README.md

metadata

base_model:
  - openai/clip-vit-base-patch32
datasets:
  - svhn
metrics:
  - accuracy

Model Card

Model Details

• Architecture: ViT-Base with patch size 32
• Training Data: SVHN

Training Details

Adam Optimizer with a constant learning rate 1e-5 for 4000 steps training (batch_size=32). Only the vision encoder is fine-tuned.

Evaluation Results

• pre-trained: 0.23536789417266846
• fine-tuned: 0.9714505076408386

Usage

load vision model

from transformers import CLIPVisionModel

vision_model = CLIPVisionModel.from_pretrained('tanganke/clip-vit-base-patch32_svhn')

substitute the vision encoder of clip

from transformers import CLIPModel

clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
clip_model.vision_model.load_state_dict(vision_model.vision_model.state_dict())