[Layer][GPU][TF FE] test_segment_sum failed with get_shape was called… #28778
Conversation
github-actions
bot
added
category: Core
|
From TF FE side, looks good. |
|
build_jenkins |
|
build_jenkins |
| const size_t num_segments_idx = 2; | ||
| TensorsContainer const_data(&impl_param.get_stream()); | ||
| if (memory_deps.count(num_segments_idx) > 0) { | ||
| const_data.emplace(3, memory_deps.at(num_segments_idx)); |
There was a problem hiding this comment.
Do I understand correctly that it uses dependency_idx=2 as num_segments input? It seems that dependency_idx=2 refers to segment_ids, doesn't it?
There was a problem hiding this comment.
Yes, dependency_idx = 2 likely refers to segment IDs, and num_segments would be the number of segments.
There was a problem hiding this comment.
@pravin25, I believe that such replacement is not correct in all cases. If an "empty segment" is the last one (and thus not present in segment_ids), this approach may lead to incorrect segment number calculations
Here is example from the specs:
<layer ... type="EmbeddingSegmentsSum" ... >
<input>
<port id="0"> <!-- emb_table value is: [[-0.2, -0.6], [-0.1, -0.4], [-1.9, -1.8], [-1., 1.5], [ 0.8, -0.7]] -->
<dim>5</dim>
<dim>2</dim>
</port>
<port id="1"> <!-- indices value is: [0, 2, 3, 4] -->
<dim>4</dim>
</port>
<port id="2"/> <!-- segment_ids value is: [0, 0, 2, 2] - second segment is empty -->
<dim>4</dim>
</port>
<port id="3"/> <!-- num_segments value is: 3 -->
<port id="4"/> <!-- default_index value is: 0 -->
<port id="5"/> <!-- per_sample_weigths value is: [0.5, 0.5, 0.5, 0.5] -->
<dim>4</dim>
</port>
</input>
<output>
<port id="6"> <!-- output value is: [[-1.05, -1.2], [-0.2, -0.6], [-0.1, 0.4]] -->
<dim>3</dim>
<dim>2</dim>
</port>
</output>
</layer>
But what if segment_ids was configured like this?
<port id="2"/> <!-- segment_ids value is: [0, 0, 1, 1] - third segment is empty -->
<dim>4</dim>
</port>
There was a problem hiding this comment.
@sshlyapn , First I want to correct this example. It should be
<port id="2"/> <!-- segment_ids value is: [0, 0, 0, 1, 1] - fourth segment is empty -->
<dim>5</dim>
</port>Because data shape dimension and segment_id_shape have the same dimension.
If it is a mismatch, OpenVINO throws an exception.
Based on segment_id, the kernel calculates segment ops.
The second thing I take care of is if the blank segment_id comes.
Here is that core/shape_inference/includes/embedding_segments_sum_shape_inference.hpp:56-63
|
@pravin25 please take a look on failed tests |
|
build_jenkins |
|
build_jenkins |
… on a descriptor Fix GPU TensorFlow SegmentSum test failure - Added dynamic shape support in GPU inference pipeline for SegmentSum ops - Updated tensorflow_tests/test_tf_SegmentSum.py to remove GPU skip - Ensured correct inputs for EmbededdingSegmentSum ops - Ensured correct handling of dynamic shapes during GPU inference for SegmentSum - Pass custom_eps via kwargs for flexible test tolerances. - Removed calculate segments_num logic from embedding_segments_sum_shape_inference.hpp Following both tickets are interrelated. https://jira.devtools.intel.com/browse/CVS-105896 https://jira.devtools.intel.com/browse/CVS-152352 https://jira.devtools.intel.com/browse/CVS-156362
Details:
custom_epsthroughkwargsto allow dynamic epsilon values for comparison in tests.This improves flexibility in handling precision variations (FP32, FP16).
Tickets: