<!--Copyright 2021 The HuggingFace Team. All rights reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the ⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be rendered properly in your Markdown viewer. specific language governing permissions and limitations under the License. -->

This model was released on 2020-06-17 and added to Hugging Face Transformers on 2021-11-18.

ImageGPT

<div class="flex flex-wrap space-x-1"> </div>

Overview

The ImageGPT model was proposed in Generative Pretraining from Pixels by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever. ImageGPT (iGPT) is a GPT-2-like model trained to predict the next pixel value, allowing for both unconditional and conditional image generation.

The abstract from the paper is the following:

Inspired by progress in unsupervised representation learning for natural language, we examine whether similar models can learn useful representations for images. We train a sequence Transformer to auto-regressively predict pixels, without incorporating knowledge of the 2D input structure. Despite training on low-resolution ImageNet without labels, we find that a GPT-2 scale model learns strong image representations as measured by linear probing, fine-tuning, and low-data classification. On CIFAR-10, we achieve 96.3% accuracy with a linear probe, outperforming a supervised Wide ResNet, and 99.0% accuracy with full fine-tuning, matching the top supervised pre-trained models. We are also competitive with self-supervised benchmarks on ImageNet when substituting pixels for a VQVAE encoding, achieving 69.0% top-1 accuracy on a linear probe of our features.

<small> Summary of the approach. Taken from the original paper. </small>

This model was contributed by nielsr, based on this issue. The original code can be found here.

Usage tips

• ImageGPT is almost exactly the same as GPT-2, with the exception that a different activation function is used (namely "quick gelu"), and the layer normalization layers don't mean center the inputs. ImageGPT also doesn't have tied input- and output embeddings.
• As the time- and memory requirements of the attention mechanism of Transformers scales quadratically in the sequence length, the authors pre-trained ImageGPT on smaller input resolutions, such as 32x32 and 64x64. However, feeding a sequence of 32x32x3=3072 tokens from 0..255 into a Transformer is still prohibitively large. Therefore, the authors applied k-means clustering to the (R,G,B) pixel values with k=512. This way, we only have a 32*32 = 1024-long sequence, but now of integers in the range 0..511. So we are shrinking the sequence length at the cost of a bigger embedding matrix. In other words, the vocabulary size of ImageGPT is 512, + 1 for a special "start of sentence" (SOS) token, used at the beginning of every sequence. One can use [ImageGPTImageProcessor] to prepare images for the model.
• Despite being pre-trained entirely unsupervised (i.e. without the use of any labels), ImageGPT produces fairly performant image features useful for downstream tasks, such as image classification. The authors showed that the features in the middle of the network are the most performant, and can be used as-is to train a linear model (such as a sklearn logistic regression model for example). This is also referred to as "linear probing". Features can be easily obtained by first forwarding the image through the model, then specifying output_hidden_states=True, and then average-pool the hidden states at whatever layer you like.
• Alternatively, one can further fine-tune the entire model on a downstream dataset, similar to BERT. For this, you can use [ImageGPTForImageClassification].
• ImageGPT comes in different sizes: there's ImageGPT-small, ImageGPT-medium and ImageGPT-large. The authors did also train an XL variant, which they didn't release. The differences in size are summarized in the following table:

Model variant	Depths	Hidden sizes	Decoder hidden size	Params (M)	ImageNet-1k Top 1
MiT-b0	[2, 2, 2, 2]	[32, 64, 160, 256]	256	3.7	70.5
MiT-b1	[2, 2, 2, 2]	[64, 128, 320, 512]	256	14.0	78.7
MiT-b2	[3, 4, 6, 3]	[64, 128, 320, 512]	768	25.4	81.6
MiT-b3	[3, 4, 18, 3]	[64, 128, 320, 512]	768	45.2	83.1
MiT-b4	[3, 8, 27, 3]	[64, 128, 320, 512]	768	62.6	83.6
MiT-b5	[3, 6, 40, 3]	[64, 128, 320, 512]	768	82.0	83.8

Resources

A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ImageGPT.

<PipelineTag pipeline="image-classification"/>

• Demo notebooks for ImageGPT can be found here.
• [ImageGPTForImageClassification] is supported by this example script and notebook.
• See also: Image classification task guide

If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.

ImageGPTConfig

[[autodoc]] ImageGPTConfig

ImageGPTImageProcessor

[[autodoc]] ImageGPTImageProcessor - preprocess

ImageGPTImageProcessorPil

[[autodoc]] ImageGPTImageProcessorPil - preprocess

ImageGPTModel

[[autodoc]] ImageGPTModel - forward

ImageGPTForCausalImageModeling

[[autodoc]] ImageGPTForCausalImageModeling - forward

ImageGPTForImageClassification

[[autodoc]] ImageGPTForImageClassification - forward