Strange predictions

[allchemist]

Hello. I'm trying neuralfingerprint, and have faced strange behaviour:
When i apply the model to csv file, which contains only smiles:
CCC
FFF
i get the result:
CCC,-3.4293943508031028
FCF,-2.6789522776231816

but when i put only CCC, i get another result.
CCC,-3.0120117325667533

If there are same molecules in input file, it gives same results for them, like:
CCC,-3.0120117325667533
CCC,-3.0120117325667533

The predictions are reproductable (dont change after another run), but the exact values depends on the contents of test csv file. I use your example.

smiles = read_smiles(task_params['experiment_data_file'])
result = predict_func(smiles)

I'm wondering if it is a bug or a feature.

[duvenaud]

This is due to us using batch normalization:
https://github.com/HIPS/neural-fingerprint/blob/master/neuralfingerprint/build_convnet.py#L69
which we found to be important for training.

I realize this is a bit annoying - there is a more sophisticated form of batch normalization that should give approximately the same results when applied to a single example, but we never implemented it. For simplicity's sake.