An op-for-op PyTorch reimplementation of DeepMind's BigGAN model with the pre-trained weights from DeepMind.
This repository contains an op-for-op PyTorch reimplementation of DeepMind's BigGAN that was released with the paper Large Scale GAN Training for High Fidelity Natural Image Synthesis by Andrew Brock, Jeff Donahue and Karen Simonyan.
This PyTorch implementation of BigGAN is provided with the pretrained 128x128, 256x256 and 512x512 models by DeepMind. We also provide the scripts used to download and convert these models from the TensorFlow Hub models.
This reimplementation was done from the raw computation graph of the Tensorflow version and behave similarly to the TensorFlow version (variance of the output difference of the order of 1e-5).
This implementation currently only contains the generator as the weights of the discriminator were not released (although the structure of the discriminator is very similar to the generator so it could be added pretty easily. Tell me if you want to do a PR on that, I would be happy to help.)
This repo was tested on Python 3.6 and PyTorch 1.0.1
PyTorch pretrained BigGAN can be installed from pip as follows:
pip install pytorch-pretrained-biggan
If you simply want to play with the GAN this should be enough.
If you want to use the conversion scripts and the imagenet utilities, additional requirements are needed, in particular TensorFlow and NLTK. To install all the requirements please use the
git clone https://github.com/huggingface/pytorch-pretrained-BigGAN.git cd pytorch-pretrained-BigGAN pip install -r full_requirements.txt
This repository provide direct and simple access to the pretrained "deep" versions of BigGAN for 128, 256 and 512 pixels resolutions as described in the associated publication. Here are some details on the models:
BigGAN-deep-128: a 50.4M parameters model generating 128x128 pixels images, the model dump weights 201 MB,
BigGAN-deep-256: a 55.9M parameters model generating 256x256 pixels images, the model dump weights 224 MB,
BigGAN-deep-512: a 56.2M parameters model generating 512x512 pixels images, the model dump weights 225 MB.
Please refer to Appendix B of the paper for details on the architectures.
All models comprise pre-computed batch norm statistics for 51 truncation values between 0 and 1 (see Appendix C.1 in the paper for details).
Here is a quick-start example using
BigGAN with a pre-trained model.
See the doc section below for details on these classes and methods.
import torch from pytorch_pretrained_biggan import (BigGAN, one_hot_from_names, truncated_noise_sample, save_as_images, display_in_terminal) # OPTIONAL: if you want to have more information on what's happening, activate the logger as follows import logging logging.basicConfig(level=logging.INFO) # Load pre-trained model tokenizer (vocabulary) model = BigGAN.from_pretrained('biggan-deep-256') # Prepare a input truncation = 0.4 class_vector = one_hot_from_names(['soap bubble', 'coffee', 'mushroom'], batch_size=3) noise_vector = truncated_noise_sample(truncation=truncation, batch_size=3) # All in tensors noise_vector = torch.from_numpy(noise_vector) class_vector = torch.from_numpy(class_vector) # If you have a GPU, put everything on cuda noise_vector = noise_vector.to('cuda') class_vector = class_vector.to('cuda') model.to('cuda') # Generate an image with torch.no_grad(): output = model(noise_vector, class_vector, truncation) # If you have a GPU put back on CPU output = output.to('cpu') # If you have a sixtel compatible terminal you can display the images in the terminal # (see https://github.com/saitoha/libsixel for details) display_in_terminal(output) # Save results as png images save_as_images(output)
To load one of DeepMind's pre-trained models, instantiate a
BigGAN model with
model = BigGAN.from_pretrained(PRE_TRAINED_MODEL_NAME_OR_PATH, cache_dir=None)
PRE_TRAINED_MODEL_NAME_OR_PATH is either:
the shortcut name of a Google AI's or OpenAI's pre-trained model selected in the list:
biggan-deep-128: 12-layer, 768-hidden, 12-heads, 110M parameters
biggan-deep-256: 24-layer, 1024-hidden, 16-heads, 340M parameters
biggan-deep-512: 12-layer, 768-hidden, 12-heads , 110M parameters
a path or url to a pretrained model archive containing:
config.json: a configuration file for the model, and
pytorch_model.bina PyTorch dump of a pre-trained instance of
BigGAN(saved with the usual
PRE_TRAINED_MODEL_NAME_OR_PATH is a shortcut name, the pre-trained weights will be downloaded from AWS S3 (see the links here) and stored in a cache folder to avoid future download (the cache folder can be found at
cache_dir can be an optional path to a specific directory to download and cache the pre-trained model weights.
BigGANConfig is a class to store and load BigGAN configurations. It's defined in
Here are some details on the attributes:
output_dim: output resolution of the GAN (128, 256 or 512) for the pre-trained models,
z_dim: size of the noise vector (128 for the pre-trained models).
class_embed_dim: size of the class embedding vectors (128 for the pre-trained models).
channel_width: size of each channel (128 for the pre-trained models).
num_classes: number of classes in the training dataset, like imagenet (1000 for the pre-trained models).
layers: A list of layers definition. Each definition for a layer is a triple of [up-sample in the layer ? (bool), number of input channels (int), number of output channels (int)]
attention_layer_position: Position of the self-attention layer in the layer hierarchy (8 for the pre-trained models).
eps: epsilon value to use for spectral and batch normalization layers (1e-4 for the pre-trained models).
n_stats: number of pre-computed statistics for the batch normalization layers associated to various truncation values between 0 and 1 (51 for the pre-trained models).
BigGAN is a PyTorch model (
torch.nn.Module) of BigGAN defined in
model.py. This model comprises the class embeddings (a linear layer) and the generator with a series of convolutions and conditional batch norms. The discriminator is currently not implemented since pre-trained weights have not been released for it.
The inputs and output are identical to the TensorFlow model inputs and outputs.
We detail them here.
BigGAN takes as inputs:
z: a torch.FloatTensor of shape [batch_size, config.z_dim] with noise sampled from a truncated normal distribution, and
class_label: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token types indices selected in [0, 1]. Type 0 corresponds to a
sentence Aand type 1 corresponds to a
sentence Btoken (see BERT paper for more details).
truncation: a float between 0 (not comprised) and 1. The truncation of the truncated normal used for creating the noise vector. This truncation value is used to selecte between a set of pre-computed statistics (means and variances) for the batch norm layers.
BigGAN outputs an array of shape [batch_size, 3, resolution, resolution] where resolution is 128, 256 or 512 depending of the model:
We provide a few utility method to use the model. They are defined in
Here are some details on these methods:
truncated_noise_sample(batch_size=1, dim_z=128, truncation=1., seed=None):
Create a truncated noise vector.
Convert an output tensor from BigGAN in a list of images.
Convert and save an output tensor from BigGAN in a list of saved images.
Convert and display an output tensor from BigGAN in the terminal. This function use
libsixel and will only work in a libsixel-compatible terminal. Please refer to https://awesomeopensource.com/project/saitoha/libsixel for more details.
Create a one-hot vector from a class index or a list of class indices.
len(int_or_list) == batch_size
Create a one-hot vector from the name of an imagenet class ('tennis ball', 'daisy', ...). We use NLTK's wordnet search to try to find the relevant synset of ImageNet and take the first one. If we can't find it direcly, we look at the hyponyms and hypernyms of the class name.
Scripts to download and convert the TensorFlow models from TensorFlow Hub are provided in ./scripts.
The scripts can be used directly as: