sushantmoon / memae-pytorch Goto Github PK

During training, the encoder and decoder are optimized to minimize the reconstruction error. The memory contents are simultaneously updated to record the prototypical elements of the encoded normal data. Given a testing sample, the model performs reconstruction merely using a restricted number of the normal patterns recorded in the memory. As a result, the reconstruction tends to be close to the normal sample, resulting in small reconstruction errors for normal samples and large
errors on anomalies, which will be used as a criterion to detect the anomalies.

Help-seeking: In the original paper, the memory updates normal data during training. I see the initialization of the weight parameter in MemoryUnit. Is this the M matrix in the paper? If so, how does the matrix ensure that it gets normal data? If yes, is the M matrix not derived from the data in the training process? I'm a little confused.

I got some errors at runtime!-RuntimeError: Found dtype Long but expected Float. I suspect it is line 78 in run.py, loss.backward(), can you tell me why? Or an upload error?Thank you very much！

Those variables are not defined in the run.py.
Is it dataloader?

Thanks for the good job！