Closed jhlee525 closed 8 years ago
Thanks for bringing this to my attention. The problem is that no appropriate defaults were set for a couple of parameters to batch_normalize and the only way to pass them in is to use defaults_scope.
The workaround for now is to set them in a defaults_scope:
with pt.defaults_scope(learned_moments_update_rate=0.0003, variance_epsilon=0.001): x = tf.placeholder(tf.float32, [None, 224, 224, 3]) net = pt.wrap(x) net = net.conv2d(7, 64, batch_normalize=True)
Defaults added and pt.BatchNormalizationArguments lets you customize the values passed through conv2d. Please reopon if this doesn't work for you
Whenever I inserted batch normalization in prettytensor's conv2d function, it fails assertion.
My simple test code is
Then the console log prints
Thank you