Inconsist results between training and inference and between model.evaluate() and model.predict()

Question

I am training a model for binary classification with cross-entropy loss in tensorflow. My training and validation datasets are built as follows:

train_ds = t