Created by: wojtuss

The fc_gru_fuse_pass breaks accuracy when running CRNN-CTC inference using the C-API.

The application https://github.com/wojtuss/models/tree/wojtuss/capi-ocr-recognition/fluid/PaddleCV/ocr_recognition/capi runs inference using the CRNN-CTC trained model https://github.com/wojtuss/models/blob/wojtuss/capi-ocr-recognition/fluid/PaddleCV/ocr_recognition/models/CRNN-CTC_model.tar.gz (with weights from Baidu, with EditDistance evaluator added) on the standard dataset.

With the fc_gru_fuse_pass disabled in infer_ocr_recognition.cc:233 the accuracy we get from the app is equal to the accuracy we get using python eval.py script: avg distance = 0.09206844, avg seq error = 0.2425. With the fc_gru_fuse_pass enabled, the accuracy gets lost: avg seq error = 1.0000.

The script https://github.com/wojtuss/models/blob/wojtuss/capi-ocr-recognition/fluid/PaddleCV/ocr_recognition/capi/infer_ocr_with_accuracy.sh provides an example of how to run the application.