TorchMultimodal 教程：微调 FLAVA

Multimodal AI has recently become very popular owing to its ubiquitous nature, from use cases like image captioning and visual search to more recent applications like image generation from text. TorchMultimodal is a library powered by Pytorch consisting of building blocks and end to end examples, aiming to enable and accelerate research in multimodality.

In this tutorial, we will demonstrate how to use a pretrained SoTA model called FLAVA from TorchMultimodal library to finetune on a multimodal task i.e. visual question answering (VQA). The model consists of two unimodal transformer based encoders for text and image and a multimodal encoder to combine the two embeddings. It is pretrained using contrastive, image text matching and text, image and multimodal masking losses.

Installation

We will use TextVQA dataset and bert tokenizer from Hugging Face for this tutorial. So you need to install datasets and transformers in addition to TorchMultimodal.

Note

When running this tutorial in Google Colab, install the required packages by creating a new cell and running the following commands:

!pip install torchmultimodal-nightly
!pip install datasets
!pip install transformers

Steps

1. Download the Hugging Face dataset to a directory on your computer by running the following command:

   wget http://dl.fbaipublicfiles.com/pythia/data/vocab.tar.gz
   tar xf vocab.tar.gz
   

   Note

   If you are running this tutorial in Google Colab, run these commands in a new cell and prepend these commands with an exclamation mark (!)

2. For this tutorial, we treat VQA as a classification task where the inputs are images and question (text) and the output is an answer class. So we need to download the vocab file with answer classes and create the answer to label mapping.

   We also load the textvqa dataset containing 34602 training samples (images,questions and answers) from Hugging Face

We see there are 3997 answer classes including a class representing unknown answers.

with open("data/vocabs/answers_textvqa_more_than_1.txt") as f:
  vocab = f.readlines()

answer_to_idx = {}
for idx, entry in enumerate(vocab):
  answer_to_idx[entry.strip("\n")] = idx
print(len(vocab))
print(vocab[:5])

from datasets import load_dataset
dataset = load_dataset("textvqa")

Lets display a sample entry from the dataset:

import matplotlib.pyplot as plt
import numpy as np
idx = 5
print("Question: ", dataset["train"][idx]["question"])
print("Answers: " ,dataset["train"][idx]["answers"])
im = np.asarray(dataset["train"][idx]["image"].resize((500,500)))
plt.imshow(im)
plt.show()

3. Next, we write the transform function to convert the image and text into Tensors consumable by our model - For images, we use the transforms from torchvision to convert to Tensor and resize to uniform sizes - For text, we tokenize (and pad) them using the BertTokenizer from Hugging Face - For answers (i.e. labels), we take the most frequently occurring answer as the label to train with:

import torch
from torchvision import transforms
from collections import defaultdict
from transformers import BertTokenizer
from functools import partial

def transform(tokenizer, input):
  batch = {}
  image_transform = transforms.Compose([transforms.ToTensor(), transforms.Resize([224,224])])
  image = image_transform(input["image"][0].convert("RGB"))
  batch["image"] = [image]

  tokenized=tokenizer(input["question"],return_tensors='pt',padding="max_length",max_length=512)
  batch.update(tokenized)


  ans_to_count = defaultdict(int)
  for ans in input["answers"][0]:
    ans_to_count[ans] += 1
  max_value = max(ans_to_count, key=ans_to_count.get)
  ans_idx = answer_to_idx.get(max_value,0)
  batch["answers"] = torch.as_tensor([ans_idx])
  return batch

tokenizer=BertTokenizer.from_pretrained("bert-base-uncased",padding="max_length",max_length=512)
transform=partial(transform,tokenizer)
dataset.set_transform(transform)

4. Finally, we import the flava_model_for_classification from torchmultimodal. It loads the pretrained FLAVA checkpoint by default and includes a classification head.

The model forward function passes the image through the visual encoder and the question through the text encoder. The image and question embeddings are then passed through the multimodal encoder. The final embedding corresponding to the CLS token is passed through a MLP head which finally gives the probability distribution over each possible answers.

from torchmultimodal.models.flava.model import flava_model_for_classification
model = flava_model_for_classification(num_classes=len(vocab))

5. We put together the dataset and model in a toy training loop to demonstrate how to train the model for 3 iterations:

from torch import nn
BATCH_SIZE = 2
MAX_STEPS = 3
from torch.utils.data import DataLoader

train_dataloader = DataLoader(dataset["train"], batch_size= BATCH_SIZE)
optimizer = torch.optim.AdamW(model.parameters())


epochs = 1
for _ in range(epochs):
  for idx, batch in enumerate(train_dataloader):
    optimizer.zero_grad()
    out = model(text = batch["input_ids"], image = batch["image"], labels = batch["answers"])
    loss = out.loss
    loss.backward()
    optimizer.step()
    print(f"Loss at step {idx} = {loss}")
    if idx >= MAX_STEPS-1:
      break

Conclusion

This tutorial introduced the basics around how to finetune on a multimodal task using FLAVA from TorchMultimodal. Please also check out other examples from the library like MDETR which is a multimodal model for object detection and Omnivore which is multitask model spanning image, video and 3d classification.