No valided stratified example error

[croshong]

I have a question related to balancing and stratify_receptor.

When I train a model with default2018, It does not have any problem, But when I tried to train a model with dense, there are following error

File "train.py", line 935, in results = train_and_test_model(args, train_test_files[i], outname, cont) File "train.py", line 502, in train_and_test_model solver.step(test_interval) ValueError: No valid stratified examples.

The setting of balancing and stratify_receptor is same between dense.model and default2018.model

and in my training data there are positive and negative labeled samples.

Any possible clue to this error?

Thanks

[dkoes]

This presumably means you have a receptor that has only all positive or all negative examples.

[croshong]

If that's the case, I think training with default2018.model also should fail, but it works well, so I guess there can be other reasson

[croshong]

When I use the data included in https://github.com/gnina/models/tree/master/data/PDBBind2016/General_types/fixed_gen_uff_completeset_*, I got the same error in dense model. Is there something that I should modify ?

[dkoes]

I think the issue is I changed this from being a silent failure to an error, which is why things that use to work no longer work. The quickest fix is to filter the input data to remove the problematic entires.

[croshong]

In my data, every entry has at least one positive or negative example so what you mean by problematic entry is the entry with too biased to positive or negative?

[dkoes]

There are quite a few examples in those files where a receptor only has one class of examples:

awk '{print $4,$1}' fixed_gen_uff_completeset_train0.types  | sort -u | awk '{print $1}' | uniq -c | sort -r -n

[francoep]

Dense net caffe model architecture does not have a RMSD column, the def2018 architecture does. You need to use different types files for the Dense net, which I suspect is the source of the error.

[Kerro-junior]

If that's the case, I think training with default2018.model also should fail, but it works well, so I guess there can be other reasson

Hi, i want to ask what is your original weightfile for default2018.model, did you use 'crossdock_2018_0.def' in gninasrc/lib/models/weights ?

[dkoes]

There is no one weight file for a given model architectures, but that would be the default model weights used when requesting the crossdock_2018 single model from gnina.

[Dadiao-shuai]

Hi, I use some processed crystal samples to train the default2018.model, so there are only samples labeled '0'. So the error rose up: No valid stratified examples

How could I fix this issue?

[dkoes]

Turn of stratification/balancing.