Axis conversion "off by one" in UFF conversion of tf.gather

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Software version: TensorRT 5.0.4 or TensorRT 5.1.2rc (same behavior on both)

Problem summary: Axis parameter conversion is “off by one” in UFF conversion of tf.gather from Tensorflow (uff.from_tensorflow_frozen_model)

  1. according to TensorRT docuementation of the IGatherLayer, “the parameter axis is zero-indexed and starts at the first non-batch dimension of data tensor”
  2. however for the Tensorflow function tf.gather, the axis parameter starts with the first dimension (including batch)
  3. thus the UFF converter should subtract 1 from the provided axis parameter when setting the axis parameter of the IGatherLayer. However it does not.
  4. I have confirmed this because running the UFF converter and executing in TensorRT the following does not generate the correct results:
    data = tf.zeros([10, 2, 5])
    g1 = tf.gather(data, 0, axis=1)
    however decrementing the axis and running under TensorRT then generates the correct result
    g1 = tf.gather(data, 0, axis=0)
    even though according to (1) above, TensorRT is not capable of gathering over the batch dimension (thus the converter should have reported an error because I used axis=0 above)
2

Hello,

we are triaging and will keep you updated.

3

I wondered if the triaging has produced any conclusions regarding this issue, and its possible resolution to allow use of the Gather operator from Tensorflow networks? Thank you in advance.

4

Could you please provide an update of the result of triaging this issue? Is there any workaround, or will it be fixed in a subsequent release? Thank you.

5

This issue (an apparent bug in the released version of TensorRT) was posted more than 5 weeks ago. It would be appreciated if you could provide an update on the triaging and resolution. Thank you.