MLatom 3.25.5

• Fixed: fine-tuning OMNI-P2x trained every layer of its networks — although the documentation says, and the code asks, that the first and third layers of each network stay fixed (fixed_layers=[[0, 4]]). The model class behind OMNI-P2x (vecmsani) and MS-ANI (msani) never applied fixed_layers in train(). Both now do, so fine-tuning OMNI-P2x, or training an MS-ANI model with fixed_layers, gives different models than before. A model saved after such training keeps those layers frozen when it is loaded again, as fine-tuned ANI models already did. By Janko Čivić (@jankocivic) in #54.
• Fixed: fine-tuning OMNI-P2x with train_osc=True trained the oscillator-strength model from scratch rather than from the pretrained one — it saved the pretrained model under one file name and loaded another — so the fine-tuned model knew only the elements in your data. It now fine-tunes the pretrained model.
• Fixed: fix_layers() and fixed_layers froze nothing in an ANI model created directly with ml.models.ani(...), such as the one in the transfer-learning tutorial, where ani.fix_layers([[0, 6]]) left every layer trainable; they worked only in ANI models that came from a pretrained method such as ANI-1ccx or AIQM1. They now freeze the requested layers in every network the model has. That includes a network just added for a new element, which starts from random weights; models from a pretrained method leave such networks trainable.
• Fixed: a flat list of layers, such as fix_layers([0, 4]), which the docstrings describe, failed with a TypeError in MS-ANI models, the one behind OMNI-P2x included, and in ANI models from a pretrained method. A flat list now fixes those layers in every element's network, in fix_layers() and in fixed_layers alike, as it already did in OMNI-P1.
• Fixed: a list set as fixed_layers in a copy of a model's own hyperparameters (hp = model.hyperparameters.copy(); hp['fixed_layers'] = [[0, 4]]) was turned into True, so train(hyperparameters=hp) froze nothing. It is now kept as given.
• Fixed: fine-tuning AIQM1, AIQM2, the ANI methods or OMNI-P2x wrote its default hyperparameters, fixed_layers=[[0, 4]] among them, into the dict passed as hyperparameters, so reusing that dict to train another model passed them on. It now leaves that dict as it was.