pshishodia/imagenet-1k-256

The dataset contains imagenet-1k resized to 256x256, along with corresponding captions generated by gemini-2.0-flash, and discrete tokens generated by NVIDIA Cosmos tokenizer.

Specifically the following steps are done in order

1. Start with hf:timm/imagenet-1k-wds
2. Generate captions for the images with gemini-2.0-flash with the following config:.

_SYSTEM_PROMPT = '''Caption the ImageNet-21K in a single, detailed paragraph, without introductory phrases like "This image showcases", "Here's a detailed description".'''
def get_messages(img, wordnet_lemmas, system_prompt):
    messages = [
        {
            "role": "system",
            "content": [
                {"type": "text", "text": system_prompt},
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/jpeg;base64,{encode_image_from_pil(img)}"
                    },
                },
                {
                    "type": "text",
                    "text": f"ImageNet-21K Wordnet Labels: {wordnet_lemmas}.",
                },
            ],
        }
    ]
    return messages

3. Center Crop and Resize images to 256x256. Discard images with min(w, h) < 256. We end up discarding X train and Y validation samples. TODO: Add x, y
4. Generate Discrete tokens using CosmosDI8x8 tokenizer with cosmos-predict1 inference code