[feat] add H0-mini model

[afilt]

This PR integrates:

• The implementation of H0-mini. The model is available on Hugging Face.
• The implementation of UNI2-h. The model is available on Hugging Face.
• The implementation of GPFM (to be discussed). The model is available on Github.

nox pipeline was run successfully on my local workspace.

Note:

• H0-mini is a ViT-B model with 768 output dimensions
• UNI2-h is a modified ViT-H model with 1,024 output dimensions

Warning:

• GPFM implementation is not stable as it relies on querying the weights path to disk (/path/to/gpfm_checkpoint/GPFM.pt). We converted the checkpoint from the original Github into a timm compatible state dict. Let me know if you want to integrate or not this model. If so, what would be your recommendations ?

To reproduce our results, you first need to be granted access to bioptimus/H0-mini on HF. Then, simply run bash train_h0_mini.sh:

EXTRACTORS=(
    "pathology/bioptimus_h0_mini"
)

######################################################################

TASKS=(
    "bach"
    "mhist"
    "crc"
    "patch_camelyon"
)

for task in ${TASKS[*]}
do
    for extractor in ${EXTRACTORS[*]}
    do
        HF_TOKEN="****" \
        OUTPUT_ROOT="/home/owkin/project/eva_benchmarks/" \
        DATA_ROOT="/home/owkin/project/eva_data/${task}" \
        EMBEDDINGS_ROOT="/home/owkin/project/eva_embeddings/" \
        MODEL_NAME=${extractor} \
        N_DATA_WORKERS=16 \
        NORMALIZE_MEAN="[0.707223, 0.578729, 0.703617]" \
        NORMALIZE_STD="[0.211883, 0.230117, 0.177517]" \
        IN_FEATURES=768 \
        eva predict_fit --config configs/vision/pathology/offline/classification/${task}.yaml
    done    
done

######################################################################

TASKS=(
    "camelyon16_small"
    "panda_small"
)

for task in ${TASKS[*]}
do
    for extractor in ${EXTRACTORS[*]}
    do
        HF_TOKEN="****" \
        OUTPUT_ROOT="/home/owkin/project/eva_benchmarks/" \
        DATA_ROOT="/home/owkin/project/eva_data/${task}" \
        EMBEDDINGS_ROOT="/home/owkin/project/eva_embeddings/${extractor}/${task}" \
        MODEL_NAME=${extractor} \
        N_DATA_WORKERS=8 \
        NORMALIZE_MEAN="[0.707223, 0.578729, 0.703617]" \
        NORMALIZE_STD="[0.211883, 0.230117, 0.177517]" \
        IN_FEATURES=768 \
        eva predict_fit --config configs/vision/pathology/offline/classification/${task}.yaml
    done    
done

#####################################################################

TASKS=(
    "consep"
    "monusac"
)

for task in ${TASKS[*]}
do
    for extractor in ${EXTRACTORS[*]}
    do
        HF_TOKEN="****" \
        OUTPUT_ROOT="/home/owkin/project/eva_benchmarks/" \
        DATA_ROOT="/home/owkin/project/eva_data/${task}" \
        MODEL_NAME=${extractor} \
        N_DATA_WORKERS=16 \
        IN_FEATURES=768 \
        NORMALIZE_MEAN="[0.707223, 0.578729, 0.703617]" \
        NORMALIZE_STD="[0.211883, 0.230117, 0.177517]" \
        eva fit --config configs/vision/pathology/online/segmentation/${task}.yaml
    done
done

cc @nkaenzig