Rishik001/ijepa_stuff/task_3_rerun_2309 - Storage Bucket

Task 3: Category-Conditioned Heterogeneity Analysis

Description

Heterogeneous failure and category-conditioned sweep across predictor layers. Evaluates whether I-JEPA inversion failures correlate with texture statistics (Laplacian variance, edge density) across ImageNet categories.

Command

python examples/ijepa/task3_new_loader.py \
    --data_dir /workspace/images \
    --weights meta \
    --num_samples 1000 \
    --n_categories 50 \
    --seed 42 \
    --top_k 6 \
    --n_steps 10 \
    --output_json results_task_3/final_run_results.json \
    --output_fig results_task_3/fig3_category_heterogeneity.png \
    --save_plots \
    2>&1 | tee results_task_3/final_run.log

Configuration

Parameter	Value
Script	examples/ijepa/task3_new_loader.py (ModelHub strict loader)
Checkpoint	vith14_in1k_ep300.pth.tar (Meta official, 9.7 GB)
Loader	ModelHub.load_checkpoint (strict: 389/389 encoder, 389/389 target, 152/152 predictor)
Architecture	ViT-H/14: d=1280, L=32, H=16 encoder; d=384, L=12 predictor; 256 patches
Samples	N=1000, balanced across 50 ImageNet categories (20 per class)
Predictor Layers	12
Visible Context	51/256 patches
IG Steps	10
Top-K Patches	6 (Jaccard overlap)
Seed	42

Key Results

Metric	Value
Texture/Inversion Correlation	r = -0.0903 (p = 4.28e-03)
Inversion Failure Group	Laplacian Var = 0.02, Edge Density = 0.16
Alignment Group	Laplacian Var = 0.03, Edge Density = 0.16

Outputs

• final_run_results.json -- Full per-category results with metadata
• final_run.log -- Runtime log
• task3_dataset_manifest.json -- Exact images used (reproducibility)

Run Date

2026-09-23