Generate text captions for images from their CLIP embeddings. Includes PyTorch model code and example training script.

Overview

clip-text-decoder

Generate text captions for images from their CLIP embeddings. Includes PyTorch model code and example training script.

Example Predictions

Example captions were computed with the pretrained model mentioned below.

"A man riding a wave on top of a surfboard."

A surfer riding a wave

A baseball player is swinging a bat at a ball.

Baseball player

"A dog running across a field with a frisbee."

Dog with frisbee

Installation

Install for easier access to the following objects/classes:

  • clip_text_decoder.datasets.ClipCocoCaptionsDataset
  • clip_text_decoder.models.ClipDecoder
  • clip_text_decoder.models.ClipDecoderInferenceModel
  • clip_text_decoder.tokenizer.Tokenizer

The train.py script will not be available in the installed package, since it's located in the root directory. To train new models, either clone this repository or recreate train.py locally.

Using pip:

pip install clip-text-decoder

From source:

git clone https://github.com/fkodom/clip-text-decoder.git
cd clip-text-decoder
pip install .

NOTE: You'll also need to install openai/CLIP to encode images with CLIP. This is also required by ClipCocoCaptionsDataset to build the captions dataset the first time (cached for subsequent calls).

pip install "clip @ git+https://github.com/openai/CLIP.git"

For technical reasons, the CLIP dependency can't be included in the PyPI package, since it's not an officially published package.

Training

Open In Colab

Launch your own training session using the provided script (train.py):

python train.py --max-epochs 5

Training CLI arguments, along with their default values:

--max-epochs 5  # (int)
--num-layers 6  # (int)
--dim-feedforward 256  # (int)
--precision 16  # (16 or 32)
--seed 0  # (int)

Inference

The training script will produce a model.zip archive, containing the Tokenizer and trained model parameters. To perform inference with it:

import clip
from PIL import Image
import torch

from clip_text_decoder.model import ClipDecoderInferenceModel

device = "cuda" if torch.cuda.is_available() else "cpu"
model = ClipDecoderInferenceModel.load("path/to/model.zip").to(device)
clip_model, clip_preprocessor = clip.load("ViT-B/32", device=device, jit=False)

# Create a blank dummy image
dummy_image = Image.new("RGB", (224, 224))
preprocessed = clip_preprocessor(dummy_image).to(device)
# Add a batch dimension using '.unsqueeze(0)'
encoded = clip_model.encode_image(preprocessed.unsqueeze(0))
text = model(encoded)

print(text)
# Probably some nonsense, because we used a dummy image.

Pretrained Models

A pretrained CLIP decoder is hosted in my Google Drive, and can easily be downloaded by:

from clip_text_decoder.model import ClipDecoderInferenceModel

model = ClipDecoderInferenceModel.download_pretrained()

To cache the pretrained model locally, so that it's not re-downloaded each time:

model = ClipDecoderInferenceModel.download_pretrained("/path/to/model.zip")

Shortcomings

  • Only works well with COCO-style images. If you go outside the distribution of COCO objects, you'll get nonsense text captions.
  • Relatively short training time. Even within the COCO domain, you'll occasionally see incorrect captions. Quite a few captions will have bad grammar, repetitive descriptors, etc.
Comments
  • Decoding Text Embeddings Coded Using Hugging Face ClipTextModel

    Decoding Text Embeddings Coded Using Hugging Face ClipTextModel

    Suppose that I have text embeddings created using Hugging Face's ClipTextModel using the following method:

    import torch
    from transformers import CLIPTokenizer, CLIPTextModel
    
    class_list = ["i love going home and playing with my wife and kids", "i love going home", "playing with my wife and kids", 
    "family", "war", "writing"]
    
    model = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14")
    tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
    
    inputs = tokenizer(class_list, padding=True, return_tensors="pt")
    outputs = model(**inputs)
    hidden_state = outputs.last_hidden_state
    embeddings = outputs.pooler_output
    

    Questions:

    1. Is It possible to use the clip-text-decoder to convert the embeddings back to text?
    2. If it is indeed possible to do so, could you provide an example of how?

    Looking forward to receiving your feedback.

    opened by mbdzi 6
  • Fix string error when loading clip models.

    Fix string error when loading clip models.

    error

    The model name string ( VIT-xxx ) in the check_vision_backbone function is not compatible with the model name string ( ViT-xxx ) of the clip repository, which will cause at least one error in check_vision_backbone function or when loading the clip model.

    solution

    In this PR, the model name string in the check_vision_backbone function is modified to ViT-xxx to make it compatible with the clip repository.

    opened by Adenialzz 1
  • BLIP vision backbone

    BLIP vision backbone

    • Added blip backbone; still cleaning up last pieces
    • Bug fixes for training script, and remove debug code.
    • Fix dependencies in test workflow; update README statistics
    • Fix test issue with CUDA device
    • Update unit tests for newer Python, torch versions
    • Test up to Python 3.10
    • Test up to Python 3.9
    • Install lavis first
    opened by fkodom 0
  • Feature: Beam Search

    Feature: Beam Search

    • Add beam search, clip dependency to setup.py
    • Fix installation instructions
    • Remove main clause
    • Add '--beam-size' option to 'train.py' script.
    • Update README; propagate the '--beam-size' arg through eval functions
    • Update setup.cfg, add pre-commit hooks
    • Reformat images
    • Remove fixed image width
    • Add detail to README; comments to call method for beam search
    • Updated README headline
    opened by fkodom 0
  • Bug Fixes for Broken Tests

    Bug Fixes for Broken Tests

    • Cache the old fashioned way :)
    • Fix silly typo in test for image caption model
    • Apply black and isort formatting
    • Install latest version of 'black', reapply formatting
    • Fix flake8 issue (duplicate function definition), and install latest patch version of pytorch for tests.
    • Skip slow tests by default, add 'slow' marker to inference model tests.
    opened by fkodom 0
  • GPT2 Decoder

    GPT2 Decoder

    • Update model to use DistilGPT2 as a pre-trained decoder.
    • Removed tokenizer (no longer used), fixed bugs in Model source file, and updated model unit tests.
    • Backwards compatibility for 'gdown.download' method.
    • Update installation requirements, caption examples in README
    opened by fkodom 0
  • Upgrade CodeSee workflow to version 2

    Upgrade CodeSee workflow to version 2

    CodeSee is a code visibility platform.

    This change updates the CodeSee workflow file to the latest version for security, maintenance, and support improvements (see changelog below).

    That workflow file:

    • runs CodeSee's code analysis on every PR push and merge
    • uploads that analysis to CodeSee.
    • It does not transmit your code.

    The code analysis is used to generate maps and insights about this codebase.

    CodeSee workflow changelog:

    • Improved security: Updates permission to be read-only.
    • Improved future maintenance: Replaces the body of the workflow with a single github action: codesee-action. This makes it significantly easier for CodeSee to introduce future improvements and fixes without requiring another PR like this.
    • Improved Python support: The action now properly supports Python 3.11, and will continue to support new Python versions as they are released.
    opened by codesee-maps[bot] 1
  • Incompatible checksum error

    Incompatible checksum error

    I see the following error when trying to load the pretrained model.

        tokenizer=pickle.loads(tokenizer_buffer.read()),
      File "stringsource", line 6, in spacy.pipeline.trainable_pipe.__pyx_unpickle_TrainablePipe
    _pickle.PickleError: Incompatible checksums (102742709 vs 0x417ddeb = (cfg, model, name, vocab))
    

    Am I missing something?

    opened by dapurv5 0
Releases(1.4.4)
  • 1.4.4(Nov 7, 2022)

    What's Changed

    • Fix string error when loading clip models. by @Adenialzz in https://github.com/fkodom/clip-text-decoder/pull/12

    New Contributors

    • @Adenialzz made their first contribution in https://github.com/fkodom/clip-text-decoder/pull/12

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.4.3...1.4.4

    Source code(tar.gz)
    Source code(zip)
  • 1.4.3(Nov 7, 2022)

    What's Changed

    • Refactor Dataset by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/11

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.4.2...1.4.3

    Source code(tar.gz)
    Source code(zip)
  • 1.4.2(Oct 26, 2022)

    What's Changed

    • Huggingface Evaluate by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/9

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.4.1...1.4.2

    Source code(tar.gz)
    Source code(zip)
  • 1.4.1(Oct 26, 2022)

    What's Changed

    • Datapipes by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/8

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.4.0...1.4.1

    Source code(tar.gz)
    Source code(zip)
  • 1.4.0(Oct 23, 2022)

    What's Changed

    • BLIP vision backbone by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/7

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.3.0...1.4.0

    Source code(tar.gz)
    Source code(zip)
  • 1.3.0(Oct 2, 2022)

    What's Changed

    • Feature: Beam Search by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/5
    • Bug Fix: PyPI Release by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/6

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.2.0...1.3.0

    Source code(tar.gz)
    Source code(zip)
  • 1.2.0(Jan 29, 2022)

    What's Changed

    • Cache CLIP embeddings for the dataset, rather than recomputing them each time.

    • Reduce model file sizes by storing at lower precision

    • Add an ImageCaptionInferenceModel class for easier out-of-the-box use

    • Fix some broken unit tests

    • Better Data Caching by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/3

    • Bug Fixes for Broken Tests by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/4

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.1.0...1.2.0

    Source code(tar.gz)
    Source code(zip)
  • 1.1.0(Dec 22, 2021)

    What's Changed

    • GPT2 Decoder by @fkodom in https://github.com/fkodom/clip-text-decoder/pull/2

    New Contributors

    • @fkodom made their first contribution in https://github.com/fkodom/clip-text-decoder/pull/2

    Full Changelog: https://github.com/fkodom/clip-text-decoder/compare/1.0.0...1.1.0

    Source code(tar.gz)
    Source code(zip)
  • 0.1.1(Nov 14, 2021)

  • 0.1.0(Nov 14, 2021)

Owner
Frank Odom
Director of Innovation at Plainsight. I like neural nets, and neural nets like me.
Frank Odom
Heterogeneous Deep Graph Infomax

Heterogeneous-Deep-Graph-Infomax Parameter Setting: HDGI-A: Node-level dimension: 16 Attention head: 4 Semantic-level attention vector: 8 learning rat

52 Oct 31, 2022
Implementation of "Semi-supervised Domain Adaptive Structure Learning"

Semi-supervised Domain Adaptive Structure Learning - ASDA This repo contains the source code and dataset for our ASDA paper. Illustration of the propo

3 Dec 13, 2021
Quickly and easily create / train a custom DeepDream model

Dream-Creator This project aims to simplify the process of creating a custom DeepDream model by using pretrained GoogleNet models and custom image dat

55 Dec 27, 2022
A foreign language learning aid using a neural network to predict probability of translating foreign words

Langy Langy is a reading-focused foreign language learning aid orientated towards young children. Reading is an activity that every child knows. It is

Shona Lowden 6 Nov 17, 2021
Self-Supervised Pre-Training for Transformer-Based Person Re-Identification

Self-Supervised Pre-Training for Transformer-Based Person Re-Identification [pdf] The official repository for Self-Supervised Pre-Training for Transfo

Hao Luo 116 Jan 04, 2023
PyTorch module to use OpenFace's nn4.small2.v1.t7 model

OpenFace for Pytorch Disclaimer: This codes require the input face-images that are aligned and cropped in the same way of the original OpenFace. * I m

Pete Tae-hoon Kim 176 Dec 12, 2022
CycleTransGAN-EVC: A CycleGAN-based Emotional Voice Conversion Model with Transformer

CycleTransGAN-EVC CycleTransGAN-EVC: A CycleGAN-based Emotional Voice Conversion Model with Transformer Demo emotion CycleTransGAN CycleTransGAN Cycle

24 Dec 15, 2022
Evidential Softmax for Sparse Multimodal Distributions in Deep Generative Models

Evidential Softmax for Sparse Multimodal Distributions in Deep Generative Models Abstract Many applications of generative models rely on the marginali

Stanford Intelligent Systems Laboratory 9 Jun 06, 2022
The PyTorch implementation of paper REST: Debiased Social Recommendation via Reconstructing Exposure Strategies

REST The PyTorch implementation of paper REST: Debiased Social Recommendation via Reconstructing Exposure Strategies. Usage Download dataset Download

DMIRLAB 2 Mar 13, 2022
Edge-oriented Convolution Block for Real-time Super Resolution on Mobile Devices, ACM Multimedia 2021

Codes for ECBSR Edge-oriented Convolution Block for Real-time Super Resolution on Mobile Devices Xindong Zhang, Hui Zeng, Lei Zhang ACM Multimedia 202

xindong zhang 236 Dec 26, 2022
Scalable implementation of Lee / Mykland (2012) and Ait-Sahalia / Jacod (2012) Jump tests for noisy high frequency data

JumpDetectR Name of QuantLet : JumpDetectR Published in : 'To be published as "Jump dynamics in high frequency crypto markets"' Description : 'Scala

LvB 12 Jan 01, 2023
Deepface is a lightweight face recognition and facial attribute analysis (age, gender, emotion and race) framework for python

deepface Deepface is a lightweight face recognition and facial attribute analysis (age, gender, emotion and race) framework for python. It is a hybrid

Kushal Shingote 2 Feb 10, 2022
A lightweight library to compare different PyTorch implementations of the same network architecture.

TorchBug is a lightweight library designed to compare two PyTorch implementations of the same network architecture. It allows you to count, and compar

Arjun Krishnakumar 5 Jan 02, 2023
ThunderSVM: A Fast SVM Library on GPUs and CPUs

What's new We have recently released ThunderGBM, a fast GBDT and Random Forest library on GPUs. add scikit-learn interface, see here Overview The miss

Xtra Computing Group 1.4k Dec 22, 2022
This is the repository for CVPR2021 Dynamic Metric Learning: Towards a Scalable Metric Space to Accommodate Multiple Semantic Scales

Intro This is the repository for CVPR2021 Dynamic Metric Learning: Towards a Scalable Metric Space to Accommodate Multiple Semantic Scales Vehicle Sam

39 Jul 21, 2022
Package to compute Mauve, a similarity score between neural text and human text. Install with `pip install mauve-text`.

MAUVE MAUVE is a library built on PyTorch and HuggingFace Transformers to measure the gap between neural text and human text with the eponymous MAUVE

Krishna Pillutla 182 Jan 02, 2023
A simple API wrapper for Discord interactions.

Your ultimate Discord interactions library for discord.py. About | Installation | Examples | Discord | PyPI About What is discord-py-interactions? dis

james 641 Jan 03, 2023
Pairwise model for commonlit competition

Pairwise model for commonlit competition To run: - install requirements - create input directory with train_folds.csv and other competition data - cd

abhishek thakur 45 Aug 31, 2022
Action Recognition for Self-Driving Cars

Action Recognition for Self-Driving Cars This repo contains the codes for the 2021 Fall semester project "Action Recognition for Self-Driving Cars" at

VITA lab at EPFL 3 Apr 07, 2022
Molecular Sets (MOSES): A benchmarking platform for molecular generation models

Molecular Sets (MOSES): A benchmarking platform for molecular generation models Deep generative models are rapidly becoming popular for the discovery

Neelesh C A 3 Oct 14, 2022