VictorSanh
commited on
Commit
•
061ffa4
1
Parent(s):
ccf5303
Update README.md
Browse files
README.md
CHANGED
@@ -8,5 +8,147 @@ tags:
|
|
8 |
- code
|
9 |
---
|
10 |
|
|
|
11 |
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
- code
|
9 |
---
|
10 |
|
11 |
+
**Try out the [demo](https://huggingface.co/spaces/HuggingFaceM4/ai_raven)!**
|
12 |
|
13 |
+
# Model Description
|
14 |
+
|
15 |
+
This model was trained to solve Raven's Progressive Matrices. It is based on an early checkpoint of our upcoming vision-language foundation model.
|
16 |
+
|
17 |
+
We use the [RAVEN](https://huggingface.co/datasets/HuggingFaceM4/RAVEN) dataset of procedurally generated Raven puzzles to train the system. On the validation set, the model would reach 91% accuracy.
|
18 |
+
|
19 |
+
# Code snippet
|
20 |
+
|
21 |
+
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.
|
22 |
+
|
23 |
+
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.
|
24 |
+
|
25 |
+
```python
|
26 |
+
import torch
|
27 |
+
import requests
|
28 |
+
|
29 |
+
from io import BytesIO
|
30 |
+
from PIL import Image
|
31 |
+
from transformers import AutoModelForCausalLM, AutoProcessor
|
32 |
+
|
33 |
+
from transformers.image_utils import to_numpy_array, PILImageResampling, ChannelDimension
|
34 |
+
from transformers.image_transforms import resize, to_channel_dimension_format
|
35 |
+
|
36 |
+
|
37 |
+
DEVICE = torch.device("cuda")
|
38 |
+
PROCESSOR = AutoProcessor.from_pretrained(
|
39 |
+
"HuggingFaceM4/tr_272_bis_opt_step_15000_merge",
|
40 |
+
token=API_TOKEN,
|
41 |
+
)
|
42 |
+
MODEL = AutoModelForCausalLM.from_pretrained(
|
43 |
+
"HuggingFaceM4/tr_272_bis_opt_step_15000_merge",
|
44 |
+
token=API_TOKEN,
|
45 |
+
trust_remote_code=True,
|
46 |
+
torch_dtype=torch.bfloat16,
|
47 |
+
).to(DEVICE)
|
48 |
+
image_seq_len = MODEL.config.perceiver_config.resampler_n_latents
|
49 |
+
BOS_TOKEN = PROCESSOR.tokenizer.bos_token
|
50 |
+
BAD_WORDS_IDS = PROCESSOR.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
|
51 |
+
|
52 |
+
|
53 |
+
def convert_to_rgb(image):
|
54 |
+
# `image.convert("RGB")` would only work for .jpg images, as it creates a wrong background
|
55 |
+
# for transparent images. The call to `alpha_composite` handles this case
|
56 |
+
if image.mode == "RGB":
|
57 |
+
return image
|
58 |
+
|
59 |
+
image_rgba = image.convert("RGBA")
|
60 |
+
background = Image.new("RGBA", image_rgba.size, (255, 255, 255))
|
61 |
+
alpha_composite = Image.alpha_composite(background, image_rgba)
|
62 |
+
alpha_composite = alpha_composite.convert("RGB")
|
63 |
+
return alpha_composite
|
64 |
+
|
65 |
+
|
66 |
+
# The processor is the same as the Idefics processor except for the BILINEAR interpolation,
|
67 |
+
# so this is a hack in order to redefine ONLY the transform method
|
68 |
+
def custom_transform(x):
|
69 |
+
x = convert_to_rgb(x)
|
70 |
+
x = to_numpy_array(x)
|
71 |
+
|
72 |
+
height, width = x.shape[:2]
|
73 |
+
aspect_ratio = width / height
|
74 |
+
if width >= height and width > 980:
|
75 |
+
width = 980
|
76 |
+
height = int(width / aspect_ratio)
|
77 |
+
elif height > width and height > 980:
|
78 |
+
height = 980
|
79 |
+
width = int(height * aspect_ratio)
|
80 |
+
width = max(width, 378)
|
81 |
+
height = max(height, 378)
|
82 |
+
|
83 |
+
x = resize(x, (height, width), resample=PILImageResampling.BILINEAR)
|
84 |
+
x = PROCESSOR.image_processor.rescale(x, scale=1 / 255)
|
85 |
+
x = PROCESSOR.image_processor.normalize(
|
86 |
+
x,
|
87 |
+
mean=PROCESSOR.image_processor.image_mean,
|
88 |
+
std=PROCESSOR.image_processor.image_std
|
89 |
+
)
|
90 |
+
x = to_channel_dimension_format(x, ChannelDimension.FIRST)
|
91 |
+
x = torch.tensor(x)
|
92 |
+
return x
|
93 |
+
|
94 |
+
|
95 |
+
# Create text token inputs
|
96 |
+
image_seq = '<image>' * image_seq_len
|
97 |
+
inputs = PROCESSOR.tokenizer(
|
98 |
+
[
|
99 |
+
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:",
|
100 |
+
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:",
|
101 |
+
],
|
102 |
+
return_tensors="pt",
|
103 |
+
add_special_tokens=False,
|
104 |
+
padding=True,
|
105 |
+
)
|
106 |
+
|
107 |
+
# Create pixel inputs
|
108 |
+
raw_images = [
|
109 |
+
[your_raven_puzzle_as_a_pil_image1],
|
110 |
+
[your_raven_puzzle_as_a_pil_image2],
|
111 |
+
]
|
112 |
+
output_images = [
|
113 |
+
[PROCESSOR.image_processor(img, transform=custom_transform) for img in img_list]
|
114 |
+
for img_list in raw_images
|
115 |
+
]
|
116 |
+
total_batch_size = len(output_images)
|
117 |
+
max_num_images = max([len(img_l) for img_l in output_images])
|
118 |
+
max_height = max([i.size(2) for img_l in output_images for i in img_l])
|
119 |
+
max_width = max([i.size(3) for img_l in output_images for i in img_l])
|
120 |
+
padded_image_tensor = torch.zeros(total_batch_size, max_num_images, 3, max_height, max_width)
|
121 |
+
padded_pixel_attention_masks = torch.zeros(
|
122 |
+
total_batch_size, max_num_images, max_height, max_width, dtype=torch.bool
|
123 |
+
)
|
124 |
+
for batch_idx, img_l in enumerate(output_images):
|
125 |
+
for img_idx, img in enumerate(img_l):
|
126 |
+
im_height, im_width = img.size()[2:]
|
127 |
+
padded_image_tensor[batch_idx, img_idx, :, :im_height, :im_width] = img
|
128 |
+
padded_pixel_attention_masks[batch_idx, img_idx, :im_height, :im_width] = True
|
129 |
+
|
130 |
+
inputs["pixel_values"] = padded_image_tensor
|
131 |
+
inputs["pixel_attention_mask"] = padded_pixel_attention_masks
|
132 |
+
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
|
133 |
+
|
134 |
+
generated_ids = MODEL.generate(**inputs, bad_words_ids=BAD_WORDS_IDS, max_new_tokens=10)
|
135 |
+
generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)
|
136 |
+
|
137 |
+
print(generated_texts)
|
138 |
+
```
|
139 |
+
|
140 |
+
# Model Details
|
141 |
+
|
142 |
+
- **Developed by:** Hugging Face
|
143 |
+
- **Model type:** Multi-modal model
|
144 |
+
- **Language(s) (NLP):** en
|
145 |
+
- **License:** see [License section](#license)
|
146 |
+
- **Parent Models:** [SigLIP](https://github.com/huggingface/transformers/pull/26522) and [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
|
147 |
+
- **Resources for more information:**
|
148 |
+
- RAVEN dataset: [Dataset card](https://huggingface.co/datasets/HuggingFaceM4/RAVEN)
|
149 |
+
|
150 |
+
# License
|
151 |
+
|
152 |
+
The model is built on top of two pre-trained models: [SigLIP](https://github.com/huggingface/transformers/pull/26522) and [mistralai/Mistral-7B-v0.1](https://huggingface.co/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.
|
153 |
+
|
154 |
+
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.
|