Performance of LVIS det & ins seg

[yhcao6]

Dear author, thanks for your excellent work. I have been trying to reproduce the performance of EVA-01 LVIS det & ins seg recently. In the table of Readme, the performance is det: 62.2, seg: 55.0. However, even though I trained the model myself twice, I got the result as follows: det: 58.1, seg: 52.3. I can't reproduce the same results.

This is my training procedures:

1. I train the model using this config: cascade_mask_rcnn_vitdet_eva.py with 64 A100 GPUs. During training, the loss can become NaN sometimes, so I set the gradient clip to be true to stabilize the training. I don't modify anything else.
2. I test the model using this config: cascade_mask_rcnn_vitdet_eva_1536, I follow the official instructions using this testing command:

python tools/lazyconfig_train_net.py --num-gpus 8 \
    --eval-only \
    --config-file projects/ViTDet/configs/LVIS/cascade_mask_rcnn_vitdet_eva_1536.py \
    "train.init_checkpoint=/path/to/eva_lvis.pth" \ # https://huggingface.co/BAAI/EVA/blob/main/eva_lvis.pth
    "dataloader.evaluator.max_dets_per_image=1000" \
    "model.roi_heads.maskness_thresh=0.5" # use maskness to calibrate mask predictions

Did I miss something important that resulted in a lower performance? Could you help me to reproduce the same performance?