Running training script without feature extractor model

[KamilaLis]

Hi!

I am trying to replicate your work. Repository does not contain feature extractor model nor model for transfer learning so I changed in settings.py:

(line 139) sel_model = Model.CLASSIFIER_MCDCNN
(line 143) sel_dataset = Dataset.BALABIT
(line 177) sel_occ_features = OCCFeatures.RAW_X_Y_DIR

but now I get dimension error ("expected input_layer to have 3 dimensions, but got array with shape (230, 256)").
How can I train PLAIN model?

Thanks!

[norbertFejer]

Hello, Sorry for the late response. For using transfer learning or a feature extractor you need to pretrain a model, save it and then define the saved model name in constants.py file (USED_MODEL_FOR_TRANSFER_LEARNING and USED_MODEL_FOR_OCC_FEATURE_EXTRACTION parameters). After that in the settings.py you need to specify the sel_method to TRANSFER_LEARNING. The feature extractor models actually are in the models folder, they differ from the models, that we use for training in only one thing: for extracting features we remove the last layer from the model. You get the dimension error, because the models accepts (n,128,2) dimension inputs, where n means the number of blocks (a block contains 128 mouse movements), 128 means the block size (defined in constants.py) and 2 means the x and y coordinates. I hope this explanation solves your problem. If you have further questions please inform me on this email. You should attach detailed error message if it will still does not work or we can hold a Teams meeting, where I can help for you. Have a nice day, Norbert.

[yuanyangwangTJ]

Hello, Sorry for the late response. For using transfer learning or a feature extractor you need to pretrain a model, save it and then define the saved model name in constants.py file (USED_MODEL_FOR_TRANSFER_LEARNING and USED_MODEL_FOR_OCC_FEATURE_EXTRACTION parameters). After that in the settings.py you need to specify the sel_method to TRANSFER_LEARNING. The feature extractor models actually are in the models folder, they differ from the models, that we use for training in only one thing: for extracting features we remove the last layer from the model. You get the dimension error, because the models accepts (n,128,2) dimension inputs, where n means the number of blocks (a block contains 128 mouse movements), 128 means the block size (defined in constants.py) and 2 means the x and y coordinates. I hope this explanation solves your problem. If you have further questions please inform me on this email. You should attach detailed error message if it will still does not work or we can hold a Teams meeting, where I can help for you. Have a nice day, Norbert.

Could you describe the details about training of feature extractor? How can I start from scratch to train an extractor?