Dataset for "Fine-Tuning Pre-trained MAE for classification"

[royleung01]

HI All, I am new here to learn MAE ! Good to see you all.

I am trying to run the "Fine-tuning Pre-trained MAE for Classification" Evaluation with ImageNet fine-tuned models.

I follow the instruction to download the "fine-tuned Checkpoint". However, I cannot give the same/nearest result of the "Reference ImageNet accuracy" (83.664, 85.952, 86.928). I put ImageNet "Train and Val" dataset to the folder.

The result in "ViT-Base" as follow.

python main_finetune.py --eval --resume ../mae_finetuned_vit_base.pth --model vit_base_patch16 --batch_size 16 --data_path ..\ILSVRC\Data\DET --device directml

[00:15:02.980749] Test: [ 0/19299] eta: 5 days, 23:09:28 loss: 9.4609 (9.4609) acc1: 0.0000 (0.0000) acc5: 0.0000 (0.0000) time: 26.7044 data: 23.3047 [00:15:17.087085] Test: [ 10/19299] eta: 19:52:32 loss: 8.7052 (8.6560) acc1: 0.0000 (0.0000) acc5: 0.0000 (1.7045) time: 3.7095 data: 2.1197 [00:15:27.845380] Test: [ 20/19299] eta: 13:08:56 loss: 8.7110 (8.6961) acc1: 0.0000 (0.0000) acc5: 0.0000 (1.4881) time: 1.2429 data: 0.0009

[05:32:32.324447] Test: [19298/19299] eta: 0:00:00 loss: 8.8026 (9.3633) acc1: 0.0000 (0.0214) acc5: 0.0000 (0.2254) time: 0.8783 data: 0.0009 [05:32:35.807836] Test: Total time: 5:17:59 (0.9886 s / it) [05:32:35.809842] * Acc@1 0.021 Acc@5 0.225 loss 9.363 [05:32:35.814854] Accuracy of the network on the 308782 test images: 0.0% As my device no Nvidia based machine, i cannot use CUDA but add a torch_directML. I am no sure if it is affected. I would like to ask if i load the right dataset ImageNet ILSVRC2017 DET Dataset "Train and Val" for this Evaluation? also can i use directML for this program? Thanks and much appreciated ! Regards, Roy Leung