Get different performance with trained model in model zoo

[Kyfafyd]

with running the following command:

python train_net.py --num-gpus 6 --config-file configs/Detic_LbaseCCimg_CLIP_R5021k_640b64_4x_ft4x_max-size.yaml --eval-only MODEL.WEIGHTS trainedmodels/Detic_LbaseCCimg_CLIP_R5021k_640b64_4x_ft4x_max-size.pth

get

[07/25 19:56:40 d2.evaluation.testing]: copypaste: Task: bbox
[07/25 19:56:40 d2.evaluation.testing]: copypaste: AP,AP50,AP75,APs,APm,APl,APr,APc,APf
[07/25 19:56:40 d2.evaluation.testing]: copypaste: 33.9075,46.6746,35.7582,25.1269,44.0678,50.2866,19.8430,34.0934,39.8814
[07/25 19:56:40 d2.evaluation.testing]: copypaste: Task: segm
[07/25 19:56:40 d2.evaluation.testing]: copypaste: AP,AP50,AP75,APs,APm,APl,APr,APc,APf
[07/25 19:56:40 d2.evaluation.testing]: copypaste: 30.4862,44.4716,32.4350,20.9736,41.7839,48.2371,18.6669,30.8249,35.3029

[xingyizhou]

Hi,

Can you show the full log? It looks like the expected results for this config are 31.0/ 18.7 mAP_mask/ mAP_mask_rare. I should have checked the results for all models in the model zoo, but I don't have a machine now to re-verify this.

Best,
Xingyi