Open lrizzello opened 5 years ago
Thank you for your post. We noticed you have not filled out the following field in the issue template. Could you update them if they are relevant in your case, or leave them as N/A? Thanks. What is the top-level directory of the model you are using Bazel version Exact command to reproduce
I updated my original post.
Hello,
I've been trying to use auto deeplab on my own dataset. The only part I touched is the stem creation. Instead of feeding a image set of dimension [?, y, x, 3] I feed data of dimension [?, y, x, 13]. The first 10 features of the last dimension correspond to character embeddings, and the last 3 are unrelated floats in the [0, 1] range. There are no NANs in this dataset.
To do this, I replace the beginning of the _nas_net function in nas_network.py. I get an array with shape [?, y, x, 4] as input, the first feature is a char ID to be replaced by a 10 dimensional embedding, the 3 other features stay as they are. This leads to a [?, y, x, 13] tensor which is then used by the rest of the code.
As my input seems sensible enough, I believe this may be a bug with the implementation. Could someone investigate this further, please?
System information
Describe the current behavior program crashes with error pasted at the bottom of this issue
Describe the expected behavior auto deeplab keeps going as expected
Code to reproduce the issue replace the content of /tensorflow/models/research/deeplab/core/nas_network.py by this code and run nas_network.py directly from the /tensorflow/research/ workspace. The only modification I brought to this file, besides the new main part at the bottom to run a simple fail case, can be found in the _nas_stem function.
Other info / logs