Question About Softmax Output Behavior in SpliceAI

[hezt]

Hi SpliceAI authors,

Thanks a lot for the great work! In a lot of cases I see softmax outputs like [0.994, 0.005, 0.00009], but the true label is the low-score class (e.g., acceptor, 0.005).

How do you recommend handling this during inference and evaluation? I’m not using the variant scoring pipeline, just raw model outputs.

My code to use SpliceAI is like the last part of this repo's README, the context is 10k:

# Run all 5 models and average
# SpliceAI models output predictions for center 5000 bp only
y = np.mean([models[m].predict(x, verbose=0) for m in range(5)], axis=0)
# y shape: [1, 5000, 3] where last dim is [index 0: neither, index 1: acceptor, index 2: donor]        
# Extract prediction for this sequence
pred = y[0]  # Shape: [5000, 3]

Thanks!