it in the coming days.

vsilier commented 3 years ago

Thank you so much for your help.

I'm not sure if the ChemCeption model has been updated, but there seems to be an error with the script.

ValueError: Input 0 of layer conv2d_63 is incompatible with the layer: expected axis -1 of input shape to have value 4 but received input with shape [32, 80, 80, 1]

I could work with the old version of ChemCeption, but if you have time in the next few days to update the script I'll gladly use it!

vsomnath commented 3 years ago

Ah I think this error must be because of the type of input used. Maybe try changing the img_spec argument to engd instead of std as default. I hope that fixes things.