Bug - SER Predictions depend on number of inputs

[av1659]

There is a bug in your SER model where the predictions change based on how many sequences the input file has. For instance, I have a sequence X in its own fasta file. For this sequence X I get predicted Tm of 56.778 with the SER model.

If my input file now has sequence X plus another sequence underneath, I get a predicted Tm for sequence X as 70.714. If the input file has four sequences, for that same sequence X, I get a predicted Tm of 95.019!

The outputs of the model shouldn't depend on the number of inputs given. BTW this issue is not there with the EMB model.

[av1659]

I believe it's because of these lines of code:

    scaler = MinMaxScaler()
    df2 = pd.DataFrame(scaler.fit_transform(df), columns=df.columns)

During the test time, the scaling should be done by the min/max values from the training dataset, not recomputed from the test dataset...