Reproducing Panoptic Seg 100-5 results

[kmn5409]

I recently tried to train the Panoptic Segmentation setting for ADE20K for the 100-5 setting. The following command I used to train the model was

python train_continual.py --dist-url auto --num-gpus 1 --config-file configs/ade20k/panoptic-segmentation/100-5.yaml \
    CONT.TASK 1 SOLVER.BASE_LR 0.0 SOLVER.MAX_ITER 2524 CONT.COLLECT_QUERY_MODE True OUTPUT_DIR ./output/ps/100-5/step1

python train_continual.py --dist-url auto --num-gpus $ngpus --config-file configs/ade20k/panoptic-segmentation/100-5.yaml \
    CONT.TASK 1 SOLVER.BASE_LR 0.0001 SOLVER.MAX_ITER 160000 CONT.COLLECT_QUERY_MODE False OUTPUT_DIR ./output/ps/100-5/step1

However when testing for the first time step I get an mIoU of 25.743. But when using the base weights from huggingface (https://huggingface.co/LightningNO1/SimCIS/tree/main/ps_100base) when I test for step 1 I get 61.7.

Can you suggest any problems I might be doing wrong?

[Guo954]

I pre-trained the source model myself, but the test results during the TTA phase are poor. Did I do something wrong? Could you please advise?
loading data from: output/fundus_src_except_REFUGEValid/model_final.pth
{"REFUGE_Valid": {"Dice Coefficient": 87.6831651760392, "Enhanced Alignment Metric": 96.38961346554501, "Structural Similarity Metric": 91.06561005643424}, "REFUGE_mean": {"Dice Coefficient": 87.6831651760392, "Enhanced Alignment Metric": 96.38961346554501, "Structural Similarity Metric": 91.06561005643424}}
loading data from: output/fundus_src_except_REFUGEValid/model_final.pth
{"REFUGE_Valid": {"Dice Coefficient": 87.71442072998099, "Enhanced Alignment Metric": 96.395248565209, "Structural Similarity Metric": 91.08409259142373}, "REFUGE_mean": {"Dice Coefficient": 87.71442072998099, "Enhanced Alignment Metric": 96.395248565209, "Structural Similarity Metric": 91.08409259142373}}