Implementation:Norrrrrrr lyn WAInjectBench OpenCLIP Init

Knowledge Sources	WAInjectBench OpenCLIP
Domains	Computer_Vision, Representation_Learning
Last Updated	2026-02-14 16:00 GMT

Overview

Concrete tool for loading the ViT-B-32 CLIP model with LAION-2B weights for image embedding, provided by the open_clip library as used in the WAInjectBench image embedding trainer.

Description

The image embedding training script uses open_clip.create_model_and_transforms to load a ViT-B-32 model with laion2b_s34b_b79k pre-trained weights. The function returns a triplet: (model, _, preprocess), where the model produces 512-dim embeddings and preprocess is the required image transform. The model is moved to the target device (CUDA or CPU).

Usage

Initialize once before processing multiple JSONL training files. The model and preprocess transform are shared across all image embedding extraction calls.

Code Reference

Source Location

• Repository: WAInjectBench
• File: train/embedding-i.py (L79-83)

Signature

model, _, preprocess = open_clip.create_model_and_transforms(
    "ViT-B-32",
    pretrained="laion2b_s34b_b79k"
)
model = model.to(device)

Import

import open_clip
import torch

I/O Contract

Inputs

Name	Type	Required	Description
model_name	str	Yes	Model architecture (hardcoded: "ViT-B-32")
pretrained	str	Yes	Pre-trained weights identifier (hardcoded: "laion2b_s34b_b79k")
device	str	No	Target device (default "cuda", falls back to "cpu")

Outputs

Name	Type	Description
model	open_clip.CLIP	Loaded CLIP model producing 512-dim image embeddings
preprocess	torchvision.transforms.Compose	Image preprocessing pipeline for the model

Usage Examples

Initializing the Image Embedder

import open_clip
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model, _, preprocess = open_clip.create_model_and_transforms(
    "ViT-B-32",
    pretrained="laion2b_s34b_b79k"
)
model = model.to(device)

# Verify: model.visual.output_dim == 512
print(f"Embedding dim: {model.visual.output_dim}")

Related Pages

Implements Principle

• Principle:Norrrrrrr_lyn_WAInjectBench_Image_Embedding_Initialization

Requires Environment

• Environment:Norrrrrrr_lyn_WAInjectBench_Conda_Python_39_CUDA_Environment