tmp/video-velocity-model/data_processing/scripts/run_umap.sh

# Run UMAP on a single latent .npz
./.venv/bin/python data_processing/umap_alg.py \
  --input /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_dinov2_wreg_base/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  --output /tmp/dinov2_umap.png

# Run UMAP and mark GOP boundaries from the source video
./.venv/bin/python data_processing/umap_alg.py \
  --input /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_dinov2_wreg_base/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  --gop-video-path /mnt/posttrain/zhaoshitian/datasets/ucf101/OpenDataLab___UCF101/raw/data/UCF-101/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01.avi \
  --output /tmp/dinov2_umap_gop.png

# Use flattened frame vectors instead of mean pooling
./.venv/bin/python data_processing/umap_alg.py \
  --input /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_dinov2_wreg_base/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  --projection flatten \
  --output /tmp/dinov2_umap_flatten.png

# Compare the same video across dinov2, mae, siglip2, and flux2_ae in one command
./.venv/bin/python data_processing/umap_alg.py \
  --input \
  /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_dinov2_wreg_base/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_mae_base_p16/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_siglip2_base_p16_i256/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/ucf101_train_split1_flux2_ae/ApplyEyeMakeup/v_ApplyEyeMakeup_g08_c01_patch_tokens.npz \
  --output-dir /mnt/posttrain/zhaoshitian/datasets/ucf101/manifold-analysis-preprocessed/video_features/umap_plots \
  --projection mean_pool



# # read a different key from an npz file
# ./.venv/bin/python data_processing/umap_alg.py --input /path/to/features.npz --input-key my_features

# # run multiple inputs and store per-model plots under one root directory
# ./.venv/bin/python data_processing/umap_alg.py --input /path/a.npz /path/b.npz --output-dir /path/to/umap_plots

# # try cosine distance instead of euclidean
# ./.venv/bin/python data_processing/umap_alg.py --input /path/to/features.npy --metric cosine

# # tune the UMAP geometry
# ./.venv/bin/python data_processing/umap_alg.py --input /path/to/features.npy --n_neighbors 30 --min_dist 0.0

# # hide the frame-index colorbar
# ./.venv/bin/python data_processing/umap_alg.py --input /path/to/features.npy --no_colorbar

# # explicitly provide the source video and mark GOP starts
# ./.venv/bin/python data_processing/umap_alg.py --input /path/to/features.npz --gop-video-path /path/to/video.mp4

# # infer source videos from a shared dataset root and mark GOP starts
# ./.venv/bin/python data_processing/umap_alg.py --input /path/a_patch_tokens.npz /path/b_patch_tokens.npz --gop-video-root /path/to/videos