Now Reading
Hands-on Guide to OpenAI’s CLIP – Connecting Text To Images

Hands-on Guide to OpenAI’s CLIP – Connecting Text To Images

Jayita Bhattacharyya

Standard computer vision datasets cannot generalize many aspects of vision-based models. Creating image datasets would be laborious and have limitations, with restrictions over only a certain range of object categories. To overcome these image label constraints, OpenAI has designed its new neural network architecture CLIP (Contrastive Language-Image Pretraining) for Learning Transferable Visual Models From Natural Language Supervision

OpenAI has already done commendable work in the world of AI and deep learning with GPT2 and GPT3. CLIP is an extension of that. It provides predictions with captions on images based on simple pre-trained models in a more robust and scalable state-of-the-art method for image recognition being built on a dataset of nearly 400M image and text pairs scraped from the internet. After pre-training the model, natural language processing is used to match learned visual concepts that can enable zero-shot transfer learning. This approach’s performance measures are done by testing and benchmarking over various existing vision datasets, including action recognition in videos, optical character recognition, geo-localization, and many more based on object classification and other tasks.

CLIP implements several existing learning visual representations from natural language supervision techniques. This involves modern and advanced architectures like the Vision and text Transformers, ICMLM, which explores masked language modelling, VirTex, which is applied to autoregressive language modelling, and ConVIRT, used in the contrastive objective that is used in CLIP for medical imaging.

How CLIP works?

Generally, the working of vision models is that they train an image feature extractor first and then train a linear classifier to predict labels for the assigned task, but CLIP has a separate and better approach but training both the image encoder and text encoder together/parallelly to predict the accurate image-label pairs for a training batch. At the time of testing the model, the learned text encoder deploys synthesis of a zero-shot linear classifier by providing embeddings with the name or some basic information of the target dataset’s classes.

A Colab notebook has been recently made available to directly run the model 

Code Snippet

     "ViT-B/32":       "",
 ! wget {MODELS["ViT-B/32"]} -O
 --2021-01-08 06:06:44-- 

Resolving (…, 2620:1ec:bdf::13

Connecting to (||:443… connected.

HTTP request sent, awaiting response… 200 OK

Length: 353976522 (338M) [application/octet-stream]

Saving to: ‘’            100%[===================>] 337.58M  65.4MB/s    in 5.3s    

2021-01-08 06:06:50 (63.8 MB/s) – ‘’ saved [353976522/353976522]

 model = torch.jit.load("").cuda().eval()
 input_resolution = model.input_resolution.item()
 context_length = model.context_length.item()
 vocab_size = model.vocab_size.item()
 print("Model parameters:", f"{np.sum([int( for p in model.parameters()]):,}")
 print("Input resolution:", input_resolution)
 print("Context length:", context_length)
 print("Vocab size:", vocab_size) 
 Model parameters: 151,277,313
 Input resolution: 224
 Context length: 77
 Vocab size: 49408 

# Image Preprocessing

 from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
 from PIL import Image
 preprocess = Compose([
     Resize(input_resolution, interpolation=Image.BICUBIC),
 image_mean = torch.tensor([0.48145466, 0.4578275, 0.40821073]).cuda()
 image_std = torch.tensor([0.26862954, 0.26130258, 0.27577711]).cuda() 

# Text Preprocessing

! pip install ftfy regex

! wget -O bpe_simple_vocab_16e6.txt.gz

Requirement already satisfied: ftfy in /usr/local/lib/python3.6/dist-packages (5.8)

Requirement already satisfied: regex in /usr/local/lib/python3.6/dist-packages (2019.12.20)

Requirement already satisfied: wcwidth in /usr/local/lib/python3.6/dist-packages (from ftfy) (0.2.5)

–2021-01-08 06:07:04–

Resolving (…, 2620:1ec:bdf::13

Connecting to (||:443… connected.

HTTP request sent, awaiting response… 200 OK

Length: 1356917 (1.3M) [application/octet-stream]

Saving to: ‘bpe_simple_vocab_16e6.txt.gz’

bpe_simple_vocab_16 100%[===================>]   1.29M  –.-KB/s    in 0.06s   

See Also

2021-01-08 06:07:04 (22.1 MB/s) – ‘bpe_simple_vocab_16e6.txt.gz’ saved [1356917/1356917]

#Returns list of utf-8 byte and a corresponding list of Unicode strings. The reversible BPE codes work on unicode strings. This means you need many unicode characters in your vocab if you want to avoid UNKs. When you’re at something like a 10B token dataset, you need around 5K for decent coverage.

 import gzip
 import html
 import os
 from functools import lru_cache
 import ftfy
 import regex as re 

#This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup tables between utf-8 bytes and Unicode strings. And avoids mapping to whitespace/control characters the bpe code barfs on.

 def bytes_to_unicode():
     bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
     cs = bs[:]
     n = 0
     for b in range(2**8):
         if b not in bs:
             n += 1
     cs = [chr(n) for n in cs]
     return dict(zip(bs, cs)) 

#Return set of symbol pairs in a word. Word is represented as a tuple of symbols (symbols being variable-length strings).

 def get_pairs(word):
     pairs = set()
     prev_char = word[0]
     for char in word[1:]:
         pairs.add((prev_char, char))
         prev_char = char
     return pairs
 def basic_clean(text):
     text = ftfy.fix_text(text)
     text = html.unescape(html.unescape(text))
     return text.strip()
 def whitespace_clean(text):
     text = re.sub(r'\s+', ' ', text)
     text = text.strip()
     return text
 class SimpleTokenizer(object):
     def __init__(self, bpe_path: str = "bpe_simple_vocab_16e6.txt.gz"):
         self.byte_encoder = bytes_to_unicode()
         self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
         merges ="utf-8").split('\n')
         merges = merges[1:49152-256-2+1]
         merges = [tuple(merge.split()) for merge in merges]
         vocab = list(bytes_to_unicode().values())
         vocab = vocab + [v+'</w>' for v in vocab]
         for merge in merges:
         vocab.extend(['<|startoftext|>', '<|endoftext|>'])
         self.encoder = dict(zip(vocab, range(len(vocab))))
         self.decoder = {v: k for k, v in self.encoder.items()}
         self.bpe_ranks = dict(zip(merges, range(len(merges))))
         self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
         self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
     def bpe(self, token):
         if token in self.cache:
             return self.cache[token]
         word = tuple(token[:-1]) + ( token[-1] + '</w>',)
         pairs = get_pairs(word)
         if not pairs:
             return token+'</w>'
         while True:
             bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
             if bigram not in self.bpe_ranks:
             first, second = bigram
             new_word = []
             i = 0
             while i < len(word):
                     j = word.index(first, i)
                     i = j
                 if word[i] == first and i < len(word)-1 and word[i+1] == second:
                     i += 2
                     i += 1
             new_word = tuple(new_word)
             word = new_word
             if len(word) == 1:
                 pairs = get_pairs(word)
         word = ' '.join(word)
         self.cache[token] = word
         return word
     def encode(self, text):
         bpe_tokens = []
         text = whitespace_clean(basic_clean(text)).lower()
         for token in re.findall(self.pat, text):
             token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
             bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
         return bpe_tokens
     def decode(self, tokens):
         text = ''.join([self.decoder[token] for token in tokens])
         text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
         return text 

# Setting up input images and texts

 import os
 import skimage
 import IPython.display
 import matplotlib.pyplot as plt
 from PIL import Image
 import numpy as np
 from collections import OrderedDict
 import torch
 %matplotlib inline
 %config InlineBackend.figure_format = 'retina'
 # images in skimage to use along with their textual descriptions
 descriptions = {
     "page": "a page of text about segmentation",
     "chelsea": "a facial photo of a tabby cat",
     "astronaut": "a portrait of an astronaut with the American flag",
     "rocket": "a rocket standing on a launchpad",
     "motorcycle_right": "a red motorcycle standing in a garage",
     "camera": "a person looking at a camera on a tripod",
     "horse": "a black-and-white silhouette of a horse", 
     "coffee": "a cup of coffee on a saucer"
 images = []
 texts = []
 plt.figure(figsize=(16, 5))
 for filename in [filename for filename in os.listdir(skimage.data_dir) if filename.endswith(".png") or filename.endswith(".jpg")]:
     name = os.path.splitext(filename)[0]
     if name not in descriptions:
     image = preprocess(, filename)).convert("RGB"))
     plt.subplot(2, 4, len(images))
     plt.imshow(image.permute(1, 2, 0))

# Building features

 image_input = torch.tensor(np.stack(images)).cuda()
 image_input -= image_mean[:, None, None]
 image_input /= image_std[:, None, None]
 tokenizer = SimpleTokenizer()
 text_tokens = [tokenizer.encode("This is " + desc + "<|endoftext|>") for desc in texts]
 text_input = torch.zeros(len(text_tokens), model.context_length, dtype=torch.long)
 for i, tokens in enumerate(text_tokens):
     text_input[i, :len(tokens)] = torch.tensor(tokens)
 text_input = text_input.cuda()
 with torch.no_grad():
     image_features = model.encode_image(image_input).float()
     text_features = model.encode_text(text_input).float() 

# Calculating cosine similarity

 image_features /= image_features.norm(dim=-1, keepdim=True)
 text_features /= text_features.norm(dim=-1, keepdim=True)
 similarity = text_features.cpu().numpy() @ image_features.cpu().numpy().T
 count = len(descriptions)
 plt.figure(figsize=(20, 14))
 plt.imshow(similarity, vmax=0.3)
 plt.yticks(range(count), texts, fontsize=18)
 for i, image in enumerate(images):
     plt.imshow(image.permute(1, 2, 0), extent=(i - 0.5, i + 0.5, -1.6, -0.6), origin="lower")
 for x in range(similarity.shape[1]):
     for y in range(similarity.shape[0]):
         plt.text(x, y, f"{similarity[y, x]:.2f}", ha="center", va="center", size=12)
 for side in ["left", "top", "right", "bottom"]:
 plt.xlim([-0.5, count - 0.5])
 plt.ylim([count + 0.5, -2])
 plt.title("Cosine similarity between text and image features", size=20)
 Text(0.5, 1.0, 'Cosine similarity between text and image features') 

Zero-Shot Image Classification

Downloading to /root/.cache/cifar-100-python.tar.gz


 from torchvision.datasets import CIFAR100
 cifar100 = CIFAR100(os.path.expanduser("~/.cache"), transform=preprocess, download=True) 

168976384/169001437 [00:17<00:00, 13148248.27it/s]

Extracting /root/.cache/cifar-100-python.tar.gz to /root/.cache

 text_descriptions = [f"This is a photo of a {label}" for label in cifar100.classes]
 text_tokens = [tokenizer.encode(desc + "<|endoftext|>") for desc in text_descriptions]
 text_input = torch.zeros(len(text_tokens), model.context_length, dtype=torch.long)
 for i, tokens in enumerate(text_tokens):
     text_input[i, :len(tokens)] = torch.tensor(tokens)
 text_input = text_input.cuda()
 torch.Size([100, 77])
 with torch.no_grad():
     text_features = model.encode_text(text_input).float()
     text_features /= text_features.norm(dim=-1, keepdim=True)
 text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)
 top_probs, top_labels = text_probs.cpu().topk(5, dim=-1)
 plt.figure(figsize=(16, 16))
 for i, image in enumerate(images):
     plt.subplot(4, 4, 2 * i + 1)
     plt.imshow(image.permute(1, 2, 0))
     plt.subplot(4, 4, 2 * i + 2)
     y = np.arange(top_probs.shape[-1])
     plt.barh(y, top_probs[i])
     plt.yticks(y, [cifar100.classes[index] for index in top_labels[i].numpy()])


CLIP is highly efficient and flexible surpassing existing benchmarked performances recognizing everyday objects. As of now, it does have certain limitations, some routine tasks, and hard in predicting images that are out-of-distribution in pre-trained datasets.

What Do You Think?

Subscribe to our Newsletter

Get the latest updates and relevant offers by sharing your email.
What's Your Reaction?
In Love
Not Sure

Copyright Analytics India Magazine Pvt Ltd

Scroll To Top