Image Feature Extraction
Transformers
Safetensors
English
keural_vision
vision
vision-encoder
image-text
contrastive-learning
knowledge-distillation
adaptive-tokenization
Eval Results (legacy)
Instructions to use mkd-hika/keural-vision-encoder-mid with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use mkd-hika/keural-vision-encoder-mid with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-feature-extraction", model="mkd-hika/keural-vision-encoder-mid")# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("mkd-hika/keural-vision-encoder-mid", device_map="auto") - Notebooks
- Google Colab
- Kaggle
| { | |
| "architectures": [ | |
| "KeuralVisionModel" | |
| ], | |
| "default_token_budget": 512, | |
| "depth": 24, | |
| "dtype": "float32", | |
| "embed_dim": 768, | |
| "keural_config": { | |
| "data": { | |
| "captions_path": "/home/work/keural-model/vision-encoder-dataset/indexes/datacomp_train.tsv", | |
| "dataset_name": "cc3m", | |
| "extra_sources": [], | |
| "image_buckets": [ | |
| [ | |
| 384, | |
| 384 | |
| ], | |
| [ | |
| 448, | |
| 448 | |
| ] | |
| ], | |
| "images_root": "/home/work/keural-model/vision-encoder-dataset/datacomp_medium", | |
| "max_caption_length": 77, | |
| "num_workers": 6, | |
| "pixel_mean": [ | |
| 0.485, | |
| 0.456, | |
| 0.406 | |
| ], | |
| "pixel_std": [ | |
| 0.229, | |
| 0.224, | |
| 0.225 | |
| ], | |
| "text_encoder_name": "openai/clip-vit-large-patch14" | |
| }, | |
| "distillation": { | |
| "enabled": true, | |
| "internvit_embed_dim": 1024, | |
| "internvit_model": "OpenGVLab/InternViT-300M-448px", | |
| "lambda_internvit": 0.5, | |
| "lambda_siglip": 0.5, | |
| "siglip_embed_dim": 1152, | |
| "siglip_model": "google/siglip-so400m-patch14-384" | |
| }, | |
| "loss": { | |
| "lambda_distill": 0.5, | |
| "lambda_hct": 1.0, | |
| "lambda_sal": 0.5, | |
| "learnable_bias_init": -10.0, | |
| "learnable_temperature_init": 2.3, | |
| "primary": "sigmoid_siglip", | |
| "saliency_beta": 0.005, | |
| "saliency_lambda_smooth": 4.0 | |
| }, | |
| "name": "keural-mid", | |
| "saliency": { | |
| "hidden_channels": 192 | |
| }, | |
| "spatial": { | |
| "depth": 24, | |
| "embed_dim": 768, | |
| "head_dim": 64, | |
| "mlp_ratio": 4, | |
| "norm_eps": 1e-06, | |
| "num_heads": 12, | |
| "rope_base": 10000.0 | |
| }, | |
| "stem": { | |
| "activation": "gelu", | |
| "channels": [ | |
| 128, | |
| 256, | |
| 512, | |
| 768 | |
| ], | |
| "kernel_size": 3 | |
| }, | |
| "tokenizer": { | |
| "budget_floors": { | |
| "detail": 0, | |
| "global": 8, | |
| "region": 16 | |
| }, | |
| "budget_split": { | |
| "detail": 0.7, | |
| "global": 0.05, | |
| "region": 0.25 | |
| }, | |
| "default_token_budget": 512, | |
| "gumbel_temperature_end": 0.05, | |
| "gumbel_temperature_start": 1.0, | |
| "max_token_budget": 2048, | |
| "nms_radius": { | |
| "detail": 1, | |
| "region": 3 | |
| }, | |
| "saliency_smooth_kernel": { | |
| "detail": 3, | |
| "region": 7 | |
| }, | |
| "temperature_anneal_steps": 10000, | |
| "warmup_steps_fixed_grid": 4000 | |
| }, | |
| "training": { | |
| "batch_size_per_device": 768, | |
| "betas": [ | |
| 0.9, | |
| 0.95 | |
| ], | |
| "gradient_accumulation": 2, | |
| "gradient_checkpointing": true, | |
| "log_every": 100, | |
| "lr": 0.0005, | |
| "lr_schedule": "cosine_with_warmup", | |
| "max_steps": 18000, | |
| "optimizer": "adamw", | |
| "precision": "bfloat16", | |
| "save_every": 1000, | |
| "seed": 42, | |
| "warmup_steps": 2000, | |
| "weight_decay": 0.05 | |
| } | |
| }, | |
| "max_token_budget": 2048, | |
| "model_type": "keural_vision", | |
| "num_heads": 12, | |
| "transformers_version": "4.57.6" | |
| } | |