mattmdjaga

/

segformer_b2_clothes

like 227

I trained the model multiple times, first main training then reducing batch size and learning rate to squeeze performance out. And the final stage metrics are in the "files and versions" tab in the "trainer_state.json" file. I'll try find time this week to put them in a nice table on the model card, but here are the raw metrics for the last epoch:
"eval_loss": 0.1483118236064911,
"eval_mean_accuracy": 0.8013773443925859,
"eval_mean_iou": 0.6942808214048504,
"eval_overall_accuracy": 0.9608961812565836,
"eval_per_category_accuracy": [
0.99286914254759,
0.7324762382810602,
0.914249064597777,
0.7332474226804123,
0.8738827255492974,
0.7648064158559259,
0.8971837406521974,
0.738701537049782,
0.34500775995861355,
0.743003663003663,
0.754480709831002,
0.9248298956640691,
0.8952934194198154,
0.896274395931225,
0.8613392395240559,
0.8215313463514903,
0.9057897449031546,
0.6298257372654156
],
"eval_per_category_iou": [
0.9855223246429033,
0.6765220512360357,
0.8160509554140127,
0.6297914482115811,
0.7767318840844862,
0.6542057486939079,
0.8407320751640647,
0.5527613058975714,
0.29650542415080916,
0.5793839404049175,
0.6019384711976142,
0.8538285772687122,
0.8154169345492815,
0.8131239148222803,
0.7433172737080993,
0.7330964117849363,
0.835010226344999,
0.2931158177110952
],

You can view this in a table format on the model card now