DanielNguyen121 commited on
Commit
ea0383f
·
verified ·
1 Parent(s): 5f300c7

DISEF CALTECH101 reproducibility deliverables: caltech101_disef_20261005T073721Z_262ae8436d11

Browse files
DISEF_CALTECH101_PATCH.diff ADDED
The diff for this file is too large to render. See raw diff
 
DISEF_CALTECH101_RUN_REPORT.md ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # DISEF Caltech101 Run Report
2
+
3
+ ## 1. Executive Summary
4
+
5
+ Execution is being performed in the new DisEF-Caltech Molab sandbox. This report is updated throughout the run.
6
+
7
+ ## 2. DISEF Method
8
+
9
+ The SPC implementation is audited against the official DISEF implementation and paper before generation. The intended path is: 16-shot real images → LLaVA semantic captions → image-to-image diffusion conditioning → DPM-Solver++ denoising → Qwen3-VL semantic filtering → validated synthetic images. The diffusion and caption/filter models are frozen for generation; no classifier training is run.
10
+
11
+ ## 3. SPC Implementation
12
+
13
+ Original SPC commit: `96834b22d5e59c852e53d713d434cbabb2f0d165`. The original `DisEF/disef_gen.py` is a Colab export with hard-coded FGVC Aircraft paths and IPython shell syntax; it is not a safe CLI entrypoint.
14
+
15
+ ## 4. Reused PETS Fixes
16
+
17
+ The completed PETS run was inspected from `DanielNguyen121/DisEF-PETS`. Generic fixes considered for reuse: import-safe CLI, deterministic filenames/seeds, validated atomic state, manifest-driven resume, bounded supervisor, immutable HF chunks, restore/bootstrap scripts, CPU-only monitor, PIL validation, and zombie-aware status. Dataset-specific PETS paths/classes are not reused.
18
+
19
+ ## 5. New Caltech101 Fixes
20
+
21
+ - Added an explicit Caltech101 source-folder mapping: output labels use SPC's 100-class names (face, leopard, motorbike, airplane, ...), while torchvision source folders retain Faces, Leopards, Motorbikes, and airplanes.
22
+ - Added official DISEF split_coop.csv as SPC/DataSplit/split_coop_Caltech101.csv.
23
+ - Added a deterministic batched Caltech101 generator with per-task seeds, cross-sample captions, validated filenames, and 96GB benchmark mode.
24
+ - Reused generic PETS atomic state/manifest/HF chunk/supervisor/restore logic after adapting repository names and state filenames.
25
+
26
+ ## 6. Molab Environment
27
+
28
+ - Host: `5c7b4d06-5fc8-4c4d-8b49-7548e4814a1e8-sq6mx`
29
+ - Working directory: `/marimo`
30
+ - Python: 3.13.11
31
+ - GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition, 97,887 MiB VRAM, driver 595.71.05, CUDA 13.2
32
+ - Base packages: torch 2.11.0, torchvision 0.26.0, transformers 5.14.1, huggingface-hub 1.24.0, Pillow 12.3.0, pandas 3.0.3; diffusers/accelerate/xformers initially missing
33
+ - HF: `DanielNguyen121/DisEF-Caltech101` created successfully.
34
+
35
+ ## 7. Caltech101 Dataset Identity
36
+
37
+ Pending Stage B smoke test. SPC/official DISEF use the Caltech101 torchvision dataset with the official DISEF `split_coop.csv`, excluding `Faces_easy`; the resulting class space is expected to be 100 classes.
38
+
39
+ ## 8. Class List / Few-Shot / Target Count
40
+
41
+ Stage B PASS. Exact output class names are the SPC/DISEF 100-class list in util_data.py: face, leopard, motorbike, accordion, airplane, anchor, ant, barrel, bass, beaver, binocular, bonsai, brain, brontosaurus, buddha, butterfly, camera, cannon, car_side, ceiling_fan, cellphone, chair, chandelier, cougar_body, cougar_face, crab, crayfish, crocodile, crocodile_head, cup, dalmatian, dollar_bill, dolphin, dragonfly, electric_guitar, elephant, emu, euphonium, ewer, ferry, flamingo, flamingo_head, garfield, gerenuk, gramophone, grand_piano, hawksbill, headphone, hedgehog, helicopter, ibis, inline_skate, joshua_tree, kangaroo, ketch, lamp, laptop, llama, lobster, lotus, mandolin, mayfly, menorah, metronome, minaret, nautilus, octopus, okapi, pagoda, panda, pigeon, pizza, platypus, pyramid, revolver, rhino, rooster, saxophone, schooner, scissors, scorpion, sea_horse, snoopy, soccer_ball, stapler, starfish, stegosaurus, stop_sign, strawberry, sunflower, tick, trilobite, umbrella, watch, water_lilly, wheelchair, wild_cat, windsor_chair, wrench, yin_yang.
42
+
43
+ The run uses 16 deterministic real training sources per class (1,600 total), split seed 1, global seed 20261005, and 64 accepted/validated synthetic images per class. Formula: 100 classes x 64 synthetic/class = 6,400 expected images. The 16-shot source pool means the nominal average is 4 synthetic images per source, but anchors are sampled repeatedly.
44
+
45
+ ## 9. Audit Table
46
+
47
+ | Component | File | Role | Status | Problem | Fix |
48
+ |---|---|---|---|---|---|
49
+ | Generation entrypoint | `DisEF/disef_gen.py` | synthesis | broken | Colab syntax, hard-coded FGVC paths | replace with Caltech CLI |
50
+ | Caltech split | `data_train.py` / official `split_coop.csv` | dataset split | suspicious | `Faces_easy` label adjustment is implicit | explicit, tested mapping |
51
+ | Diffusion | `DisEF/disef_gen.py` | synthesis | incompatible | hard-coded SD2.1/768/FGVC | validated DISEF PETS-compatible configuration |
52
+ | Captions | `DisEF/disef_gen.py` | semantic conditioning | suspicious | prompt pairing and caching not resume-safe | deterministic distinct same-class captions |
53
+ | State/resume | absent | persistence | missing | no atomic state/manifest | reuse validated PETS architecture |
54
+ | Remote recovery | absent | sandbox durability | missing | no Caltech HF progress | implement against DisEF-Caltech101 |
55
+ | Optimization | absent | throughput | missing | no 96GB benchmark | benchmark operational batch sizes |
56
+
57
+ ## 10. Validation and Optimization
58
+
59
+ Stages A–G and the 96GB throughput benchmark: PASS. Recovery bootstrap was uploaded and verified in DanielNguyen121/DisEF-Caltech101 before detached full generation.
60
+
61
+ ## 11. Generation Configuration
62
+
63
+ Frozen after model smoke and optimization benchmark; fingerprint will be recorded before launch.
64
+
65
+ ## 12. Recovery / Monitoring / Packaging
66
+
67
+ Caltech-specific detached scripts and the lightweight read-only monitor are being prepared. All recovery data will use DanielNguyen121/DisEF-Caltech101.
68
+
69
+ ## 13. Final Result
70
+
71
+ Pending.
72
+
73
+ ## 14. Smoke and Recovery Validation
74
+
75
+ - Stage A static CLI/import/compile checks: PASS.
76
+ - Stage B dataset smoke: PASS; official DISEF CoOp split has 4,128 train, 1,649 validation, and 2,465 test records across 100 target classes.
77
+ - Stage C model smoke: PASS; Realistic Vision, LLaVA 1.5 7B, and Qwen3-VL Embedding loaded on CUDA in bfloat16.
78
+ - Stage D single image: PASS; valid 512x512 JPEG saved under the face class, reopened with PIL, and remotely chunked.
79
+ - Stage E mini run: PASS; two classes and six images, class directories/metadata/deterministic names validated.
80
+ - Stage F local resume: PASS; an intentional stop after 3/8 images resumed to 8/8; the three pre-existing image hashes were unchanged and five only missing images were added.
81
+ - Stage G remote recovery: PASS; restored 8/8 images from five immutable chunks in a clean directory, every JPEG validated, and source/restored SHA256 maps matched exactly.
82
+ - Important bug fixed: max-new did not count successful saves or cap the final batch; it now supports exact interruption testing. Zombie PIDs are treated as stale locks.
83
+
84
+ ## 15. Hardware Optimization Benchmark
85
+
86
+ | Batch | Peak VRAM MiB | Images/min | Status |
87
+ |---:|---:|---:|---|
88
+ | 1 | 7,338.9 | 266.1 | PASS |
89
+ | 2 | 7,754.1 | 430.3 | PASS |
90
+ | 4 | 8,779.4 | 501.0 | PASS |
91
+ | 8 | 10,830.1 | 487.4 | PASS |
92
+ | 16 | 14,931.3 | 503.2 | PASS |
93
+ | 32 | 23,133.6 | 481.4 | PASS |
94
+
95
+ ## 16. Selected Operational Configuration
96
+
97
+ Selection: batch size 16, bfloat16, PyTorch SDPA. It achieved the highest measured short-run throughput (503.2 images/min) while leaving substantial transient-memory headroom. Batch 32 used more VRAM but was slower; no scientific parameters were changed.
98
+
99
+ ## 17. Full Run Started
100
+
101
+ Detached supervisor PID 6903 started generator PID 6945 on 2026-10-05T07:42Z. The run first builds/reuses the LLaVA caption cache for 1,568 selected support images, then generates 6,400 images in batch-16 DISEF diffusion/filtering steps. State, manifest, heartbeat, logs, figures, and immutable HF chunks are under /marimo/runs/disef_caltech101; remote progress is in DanielNguyen121/DisEF-Caltech101 under the run-specific progress prefix.
caltech101_disef_artifacts.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7441a7f07af62c11af0d2bab7c903f7f86941a2338a0ab1c2e8a8f3c0136e186
3
+ size 4305341
caltech101_disef_checksums.txt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ d8647f2d96bad27d5da67a6703ed49d861d90da22c3d49b31f78e53734fd0674 caltech101_disef_images.zip
2
+ 7441a7f07af62c11af0d2bab7c903f7f86941a2338a0ab1c2e8a8f3c0136e186 caltech101_disef_artifacts.zip
3
+ ae3177e60e924a8eab3f6827408030ecbb9109860fd018b9a733315a215ac3cf caltech101_disef_code.zip
caltech101_disef_code.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ae3177e60e924a8eab3f6827408030ecbb9109860fd018b9a733315a215ac3cf
3
+ size 141498
caltech101_disef_manifest.csv ADDED
The diff for this file is too large to render. See raw diff
 
generation_config.json ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "attention_backend": "diffusers default SDPA",
3
+ "batch_size": 16,
4
+ "caption_model": "llava-hf/llava-1.5-7b-hf",
5
+ "caption_temperature": 0.2,
6
+ "classes": [
7
+ "face",
8
+ "leopard",
9
+ "motorbike",
10
+ "accordion",
11
+ "airplane",
12
+ "anchor",
13
+ "ant",
14
+ "barrel",
15
+ "bass",
16
+ "beaver",
17
+ "binocular",
18
+ "bonsai",
19
+ "brain",
20
+ "brontosaurus",
21
+ "buddha",
22
+ "butterfly",
23
+ "camera",
24
+ "cannon",
25
+ "car_side",
26
+ "ceiling_fan",
27
+ "cellphone",
28
+ "chair",
29
+ "chandelier",
30
+ "cougar_body",
31
+ "cougar_face",
32
+ "crab",
33
+ "crayfish",
34
+ "crocodile",
35
+ "crocodile_head",
36
+ "cup",
37
+ "dalmatian",
38
+ "dollar_bill",
39
+ "dolphin",
40
+ "dragonfly",
41
+ "electric_guitar",
42
+ "elephant",
43
+ "emu",
44
+ "euphonium",
45
+ "ewer",
46
+ "ferry",
47
+ "flamingo",
48
+ "flamingo_head",
49
+ "garfield",
50
+ "gerenuk",
51
+ "gramophone",
52
+ "grand_piano",
53
+ "hawksbill",
54
+ "headphone",
55
+ "hedgehog",
56
+ "helicopter",
57
+ "ibis",
58
+ "inline_skate",
59
+ "joshua_tree",
60
+ "kangaroo",
61
+ "ketch",
62
+ "lamp",
63
+ "laptop",
64
+ "llama",
65
+ "lobster",
66
+ "lotus",
67
+ "mandolin",
68
+ "mayfly",
69
+ "menorah",
70
+ "metronome",
71
+ "minaret",
72
+ "nautilus",
73
+ "octopus",
74
+ "okapi",
75
+ "pagoda",
76
+ "panda",
77
+ "pigeon",
78
+ "pizza",
79
+ "platypus",
80
+ "pyramid",
81
+ "revolver",
82
+ "rhino",
83
+ "rooster",
84
+ "saxophone",
85
+ "schooner",
86
+ "scissors",
87
+ "scorpion",
88
+ "sea_horse",
89
+ "snoopy",
90
+ "soccer_ball",
91
+ "stapler",
92
+ "starfish",
93
+ "stegosaurus",
94
+ "stop_sign",
95
+ "strawberry",
96
+ "sunflower",
97
+ "tick",
98
+ "trilobite",
99
+ "umbrella",
100
+ "watch",
101
+ "water_lilly",
102
+ "wheelchair",
103
+ "wild_cat",
104
+ "windsor_chair",
105
+ "wrench",
106
+ "yin_yang"
107
+ ],
108
+ "config_fingerprint": "262ae8436d11af1a1d6a0bda4945b2c58ae9e156b7ed39dd62a4887e98448e30",
109
+ "created_at": "2026-10-05T07:37:21.636479+00:00",
110
+ "cross_sample_caption": true,
111
+ "dataset": "Caltech101",
112
+ "dataset_version": "torchvision Caltech101 / official DISEF CoOp split",
113
+ "diffusion_model": "SG161222/Realistic_Vision_V2.0",
114
+ "dtype": "bfloat16",
115
+ "expected_total": 6400,
116
+ "fewshot": 16,
117
+ "filter_backend": "Qwen3-VL-Embedding-2B (SPC adaptation)",
118
+ "filter_model": "Qwen/Qwen3-VL-Embedding-2B",
119
+ "global_seed": 20261005,
120
+ "guidance_scale": 8.0,
121
+ "images_per_class": 64,
122
+ "jpeg_quality": 95,
123
+ "model_revisions": {
124
+ "caption": "b234b804b114d9e37bb655e11cbbb5f5e971b7a9",
125
+ "diffusion": "aaa388343c7f35f6416eea9d29574a4d1ad59817",
126
+ "filter": "9f2f7e710d6d81056aa5c0a4f04764fec6bb7bda"
127
+ },
128
+ "num_inference_steps": 20,
129
+ "resolution": 512,
130
+ "run_id": "caltech101_disef_20261005T073721Z_262ae8436d11",
131
+ "scheduler": "DPMSolverMultistepScheduler / DPM-Solver++",
132
+ "selected_sources_sha256": "2485b4a8b73946ebe11201185e99793660870dca862c05c73df902a1ca6f3942",
133
+ "spc_commit": "96834b22d5e59c852e53d713d434cbabb2f0d165",
134
+ "split_seed": 1,
135
+ "split_sha256": "8d52725a2e4273ffbc03a1f6d923e5c01baf1595f9572b4f14f63046ada21110",
136
+ "starting_step": 15,
137
+ "strength": 0.25
138
+ }