Reduce usage of tf.contrib.layers #1350
Conversation
lgeiger
force-pushed
the
less-tf-contrib-layers
branch
from
bb28e04 to
046329f
Compare
| net = tf.contrib.layers.conv2d_transpose( | ||
| net, num_outputs, kernel_size=[3, 3], stride=stride, padding="valid") | ||
| net = tf.layers.conv2d_transpose( | ||
| net, num_outputs, (3, 3), strides=stride, activation=tf.nn.relu) |
There was a problem hiding this comment.
This change makes the CycleGANUpsampleConv2dTranspose test fail with AttributeError: 'tuple' object has no attribute 'ndims'.
In the unit test cyclegan_upsample gets called with a numpy input, tensorflow propably tries to interpret it as a list, which makes the test fail. Using a tensor as input would fix the unit test.
# this fails
upsampled_output = common_layers.cyclegan_upsample(random_input,
output_filters, stride, "conv2d_transpose")
# this works
upsampled_output = common_layers.cyclegan_upsample(tf.convert_to_tensor(random_input),
output_filters, stride, "conv2d_transpose")@afrozenator Is this a bug that should be reported to the Tensorflow team, or is this expected behavior with tf.layers?
There was a problem hiding this comment.
Let's just fix the test?
There was a problem hiding this comment.
👍 fixed it
I just wanted to bring this to your attention since the failure doesn't happen in the other upsampling tests.
There was a problem hiding this comment.
Lets just fix the test and this will be good to go. Many many thanks for doing this! @lgeiger
|
Thanks a lot of this and other PRs @lgeiger - much appreciated! |
PiperOrigin-RevId: 228806666
This reverts commit 57a9720.
This reverts commit 57a9720.
* Replace tf.contrib.layers initializers with tf.initializers * Remove unused imports of tf.contrib.layers * tf.contrib.layers.conv2d -> tf.layers.conv2d * tf.contrib.layers.flatten -> tf.layers.flatten * tf.contrib.layers.fully_connected -> tf.layers.dense * Fix cyclegan_upsample unit test See tensorflow#1350 (comment)
PiperOrigin-RevId: 228806666
This PR replaces
tf.contrib.layerswith their stabletf.layersandtf.initializersequivalents.