tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
66f4e01 verified | { | |
| "mic": "0.3", | |
| "name": "SAM2 promptable image segmentation", | |
| "usage": "Sam2Model + Sam2Processor, point-prompted segmentation of synthetic images, one fine-tuning step with a user-side BCE mask loss", | |
| "architecture": "Sam2Model", | |
| "checkpoints": { | |
| "facebook/sam2.1-hiera-tiny": { | |
| "fixture": "huggingface-mic/sam2.1-hiera-tiny-mic" | |
| } | |
| }, | |
| "fixtures": { | |
| "huggingface-mic/sam2.1-hiera-tiny-mic": { | |
| "built_from": "facebook/sam2.1-hiera-tiny", | |
| "covers": [ | |
| "facebook/sam2.1-hiera-tiny" | |
| ], | |
| "build": { | |
| "seed": 17, | |
| "model_class": "Sam2Model", | |
| "config_overrides": { | |
| "fields": { | |
| "vision_config.backbone_config.hidden_size": 32, | |
| "vision_config.backbone_config.embed_dim_per_stage": [ | |
| 32, | |
| 64, | |
| 128, | |
| 256 | |
| ], | |
| "vision_config.backbone_config.blocks_per_stage": [ | |
| 1, | |
| 1, | |
| 2, | |
| 1 | |
| ], | |
| "vision_config.backbone_config.global_attention_blocks": [ | |
| 3 | |
| ], | |
| "vision_config.backbone_channel_list": [ | |
| 256, | |
| 128, | |
| 64, | |
| 32 | |
| ], | |
| "vision_config.fpn_hidden_size": 128, | |
| "mask_decoder_config.hidden_size": 128, | |
| "mask_decoder_config.mlp_dim": 256, | |
| "mask_decoder_config.iou_head_hidden_dim": 128, | |
| "prompt_encoder_config.hidden_size": 128 | |
| }, | |
| "reason": "The generic shrink leaves SAM2 inconsistent: Hiera hidden_size=64 vs embed_dim_per_stage[0]=96 (stage-0 LayerNorm fails), and the mask decoder/prompt encoder at 64 vs the FPN at 256. Shrink the Hiera stages together (32..256, backbone_channel_list reversed; one global block mid-stage since global blocks have window_size 0 and fail at query-pool boundaries) and use one width of 128 for the FPN, mask decoder and prompt encoder. The checkpoint's video-memory fields are unused by Sam2Model and left as shrunk." | |
| }, | |
| "accept_insensitive": { | |
| "pixel_values:swap_channels": "Reviewed 2026-10-27: the 1.9M-parameter Hiera fixture, overfit on two images, separates them by object position (roll_batch and flip move outputs by 1.7) but not by colour; channel swap moves it <1e-5 for random and overfit weights. The real checkpoint moves pred_masks by 2.6-23 under the same perturbation, so the checkpoint target keeps colour coverage.", | |
| "pixel_values:halve": "Reviewed 2026-10-27: same fixture, intensity halving moves outputs <1e-6 for random and overfit weights (position, not brightness, is what two training images teach a fixture this size). The real checkpoint moves pred_masks by 2.6-23, so the checkpoint target keeps intensity coverage." | |
| } | |
| } | |
| } | |
| }, | |
| "integrations": { | |
| "transformers": { | |
| "entry_point": "integrations/transformers/usage.md", | |
| "format": "runnable-markdown", | |
| "requires": [ | |
| "transformers>=5", | |
| "torch", | |
| "Pillow", | |
| "numpy" | |
| ], | |
| "targets": [ | |
| "fixtures", | |
| "checkpoints" | |
| ], | |
| "variants": { | |
| "default": {}, | |
| "eager": { | |
| "attn_implementation": "eager" | |
| }, | |
| "bfloat16": { | |
| "dtype": "bfloat16" | |
| } | |
| }, | |
| "training": true, | |
| "resources": { | |
| "fixture": { | |
| "memory_gb": 3, | |
| "download_mb": 10 | |
| }, | |
| "facebook/sam2.1-hiera-tiny": { | |
| "memory_gb": 6, | |
| "download_mb": 160 | |
| } | |
| } | |
| } | |
| }, | |
| "outputs": { | |
| "runtime": { | |
| "compare": "exact" | |
| }, | |
| "pred_masks": { | |
| "compare": "allclose", | |
| "atol": 0.0001, | |
| "rtol": 0.0001 | |
| }, | |
| "iou_scores": { | |
| "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" | |
| } | |
| } | |
| } | |