Hi, I am trying to evaluate the released EfficientNet-B3 checkpoint on PhraseCut. Box assertion error occurs due to overflow of float32 tensors in MultiheadAttention. Use float64 precision in MultiheadAttention can solve the assertion error. However, the box_acc, mean_box_iou and mean_mask_iou are all zero. Similar phenomenon occurs while finetuning the detector from the pretrained ENB3 checkpoint.
Hi, I am trying to evaluate the released EfficientNet-B3 checkpoint on PhraseCut. Box assertion error occurs due to overflow of float32 tensors in MultiheadAttention. Use float64 precision in MultiheadAttention can solve the assertion error. However, the box_acc, mean_box_iou and mean_mask_iou are all zero. Similar phenomenon occurs while finetuning the detector from the pretrained ENB3 checkpoint.