jerome
commited on
Commit
·
a0446c6
1
Parent(s):
c671ac3
initial commit
Browse files- README.md +1 -0
- flava_finetuning_tutorial.ipynb +325 -0
README.md
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
FLAVA fine-tuning
|
flava_finetuning_tutorial.ipynb
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "oUL6DV1zCIlB"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"%matplotlib inline\n",
|
12 |
+
"!nvidia-smi"
|
13 |
+
]
|
14 |
+
},
|
15 |
+
{
|
16 |
+
"cell_type": "markdown",
|
17 |
+
"metadata": {
|
18 |
+
"id": "WmJySTGXCIlD"
|
19 |
+
},
|
20 |
+
"source": [
|
21 |
+
"\n",
|
22 |
+
"# TorchMultimodal Tutorial: Finetuning FLAVA\n"
|
23 |
+
]
|
24 |
+
},
|
25 |
+
{
|
26 |
+
"cell_type": "markdown",
|
27 |
+
"metadata": {
|
28 |
+
"id": "ZJCb2uRyCIlE"
|
29 |
+
},
|
30 |
+
"source": [
|
31 |
+
"Multimodal AI has recently become very popular owing to its ubiquitous\n",
|
32 |
+
"nature, from use cases like image captioning and visual search to more\n",
|
33 |
+
"recent applications like image generation from text. **TorchMultimodal\n",
|
34 |
+
"is a library powered by Pytorch consisting of building blocks and end to\n",
|
35 |
+
"end examples, aiming to enable and accelerate research in\n",
|
36 |
+
"multimodality**.\n",
|
37 |
+
"\n",
|
38 |
+
"In this tutorial, we will demonstrate how to use a **pretrained SoTA\n",
|
39 |
+
"model called** [FLAVA](https://arxiv.org/pdf/2112.04482.pdf)_ **from\n",
|
40 |
+
"TorchMultimodal library to finetune on a multimodal task i.e. visual\n",
|
41 |
+
"question answering** (VQA). The model consists of two unimodal transformer\n",
|
42 |
+
"based encoders for text and image and a multimodal encoder to combine\n",
|
43 |
+
"the two embeddings. It is pretrained using contrastive, image text matching and \n",
|
44 |
+
"text, image and multimodal masking losses.\n",
|
45 |
+
"\n"
|
46 |
+
]
|
47 |
+
},
|
48 |
+
{
|
49 |
+
"cell_type": "markdown",
|
50 |
+
"metadata": {
|
51 |
+
"id": "0TjU3iQgCIlE"
|
52 |
+
},
|
53 |
+
"source": [
|
54 |
+
"## Installation\n",
|
55 |
+
"We will use TextVQA dataset and bert tokenizer from HuggingFace for this\n",
|
56 |
+
"tutorial. So you need to install datasets and transformers in addition to TorchMultimodal.\n",
|
57 |
+
"\n",
|
58 |
+
"<div class=\"alert alert-info\"><h4>Note</h4><p>When running this tutorial in Google Colab, install the required packages by\n",
|
59 |
+
" creating a new cell and running the following commands:\n",
|
60 |
+
"\n",
|
61 |
+
"```\n",
|
62 |
+
"!pip install torchmultimodal-nightly\n",
|
63 |
+
"!pip install datasets\n",
|
64 |
+
"!pip install transformers</p></div>\n",
|
65 |
+
"```\n"
|
66 |
+
]
|
67 |
+
},
|
68 |
+
{
|
69 |
+
"cell_type": "markdown",
|
70 |
+
"metadata": {
|
71 |
+
"id": "LGuYfyaJCIlE"
|
72 |
+
},
|
73 |
+
"source": [
|
74 |
+
"## Steps \n",
|
75 |
+
"\n",
|
76 |
+
"1. Download the HuggingFace dataset to a directory on your computer by running the following command:\n",
|
77 |
+
"\n",
|
78 |
+
"```\n",
|
79 |
+
"wget http://dl.fbaipublicfiles.com/pythia/data/vocab.tar.gz \n",
|
80 |
+
"tar xf vocab.tar.gz\n",
|
81 |
+
"```\n",
|
82 |
+
" .. note:: \n",
|
83 |
+
" If you are running this tutorial in Google Colab, run these commands\n",
|
84 |
+
" in a new cell and prepend these commands with an exclamation mark (!)\n",
|
85 |
+
"\n",
|
86 |
+
"\n",
|
87 |
+
"2. For this tutorial, we treat VQA as a classification task where\n",
|
88 |
+
" the inputs are images and question (text) and the output is an answer class. \n",
|
89 |
+
" So we need to download the vocab file with answer classes and create the answer to\n",
|
90 |
+
" label mapping.\n",
|
91 |
+
"\n",
|
92 |
+
" We also load the [textvqa\n",
|
93 |
+
" dataset](https://arxiv.org/pdf/1904.08920.pdf)_ containing 34602 training samples\n",
|
94 |
+
" (images,questions and answers) from HuggingFace\n",
|
95 |
+
"\n",
|
96 |
+
"We see there are 3997 answer classes including a class representing\n",
|
97 |
+
"unknown answers.\n",
|
98 |
+
"\n",
|
99 |
+
"\n"
|
100 |
+
]
|
101 |
+
},
|
102 |
+
{
|
103 |
+
"cell_type": "code",
|
104 |
+
"execution_count": null,
|
105 |
+
"metadata": {
|
106 |
+
"id": "b6c1oq0lCIlF"
|
107 |
+
},
|
108 |
+
"outputs": [],
|
109 |
+
"source": [
|
110 |
+
"with open(\"data/vocabs/answers_textvqa_more_than_1.txt\") as f:\n",
|
111 |
+
" vocab = f.readlines()\n",
|
112 |
+
"\n",
|
113 |
+
"answer_to_idx = {}\n",
|
114 |
+
"for idx, entry in enumerate(vocab):\n",
|
115 |
+
" answer_to_idx[entry.strip(\"\\n\")] = idx\n",
|
116 |
+
"print(len(vocab))\n",
|
117 |
+
"print(vocab[:5])\n",
|
118 |
+
"\n",
|
119 |
+
"from datasets import load_dataset\n",
|
120 |
+
"dataset = load_dataset(\"textvqa\")"
|
121 |
+
]
|
122 |
+
},
|
123 |
+
{
|
124 |
+
"cell_type": "markdown",
|
125 |
+
"metadata": {
|
126 |
+
"id": "kGCla9GgCIlF"
|
127 |
+
},
|
128 |
+
"source": [
|
129 |
+
"Lets display a sample entry from the dataset:\n",
|
130 |
+
"\n",
|
131 |
+
"\n"
|
132 |
+
]
|
133 |
+
},
|
134 |
+
{
|
135 |
+
"cell_type": "code",
|
136 |
+
"execution_count": null,
|
137 |
+
"metadata": {
|
138 |
+
"id": "GLS8HGYtCIlF"
|
139 |
+
},
|
140 |
+
"outputs": [],
|
141 |
+
"source": [
|
142 |
+
"import matplotlib.pyplot as plt\n",
|
143 |
+
"import numpy as np \n",
|
144 |
+
"idx = 5 \n",
|
145 |
+
"print(\"Question: \", dataset[\"train\"][idx][\"question\"]) \n",
|
146 |
+
"print(\"Answers: \" ,dataset[\"train\"][idx][\"answers\"])\n",
|
147 |
+
"im = np.asarray(dataset[\"train\"][idx][\"image\"].resize((500,500)))\n",
|
148 |
+
"plt.imshow(im)\n",
|
149 |
+
"plt.show()"
|
150 |
+
]
|
151 |
+
},
|
152 |
+
{
|
153 |
+
"cell_type": "markdown",
|
154 |
+
"metadata": {
|
155 |
+
"id": "J1UO_daoCIlG"
|
156 |
+
},
|
157 |
+
"source": [
|
158 |
+
"3. Next, we write the transform function to convert the image and text into\n",
|
159 |
+
"Tensors consumable by our model - For images, we use the transforms from\n",
|
160 |
+
"torchvision to convert to Tensor and resize to uniform sizes - For text,\n",
|
161 |
+
"we tokenize (and pad) them using the BertTokenizer from HuggingFace -\n",
|
162 |
+
"For answers (i.e. labels), we take the most frequently occuring answer\n",
|
163 |
+
"as the label to train with:\n",
|
164 |
+
"\n",
|
165 |
+
"\n"
|
166 |
+
]
|
167 |
+
},
|
168 |
+
{
|
169 |
+
"cell_type": "code",
|
170 |
+
"execution_count": null,
|
171 |
+
"metadata": {
|
172 |
+
"id": "rO7lCn4DCIlG"
|
173 |
+
},
|
174 |
+
"outputs": [],
|
175 |
+
"source": [
|
176 |
+
"import torch\n",
|
177 |
+
"from torchvision import transforms\n",
|
178 |
+
"from collections import defaultdict\n",
|
179 |
+
"from transformers import BertTokenizer\n",
|
180 |
+
"from functools import partial\n",
|
181 |
+
"\n",
|
182 |
+
"def transform(tokenizer, input):\n",
|
183 |
+
" batch = {}\n",
|
184 |
+
" image_transform = transforms.Compose([transforms.ToTensor(), transforms.Resize([224,224])])\n",
|
185 |
+
" image = image_transform(input[\"image\"][0].convert(\"RGB\"))\n",
|
186 |
+
" batch[\"image\"] = [image]\n",
|
187 |
+
"\n",
|
188 |
+
" tokenized=tokenizer(input[\"question\"],return_tensors='pt',padding=\"max_length\",max_length=512)\n",
|
189 |
+
" batch.update(tokenized)\n",
|
190 |
+
"\n",
|
191 |
+
"\n",
|
192 |
+
" ans_to_count = defaultdict(int)\n",
|
193 |
+
" for ans in input[\"answers\"][0]:\n",
|
194 |
+
" ans_to_count[ans] += 1\n",
|
195 |
+
" max_value = max(ans_to_count, key=ans_to_count.get)\n",
|
196 |
+
" ans_idx = answer_to_idx.get(max_value,0)\n",
|
197 |
+
" batch[\"answers\"] = torch.as_tensor([ans_idx])\n",
|
198 |
+
" return batch\n",
|
199 |
+
"\n",
|
200 |
+
"tokenizer=BertTokenizer.from_pretrained(\"bert-base-uncased\",padding=\"max_length\",max_length=512)\n",
|
201 |
+
"transform=partial(transform,tokenizer)\n",
|
202 |
+
"dataset.set_transform(transform)"
|
203 |
+
]
|
204 |
+
},
|
205 |
+
{
|
206 |
+
"cell_type": "markdown",
|
207 |
+
"metadata": {
|
208 |
+
"id": "LOMy3UbpCIlG"
|
209 |
+
},
|
210 |
+
"source": [
|
211 |
+
"4. Finally, we import the flava_model_for_classification from\n",
|
212 |
+
"torchmultimodal. It loads the pretrained flava checkpoint by default and\n",
|
213 |
+
"includes a classification head.\n",
|
214 |
+
"\n",
|
215 |
+
"The model forward function passes the image through the visual encoder\n",
|
216 |
+
"and the question through the text encoder. The image and question\n",
|
217 |
+
"embeddings are then passed through the multimodal encoder. The final\n",
|
218 |
+
"embedding corresponding to the CLS token is passed through a MLP head\n",
|
219 |
+
"which finally gives the probability distribution over each possible\n",
|
220 |
+
"answers.\n",
|
221 |
+
"\n",
|
222 |
+
"\n"
|
223 |
+
]
|
224 |
+
},
|
225 |
+
{
|
226 |
+
"cell_type": "code",
|
227 |
+
"execution_count": null,
|
228 |
+
"metadata": {
|
229 |
+
"id": "drSfcYNCCIlG"
|
230 |
+
},
|
231 |
+
"outputs": [],
|
232 |
+
"source": [
|
233 |
+
"from torchmultimodal.models.flava.model import flava_model_for_classification\n",
|
234 |
+
"model = flava_model_for_classification(num_classes=len(vocab))"
|
235 |
+
]
|
236 |
+
},
|
237 |
+
{
|
238 |
+
"cell_type": "markdown",
|
239 |
+
"metadata": {
|
240 |
+
"id": "976mlWvaCIlG"
|
241 |
+
},
|
242 |
+
"source": [
|
243 |
+
"5. We put together the dataset and model in a toy training loop to\n",
|
244 |
+
"demonstrate how to train the model for 3 iterations:\n",
|
245 |
+
"\n",
|
246 |
+
"\n"
|
247 |
+
]
|
248 |
+
},
|
249 |
+
{
|
250 |
+
"cell_type": "code",
|
251 |
+
"execution_count": null,
|
252 |
+
"metadata": {
|
253 |
+
"id": "0KvxQ4xaCIlH"
|
254 |
+
},
|
255 |
+
"outputs": [],
|
256 |
+
"source": [
|
257 |
+
"from torch import nn\n",
|
258 |
+
"BATCH_SIZE = 2\n",
|
259 |
+
"MAX_STEPS = 3\n",
|
260 |
+
"from torch.utils.data import DataLoader\n",
|
261 |
+
"\n",
|
262 |
+
"train_dataloader = DataLoader(dataset[\"train\"], batch_size= BATCH_SIZE)\n",
|
263 |
+
"optimizer = torch.optim.AdamW(model.parameters())\n",
|
264 |
+
"\n",
|
265 |
+
"\n",
|
266 |
+
"epochs = 1\n",
|
267 |
+
"for _ in range(epochs):\n",
|
268 |
+
" for idx, batch in enumerate(train_dataloader):\n",
|
269 |
+
" optimizer.zero_grad()\n",
|
270 |
+
" out = model(text = batch[\"input_ids\"], image = batch[\"image\"], labels = batch[\"answers\"])\n",
|
271 |
+
" loss = out.loss\n",
|
272 |
+
" loss.backward()\n",
|
273 |
+
" optimizer.step()\n",
|
274 |
+
" print(f\"Loss at step {idx} = {loss}\")\n",
|
275 |
+
" if idx > MAX_STEPS-1:\n",
|
276 |
+
" break"
|
277 |
+
]
|
278 |
+
},
|
279 |
+
{
|
280 |
+
"cell_type": "markdown",
|
281 |
+
"metadata": {
|
282 |
+
"id": "A7An1sjZCIlH"
|
283 |
+
},
|
284 |
+
"source": [
|
285 |
+
"## Conclusion\n",
|
286 |
+
"\n",
|
287 |
+
"This tutorial introduced the basics around how to finetune on a\n",
|
288 |
+
"multimodal task using FLAVA from TorchMultimodal. Please also check out\n",
|
289 |
+
"other examples from the library like\n",
|
290 |
+
"[MDETR](https://github.com/facebookresearch/multimodal/tree/main/torchmultimodal/models/mdetr)_\n",
|
291 |
+
"which is a multimodal model for object detection and\n",
|
292 |
+
"[Omnivore](https://github.com/facebookresearch/multimodal/blob/main/torchmultimodal/models/omnivore.py)_\n",
|
293 |
+
"which is multitask model spanning image, video and 3d classification.\n",
|
294 |
+
"\n",
|
295 |
+
"\n"
|
296 |
+
]
|
297 |
+
}
|
298 |
+
],
|
299 |
+
"metadata": {
|
300 |
+
"kernelspec": {
|
301 |
+
"display_name": "Python 3",
|
302 |
+
"language": "python",
|
303 |
+
"name": "python3"
|
304 |
+
},
|
305 |
+
"language_info": {
|
306 |
+
"codemirror_mode": {
|
307 |
+
"name": "ipython",
|
308 |
+
"version": 3
|
309 |
+
},
|
310 |
+
"file_extension": ".py",
|
311 |
+
"mimetype": "text/x-python",
|
312 |
+
"name": "python",
|
313 |
+
"nbconvert_exporter": "python",
|
314 |
+
"pygments_lexer": "ipython3",
|
315 |
+
"version": "3.10.9"
|
316 |
+
},
|
317 |
+
"colab": {
|
318 |
+
"provenance": []
|
319 |
+
},
|
320 |
+
"accelerator": "GPU",
|
321 |
+
"gpuClass": "standard"
|
322 |
+
},
|
323 |
+
"nbformat": 4,
|
324 |
+
"nbformat_minor": 0
|
325 |
+
}
|