Try out the demo!

Model Description

This model was trained to solve Raven's Progressive Matrices. It is based on an early checkpoint of our upcoming vision-language foundation model.

We use the RAVEN dataset of procedurally generated Raven puzzles to train the system. On the validation set, the model would reach 91% accuracy.

Code snippet

The model has been specifically fine-tuned for solving Raven puzzles and we cannot guarantee that it will behave accurately outside of this use-case with no proper adaptation.

The code snippet how to do batch inference with the model. A lot of the input preparation will be encapsulated once we integrate the model into HF Transformers.

import torch
import requests

from io import BytesIO
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor

from transformers.image_utils import to_numpy_array, PILImageResampling, ChannelDimension
from transformers.image_transforms import resize, to_channel_dimension_format


DEVICE = torch.device("cuda")
PROCESSOR = AutoProcessor.from_pretrained(
    "HuggingFaceM4/tr_272_bis_opt_step_15000_merge",
    token=API_TOKEN,
)
MODEL = AutoModelForCausalLM.from_pretrained(
    "HuggingFaceM4/tr_272_bis_opt_step_15000_merge",
    token=API_TOKEN,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to(DEVICE)
image_seq_len = MODEL.config.perceiver_config.resampler_n_latents
BOS_TOKEN = PROCESSOR.tokenizer.bos_token
BAD_WORDS_IDS = PROCESSOR.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids


def convert_to_rgb(image):
    # `image.convert("RGB")` would only work for .jpg images, as it creates a wrong background
    # for transparent images. The call to `alpha_composite` handles this case
    if image.mode == "RGB":
        return image

    image_rgba = image.convert("RGBA")
    background = Image.new("RGBA", image_rgba.size, (255, 255, 255))
    alpha_composite = Image.alpha_composite(background, image_rgba)
    alpha_composite = alpha_composite.convert("RGB")
    return alpha_composite


# The processor is the same as the Idefics processor except for the BILINEAR interpolation,
# so this is a hack in order to redefine ONLY the transform method
def custom_transform(x):
    x = convert_to_rgb(x)
    x = to_numpy_array(x)

    height, width = x.shape[:2]
    aspect_ratio = width / height
    if width >= height and width > 980:
        width = 980
        height = int(width / aspect_ratio)
    elif height > width and height > 980:
        height = 980
        width = int(height * aspect_ratio)
    width = max(width, 378)
    height = max(height, 378)

    x = resize(x, (height, width), resample=PILImageResampling.BILINEAR)
    x = PROCESSOR.image_processor.rescale(x, scale=1 / 255)
    x = PROCESSOR.image_processor.normalize(
        x,
        mean=PROCESSOR.image_processor.image_mean,
        std=PROCESSOR.image_processor.image_std
    )
    x = to_channel_dimension_format(x, ChannelDimension.FIRST)
    x = torch.tensor(x)
    return x


# Create text token inputs
image_seq = '<image>' * image_seq_len
inputs = PROCESSOR.tokenizer(
    [
        f"{BOS_TOKEN}User:<fake_token_around_image>{image_seq}<fake_token_around_image>Which figure should complete the logical sequence?<end_of_utterance>\nAssistant:",
        f"{BOS_TOKEN}User:<fake_token_around_image>{image_seq}<fake_token_around_image>Which figure should complete the logical sequence?<end_of_utterance>\nAssistant:",
    ],
    return_tensors="pt",
    add_special_tokens=False,
    padding=True,
)

# Create pixel inputs
raw_images = [
    [your_raven_puzzle_as_a_pil_image1],
    [your_raven_puzzle_as_a_pil_image2],
]
output_images = [
    [PROCESSOR.image_processor(img, transform=custom_transform) for img in img_list]
    for img_list in raw_images
]
total_batch_size = len(output_images)
max_num_images = max([len(img_l) for img_l in output_images])
max_height = max([i.size(2) for img_l in output_images for i in img_l])
max_width = max([i.size(3) for img_l in output_images for i in img_l])
padded_image_tensor = torch.zeros(total_batch_size, max_num_images, 3, max_height, max_width)
padded_pixel_attention_masks = torch.zeros(
    total_batch_size, max_num_images, max_height, max_width, dtype=torch.bool
)
for batch_idx, img_l in enumerate(output_images):
    for img_idx, img in enumerate(img_l):
        im_height, im_width = img.size()[2:]
        padded_image_tensor[batch_idx, img_idx, :, :im_height, :im_width] = img
        padded_pixel_attention_masks[batch_idx, img_idx, :im_height, :im_width] = True

inputs["pixel_values"] = padded_image_tensor
inputs["pixel_attention_mask"] = padded_pixel_attention_masks
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}

generated_ids = MODEL.generate(**inputs, bad_words_ids=BAD_WORDS_IDS, max_new_tokens=10)
generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)

print(generated_texts)

Model Details

• Developed by: Hugging Face
• Model type: Multi-modal model
• Language(s) (NLP): en
• License: see License section
• Parent Models: SigLIP and mistralai/Mistral-7B-v0.1
• Resources for more information:
  • RAVEN dataset: Dataset card

License

The model is built on top of two pre-trained models: SigLIP and mistralai/Mistral-7B-v0.1, which are delivered under an Apache-2.0 license. As such, users should comply with the licenses of these models.

The two pre-trained models are connected to each other with newly initialized parameters that we train. These are not based on any of the two base frozen models forming the composite model. We release the additional weights we trained under an Apache-2.0 license.