sync run artifacts: inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cfe434d017e84085b955e7fce6a4dd28acb5dcda6aa5df2cbe78e27acea9787
+size 5114658721

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/e116_best_snapshot.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:554b01d24c45b26a0071b4d8e7b24b3ffefa65c06c9789435aabfd301649bf28
+size 5114658721

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/e150_snapshot.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:115962937c75e7eb1d496602782262e4784d3c378d047680ca8c9c813f8b3845
+size 5114754241

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/git-info.txt

@@ -0,0 +1,2 @@
+branch: main
+commit: e82d40bb2e9210dc731666848badd739c694f1ee

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_0_log.err

@@ -0,0 +1,102 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run vfclercb
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260611_004330-vfclercb
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Syncing run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ablation
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+[01:21:02] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[12:02:17] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: uploading output.log; uploading config.yaml
+wandb: uploading history steps 930-930, summary
+wandb: 
+wandb: Run history:
+wandb:                                     epoch ▁▁▁▁▁▁▁▁▂▂▂▂▃▃▃▃▄▄▄▄▅▅▅▅▅▅▅▆▆▆▆▆▆▆▆▇▇▇▇█
+wandb:                    epoch/avg_data_time_ms ▄▅▅▄▄▄▄█▄▆▄▆▆▃▅▅▇▆▄▄▅▁▄▄▅▂▆▃▃▅
+wandb:                     epoch/avg_gpu_time_ms ▄▄▃▅▅▆▆▅▃▃▃▅▄█▄▃▆▆▅▆█▁▂▂▆▄▄▃▇▇
+wandb:                    epoch/avg_iter_time_ms ▄▄▄▅▅▆▆▆▃▄▃▅▄█▄▃▇▆▅▆█▁▂▂▆▄▅▃▇▇
+wandb:                            epoch/avg_loss █▆▅▄▃▃▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 ▁▂▃▄▆▇█
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 ▁▃▄▅▅▇█
+wandb: linear/ssv2_coarse10_temporal_concat_top1 ▁▁▄▅▆▆█
+wandb: linear/ssv2_coarse10_temporal_concat_top5 ▁▃▅▆▆██
+wandb:                         train/ema/cos_sim ▁▃▄▄▄▆▆▇▇███████████████████████████████
+wandb:                                       +34 ...
+wandb: 
+wandb: Run summary:
+wandb:                                     epoch 30
+wandb:                    epoch/avg_data_time_ms 6.85255
+wandb:                     epoch/avg_gpu_time_ms 4683.10447
+wandb:                    epoch/avg_iter_time_ms 4696.56027
+wandb:                            epoch/avg_loss 0.51306
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 41.16466
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 83.73494
+wandb: linear/ssv2_coarse10_temporal_concat_top1 56.1245
+wandb: linear/ssv2_coarse10_temporal_concat_top5 91.96787
+wandb:                         train/ema/cos_sim 0.79508
+wandb:                                       +34 ...
+wandb: 
+wandb: 🚀 View run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow at: https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+wandb: ⭐️ View project at: https://wandb.ai/uvjepa/vjepa_ablation
+wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
+wandb: Find logs at: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260611_004330-vfclercb/logs
+[rank0]:[W611 12:55:44.290842283 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_1_log.err

@@ -0,0 +1,42 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+[12:37:51] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+[rank1]:[W611 12:55:28.432416152 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_1_log.out

@@ -0,0 +1,212 @@
+submitit INFO (2026-06-11 00:43:17,122) - Starting with JobEnvironment(job_id=25163, hostname=node407, local_rank=1(3), node=0(2), global_rank=1(6))
+submitit INFO (2026-06-11 00:43:17,122) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 30,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 00:43:28][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/6, tasks_per_node=3
+[WARNING ][2026-06-11 00:43:34][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 00:43:40][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 00:43:43][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 00:43:44][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 00:43:44][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 00:43:47][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 6
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-06-11 00:45:06][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-11 12:55:27,874) - Job completed successfully
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-11 12:55:27,880) - Exiting after successful completion
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_2_log.err

@@ -0,0 +1,36 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[07:07:33] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[09:22:17] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[10:48:59] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[rank2]:[W611 12:55:28.433278605 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_2_log.out

@@ -0,0 +1,212 @@
+submitit INFO (2026-06-11 00:43:17,122) - Starting with JobEnvironment(job_id=25163, hostname=node407, local_rank=2(3), node=0(2), global_rank=2(6))
+submitit INFO (2026-06-11 00:43:17,122) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 30,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 00:43:28][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 2/6, tasks_per_node=3
+[WARNING ][2026-06-11 00:43:34][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 00:43:40][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 00:43:44][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 00:43:44][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 00:43:44][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 00:43:47][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 2 / 6
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 2)...
+[INFO    ][2026-06-11 00:45:06][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-11 12:55:27,888) - Job completed successfully
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-11 12:55:27,890) - Exiting after successful completion
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_3_log.err

@@ -0,0 +1,48 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+[04:07:16] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank3]:[W611 12:55:28.458244544 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_3_log.out

@@ -0,0 +1,212 @@
+submitit INFO (2026-06-11 00:43:17,099) - Starting with JobEnvironment(job_id=25163, hostname=node411, local_rank=0(3), node=1(2), global_rank=3(6))
+submitit INFO (2026-06-11 00:43:17,099) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 30,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 00:43:28][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 3/6, tasks_per_node=3
+[WARNING ][2026-06-11 00:43:35][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 00:43:40][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 00:43:43][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 00:43:44][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 00:43:44][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 00:43:47][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 3 / 6
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 3)...
+[INFO    ][2026-06-11 00:45:06][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-11 12:55:27,844) - Job completed successfully
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-11 12:55:27,849) - Exiting after successful completion
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_4_log.err

@@ -0,0 +1,50 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+[11:33:27] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank4]:[W611 12:55:28.447750472 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_4_log.out

@@ -0,0 +1,212 @@
+submitit INFO (2026-06-11 00:43:17,099) - Starting with JobEnvironment(job_id=25163, hostname=node411, local_rank=1(3), node=1(2), global_rank=4(6))
+submitit INFO (2026-06-11 00:43:17,099) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 30,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 00:43:28][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 4/6, tasks_per_node=3
+[WARNING ][2026-06-11 00:43:35][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 00:43:40][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 00:43:43][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 00:43:43][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 00:43:44][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 00:43:47][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 4 / 6
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 4)...
+[INFO    ][2026-06-11 00:45:06][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-11 12:55:27,846) - Job completed successfully
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-11 12:55:27,853) - Exiting after successful completion
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_5_log.err

@@ -0,0 +1,45 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+[02:13:55] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+[04:34:17] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank5]:[W611 12:55:28.438561007 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_5_log.out

@@ -0,0 +1,212 @@
+submitit INFO (2026-06-11 00:43:17,099) - Starting with JobEnvironment(job_id=25163, hostname=node411, local_rank=2(3), node=1(2), global_rank=5(6))
+submitit INFO (2026-06-11 00:43:17,100) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25163/25163_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 30,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 00:43:25][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 00:43:28][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 5/6, tasks_per_node=3
+[WARNING ][2026-06-11 00:43:35][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 00:43:40][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 00:43:44][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 00:43:44][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 00:43:44][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 00:43:46][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 00:43:47][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 5 / 6
+[INFO    ][2026-06-11 00:43:47][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 00:43:47][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 5)...
+[INFO    ][2026-06-11 00:45:06][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-11 12:55:27,847) - Job completed successfully
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-11 12:55:27,853) - Exiting after successful completion
+[INFO    ][2026-06-11 12:55:27][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_0_log.err

@@ -0,0 +1,124 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run vfclercb
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260611_195811-vfclercb
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Resuming run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ablation
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: WARNING Tried to log to step 0 that is less than the current step 9001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+wandb: WARNING Tried to log to step 9000 that is less than the current step 9001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+[20:06:18] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[22:04:49] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+[03:57:12] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[05:07:48] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: uploading output.log; uploading wandb-summary.json
+wandb: uploading history steps 1860-1860, summary, console lines 1357-1366
+wandb: 
+wandb: Run history:
+wandb:                                     epoch ▁▁▁▁▁▂▂▂▂▃▃▃▃▄▄▄▄▅▅▅▅▅▅▅▆▆▆▆▆▆▇▇▇███████
+wandb:                    epoch/avg_data_time_ms ▃▃▃▅▇▁▇▇▇▇▃▃▅▃▂▄▄▅▃▅▅▅█▆▇▄▆██▇
+wandb:                     epoch/avg_gpu_time_ms ▅▇▃▄▄▃▇▅▁▃▄▅▁▃▁▅▃▂▄▄▅▅▇█▆▇▃▅▄▁
+wandb:                    epoch/avg_iter_time_ms ▅▇▃▄▄▃▇▅▁▄▄▅▂▃▁▅▃▂▄▄▅▅▇█▆▇▃▆▄▁
+wandb:                            epoch/avg_loss ▁▁▂▃▃▃▃▄▄▄▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇▇▇██
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 ▁▂▅▅█▆
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 ▁▄▅▃██
+wandb: linear/ssv2_coarse10_temporal_concat_top1 ▁▃█▂▇▅
+wandb: linear/ssv2_coarse10_temporal_concat_top5 ▂▂▁▃▅█
+wandb:                         train/ema/cos_sim █▅▆█▄▅▄▅▄▅▅▆▂▃▅▄▄▅▃▄▁▃▃▃▅▃▃▄▃▃▂▂▃▂▂▃▁▃▄▁
+wandb:                                       +34 ...
+wandb: 
+wandb: Run summary:
+wandb:                                     epoch 60
+wandb:                    epoch/avg_data_time_ms 7.41441
+wandb:                     epoch/avg_gpu_time_ms 4642.75277
+wandb:                    epoch/avg_iter_time_ms 4656.73333
+wandb:                            epoch/avg_loss 0.53577
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 43.4739
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 86.44578
+wandb: linear/ssv2_coarse10_temporal_concat_top1 57.12851
+wandb: linear/ssv2_coarse10_temporal_concat_top5 93.07229
+wandb:                         train/ema/cos_sim 0.76809
+wandb:                                       +34 ...
+wandb: 
+wandb: 🚀 View run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow at: https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+wandb: ⭐️ View project at: https://wandb.ai/uvjepa/vjepa_ablation
+wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
+wandb: Find logs at: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260611_195811-vfclercb/logs
+[rank0]:[W612 08:09:43.865534111 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_1_log.err

@@ -0,0 +1,41 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[20:15:37] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+[20:50:50] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank1]:[W612 08:09:41.224472889 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_1_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-06-11 19:57:57,500) - Starting with JobEnvironment(job_id=25214, hostname=node404, local_rank=1(3), node=0(2), global_rank=1(6))
+submitit INFO (2026-06-11 19:57:57,500) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 19:58:10][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/6, tasks_per_node=3
+[WARNING ][2026-06-11 19:58:23][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 19:58:28][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 19:58:32][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 19:58:32][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 19:58:32][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 19:58:37][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 6
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-06-11 19:58:38][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded optimizers from epoch 30
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 20:00:08][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-12 08:09:41,007) - Job completed successfully
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-12 08:09:41,009) - Exiting after successful completion
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_2_log.err

@@ -0,0 +1,61 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+[21:23:05] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+[01:08:05] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+[01:19:45] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+[06:15:56] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank2]:[W612 08:09:41.221863207 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_2_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-06-11 19:57:57,500) - Starting with JobEnvironment(job_id=25214, hostname=node404, local_rank=2(3), node=0(2), global_rank=2(6))
+submitit INFO (2026-06-11 19:57:57,500) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 19:58:10][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 2/6, tasks_per_node=3
+[WARNING ][2026-06-11 19:58:23][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 19:58:28][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 19:58:33][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 19:58:33][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 19:58:33][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 19:58:37][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 2 / 6
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 2)...
+[INFO    ][2026-06-11 19:58:38][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded optimizers from epoch 30
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 20:00:08][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-12 08:09:41,012) - Job completed successfully
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-12 08:09:41,014) - Exiting after successful completion
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_3_log.err

@@ -0,0 +1,19 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[21:06:09] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+[03:45:16] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+[rank3]:[W612 08:09:42.670453105 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_3_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-06-11 19:57:57,500) - Starting with JobEnvironment(job_id=25214, hostname=node406, local_rank=0(3), node=1(2), global_rank=3(6))
+submitit INFO (2026-06-11 19:57:57,500) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 19:58:09][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 3/6, tasks_per_node=3
+[WARNING ][2026-06-11 19:58:23][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 19:58:28][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 19:58:32][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 19:58:32][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 19:58:32][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 19:58:37][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 3 / 6
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 3)...
+[INFO    ][2026-06-11 19:58:38][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded optimizers from epoch 30
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 20:00:08][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-12 08:09:41,170) - Job completed successfully
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-12 08:09:41,173) - Exiting after successful completion
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_4_log.err

@@ -0,0 +1,33 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank4]:[W612 08:09:41.289808277 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_4_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-06-11 19:57:57,500) - Starting with JobEnvironment(job_id=25214, hostname=node406, local_rank=1(3), node=1(2), global_rank=4(6))
+submitit INFO (2026-06-11 19:57:57,500) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 19:58:09][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 4/6, tasks_per_node=3
+[WARNING ][2026-06-11 19:58:23][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 19:58:28][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 19:58:32][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 19:58:32][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 19:58:32][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 19:58:37][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 4 / 6
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 4)...
+[INFO    ][2026-06-11 19:58:38][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded optimizers from epoch 30
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 20:00:08][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-12 08:09:41,141) - Job completed successfully
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-12 08:09:41,144) - Exiting after successful completion
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_5_log.err

@@ -0,0 +1,47 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+[23:40:54] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+[00:15:42] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+[02:08:41] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+[03:06:55] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank5]:[W612 08:09:42.656736785 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_5_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-06-11 19:57:57,500) - Starting with JobEnvironment(job_id=25214, hostname=node406, local_rank=2(3), node=1(2), global_rank=5(6))
+submitit INFO (2026-06-11 19:57:57,500) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_25214/25214_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 32,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 2,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-11 19:58:06][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-11 19:58:09][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 5/6, tasks_per_node=3
+[WARNING ][2026-06-11 19:58:23][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-11 19:58:28][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-11 19:58:32][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-11 19:58:32][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-11 19:58:32][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-11 19:58:36][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-11 19:58:37][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 5 / 6
+[INFO    ][2026-06-11 19:58:37][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-11 19:58:37][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 5)...
+[INFO    ][2026-06-11 19:58:38][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 30 with msg: <All keys matched successfully>
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] loaded optimizers from epoch 30
+[INFO    ][2026-06-11 19:58:48][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-11 20:00:08][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-06-12 08:09:41,163) - Job completed successfully
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-06-12 08:09:41,166) - Exiting after successful completion
+[INFO    ][2026-06-12 08:09:41][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_0_log.err

@@ -0,0 +1,18 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run vfclercb
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260623_190228-vfclercb
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Resuming run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ablation
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+srun: Job step aborted: Waiting up to 32 seconds for job step to finish.
+slurmstepd: error: *** STEP 26022.0 ON node403 CANCELLED AT 2026-06-23T19:04:10 ***
+slurmstepd: error: *** JOB 26022 ON node403 CANCELLED AT 2026-06-23T19:04:10 ***
+submitit WARNING (2026-06-23 19:04:10,889) - Bypassing signal SIGTERM
+submitit WARNING (2026-06-23 19:04:10,890) - Bypassing signal SIGCONT
+slurmstepd: error: Failed to send MESSAGE_TASK_EXIT: Connection refused

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_0_log.out

@@ -0,0 +1,241 @@
+submitit INFO (2026-06-23 19:01:59,553) - Starting with JobEnvironment(job_id=26022, hostname=node403, local_rank=0(3), node=0(1), global_rank=0(3))
+submitit INFO (2026-06-23 19:01:59,553) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-23 19:02:26][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 0/3, tasks_per_node=3
+[INFO    ][2026-06-23 19:02:27][app.vjepa.train     ][main                     ] Resuming wandb run: vfclercb
+[INFO    ][2026-06-23 19:02:34][app.vjepa.train     ][main                     ] Initialized wandb run: d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow (id=vfclercb)
+[WARNING ][2026-06-23 19:03:08][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-23 19:03:19][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-23 19:03:21][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-23 19:03:22][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-23 19:03:24][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-23 19:03:29][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-23 19:03:29][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 0 / 3
+[INFO    ][2026-06-23 19:03:29][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-23 19:03:29][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-23 19:03:29][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 0)...
+[INFO    ][2026-06-23 19:03:31][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] loaded optimizers from epoch 60
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-23 19:03:42][app.vjepa.train     ][main                     ] Running initial KNN evaluation (epoch 0)...
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Starting ssv2_coarse10 KNN evaluation (frames_per_clip=8)...
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Pool type: temporal_concat, CLS token: False
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Using fps=2 (dynamic frame_step per video)
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Presampled train subset with seed=0: 5000 videos across 10 classes
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Presampled val subset with seed=0: 996 videos across 10 classes
+[INFO    ][2026-06-23 19:03:43][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-23 19:03:43][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-23 19:03:43][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-23 19:03:43][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Train loader: 209 batches, Val loader: 42 batches
+[INFO    ][2026-06-23 19:03:43][src.utils.knn_eval  ][run_knn_eval             ] [KNN Eval] Extracting training features...
+[INFO    ][2026-06-23 19:03:56][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 0/209 batches
+[INFO    ][2026-06-23 19:04:00][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 20/209 batches
+[INFO    ][2026-06-23 19:04:05][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 40/209 batches
+[INFO    ][2026-06-23 19:04:10][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 60/209 batches
+submitit WARNING (2026-06-23 19:04:10,889) - Bypassing signal SIGTERM
+[WARNING ][2026-06-23 19:04:10][submitit            ][bypass                   ] Bypassing signal SIGTERM
+submitit WARNING (2026-06-23 19:04:10,890) - Bypassing signal SIGCONT
+[WARNING ][2026-06-23 19:04:10][submitit            ][bypass                   ] Bypassing signal SIGCONT
+[INFO    ][2026-06-23 19:04:14][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 80/209 batches
+[INFO    ][2026-06-23 19:04:19][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 100/209 batches
+[INFO    ][2026-06-23 19:04:24][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 120/209 batches
+[INFO    ][2026-06-23 19:04:28][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 140/209 batches
+[INFO    ][2026-06-23 19:04:33][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 160/209 batches
+[INFO    ][2026-06-23 19:04:37][src.utils.knn_eval  ][_extract_features        ] [KNN Eval]   Extracted 180/209 batches

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_1_log.err

@@ -0,0 +1,6 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+submitit WARNING (2026-06-23 19:04:10,923) - Bypassing signal SIGTERM
+submitit WARNING (2026-06-23 19:04:10,924) - Bypassing signal SIGCONT

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_1_log.out

@@ -0,0 +1,217 @@
+submitit INFO (2026-06-23 19:01:59,553) - Starting with JobEnvironment(job_id=26022, hostname=node403, local_rank=1(3), node=0(1), global_rank=1(3))
+submitit INFO (2026-06-23 19:01:59,553) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-23 19:02:26][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/3, tasks_per_node=3
+[WARNING ][2026-06-23 19:03:08][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-23 19:03:19][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-23 19:03:22][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-23 19:03:22][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-23 19:03:24][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-23 19:03:29][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-23 19:03:29][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 3
+[INFO    ][2026-06-23 19:03:29][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-23 19:03:29][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-23 19:03:29][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-06-23 19:03:31][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] loaded optimizers from epoch 60
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+submitit WARNING (2026-06-23 19:04:10,923) - Bypassing signal SIGTERM
+[WARNING ][2026-06-23 19:04:10][submitit            ][bypass                   ] Bypassing signal SIGTERM
+submitit WARNING (2026-06-23 19:04:10,924) - Bypassing signal SIGCONT
+[WARNING ][2026-06-23 19:04:10][submitit            ][bypass                   ] Bypassing signal SIGCONT

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_2_log.err

@@ -0,0 +1,6 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+submitit WARNING (2026-06-23 19:04:10,794) - Bypassing signal SIGTERM
+submitit WARNING (2026-06-23 19:04:10,794) - Bypassing signal SIGCONT

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_2_log.out

@@ -0,0 +1,217 @@
+submitit INFO (2026-06-23 19:01:59,553) - Starting with JobEnvironment(job_id=26022, hostname=node403, local_rank=2(3), node=0(1), global_rank=2(3))
+submitit INFO (2026-06-23 19:01:59,553) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26022/26022_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 60,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-06-23 19:02:23][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-06-23 19:02:26][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 2/3, tasks_per_node=3
+[WARNING ][2026-06-23 19:03:08][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-06-23 19:03:19][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-06-23 19:03:21][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-06-23 19:03:22][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-06-23 19:03:24][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-06-23 19:03:28][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-06-23 19:03:29][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-06-23 19:03:29][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 2 / 3
+[INFO    ][2026-06-23 19:03:29][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-06-23 19:03:29][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-06-23 19:03:29][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 2)...
+[INFO    ][2026-06-23 19:03:31][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:41][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] loaded optimizers from epoch 60
+[INFO    ][2026-06-23 19:03:42][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+submitit WARNING (2026-06-23 19:04:10,794) - Bypassing signal SIGTERM
+[WARNING ][2026-06-23 19:04:10][submitit            ][bypass                   ] Bypassing signal SIGTERM
+submitit WARNING (2026-06-23 19:04:10,794) - Bypassing signal SIGCONT
+[WARNING ][2026-06-23 19:04:10][submitit            ][bypass                   ] Bypassing signal SIGCONT

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_0_log.err

@@ -0,0 +1,148 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run vfclercb
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260707_132026-vfclercb
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Resuming run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ablation
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: WARNING Tried to log to step 0 that is less than the current step 18001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+wandb: WARNING Tried to log to step 18000 that is less than the current step 18001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+[15:57:54] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[17:09:51] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+[04:40:06] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[05:19:11] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[12:42:10] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: updating run metadata
+wandb: uploading output.log; uploading wandb-summary.json
+wandb: uploading history steps 2790-2790, summary, console lines 1360-1367
+wandb: 
+wandb: Run history:
+wandb:                                     epoch ▁▁▁▁▁▁▂▂▂▂▂▃▃▃▃▄▅▅▅▅▅▅▅▅▅▅▅▅▆▆▇▇▇▇▇█████
+wandb:                    epoch/avg_data_time_ms ▆▃▄▄▁▅▃▅▅▅▃▅▄▃▃▃█▃▅▅▂▆▄▆▁▆▅▃▅▄
+wandb:                     epoch/avg_gpu_time_ms ▃▄▃▂▂▃▂▇▇▅▆▇▄█▅▄▄▁▂▃▇▄▇▃▂▆▆▅▇▆
+wandb:                    epoch/avg_iter_time_ms ▃▄▃▂▂▃▂▇▇▅▆▇▄█▅▄▄▁▃▄▇▅▇▄▂▆▆▅▇▆
+wandb:                            epoch/avg_loss ▁▁▁▂▂▂▃▃▃▄▄▄▄▄▅▅▅▆▆▆▆▇▇▇▇▇▇███
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 ▁█▇▄▄▂
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 █▇▂▃▁▆
+wandb: linear/ssv2_coarse10_temporal_concat_top1 ▁▂▃█▆█
+wandb: linear/ssv2_coarse10_temporal_concat_top5 ▄▄▁▃█▄
+wandb:                         train/ema/cos_sim ▆▇█▅▅▆▅▃▄▅▅▂▅▄▄▆▅▄▃▅▄▄▆▄▄▃▃█▃▃▄▂▂▂▃▄▄▄▃▁
+wandb:                                       +34 ...
+wandb: 
+wandb: Run summary:
+wandb:                                     epoch 90
+wandb:                    epoch/avg_data_time_ms 15.44802
+wandb:                     epoch/avg_gpu_time_ms 9210.69954
+wandb:                    epoch/avg_iter_time_ms 9238.44177
+wandb:                            epoch/avg_loss 0.54822
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 45.08032
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 87.249
+wandb: linear/ssv2_coarse10_temporal_concat_top1 58.93574
+wandb: linear/ssv2_coarse10_temporal_concat_top5 92.67068
+wandb:                         train/ema/cos_sim 0.74331
+wandb:                                       +34 ...
+wandb: 
+wandb: 🚀 View run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow at: https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+wandb: ⭐️ View project at: https://wandb.ai/uvjepa/vjepa_ablation
+wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
+wandb: Find logs at: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260707_132026-vfclercb/logs
+[rank0]:[W708 12:54:53.410661669 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_1_log.err

@@ -0,0 +1,82 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[15:05:40] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+[00:44:39] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+[10:33:29] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank1]:[W708 12:54:52.177463153 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_1_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-07-07 13:20:01,953) - Starting with JobEnvironment(job_id=26986, hostname=node404, local_rank=1(3), node=0(1), global_rank=1(3))
+submitit INFO (2026-07-07 13:20:01,953) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': -1,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 90,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-07-07 13:20:22][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-07-07 13:20:22][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-07-07 13:20:22][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-07-07 13:20:25][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/3, tasks_per_node=3
+[WARNING ][2026-07-07 13:21:13][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-07-07 13:21:23][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-07-07 13:21:26][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-07-07 13:21:26][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-07-07 13:21:27][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-07-07 13:21:33][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-07-07 13:21:33][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 3
+[INFO    ][2026-07-07 13:21:33][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-07-07 13:21:33][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-07-07 13:21:33][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-07-07 13:21:34][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-07 13:21:44][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-07-07 13:21:44][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-07-07 13:21:45][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-07-07 13:21:45][root                ][load_checkpoint          ] loaded optimizers from epoch 60
+[INFO    ][2026-07-07 13:21:45][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-07 13:23:07][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-07-08 12:54:51,517) - Job completed successfully
+[INFO    ][2026-07-08 12:54:51][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-07-08 12:54:51,520) - Exiting after successful completion
+[INFO    ][2026-07-08 12:54:51][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_2_log.err

@@ -0,0 +1,79 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+[17:04:08] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+[04:14:21] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+[09:40:33] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+[11:51:11] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank2]:[W708 12:54:52.180945773 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_2_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-07-07 13:20:01,953) - Starting with JobEnvironment(job_id=26986, hostname=node404, local_rank=2(3), node=0(1), global_rank=2(3))
+submitit INFO (2026-07-07 13:20:01,953) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_26986/26986_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': -1,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 90,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-07-07 13:20:22][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-07-07 13:20:22][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-07-07 13:20:22][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-07-07 13:20:25][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 2/3, tasks_per_node=3
+[WARNING ][2026-07-07 13:21:13][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-07-07 13:21:22][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-07-07 13:21:25][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-07-07 13:21:25][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-07-07 13:21:27][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-07-07 13:21:32][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-07-07 13:21:33][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-07-07 13:21:33][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 2 / 3
+[INFO    ][2026-07-07 13:21:33][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-07-07 13:21:33][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-07-07 13:21:33][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 2)...
+[INFO    ][2026-07-07 13:21:34][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-07 13:21:44][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-07-07 13:21:44][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-07-07 13:21:45][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 60 with msg: <All keys matched successfully>
+[INFO    ][2026-07-07 13:21:45][root                ][load_checkpoint          ] loaded optimizers from epoch 60
+[INFO    ][2026-07-07 13:21:45][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-07 13:23:07][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-07-08 12:54:51,527) - Job completed successfully
+[INFO    ][2026-07-08 12:54:51][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-07-08 12:54:51,529) - Exiting after successful completion
+[INFO    ][2026-07-08 12:54:51][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_0_log.err

@@ -0,0 +1,161 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run vfclercb
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260708_144006-vfclercb
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Resuming run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ablation
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: WARNING Tried to log to step 0 that is less than the current step 27001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+wandb: WARNING Tried to log to step 27000 that is less than the current step 27001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+[15:19:04] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+[21:50:35] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+[22:45:51] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+[00:28:09] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+[01:44:13] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[08:00:09] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+[11:51:53] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+[14:05:19] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: updating run metadata
+wandb: updating run metadata; uploading output.log; uploading wandb-summary.json
+wandb: uploading output.log; uploading config.yaml
+wandb: 
+wandb: Run history:
+wandb:                                     epoch ▁▁▁▁▂▂▂▂▂▂▂▂▃▃▃▃▄▄▄▄▄▄▄▄▅▅▅▅▅▆▆▆▆▇▇▇▇▇▇█
+wandb:                    epoch/avg_data_time_ms ▆█▇▂▇▆▄█▇▄▄▅▇▅▆█▆▆▇▆▄▅▂▅▇▃▁▅▂▆
+wandb:                     epoch/avg_gpu_time_ms █▆▃▃▇▃█▆▅▃▃▂▃▆▇▂▂▄▆▆▇▅▂█▆▆▂▁▅▄
+wandb:                    epoch/avg_iter_time_ms █▆▃▃▇▃█▆▅▃▃▂▃▆▇▂▂▄▆▆▇▅▂█▆▆▂▁▅▄
+wandb:                            epoch/avg_loss ▁▂▄▅▄▅▄▅▆▇▇█▇▆▆█▇▅▄▅▅▅▅▂▃▄▄▂▂▂
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 ▃▁▄█▆▅
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 ▇█▇▄▅▁
+wandb: linear/ssv2_coarse10_temporal_concat_top1 ▄▁▂▆█▅
+wandb: linear/ssv2_coarse10_temporal_concat_top5 ▁▆▆▇█▇
+wandb:                         train/ema/cos_sim ▄▇▅▆▄▃▂▂▅▅▇▄▄▆▃▄▃▂▄▃▄▃▃▅█▆▂▄▃▄▆▅▆▂▅▄▁▄▆█
+wandb:                                       +34 ...
+wandb: 
+wandb: Run summary:
+wandb:                                     epoch 120
+wandb:                    epoch/avg_data_time_ms 15.90538
+wandb:                     epoch/avg_gpu_time_ms 9191.97061
+wandb:                    epoch/avg_iter_time_ms 9220.20652
+wandb:                            epoch/avg_loss 0.54899
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 46.98795
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 86.14458
+wandb: linear/ssv2_coarse10_temporal_concat_top1 59.13655
+wandb: linear/ssv2_coarse10_temporal_concat_top5 93.4739
+wandb:                         train/ema/cos_sim 0.75001
+wandb:                                       +34 ...
+wandb: 
+wandb: 🚀 View run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow at: https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+wandb: ⭐️ View project at: https://wandb.ai/uvjepa/vjepa_ablation
+wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
+wandb: Find logs at: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260708_144006-vfclercb/logs
+[rank0]:[W709 14:15:10.545157731 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_1_log.err

@@ -0,0 +1,102 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+[16:14:19] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[18:11:00] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+[20:03:38] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+[22:11:43] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[03:54:03] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[06:54:57] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[11:37:52] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank1]:[W709 14:15:08.752473589 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_1_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-07-08 14:39:49,900) - Starting with JobEnvironment(job_id=27036, hostname=node404, local_rank=1(3), node=0(1), global_rank=1(3))
+submitit INFO (2026-07-08 14:39:49,900) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': -1,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 120,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-07-08 14:40:02][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-07-08 14:40:02][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-07-08 14:40:02][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-07-08 14:40:05][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/3, tasks_per_node=3
+[WARNING ][2026-07-08 14:40:37][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-07-08 14:40:44][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-07-08 14:40:47][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-07-08 14:40:48][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-07-08 14:40:48][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-07-08 14:40:51][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-07-08 14:40:51][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 3
+[INFO    ][2026-07-08 14:40:51][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-07-08 14:40:51][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-07-08 14:40:51][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-07-08 14:40:52][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-08 14:41:00][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 90 with msg: <All keys matched successfully>
+[INFO    ][2026-07-08 14:41:00][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 90 with msg: <All keys matched successfully>
+[INFO    ][2026-07-08 14:41:00][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 90 with msg: <All keys matched successfully>
+[INFO    ][2026-07-08 14:41:01][root                ][load_checkpoint          ] loaded optimizers from epoch 90
+[INFO    ][2026-07-08 14:41:01][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-08 14:42:20][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-07-09 14:15:07,327) - Job completed successfully
+[INFO    ][2026-07-09 14:15:07][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-07-09 14:15:07,329) - Exiting after successful completion
+[INFO    ][2026-07-09 14:15:07][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_2_log.err

@@ -0,0 +1,103 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+[15:09:37] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+[16:15:57] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+[17:10:23] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+[18:14:21] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[19:11:49] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+[19:24:06] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+[22:28:03] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+[03:32:39] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+[07:47:35] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+[09:49:39] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank2]:[W709 14:15:08.751823810 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_2_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-07-08 14:39:49,900) - Starting with JobEnvironment(job_id=27036, hostname=node404, local_rank=2(3), node=0(1), global_rank=2(3))
+submitit INFO (2026-07-08 14:39:49,900) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27036/27036_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': -1,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 120,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-07-08 14:40:02][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-07-08 14:40:02][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-07-08 14:40:02][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-07-08 14:40:05][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 2/3, tasks_per_node=3
+[WARNING ][2026-07-08 14:40:37][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-07-08 14:40:43][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-07-08 14:40:47][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-07-08 14:40:47][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-07-08 14:40:47][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-07-08 14:40:50][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-07-08 14:40:51][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-07-08 14:40:51][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 2 / 3
+[INFO    ][2026-07-08 14:40:51][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-07-08 14:40:51][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-07-08 14:40:51][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 2)...
+[INFO    ][2026-07-08 14:40:52][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-08 14:41:00][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 90 with msg: <All keys matched successfully>
+[INFO    ][2026-07-08 14:41:00][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 90 with msg: <All keys matched successfully>
+[INFO    ][2026-07-08 14:41:00][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 90 with msg: <All keys matched successfully>
+[INFO    ][2026-07-08 14:41:01][root                ][load_checkpoint          ] loaded optimizers from epoch 90
+[INFO    ][2026-07-08 14:41:01][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-08 14:42:20][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-07-09 14:15:07,350) - Job completed successfully
+[INFO    ][2026-07-09 14:15:07][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-07-09 14:15:07,355) - Exiting after successful completion
+[INFO    ][2026-07-09 14:15:07][submitit            ][process_job              ] Exiting after successful completion

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27070/27070_0_log.err

@@ -0,0 +1,167 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run vfclercb
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260709_143819-vfclercb
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Resuming run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ablation
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: WARNING Tried to log to step 0 that is less than the current step 36001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+wandb: WARNING Tried to log to step 36000 that is less than the current step 36001. Steps must be monotonically increasing, so this data will be ignored. See https://wandb.me/define-metric to log data out of order.
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+[15:20:37] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+[17:06:20] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+[17:09:20] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+[19:44:41] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/water_skiing/4bhvIHWADVU_000022_000032.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+[00:47:16] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+[06:08:27] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/sailing/ZolXG_-RgIs_000970_000980.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/dancing_macarena/uGQgxFHemyA_000059_000069.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+wandb: updating run metadata; uploading wandb-summary.json; uploading output.log
+wandb: updating run metadata; uploading output.log
+wandb: updating run metadata
+wandb: uploading config.yaml
+wandb: uploading history steps 4650-4650, summary, console lines 1356-1364
+wandb: 
+wandb: Run history:
+wandb:                                     epoch ▁▁▁▁▂▂▃▃▃▃▃▃▃▃▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▆▇▇▇▇▇▇████
+wandb:                    epoch/avg_data_time_ms ▄▄▄▂▃▃▂▂▂▄█▄▂▁▃▂▆▄▅▃▃▄▂▄▃▂▆▇▁▅
+wandb:                     epoch/avg_gpu_time_ms ▃▂▅▄▂▃▅▂▆▃▃▃▆▄▃▂▄▁▅▅▃▂▃▂▄▁▁▁█▂
+wandb:                    epoch/avg_iter_time_ms ▃▂▅▄▂▃▅▂▆▃▃▃▆▄▃▂▄▁▅▅▃▂▃▂▄▁▁▁█▂
+wandb:                            epoch/avg_loss ██▇▇██▇▇▆▇▇▇▆▆▆▆▆▇▅▆▅▅▅▆▄▃▄▃▁▂
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 █▄▆▇▁▁
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 ██▇█▁▅
+wandb: linear/ssv2_coarse10_temporal_concat_top1 █▆▆▅▁▂
+wandb: linear/ssv2_coarse10_temporal_concat_top5 ▂█▇▄▁▃
+wandb:                         train/ema/cos_sim ▅▅▃▄▅▄▅▁▄▇▅▆▃▅█▄▄▄▆█▃▃▅▂▅▄▃▄▄▄▆▆▄▅▅▅▃▇▅▅
+wandb:                                       +34 ...
+wandb: 
+wandb: Run summary:
+wandb:                                     epoch 150
+wandb:                    epoch/avg_data_time_ms 15.81663
+wandb:                     epoch/avg_gpu_time_ms 9190.28191
+wandb:                    epoch/avg_iter_time_ms 9218.5123
+wandb:                            epoch/avg_loss 0.54449
+wandb:    knn/ssv2_coarse10_temporal_concat_top1 44.57831
+wandb:    knn/ssv2_coarse10_temporal_concat_top5 86.84739
+wandb: linear/ssv2_coarse10_temporal_concat_top1 58.33333
+wandb: linear/ssv2_coarse10_temporal_concat_top5 92.87149
+wandb:                         train/ema/cos_sim 0.74967
+wandb:                                       +34 ...
+wandb: 
+wandb: 🚀 View run d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow at: https://wandb.ai/uvjepa/vjepa_ablation/runs/vfclercb
+wandb: ⭐️ View project at: https://wandb.ai/uvjepa/vjepa_ablation
+wandb: Synced 5 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
+wandb: Find logs at: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/wandb/run-20260709_143819-vfclercb/logs
+[rank0]:[W710 14:14:36.404469477 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27070/27070_1_log.err

@@ -0,0 +1,91 @@
+/var/scratch/dcanezil/.cache/uv/virtualenvs/vd/lib/python3.12/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/utils.py:888: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/bowling/f4eb0wOlspM_000053_000063.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[22:34:34] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/climbing_tree/QX5IzK4iFCM_000002_000012.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/robot_dancing/J1x0XlWa6HM_000147_000157.avi'
+  warnings.warn(f"video path not found {fname=}")
+[22:58:07] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/making_tea/mtYFNsRcxY4_000063_000073.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+[02:07:20] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/-jGxlNQKkeo_000092_000102.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_poker/6-DyF8umej8_000097_000107.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+[06:39:37] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/long_jump/jsYiitkbr5w_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+[07:56:42] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_monopoly/NLL667uPWVA_000066_000076.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/eating_carrots/eiZ8Hzc7FPU_000080_000090.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/windsurfing/i-gzh_BPDa8_000154_000164.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/waxing_legs/cqUusXBODuw_001046_001056.avi'
+  warnings.warn(f"video path not found {fname=}")
+[10:02:45] /github/workspace/src/video/video_reader.cc:83: ERROR opening: /local/dcanezil/data/kinetics/VideoData/trimming_trees/T6dqaZioaXs_000004_000014.avi, Invalid data found when processing input
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/app/vjepa/train.py:983: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_saxophone/Zs0_2GMEPXo_000054_000064.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/surfing_water/5l5Pdd96Pao_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/baking_cookies/Q3XGZmqk1Q0_000296_000306.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/pumping_fist/wKsuyr6Xk30_000094_000104.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/picking_fruit/3VvkoFtPCCU_000045_000055.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/playing_flute/co50KUHacYw_000005_000015.avi'
+  warnings.warn(f"video path not found {fname=}")
+/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/code/src/datasets/video_dataset.py:418: UserWarning: video path not found fname='/local/dcanezil/data/kinetics/VideoData/frying_vegetables/Nr-zCXgW2Sw_000161_000171.avi'
+  warnings.warn(f"video path not found {fname=}")
+[rank1]:[W710 14:14:33.737608791 ProcessGroupNCCL.cpp:1553] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27070/27070_1_log.out

@@ -0,0 +1,218 @@
+submitit INFO (2026-07-09 14:38:05,873) - Starting with JobEnvironment(job_id=27070, hostname=node404, local_rank=1(3), node=0(1), global_rank=1(3))
+submitit INFO (2026-07-09 14:38:05,873) - Loading pickle: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/job_27070/27070_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [8, 8],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/local/dcanezil/data/kinetics/train.csv',
+                                '/local/dcanezil/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 2,
+                'num_workers': 4,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.05, 0.05],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.4, 0.4],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '40G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': True,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 8,
+                                                          'dataset_train': '/local/dcanezil/data/ssv2/train_coarse10.csv',
+                                                          'dataset_val': '/local/dcanezil/data/ssv2/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 2,
+                                                          'pool_type': 'temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 10,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ablation'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'class_token': True,
+                 'freeze_backbone': False,
+                 'is_causal': True,
+                 'model_name': 'vit_large_patch14_dinov2_lvd142m',
+                 'pe_type': '2d',
+                 'pred_depth': 12,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 6,
+                 'predictor': 'v2_cross',
+                 'stem_type': '2d',
+                 'target_embed_dim': 1024,
+                 'target_kind': 'ema',
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_alibi': False,
+                 'use_lora': False,
+                 'use_mask_tokens': True,
+                 'use_rope': False,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 1,
+    'optimization': {   'clip_grad': None,
+                        'ema': [1.0, 0.99925],
+                        'ema_hold_epochs': 0,
+                        'ema_ramp_epochs': 5,
+                        'encoder_layer_decay': 0.9,
+                        'epochs': 150,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0001,
+                        'predictor_lr_scale': 3.0,
+                        'start_lr': 1e-05,
+                        'warmup': 2,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 3}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-07-09 14:38:14][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-07-09 14:38:14][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-07-09 14:38:14][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=None
+[INFO    ][2026-07-09 14:38:17][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/3, tasks_per_node=3
+[WARNING ][2026-07-09 14:38:56][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] DINOv2 is an image model — fwd_chunk is unset so all frames will be processed jointly. Set fwd_chunk=1 for per-frame (standard) processing.
+[INFO    ][2026-07-09 14:39:01][experiments.stmodels.vision_transformers_v3][_build_dinov2_vit        ] Loading pretrained weights for vit_large_patch14_dinov2_lvd142m from timm (vit_large_patch14_dinov2)
+[INFO    ][2026-07-09 14:39:04][timm.models._builder][load_pretrained          ] Loading pretrained weights from Hugging Face hub (timm/vit_large_patch14_dinov2.lvd142m)
+[INFO    ][2026-07-09 14:39:04][timm.models._hub    ][load_state_dict_from_hf  ] [timm/vit_large_patch14_dinov2.lvd142m] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
+[INFO    ][2026-07-09 14:39:04][timm.layers.pos_embed][resample_abs_pos_embed   ] Resized position embedding: (37, 37) to (16, 16).
+[INFO    ][2026-07-09 14:39:07][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-09 14:39:07][root                ][init_video_model         ] VJEPAPredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-11): 12 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-07-09 14:39:07][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): VisionTransformer(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (norm2): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=True)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=True)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+  )
+)
+[INFO    ][2026-07-09 14:39:07][root                ][init_video_model         ] Encoder number of parameters: 302964736
+[INFO    ][2026-07-09 14:39:07][root                ][init_video_model         ] Predictor number of parameters: 22042240
+[INFO    ][2026-07-09 14:39:07][root                ][init_video_model         ] Target encoder number of parameters: 302964736
+[INFO    ][2026-07-09 14:39:08][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-07-09 14:39:08][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 3
+[INFO    ][2026-07-09 14:39:08][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-07-09 14:39:08][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/2103
+[INFO    ][2026-07-09 14:39:08][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-07-09 14:39:09][root                ][load_checkpoint          ] Loading checkpoint from /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-09 14:39:17][root                ][load_checkpoint          ] loaded pretrained encoder from epoch 120 with msg: <All keys matched successfully>
+[INFO    ][2026-07-09 14:39:17][root                ][load_checkpoint          ] loaded pretrained predictor from epoch 120 with msg: <All keys matched successfully>
+[INFO    ][2026-07-09 14:39:17][root                ][load_checkpoint          ] loaded pretrained target encoder from epoch 120 with msg: <All keys matched successfully>
+[INFO    ][2026-07-09 14:39:18][root                ][load_checkpoint          ] loaded optimizers from epoch 120
+[INFO    ][2026-07-09 14:39:18][root                ][load_checkpoint          ] read-path: /var/scratch/dcanezil/runs/vjepa/inflated/d073_full_ema_dinov2l_ssv2k400_8f_emainv_ramp5_predlr3_llrd0p9_masklow/latest.pt
+[INFO    ][2026-07-09 14:40:37][app.vjepa.train     ][main                     ] Initializing loader...
+submitit INFO (2026-07-10 14:14:32,240) - Job completed successfully
+[INFO    ][2026-07-10 14:14:32][submitit            ][process_job              ] Job completed successfully
+submitit INFO (2026-07-10 14:14:32,242) - Exiting after successful completion
+[INFO    ][2026-07-10 14:14:32][submitit            ][process_job              ] Exiting after successful completion