Support for ragged range operator

[buddhapuneeth]

Describe the bug
I am not able to convert TF model with ragged.range operator.

Urgency
We are experimenting few approaches for our usecase. One of the way is using ragged range operator. This approach is working fine in TF. We would like to see can it be support with minimal changes on tf2onnx

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Mac OS Catalina 10.15.7
• Tensorflow Version: 2.0
• Python version: 3.6

To Reproduce
sample code used:

import tensorflow as tf
import numpy as np
ip = np.array([3,5,2])
x = tf.placeholder(tf.float32, ip.shape, name="input")
d = tf.ragged.range(x, name="output",row_splits_dtype=tf.dtypes.int32)
sess = tf.Session()
feed_dict = {"input:0": ip}
output = sess.run(sess.graph.get_tensor_by_name("output:0"), feed_dict=feed_dict);
print(output)
sess.close()

rr_test2.txt

You can convert the type to .pb from .txt (I modified to upload here)

Command
python -m tf2onnx.convert --input rr_test2.pb --inputs input:0 --output rr_test2.onnx --outputs output:0 --opset 11 --fold_const

Error

Value "<dtype: 'int32'>" is not valid attribute data type.
2020-12-31 09:23:49,363 - ERROR - pass1 convert failed for name: "output"
op: "RaggedRange"
input: "output/Cast"
input: "input"
input: "output/Cast_1"
attr {
  key: "T"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "Tsplits"
  value {
    type: DT_INT32
  }
}
, ex=Not a cmessage

Expected behavior
Able to convert to ONNX

Screenshots
If applicable, add screenshots to help explain your problem.

Additional context
Add any other context about the problem here. If the issue is about a particular model, please share the model details as well to facilitate debugging.

[TomWildenhain-Microsoft]

Sure, I can look into adding this. Does the model have other ragged ops, or is this the only one? Also, is the ragged range you are using "mostly dense" (do most of the ranges have the same number of elements?) This might inform how I implement the conversion.

[TomWildenhain-Microsoft]

Also is opset 12 support sufficient, or do you need support for a lower opset?

[buddhapuneeth]

@TomWildenhain-Microsoft Thanks for quick response!

• we have only one ragged op in our model
• we can't assume it as dense, it will be variable. But variable with in a small range. For example between 2 to 4 like 2,4,3,2,4.
• Opset 12 is fine
• the output of ragged op (i.e. ragged tensor) is used as input to gather_nd operator (pass as indicies). In our case, I believe this might be fine during conversion. Please correct me if I am wrong.

[TomWildenhain-Microsoft]

Great, that all sounds good. If the length of the ranges is upper-bounded by 4, then an implementation that is inefficient in that dimension should be fine. I'll give it a shot.

[TomWildenhain-Microsoft]

I made a PR for this. You will need to install from GitHub to use it before we make a release.

pip uninstall tf2onnx
pip install git+https://github.com/onnx/tensorflow-onnx

Or before the branch merges into master:
pip install git+https://github.com/onnx/tensorflow-onnx/pull/1256

[buddhapuneeth]

@TomWildenhain-Microsoft Awesome! I verified with your changes and my sanity checks are passed. Thanks a lot!