corochann
Thank you for report. @Minys233 I think you are right. Seems we need to change this line into `splitter = split_method_dict[split](seed=seed)`. https://github.com/chainer/chainer-chemistry/blob/56e83dedb5de9dc9eb08ebf292be9ba76a4883ba/chainer_chemistry/datasets/molnet/molnet.py#L108 JFYI: sorry that we moved to maintenance phase...
I think it's okay if you want to start from the pre-trained weight of `predictor` while want to train `mlp` part from scratch!
I think this is okay to start trying transfer learning in this way.
Sorry for confusion for remaining the old legacy code, and thank you for report.
I think you can multiply weight to the loss. what is your task, regression or classification? what kind of loss function currently you are using?
## performance comparison with tox21 PR #241 : v0.4.0  PR #302 : after module refactoring, before input size invariant.  This PR #308 : input size invariant 
## performance comparison with qm9 It is difficult to see the difference after scaling... ``` In [1]: scaler.mean_ Out[1]: array([ 9.81423703e+00, 1.40609910e+00, 1.12491609e+00, 2.70606047e+00, 7.51916720e+01, -2.39976429e-01, 1.11237232e-02, 2.51099679e-01, 1.18952558e+03, 1.48523891e-01,...
Which model (network) you are using? Can you share your prediction code? Please refer typical usage for `predict` method. https://github.com/pfnet-research/chainer-chemistry/blob/master/examples/qm9/predict_qm9.py#L146