How to use pretrained models.

I try to load pretrained model but get zero accuracy in the begining. In train.py I am doing the following way:

` sym, arg_params, aux_params = mx.model.load_checkpoint(os.path.join(args.pretrained, “model”), args.pretrained_epoch) sym = get_symbol(args) model = mx.mod.Module( context=ctx, symbol=sym, ) model.bind(data_shapes=train_dataiter.provide_data, label_shapes=train_dataiter.provide_label) model.set_params(arg_params, aux_params)

model.fit(train_dataiter,
    num_epoch=999999,
    eval_data=val_dataiter,
    eval_metric=eval_metrics,
    kvstore=args.kvstore,
    optimizer=opt,
    initializer=initializer,
    allow_missing=True,
    batch_end_callback=_batch_callback,
    arg_params=arg_params,
    aux_params=aux_params)`

[JokaHD]

Hi @vkasojhaa, I think it’s ok. But the default output of the model with fc7 layer is classification results. I guess it is necessary to get internal layer after load checkpoint for extracting features. see Extract Features in https://github.com/dmlc/mxnet-notebooks/blob/master/python/how_to/predict.ipynb

[JokaHD]

Hi @vkasojhaa, Generally, we take the output of embedding layer as feature of an image. Then calculate cosine similarity between two features as their similarity. fc7_weight is only used while training and validating. It is used for classifying an input image to a certain class. If the person is not in your training set, it impossible to get correct result. That’s why we use cosine similarity.