PubMedCLIP + GPT-2 Dermatology Captioner

A dermatology image captioning model combining PubMedCLIP vision encoder with gpt2-medium language model. Trained on dermatological images for generating clinical descriptions of skin lesions.

Architecture: PubMedCLIP (ViT-B/32) → learnable prefix → GPT-2 (gpt2-medium). Trained in two stages: Stage A (META) for generalization and Stage B (SkinCAP) for style/terminology.

Metrics

Stage A (META)
val_loss=1.1222 • PPL=3.07
BLEU=36.6 • ROUGE-L=0.521 • CIDEr-D=0.10 • CLIP=34.7 • BERT_F1=0.526

Stage B (SKINCAP)
val_loss=1.1997 • PPL=3.32
BLEU=9.3 • ROUGE-L=0.267 • CIDEr-D=0.12 • CLIP=40.5 • BERT_F1=0.348

Inference

Minimal example uses inference_min.py included in this repo.
Requires: pip install torch transformers open_clip_torch pillow huggingface_hub

from huggingface_hub import snapshot_download
from inference_min import load_model, generate

# 1) download repo snapshot
repo_dir = snapshot_download("moxeeeem/pubmedclip-gpt2-captioner", allow_patterns=["*.pt","*.json","inference_min.py"])

# 2) load model from saved config/weights
model = load_model(repo_dir)  # builds CLIP backend + GPT-2 + prefix projector

# 3) run generation
img_paths = ["/path/to/derma_image.jpg"]  # local test images
caps = generate(model, img_paths, prompt="Describe the skin lesion concisely (morphology, color, scale, border, location) in one sentence.Conclude with the most likely diagnosis (1\u20133 words).")
for c in caps:
    print(c)

Files

File	Size	Check
`best_stageA.pt`	2 GB	sha256[:12]=4c3f773c92f9
`best_stageB.pt`	2 GB	sha256[:12]=2b1b7e6aaa8f
`final_captioner_gpt2-medium_TimmModel.json`	899 B	sha256[:12]=40f3fb250eab
`final_captioner_gpt2-medium_TimmModel.pt`	2 GB	sha256[:12]=799a4d1b831d
`loss_biomedclip_pubmedbert.png`	112 KB	sha256[:12]=92c3b8dfdf7f

Details

Vision Encoder: PubMedCLIP (ViT-B/32)
Language Model: GPT-2 (gpt2-medium)
CLIP weights: hf-hub:microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224
Prefix tokens: 32
Training prompt: Describe the skin lesion concisely (morphology, color, scale, border, location) in one sentence.Conclude with the most likely diagnosis (1–3 words).

Model Type Detection

Detected as: pubmedclip
Repository: moxeeeem/pubmedclip-gpt2-captioner

Auto-generated on 2025-08-30 06:06 UTC.