[Bug] [Relay][Keras] attribute go_backwards in LSTM converter function not be considered lead to wrong inference results

[jikechao]

The following test cases produced different inference results with Keras.
The TVM frontend didn't consider the go_backwards=True in function _convert_lstm (Line 985-1030) in file relay/frontend/keras.py

keras.layers.LSTM(units=32, go_backwards=True)
input_shape=[1, 2, 64]

Actual behavior

Triage

• frontend:keras
• needs-triage

cc @shingjan

[jikechao]

This bug has been solved by PR #15261.