Efficientnet-b7 unsatisfactory result

[keithyh]

I use efficientnet-b7 for balanced binary classification with 10000 positive and 10000 negtive samples.
At first,I just use the default setting of 1000 classes and trained 10 epochs,then I set num-classes=2 and trained 4 epochs.
However both checkpoints give unsatisfactory results,that is both predict all samples to be positive.

How could this unsatisfactory result occur?

On inference,I have modified lines 120 and 121 in inference.py,changed them from:

      out_file.write('{0},{1},{2},{3},{4},{5}\n'.format(
            filename, label[0], label[1], label[2], label[3], label[4]))

to

     out_file.write('{0},{1},{2},{3},{4},{5}\n'.format(
           filename, label))

,since there are only two classes.