IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[ MO ] Turning GNMT KSO OFF (#1718)

Browse Source
This commit is contained in:
Evgenya Stepyreva
2020-08-11 18:47:27 +03:00
committed by GitHub
parent abab645c42
commit 9c1f479a61
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
model-optimizer/extensions/front/tf/GNMT_DynamicSequenceLengths.py
View File

@@ -31,7 +31,7 @@ class GNMT_sequence_lengths(FrontReplacementPattern):
Seq_len_data -> Max -> Cast -> Mul -> Round -> Cast.
After matching this pattern:
1. This replacer looking for encoder sequence lengths node (using information about encoder condition stucture)
1. This replacer looking for encoder sequence lengths node (using information about encoder condition structure)
2. Create node for multiplying Encoder sequence lengths by 2 (as it works in GNMT).
3. Connect Encoder sequence lengths value multiplied by 2 with decoder TensorArrays as size.
"""
@@ -141,3 +141,11 @@ class GNMT_sequence_lengths(FrontReplacementPattern):
ta.in_port(0).disconnect()
ta.in_port(0).get_connection().set_source(mul_op.out_port(0))
if graph.graph['cmd_params'].keep_shape_ops:
log.error(
"Model can not be translated in a reshape-able way.\n"
"Model Optimizer key keep_shape_ops was turned off to prevent related errors.\n"
"There will be no success changing input shapes of the model with the help of "
"InferenceEngine reshape method", extra={'is_warning': True})
graph.graph['cmd_params'].keep_shape_ops = False