Open rak5216 opened 4 years ago
Here is my proof that (as you sniffed out) there is a subtle issue in the 1H code only for class0, and that for class metrics 1H = non1H and therefore i can optimize by non1H iou until we fix the code:
this is where it happens in tf.keras https://github.com/tensorflow/tensorflow/blob/e5bf8de410005de06a7ff5393fafdf832ef1d4ad/tensorflow/python/keras/utils/metrics_utils.py#L341-L343
but i believe issue is on our end with 1H implementation somehow in metrics_utils.py
trace back to tf1 vs. tf2 upgrade #48