pix2struct-textcaps

Visually-situated language is ubiquitous—sources range from textbooks with diagrams to web pages with images and tables, to mobile apps with buttons and forms. Perhaps due to this diversity, previous work has typically relied on domainspecific recipes with limited sharing of the underlying data, model architectures, and objectives. We present Pix2Struct, a pretrained image-to-text model for purely visual language understanding, which can be finetuned on tasks containing visually-situated language. Pix2Struct is pretrained by learning to parse masked screenshots of web pages into simplified HTML. The web, with its richness of visual elements cleanly reflected in the HTML structure, provides a large source of pretraining data well suited to the diversity of downstream tasks. Intuitively, this objective subsumes common pretraining signals such as OCR, language modeling, image captioning. In addition to the novel pretraining strategy, we introduce a variable-resolution input representation and a more flexible integration of language and vision inputs, where language prompts such as questions are rendered directly on top of the input image. For the first time, we show that a single pretrained model can achieve state-of-the-art results in six out of nine tasks across four domains: documents, illustrations, user interfaces, and natural images.

Using the model

Converting from T5x to huggingface

You can use the convert_pix2struct_checkpoint_to_pytorch.py script as follows:

python convert_pix2struct_checkpoint_to_pytorch.py --t5x_checkpoint_path PATH_TO_T5X_CHECKPOINTS --pytorch_dump_path PATH_TO_SAVE

if you are converting a large model, run:

python convert_pix2struct_checkpoint_to_pytorch.py --t5x_checkpoint_path PATH_TO_T5X_CHECKPOINTS --pytorch_dump_path PATH_TO_SAVE --use-large

Once saved, you can push your converted model with the following snippet:

from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor

model = Pix2StructForConditionalGeneration.from_pretrained(PATH_TO_SAVE)
processor = Pix2StructProcessor.from_pretrained(PATH_TO_SAVE)

model.push_to_hub("USERNAME/MODEL_NAME")
processor.push_to_hub("USERNAME/MODEL_NAME")

Running the model

In full precision, on CPU:

You can run the model in full precision on CPU:

import requests
from PIL import Image
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor

url = "https://www.ilankelman.org/stopsigns/australia.jpg"
image = Image.open(requests.get(url, stream=True).raw)

model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")

# image only
inputs = processor(images=image, return_tensors="pt")

predictions = model.generate(**inputs)
print(processor.decode(predictions[0], skip_special_tokens=True))
>>> A stop sign is on a street corner.

In full precision, on GPU:

You can run the model in full precision on CPU:

import requests
from PIL import Image
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor

url = "https://www.ilankelman.org/stopsigns/australia.jpg"
image = Image.open(requests.get(url, stream=True).raw)

model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base").to("cuda")
processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")

# image only
inputs = processor(images=image, return_tensors="pt").to("cuda")

predictions = model.generate(**inputs)
print(processor.decode(predictions[0], skip_special_tokens=True))
>>> A stop sign is on a street corner.

In half precision, on GPU:

You can run the model in full precision on CPU:

import requests
import torch

from PIL import Image
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor

url = "https://www.ilankelman.org/stopsigns/australia.jpg"
image = Image.open(requests.get(url, stream=True).raw)

model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base", torch_dtype=torch.bfloat16).to("cuda")
processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")

# image only
inputs = processor(images=image, return_tensors="pt").to("cuda", torch.bfloat16)

predictions = model.generate(**inputs)
print(processor.decode(predictions[0], skip_special_tokens=True))
>>> A stop sign is on a street corner.

Use different sequence length

This model has been trained on a sequence length of 2048. You can try to reduce the sequence length for a more memory efficient inference but you may observe some performance degradation for small sequence length (<512). Just pass max_patches when calling the processor:

inputs = processor(images=image, return_tensors="pt", max_patches=512)

Conditional generation

You can also pre-pend some input text to perform conditional generation:

import requests
from PIL import Image
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor

url = "https://www.ilankelman.org/stopsigns/australia.jpg"
image = Image.open(requests.get(url, stream=True).raw)
text = "A picture of"

model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
processor = Pix2StructProcessor.from_pretrained("google/pix