tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
fb35353 verified | { | |
| "mic": "0.3", | |
| "name": "Grounding DINO zero-shot object detection", | |
| "usage": "AutoProcessor + AutoModelForZeroShotObjectDetection, batched inference with Hungarian-matcher training step", | |
| "architecture": "GroundingDinoForObjectDetection", | |
| "checkpoints": { | |
| "IDEA-Research/grounding-dino-tiny": { | |
| "fixture": "huggingface-mic/grounding-dino-tiny-mic" | |
| } | |
| }, | |
| "fixtures": { | |
| "huggingface-mic/grounding-dino-tiny-mic": { | |
| "built_from": "IDEA-Research/grounding-dino-tiny", | |
| "covers": [ | |
| "IDEA-Research/grounding-dino-tiny" | |
| ], | |
| "build": { | |
| "seed": 17, | |
| "config_overrides": { | |
| "fields": { | |
| "backbone_config.num_heads": [ | |
| 2, | |
| 4, | |
| 8, | |
| 16 | |
| ] | |
| }, | |
| "reason": "The generic rule shrinks the Swin backbone's embed_dim to 64 (stage widths 64/128/256/512) but keeps num_heads [3,6,12,24]; 64 channels over 3 heads gives a 21-wide head and the window attention reshape fails. [2,4,8,16] keeps the per-stage head doubling with a 32-wide head." | |
| } | |
| } | |
| } | |
| }, | |
| "integrations": { | |
| "transformers": { | |
| "entry_point": "integrations/transformers/usage.md", | |
| "format": "runnable-markdown", | |
| "requires": [ | |
| "transformers>=5", | |
| "torch" | |
| ], | |
| "targets": [ | |
| "fixtures", | |
| "checkpoints" | |
| ], | |
| "variants": { | |
| "default": {}, | |
| "eager": { | |
| "attn_implementation": "eager" | |
| }, | |
| "bfloat16": { | |
| "dtype": "bfloat16" | |
| } | |
| }, | |
| "training": true, | |
| "resources": { | |
| "fixture": { | |
| "memory_gb": 1, | |
| "download_mb": 50 | |
| }, | |
| "IDEA-Research/grounding-dino-tiny": { | |
| "memory_gb": 4, | |
| "download_mb": 770, | |
| "gpu_memory_gb": 2, | |
| "variants": { | |
| "bfloat16": { | |
| "memory_gb": 2, | |
| "gpu_memory_gb": 1 | |
| } | |
| }, | |
| "devices": [ | |
| "cuda" | |
| ] | |
| } | |
| } | |
| } | |
| }, | |
| "outputs": { | |
| "runtime": { | |
| "compare": "exact" | |
| }, | |
| "class_scores": { | |
| "compare": "allclose", | |
| "atol": 0.0001, | |
| "rtol": 0.0001 | |
| }, | |
| "box_preds": { | |
| "compare": "allclose", | |
| "atol": 0.0001, | |
| "rtol": 0.0001 | |
| }, | |
| "training.loss": { | |
| "compare": "allclose", | |
| "atol": 0.0001, | |
| "rtol": 0.0001 | |
| }, | |
| "training.gradients": { | |
| "compare": "all_true" | |
| }, | |
| "training.updated": { | |
| "compare": "true" | |
| } | |
| } | |
| } | |