issue with finetune_mpp.py

[FanzhouKong]

Hi:

I was trying to fine tune the pretrained model on my data. however, when i was trying to run the finetune_mpp.py and loads the pretrained model from weights/pretrain/mpp, I couldn't load it since the Scage class, has no attribute model, while the original load functions is as follows:

model_path = os.path.join(self.config['pretrain_model_path'], model_name)
state_dict = torch.load(model_path, map_location='cuda')
model.model.load_model_state_dict(state_dict['model'])

Of course, I also tried model.load_state_dict directly, as everyone else would usually do for loading weights for pretrained data, but i got the following error:

RuntimeError: Error(s) in loading state_dict for Scage:
Missing key(s) in state_dict: "dist_bar", "head_Graph.linear_seq.0.weight", "head_Graph.linear_seq.0.bias", "head_Graph.linear_seq.3.weight", "head_Graph.linear_seq.3.bias", "head_finger_keeping_bond.linear_seq.0.weight", "head_finger_keeping_bond.linear_seq.0.bias", "head_finger_keeping_bond.linear_seq.3.weight", "head_finger_keeping_bond.linear_seq.3.bias".
Unexpected key(s) in state_dict: "bond_feature.embed_list.0.weight", "bond_feature.embed_list.1.weight", "bond_feature.embed_list.2.weight", "bond_feature.embed_list.3.weight", "bond_feature.embed_list.4.weight", "bond_feature.edge_type_embed.weight", "bond_feature.graph_embedding.weight", "bond_feature.distance_embedding.0.mean", "bond_feature.distance_embedding.0.std", "bond_feature.distance_embedding.0.mul.weight", "bond_feature.distance_embedding.0.bias.weight", "bond_feature.distance_embedding.1.weight", "bond_feature.distance_embedding.1.bias", "head_pair_distances.linear_in.0.weight", "head_pair_distances.linear_in.0.bias", "head_pair_distances.linear_in.2.weight", "head_pair_distances.linear_in.2.bias", "head_pair_distances.linear_out1.weight", "head_pair_distances.linear_out1.bias", "head_pair_distances.linear_out2.weight", "head_pair_distances.linear_out2.bias", "head_angle.linear_seq.0.weight", "head_angle.linear_seq.0.bias", "head_angle.linear_seq.1.weight", "head_angle.linear_seq.1.bias", "head_angle.linear_seq.1.running_mean", "head_angle.linear_seq.1.running_var", "head_angle.linear_seq.1.num_batches_tracked", "head_angle.linear_seq.4.weight", "head_angle.linear_seq.4.bias".

Looks like the design in pretrain.pth is not exactly the same as scage model itself? or there is some thing I did wrong?

Thanks for your time!

Sincerely,

Fanzhou

[KazeDog]

Dear Fanzhou,

We apologize for the delay. Due to code compatibility issues, some of the model weights we uploaded are no longer used parameters. We have redefined the model loading function in the code's runtime file to ensure compatibility with these weights.

Best regards,
Jianbo