jtuyls opened a new pull request #6589:
URL: https://github.com/apache/incubator-tvm/pull/6589
Fix for Tensorflow 1.15.0 conv2d_transpose parsing. Specifically,
attr["_output_shapes"][0] can be None.
Following snippet can be used to create a model for reproducing the bug (run
with TF 1.15.0):
```
import tensorflow as tf
output_shape = [1, 8, 8, 128]
strides = [1, 2, 2, 1]
x = tf.compat.v1.placeholder(tf.float32, shape=[1, 32, 32, 4])
w = tf.constant(0.1, shape=[7, 7, 128, 4])
ct = tf.nn.conv2d_transpose(x, w, output_shape=output_shape,
strides=strides, padding='SAME')
graph_def = tf.compat.v1.get_default_graph().as_graph_def()
with tf.io.gfile.GFile('conv2d_transpose_tf1_15.pb', "wb") as f:
f.write(graph_def.SerializeToString())
```
In general, I think it might be useful to adjust the CI to run the frontend
tests with multiple versions of the the external frameworks to verify
compatibility with the versions we want to support?
@siju-samuel @masahi, could you have a look?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
