Base command
python ./eval_linear.py --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --evaluate
python ./eval_linear.py --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --evaluate --patch_size=8Model vit_small built.
We load the reference pretrained linear weights.
Downloading: "https://dl.fbaipublicfiles.com/dino/dino_deitsmall16_pretrain/dino_deitsmall16_linearweights.pth" to /home/vihang/.cache/torch/hub/checkpoints/dino_deitsmall16_linearweights.pth
100.0%
Test: [ 0/391] eta: 0:17:39 loss: 0.380281 (0.380281) acc1: 93.750000 (93.750000) acc5: 98.437500 (98.437500) time: 2.708638 data: 1.828419 max mem: 846
Test: [ 20/391] eta: 0:01:20 loss: 0.559775 (0.560542) acc1: 84.375000 (85.156250) acc5: 96.875000 (96.837798) time: 0.091588 data: 0.000287 max mem: 993
Test: [ 40/391] eta: 0:00:53 loss: 0.504800 (0.601536) acc1: 84.375000 (84.032012) acc5: 97.656250 (96.455793) time: 0.083965 data: 0.000281 max mem: 993
Test: [ 60/391] eta: 0:00:42 loss: 0.502291 (0.564686) acc1: 87.500000 (85.245902) acc5: 97.656250 (96.695697) time: 0.081432 data: 0.000253 max mem: 993
Test: [ 80/391] eta: 0:00:36 loss: 0.716538 (0.618657) acc1: 79.687500 (83.516590) acc5: 96.093750 (96.508488) time: 0.083536 data: 0.000242 max mem: 993
Test: [100/391] eta: 0:00:32 loss: 0.623209 (0.628215) acc1: 85.156250 (83.408106) acc5: 96.093750 (96.441832) time: 0.083887 data: 0.000238 max mem: 993
Test: [120/391] eta: 0:00:28 loss: 0.531973 (0.632460) acc1: 83.593750 (83.316116) acc5: 96.875000 (96.526343) time: 0.083836 data: 0.000224 max mem: 993
Test: [140/391] eta: 0:00:25 loss: 0.435032 (0.619030) acc1: 89.843750 (83.626995) acc5: 97.656250 (96.636746) time: 0.083697 data: 0.000240 max mem: 993
Test: [160/391] eta: 0:00:23 loss: 0.679810 (0.630977) acc1: 81.250000 (83.448175) acc5: 95.312500 (96.462539) time: 0.084011 data: 0.000206 max mem: 993
Test: [180/391] eta: 0:00:20 loss: 1.136989 (0.689275) acc1: 69.531250 (81.979454) acc5: 90.625000 (95.795925) time: 0.083333 data: 0.000245 max mem: 993
Test: [200/391] eta: 0:00:18 loss: 1.277641 (0.747734) acc1: 64.843750 (80.612562) acc5: 90.625000 (95.114272) time: 0.082513 data: 0.000226 max mem: 993
Test: [220/391] eta: 0:00:16 loss: 1.096730 (0.774484) acc1: 73.437500 (80.062217) acc5: 91.406250 (94.768100) time: 0.081201 data: 0.000229 max mem: 993
Test: [240/391] eta: 0:00:14 loss: 0.753314 (0.788769) acc1: 78.906250 (79.839860) acc5: 92.968750 (94.508558) time: 0.083774 data: 0.000209 max mem: 993
Test: [260/391] eta: 0:00:12 loss: 1.213251 (0.828748) acc1: 64.843750 (78.747605) acc5: 90.625000 (94.064296) time: 0.083779 data: 0.000236 max mem: 993
Test: [280/391] eta: 0:00:10 loss: 1.027571 (0.848783) acc1: 74.218750 (78.386343) acc5: 92.187500 (93.808385) time: 0.083863 data: 0.000228 max mem: 993
Test: [300/391] eta: 0:00:08 loss: 1.108118 (0.867847) acc1: 72.656250 (77.974460) acc5: 89.843750 (93.513808) time: 0.083811 data: 0.000254 max mem: 993
Test: [320/391] eta: 0:00:06 loss: 1.179348 (0.883526) acc1: 71.875000 (77.587130) acc5: 88.281250 (93.251071) time: 0.083823 data: 0.000246 max mem: 993
Test: [340/391] eta: 0:00:04 loss: 1.224718 (0.904226) acc1: 67.187500 (77.057368) acc5: 89.062500 (93.032900) time: 0.084049 data: 0.000241 max mem: 993
Test: [360/391] eta: 0:00:02 loss: 0.979421 (0.914856) acc1: 75.000000 (76.820031) acc5: 91.406250 (92.949273) time: 0.083515 data: 0.000244 max mem: 993
Test: [380/391] eta: 0:00:00 loss: 0.814474 (0.913746) acc1: 76.562500 (76.851624) acc5: 93.750000 (92.950295) time: 0.083795 data: 0.000182 max mem: 993
Test: [390/391] eta: 0:00:00 loss: 0.723986 (0.911126) acc1: 76.562500 (76.908000) acc5: 93.750000 (92.980000) time: 0.082328 data: 0.000111 max mem: 993
Test: Total time: 0:00:35 (0.090729 s / it)
* Acc@1 76.908 Acc@5 92.980 loss 0.911
Accuracy of the network on the 50000 test images: 76.9%
python eval_knn.py --arch=vit_small --patch_size=8 --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC --dump_features eval_knn_vits_8_featuresWill run the code on one GPU.
| distributed init (rank 0): env://
git:
sha: 66febe141b5fc6ba5cdcf33407b1049ec3c6c462, status: has uncommited changes, branch: gimlet
arch: vit_small
batch_size_per_gpu: 128
checkpoint_key: teacher
data_path: /datasets/imagenet/ILSVRC/Data/CLS-LOC
dist_url: env://
dump_features: eval_knn_vits_8_features
gpu: 0
half: False
load_features: None
local_rank: 0
nb_knn: [10, 20, 100, 200]
num_workers: 10
patch_size: 8
pretrained_weights:
rank: 0
temperature: 0.07
use_cuda: True
world_size: 1
Data loaded with 1281167 train and 50000 val imgs.
Model vit_small 8x8 built.
Please use the `--pretrained_weights` argument to indicate the path of the checkpoint to evaluate.
Since no pretrained weights have been provided, we load the reference pretrained DINO weights.
Extracting features for train set...
Storing features into tensor of shape torch.Size([1281167, 384])
[ 0/10010] eta: 8:41:41 time: 3.126985 data: 1.496408 max mem: 4515
<REMOVED>
[10009/10010] eta: 0:00:00 time: 0.636899 data: 0.000107 max mem: 6392
Total time: 1:28:46 (0.532138 s / it)
Extracting features for val set...
Storing features into tensor of shape torch.Size([50000, 384])
[ 0/391] eta: 0:10:42 time: 1.642119 data: 1.128212 max mem: 6392
<REMOVED>
[390/391] eta: 0:00:00 time: 0.636005 data: 0.000107 max mem: 6465
Total time: 0:03:28 (0.534096 s / it)
Features are ready!
Start the k-NN classification.
10-NN classifier result: Top1: 78.33, Top5: 90.896
20-NN classifier result: Top1: 78.33, Top5: 92.4
100-NN classifier result: Top1: 76.916, Top5: 93.662
200-NN classifier result: Top1: 76.096, Top5: 93.792
python ./eval_linear.py --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --evaluate --patch_size=16Model vit_small built.
We load the reference pretrained linear weights.
Test: [ 0/391] eta: 0:22:05 loss: 0.312423 (0.312423) acc1: 95.312500 (95.312500) acc5: 98.437500 (98.437500) time: 3.391267 data: 1.672790 max mem: 4970
Test: [ 20/391] eta: 0:03:44 loss: 0.452702 (0.475851) acc1: 86.718750 (86.904762) acc5: 97.656250 (97.693452) time: 0.464485 data: 0.000323 max mem: 5559
Test: [ 40/391] eta: 0:03:07 loss: 0.463861 (0.515941) acc1: 85.156250 (85.575457) acc5: 98.437500 (97.351372) time: 0.463942 data: 0.000553 max mem: 5559
Test: [ 60/391] eta: 0:02:49 loss: 0.441266 (0.489972) acc1: 87.500000 (86.680328) acc5: 96.875000 (97.297643) time: 0.467964 data: 0.000543 max mem: 5559
Test: [ 80/391] eta: 0:02:33 loss: 0.650984 (0.539812) acc1: 82.812500 (85.310571) acc5: 96.875000 (97.077546) time: 0.434674 data: 0.000379 max mem: 5559
Test: [100/391] eta: 0:02:22 loss: 0.532990 (0.551392) acc1: 86.718750 (85.202661) acc5: 96.875000 (96.975557) time: 0.471931 data: 0.000890 max mem: 5559
Test: [120/391] eta: 0:02:10 loss: 0.470687 (0.562025) acc1: 84.375000 (85.156250) acc5: 97.656250 (97.068698) time: 0.433611 data: 0.000233 max mem: 5559
Test: [140/391] eta: 0:01:58 loss: 0.380097 (0.549805) acc1: 88.281250 (85.328014) acc5: 98.437500 (97.212988) time: 0.431170 data: 0.000219 max mem: 5559
Test: [160/391] eta: 0:01:48 loss: 0.662740 (0.561041) acc1: 83.593750 (85.161102) acc5: 96.875000 (97.069099) time: 0.430035 data: 0.000231 max mem: 5559
Test: [180/391] eta: 0:01:37 loss: 0.863500 (0.605005) acc1: 75.781250 (84.016747) acc5: 93.750000 (96.620338) time: 0.430917 data: 0.000213 max mem: 5559
Test: [200/391] eta: 0:01:28 loss: 1.101570 (0.655729) acc1: 67.968750 (82.812500) acc5: 92.187500 (96.120958) time: 0.433986 data: 0.000225 max mem: 5559
Test: [220/391] eta: 0:01:18 loss: 0.844362 (0.674672) acc1: 78.906250 (82.356476) acc5: 94.531250 (95.927602) time: 0.430764 data: 0.000229 max mem: 5559
Test: [240/391] eta: 0:01:08 loss: 0.683789 (0.683675) acc1: 82.031250 (82.277619) acc5: 94.531250 (95.753371) time: 0.431250 data: 0.000218 max mem: 5559
Test: [260/391] eta: 0:00:59 loss: 0.949632 (0.718124) acc1: 73.437500 (81.324832) acc5: 92.187500 (95.402299) time: 0.433487 data: 0.000230 max mem: 5559
Test: [280/391] eta: 0:00:50 loss: 0.819773 (0.732555) acc1: 77.343750 (81.010899) acc5: 93.750000 (95.231873) time: 0.430006 data: 0.000199 max mem: 5559
Test: [300/391] eta: 0:00:41 loss: 0.940584 (0.748425) acc1: 75.000000 (80.658223) acc5: 92.968750 (95.021802) time: 0.430271 data: 0.000224 max mem: 5559
Test: [320/391] eta: 0:00:31 loss: 0.984055 (0.761864) acc1: 76.562500 (80.332457) acc5: 90.625000 (94.818438) time: 0.429488 data: 0.000225 max mem: 5559
Test: [340/391] eta: 0:00:22 loss: 1.017162 (0.777974) acc1: 72.656250 (79.928061) acc5: 92.968750 (94.680169) time: 0.430123 data: 0.000207 max mem: 5559
Test: [360/391] eta: 0:00:13 loss: 0.895487 (0.788876) acc1: 76.562500 (79.655038) acc5: 93.750000 (94.587517) time: 0.434021 data: 0.000202 max mem: 5559
Test: [380/391] eta: 0:00:04 loss: 0.642365 (0.789293) acc1: 77.343750 (79.615732) acc5: 95.312500 (94.582513) time: 0.429795 data: 0.000159 max mem: 5559
Test: [390/391] eta: 0:00:00 loss: 0.642365 (0.787828) acc1: 78.906250 (79.612000) acc5: 95.312500 (94.622000) time: 0.426404 data: 0.000113 max mem: 5559
Test: Total time: 0:02:54 (0.446400 s / it)
* Acc@1 79.612 Acc@5 94.622 loss 0.788
Accuracy of the network on the 50000 test images: 79.6%
python eval_knn.py --arch=vit_small --patch_size=16 --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC --dump_features eval_knn_vits_16_featuresWill run the code on one GPU.
| distributed init (rank 0): env://
git:
sha: 314dd801c98fc4098afc70abbeafcb506b2c907e, status: has uncommited changes, branch: gimlet
arch: vit_small
batch_size_per_gpu: 128
checkpoint_key: teacher
data_path: /datasets/imagenet/ILSVRC/Data/CLS-LOC
dist_url: env://
dump_features: eval_knn_vits_16_features
gpu: 0
load_features: None
local_rank: 0
nb_knn: [10, 20, 100, 200]
num_workers: 10
patch_size: 16
pretrained_weights:
rank: 0
temperature: 0.07
use_cuda: True
world_size: 1
Data loaded with 1281167 train and 50000 val imgs.
Model vit_small 16x16 built.
Please use the `--pretrained_weights` argument to indicate the path of the checkpoint to evaluate.
Since no pretrained weights have been provided, we load the reference pretrained DINO weights.
Extracting features for train set...
Storing features into tensor of shape torch.Size([1281167, 384])
<REMOVED>
[10009/10010] eta: 0:00:00 time: 0.106392 data: 0.000068 max mem: 2600
Total time: 0:14:44 (0.088400 s / it)
Extracting features for val set...
Storing features into tensor of shape torch.Size([50000, 384])
<REMOVED>
[390/391] eta: 0:00:00 time: 0.106757 data: 0.000093 max mem: 2673
Total time: 0:00:35 (0.091346 s / it)
Features are ready!
Start the k-NN classification.
10-NN classifier result: Top1: 74.402, Top5: 88.422
20-NN classifier result: Top1: 74.444, Top5: 90.02
100-NN classifier result: Top1: 72.902, Top5: 91.482
200-NN classifier result: Top1: 72.062, Top5: 91.546
git checkout vihang/eval_half
python ./eval_linear.py --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --evaluate --patch_size=16Model vit_small built.
We load the reference pretrained linear weights.
Test: [ 0/391] eta: 0:12:20 loss: 0.380615 (0.380615) acc1: 92.968750 (92.968750) acc5: 98.437500 (98.437500) time: 1.892996 data: 1.365466 max mem: 428
Test: [ 20/391] eta: 0:00:56 loss: 0.560059 (0.560416) acc1: 84.375000 (85.193452) acc5: 96.875000 (96.837798) time: 0.065381 data: 0.033924 max mem: 502
Test: [ 40/391] eta: 0:00:43 loss: 0.504395 (0.601469) acc1: 84.375000 (84.032012) acc5: 97.656250 (96.455793) time: 0.092914 data: 0.063916 max mem: 502
Test: [ 60/391] eta: 0:00:37 loss: 0.502441 (0.564628) acc1: 87.500000 (85.245902) acc5: 97.656250 (96.695697) time: 0.088490 data: 0.056156 max mem: 502
Test: [ 80/391] eta: 0:00:32 loss: 0.716309 (0.618588) acc1: 79.687500 (83.497299) acc5: 96.093750 (96.508488) time: 0.087520 data: 0.056728 max mem: 502
Test: [100/391] eta: 0:00:32 loss: 0.624023 (0.628162) acc1: 85.156250 (83.369431) acc5: 96.093750 (96.449567) time: 0.142970 data: 0.106644 max mem: 502
Test: [120/391] eta: 0:00:38 loss: 0.532715 (0.632407) acc1: 83.593750 (83.283833) acc5: 96.875000 (96.532800) time: 0.290818 data: 0.251886 max mem: 502
Test: [140/391] eta: 0:00:33 loss: 0.434326 (0.618959) acc1: 89.843750 (83.610372) acc5: 97.656250 (96.642287) time: 0.084812 data: 0.056365 max mem: 502
Test: [160/391] eta: 0:00:29 loss: 0.679688 (0.630884) acc1: 81.250000 (83.438470) acc5: 95.312500 (96.472244) time: 0.090619 data: 0.059577 max mem: 502
Test: [180/391] eta: 0:00:26 loss: 1.135742 (0.689190) acc1: 69.531250 (81.970822) acc5: 90.625000 (95.804558) time: 0.091500 data: 0.063739 max mem: 502
Test: [200/391] eta: 0:00:23 loss: 1.276367 (0.747667) acc1: 64.062500 (80.600902) acc5: 90.625000 (95.122046) time: 0.100896 data: 0.070123 max mem: 502
Test: [220/391] eta: 0:00:20 loss: 1.095703 (0.774410) acc1: 73.437500 (80.048077) acc5: 91.406250 (94.771635) time: 0.092003 data: 0.062661 max mem: 502
Test: [240/391] eta: 0:00:17 loss: 0.752930 (0.788651) acc1: 79.687500 (79.830135) acc5: 92.968750 (94.511800) time: 0.094443 data: 0.063852 max mem: 502
Test: [260/391] eta: 0:00:15 loss: 1.212891 (0.828628) acc1: 64.843750 (78.744612) acc5: 90.625000 (94.064296) time: 0.093591 data: 0.065633 max mem: 502
Test: [280/391] eta: 0:00:12 loss: 1.028320 (0.848688) acc1: 74.218750 (78.380783) acc5: 92.187500 (93.808385) time: 0.107170 data: 0.071783 max mem: 502
Test: [300/391] eta: 0:00:10 loss: 1.108398 (0.867766) acc1: 72.656250 (77.966674) acc5: 89.843750 (93.513808) time: 0.096569 data: 0.064147 max mem: 502
Test: [320/391] eta: 0:00:08 loss: 1.179688 (0.883463) acc1: 71.875000 (77.579829) acc5: 88.281250 (93.251071) time: 0.105090 data: 0.072960 max mem: 502
Test: [340/391] eta: 0:00:05 loss: 1.224609 (0.904156) acc1: 67.187500 (77.050495) acc5: 89.062500 (93.032900) time: 0.094860 data: 0.061611 max mem: 502
Test: [360/391] eta: 0:00:03 loss: 0.979492 (0.914785) acc1: 75.000000 (76.813539) acc5: 91.406250 (92.949273) time: 0.096149 data: 0.064149 max mem: 502
Test: [380/391] eta: 0:00:01 loss: 0.813477 (0.913679) acc1: 76.562500 (76.849573) acc5: 93.750000 (92.950295) time: 0.099188 data: 0.069708 max mem: 502
Test: [390/391] eta: 0:00:00 loss: 0.726074 (0.911055) acc1: 76.562500 (76.906000) acc5: 93.750000 (92.980000) time: 0.096989 data: 0.064411 max mem: 502
Test: Total time: 0:00:43 (0.110753 s / it)
* Acc@1 76.906 Acc@5 92.980 loss 0.911
Accuracy of the network on the 50000 test images: 76.9%
git checkout vihang/eval_v2
python ./eval_linear.py --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --evaluateModel vit_small built.
Test: [ 0/391] eta: 0:18:37 loss: 0.287185 (0.287185) acc1: 93.750000 (93.750000) acc5: 98.437500 (98.437500) time: 2.858557 data: 1.847843 max mem: 965
Test: [ 20/391] eta: 0:01:33 loss: 0.449175 (0.454168) acc1: 87.500000 (87.723214) acc5: 97.656250 (97.953869) time: 0.121201 data: 0.000267 max mem: 966
Test: [ 40/391] eta: 0:01:03 loss: 0.399514 (0.491631) acc1: 87.500000 (86.432927) acc5: 97.656250 (97.618140) time: 0.107072 data: 0.000223 max mem: 966
Test: [ 60/391] eta: 0:00:51 loss: 0.369379 (0.460332) acc1: 89.843750 (87.435963) acc5: 98.437500 (97.745902) time: 0.107282 data: 0.000181 max mem: 966
Test: [ 80/391] eta: 0:00:44 loss: 0.691794 (0.530860) acc1: 81.250000 (85.648148) acc5: 96.093750 (97.289738) time: 0.106931 data: 0.000173 max mem: 966
Test: [100/391] eta: 0:00:39 loss: 0.607871 (0.549595) acc1: 85.937500 (85.295483) acc5: 96.875000 (97.145730) time: 0.107492 data: 0.000128 max mem: 966
Test: [120/391] eta: 0:00:35 loss: 0.616540 (0.561549) acc1: 83.593750 (85.040031) acc5: 97.656250 (97.113895) time: 0.107259 data: 0.000171 max mem: 966
Test: [140/391] eta: 0:00:32 loss: 0.387987 (0.549117) acc1: 87.500000 (85.250443) acc5: 98.437500 (97.279477) time: 0.107170 data: 0.000180 max mem: 966
Test: [160/391] eta: 0:00:29 loss: 0.590641 (0.559286) acc1: 84.375000 (85.170807) acc5: 96.875000 (97.190411) time: 0.107158 data: 0.000197 max mem: 966
Test: [180/391] eta: 0:00:26 loss: 0.819142 (0.592688) acc1: 77.343750 (84.189399) acc5: 94.531250 (96.900898) time: 0.111123 data: 0.000660 max mem: 966
Test: [200/391] eta: 0:00:24 loss: 0.910006 (0.629575) acc1: 71.093750 (83.290578) acc5: 92.968750 (96.536847) time: 0.167249 data: 0.050212 max mem: 966
Test: [220/391] eta: 0:00:23 loss: 0.771277 (0.641363) acc1: 78.906250 (83.070560) acc5: 95.312500 (96.440187) time: 0.251233 data: 0.136657 max mem: 966
Test: [240/391] eta: 0:00:22 loss: 0.604386 (0.646282) acc1: 82.031250 (82.977827) acc5: 96.875000 (96.297977) time: 0.215593 data: 0.095335 max mem: 966
Test: [260/391] eta: 0:00:18 loss: 0.910095 (0.676159) acc1: 74.218750 (82.100096) acc5: 93.750000 (96.012931) time: 0.113562 data: 0.000376 max mem: 966
Test: [280/391] eta: 0:00:15 loss: 0.777899 (0.686569) acc1: 78.906250 (81.897798) acc5: 94.531250 (95.915814) time: 0.122285 data: 0.000526 max mem: 966
Test: [300/391] eta: 0:00:12 loss: 0.817981 (0.694610) acc1: 77.343750 (81.701620) acc5: 94.531250 (95.790075) time: 0.117873 data: 0.000669 max mem: 966
Test: [320/391] eta: 0:00:10 loss: 0.937429 (0.705814) acc1: 76.562500 (81.466608) acc5: 92.968750 (95.614291) time: 0.191644 data: 0.070490 max mem: 966
Test: [340/391] eta: 0:00:07 loss: 0.908184 (0.720872) acc1: 76.562500 (81.052969) acc5: 92.968750 (95.456837) time: 0.126619 data: 0.001506 max mem: 966
Test: [360/391] eta: 0:00:04 loss: 0.798275 (0.728798) acc1: 78.906250 (80.825831) acc5: 94.531250 (95.409886) time: 0.213799 data: 0.093423 max mem: 966
Test: [380/391] eta: 0:00:01 loss: 0.649764 (0.727318) acc1: 82.812500 (80.868602) acc5: 96.093750 (95.421178) time: 0.134691 data: 0.009982 max mem: 966
Test: [390/391] eta: 0:00:00 loss: 0.589799 (0.725904) acc1: 80.468750 (80.892000) acc5: 96.875000 (95.444000) time: 0.129553 data: 0.002552 max mem: 966
Test: Total time: 0:00:57 (0.146009 s / it)
* Acc@1 80.892 Acc@5 95.444 loss 0.726
Accuracy of the network on the 50000 test images: 80.9%
Trained manually on TinyImageNet and separately on ImageNet.
To train the Linear Classifier head:
python ./eval_linear.py --pretrained_weights=weights/dino_vitti_tinyimageimagenet.pth --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --arch=vit_tinyTest: [ 0/391] eta: 0:07:00 loss: 0.651548 (0.651548) acc1: 85.937500 (85.937500) acc5: 95.312500 (95.312500) time: 1.076122 data: 1.051785 max mem: 532
Test: [ 20/391] eta: 0:00:44 loss: 1.109726 (1.040032) acc1: 70.312500 (74.776786) acc5: 91.406250 (91.592262) time: 0.071502 data: 0.047663 max mem: 532
Test: [ 40/391] eta: 0:00:33 loss: 0.954309 (1.054288) acc1: 75.000000 (73.875762) acc5: 92.187500 (91.310976) time: 0.070326 data: 0.046484 max mem: 532
Test: [ 60/391] eta: 0:00:28 loss: 0.960422 (1.033441) acc1: 77.343750 (74.871926) acc5: 91.406250 (91.214139) time: 0.067605 data: 0.043781 max mem: 532
Test: [ 80/391] eta: 0:00:25 loss: 1.427128 (1.137124) acc1: 64.843750 (71.595293) acc5: 89.843750 (90.509259) time: 0.063824 data: 0.039945 max mem: 532
Test: [100/391] eta: 0:00:22 loss: 1.254161 (1.151063) acc1: 67.187500 (71.008663) acc5: 91.406250 (90.756498) time: 0.062968 data: 0.039105 max mem: 532
Test: [120/391] eta: 0:00:20 loss: 1.045764 (1.137343) acc1: 72.656250 (71.222882) acc5: 92.968750 (90.993027) time: 0.075784 data: 0.051893 max mem: 532
Test: [140/391] eta: 0:00:18 loss: 0.839696 (1.111849) acc1: 78.906250 (71.875000) acc5: 93.750000 (91.240027) time: 0.067497 data: 0.043587 max mem: 532
Test: [160/391] eta: 0:00:17 loss: 1.228135 (1.119752) acc1: 69.531250 (71.841033) acc5: 89.062500 (91.071429) time: 0.083286 data: 0.059484 max mem: 532
Test: [180/391] eta: 0:00:15 loss: 1.924166 (1.208384) acc1: 53.906250 (70.057838) acc5: 80.468750 (89.796271) time: 0.061169 data: 0.037340 max mem: 532
Test: [200/391] eta: 0:00:14 loss: 2.029323 (1.289078) acc1: 50.000000 (68.547886) acc5: 76.562500 (88.576648) time: 0.059261 data: 0.035410 max mem: 532
Test: [220/391] eta: 0:00:12 loss: 1.574746 (1.330794) acc1: 58.593750 (67.799067) acc5: 82.031250 (87.878252) time: 0.067436 data: 0.043592 max mem: 532
Test: [240/391] eta: 0:00:10 loss: 1.518941 (1.361375) acc1: 65.625000 (67.307443) acc5: 82.812500 (87.344398) time: 0.068633 data: 0.044742 max mem: 532
Test: [260/391] eta: 0:00:09 loss: 1.940958 (1.416640) acc1: 54.687500 (66.133860) acc5: 79.687500 (86.599018) time: 0.064766 data: 0.040891 max mem: 532
Test: [280/391] eta: 0:00:07 loss: 1.810320 (1.449457) acc1: 54.687500 (65.474867) acc5: 81.250000 (86.129337) time: 0.065093 data: 0.041177 max mem: 532
Test: [300/391] eta: 0:00:06 loss: 1.754879 (1.483069) acc1: 54.687500 (64.934593) acc5: 79.687500 (85.623443) time: 0.060532 data: 0.036658 max mem: 532
Test: [320/391] eta: 0:00:04 loss: 2.125113 (1.509535) acc1: 53.906250 (64.449474) acc5: 76.562500 (85.197625) time: 0.060982 data: 0.037146 max mem: 532
Test: [340/391] eta: 0:00:03 loss: 1.971048 (1.541902) acc1: 50.781250 (63.744043) acc5: 73.437500 (84.702621) time: 0.067219 data: 0.043330 max mem: 532
Test: [360/391] eta: 0:00:02 loss: 1.771336 (1.558294) acc1: 54.687500 (63.408934) acc5: 79.687500 (84.452909) time: 0.064617 data: 0.040770 max mem: 532
Test: [380/391] eta: 0:00:00 loss: 1.410041 (1.554849) acc1: 64.062500 (63.447343) acc5: 88.281250 (84.524688) time: 0.069729 data: 0.045909 max mem: 532
Test: [390/391] eta: 0:00:00 loss: 1.231684 (1.546354) acc1: 69.531250 (63.666000) acc5: 89.062500 (84.648000) time: 0.060161 data: 0.036815 max mem: 532
Test: Total time: 0:00:27 (0.069449 s / it)
* Acc@1 63.666 Acc@5 84.648 loss 1.546
Accuracy at epoch 99 of the network on the 50000 test images: 63.7%
Max accuracy so far: 63.74%
Training of the supervised linear classifier on frozen features completed.
Top-1 test accuracy: 63.7
python ./eval_linear.py --arch=efficientnet_b0 --data_path=/datasets/imagenet/ILSVRC/Data/CLS-LOC/ --output_dir=out/eval_linear/efficientnetb0_16_imagenet1k --pretrained_weights=out/dino_training/efficientnetb0_16_imagenet1k/checkpoint.pthTest: [ 0/131] eta: 0:09:31 loss: 1.065596 (1.065596) acc1: 75.000000 (75.000000) acc5: 92.187500 (92.187500) time: 4.362770 data: 4.268591 max mem: 4362
Test: [ 20/131] eta: 0:00:32 loss: 1.234613 (1.238226) acc1: 70.312500 (69.072423) acc5: 88.541672 (89.471728) time: 0.093271 data: 0.000208 max mem: 4362
Test: [ 40/131] eta: 0:00:17 loss: 1.640278 (1.417445) acc1: 57.552086 (63.744921) acc5: 85.677086 (87.950968) time: 0.093349 data: 0.000271 max mem: 4362
Test: [ 60/131] eta: 0:00:11 loss: 1.591256 (1.449136) acc1: 62.239586 (63.618513) acc5: 83.593750 (87.218240) time: 0.093471 data: 0.000286 max mem: 4362
Test: [ 80/131] eta: 0:00:07 loss: 1.924977 (1.561640) acc1: 54.427086 (61.905223) acc5: 78.125000 (85.169113) time: 0.093379 data: 0.000252 max mem: 4362
Test: [100/131] eta: 0:00:04 loss: 2.067794 (1.660396) acc1: 52.083336 (60.205241) acc5: 77.083336 (83.632428) time: 0.093135 data: 0.000119 max mem: 4362
Test: [120/131] eta: 0:00:01 loss: 1.980305 (1.712500) acc1: 53.906250 (59.276000) acc5: 76.041672 (82.741480) time: 0.093003 data: 0.000050 max mem: 4362
Test: [130/131] eta: 0:00:00 loss: 1.825498 (1.703395) acc1: 58.072918 (59.678002) acc5: 81.510422 (83.006003) time: 0.089193 data: 0.000046 max mem: 4362
Test: Total time: 0:00:16 (0.126527 s / it)
* Acc@1 59.678 Acc@5 83.006 loss 1.703
Accuracy at epoch 99 of the network on the 50000 test images: 59.7%
Max accuracy so far: 59.69%
Training of the supervised linear classifier on frozen features completed.
Top-1 test accuracy: 59.7