Degraded LSTM Model Performance

The table lists the issue, the root cause, and the corresponding workaround.

Issue	Description
Observation	The .h5 model contains LSTM layers and shows significant performance degradation in both FLOAT32 and INT8x8 formats after conversion. In the Code Generation Report, for a pure classification model this typically appears as: Classification Metrics (All results are relative to the original model) Argmax Errors: 2418/10000 samples Argmax Similarity: 75.82% Key Indicators: - Argmax Similarity < 85% for classification models - Performance degradation in both FLOAT32 and INT8x8, an issue beyond quantization - Model contains LSTM or bidirectional LSTM layers - For regression models, look for degradation in relevant metrics other than Argmax Similarity
Root Cause	LSTM layers in Keras models have unroll=True set for optimal TFLite Micro deployment. Without this setting, the recurrent operations may not convert correctly to TFLite Micro's expected format, resulting in poor inference performance. We are investigating a proper fix for unroll=False, in the meantime we offer a workaround.
Solution	Use the provided python script to rebuild the model while allowing unrolling of the LSTM layers and transferring the weights. A README.md with instructions is provided with the script. Download the script here.