Torchélie

Torchelie is currently very API unstable. Test at your own risks, and you should expect to update often and have many breaking changes, especially in high level features.

Feedback is absolutely welcome.

You may want to read the detailed docs

Installation

pip install git+https://github.com/vermeille/Torchelie

torchelie.recipes

Classes implementing full algorithms, from training to usage

• NeuralStyleRecipe implements Gatys' Neural Artistic Style. Also directly usable with commandline with python3 -m torchelie.recipes.neural_style

• FeatureVisRecipe implements feature visualization through backprop. The image is implemented in Fourier space which makes it powerful (see this and that ). Usable as commandline as well with python -m torchelie.recipes.feature_vis.

• DeepDreamRecipe implements something close to Deep Dream. python -m torchelie.recipes.deepdream works.

• Classification trains a model for image classification. It provides logging of loss and accuracy. It has a commandline interface with python3 -m torchelie.recipes.classification to quickly train a classifier on an image folder with train images and another with test images.

torchelie.utils

Functions:

• freeze and unfreeze that changes requires_grad for all tensor in a module.

• entropy(x, dim, reduce) computes the entropy of x along dimension dim, assuming it represents the unnormalized probabilities of a categorial distribution.

• kaiming(m) / xavier(m) returns m after a kaiming / xavier initialization of m.weight

• nb_parameters returns the number of trainables parameters in a module

• layer_by_name finds a module by its (instance) name in a module

• gram / bgram compute gram and batched gam matrices.

• DetachedModule wraps a module so that it's not detected by recursive module functions.

• FrozenModule wraps a module, freezes it and sets it to eval mode. All calls to .train() (even those made from enclosing modules) will be ignored.

torchelie.nn

Debug modules:

• Dummy does nothing to its input.

• Debug doesn't modify its input but prints some statistics. Easy to spot exploding or vanishing values.

Normalization modules:

• ImageNetInputNorm for normalizing images like torchvision.model wants them.

• MovingAverageBN2d, NoAffineMABN2d and ConditionalMABN2d are the same as above, except they also use moving average of the statistics at train time for greater stability. Useful ie for GANs if you can't use a big ass batch size and BN introduces too much noise.

• AdaIN2d is adaptive instancenorm for style transfer and stylegan.

• Spade2d / MovingAverageSpade2d, for GauGAN.

• PixelNorm from ProGAN and StyleGAN.

• BatchNorm2d, NoAffineBatchNorm2d should be strictly equivalent to Pytorch's, and ConditionalBN2d gets its weight and bias parameter from a linear projection of a z vector.

• AttenNorm2d BN with attention (Attentive Normalization, Li et al, 2019)

Misc modules:

• FiLM2d is affine conditioning f(z) * x + g(z).

• Noise returns x + a * z where a is a learnable scalar, and z is a gaussian noise of the same shape of x

• Reshape(*shape) applies x.view(x.shape[0], *shape).

• VQ is a VectorQuantization layer, embedding the VQ-VAE loss in its backward pass for a great ease of use.

Container modules:

• ConditionalSequential is an extension of nn.Sequential that also applies a second input on the layers having condition()

Model manipulation modules:

• WithSavedActivations(model, types) saves all activations of model for its layers of instance types and returns a dict of activations in the forward pass instead of just the last value. Forward takes a detach boolean arguments if the activations must be detached or not.

Net Blocks:

• MaskedConv2d is a masked convolution for PixelCNN

• TopLeftConv2d is the convolution from PixelCNN made of two conv blocks: one on top, another on the left.

• Conv2d, Conv3x3, Conv1x1, Conv2dBNReLU, Conv2dCondBNReLU, etc. Many different convenience blocks in torchelie.nn.blocks.py

• ResNetBlock, PreactResNetBlock

• ResBlock is a classical residual block with batchnorm

• ClassConditionalResBlock

• ConditionalResBlock instead uses ConditionalBN2d

• SpadeResBlock instead uses Spade2d

• AutoGANGenBlock is a block for AutoGAN

• SNResidualDiscrBlock is a residual block with spectral normalization

torchelie.models

• VggBNBone is a parameterizable stack of convs vgg style. Look at VggDebug for its usage.

• ResNetBone for resnet style bone.

• Classifier adds two linear layers to a bone for classification.

• Patch16, Patch32, Patch70, Patch286 are Pix2Pix's PatchGAN's discriminators

• UNet for image segmentation

• AutoGAN generator from the paper AutoGAN: Neural Architecture Search for Generative Adversarial Networks

• ResNet discriminator with spectral normalization

• PerceptualNet is a VGG16 with correctly named layers for more convenient use with WithSavedActivations

Debug models:

• VggDebug

• ResNetDebug

• PreactResNetDebug

torchelie.loss

Modules:

• PerceptualLoss(l) is a vgg16 based perceptual loss up to layer number l. Sum of L1 distances between x's and y's activations in vgg. Only x is backproped.

• NeuralStyleLoss

• OrthoLoss orthogonal loss.

• TotalVariationLoss TV prior on 2D images.

• ContinuousCEWithLogits is a Cross Entropy loss that allows non categorical targets.

• TemperedCrossEntropyLoss from Robust Bi-Tempered Logistic Loss Based on Bregman Divergences (Amid et al, 2019)

Functions (torchelie.loss.functional):

• ortho(x) applies an orthogonal regularizer as in Brock et al (2018) (BigGAN)

• total_variation(x) applies a spatial L1 loss on 2D tensors

• continuous_cross_entropy

• tempered_cross_entropy from Robust Bi-Tempered Logistic Loss Based on Bregman Divergences (Amid et al, 2019)

torchelie.loss.gan

Each submodule is a GAN loss function. They all contain three methods: real(x) and fake(x) to train the discriminator, and ŋenerated(x) to improve the Generator.

Available:

• Standard loss (BCE)

• Hinge

torchelie.transforms

Torchvision-like transforms:

• ResizeNoCrop resizes the longest border of an image ot a given size, instead of torchvision that resize the smallest side. The image is then smaller than the given size and needs padding for batching.

• AdaptPad pads an image so that it fits the target size.

• Canny runs canny edge detector (requires OpenCV)

• MultiBranch allows different transformations branches in order to transform the same image in different ways. Useful for self supervision tasks for instance.

torchelie.transforms.differentiable

Contains some transforms that can be backpropagated through. Its API is unstable now.

torchelie.lr_scheduler

Classes:

• CurriculumScheduler takes a lr schedule and an optimizer as argument. Call sched.step() on each batch. The lr will be interpolated linearly between keypoints.

torchelie.datasets

• HorizontalConcatDataset concatenates multiple datasets. However, while torchvision's ConcatDataset just concatenates samples, torchelie's also relabels classes. While a vertical concat like torchvision's is useful to add more examples per class, an horizontal concat merges datasets to more classes.

• PairedDataset takes to datasets and returns the cartesian products of its samples.

• MixUpDataset takes a dataset, sample all pairs and interpolates samples and labels with a random mixing value.

• NoexceptDataset wraps a dataset and suppresses the exceptions raised while loading samples. Useful in case of a big downloaded dataset with corrupted samples for instance.

torchelie.datasets.debug

• ColoredColumns / ColoredRows are datasets of precedurally generated images of rows / columns randomly colorized.

torchelie.metrics

• WindowAvg: averages measures over a k-long sequence

• ExponentialAvg: applies an exponential averaging method over measures

• RunningAvg: accumulates total number of items and sum to provide an accurate average estimation

torchelie.opt

• DeepDreamOptim is the optimizer used by DeepDream

• AddSign from Neural Optimiser search with Reinforcment learning

• RAdamW from On the Variance of the Adaptive Learning Rate and Beyond, with AdamW weight decay fix.

• Lookahead from Lookahead Optimizer: k steps forward, 1 step back

torchelie.data_learning

Data parameterization for optimization, like neural style or feature viz.

Modules:

• PixelImage an image to be optimized.

• SpectralImage an image Fourier-parameterized to ease optimization.

• CorrelateColors assumes the input is an image with decorrelated color components. It correlates back the color using some ImageNet precomputed correlation statistics to ease optimization.

torchelie.industry

• ONNXModel wraps the boilerplate code to load an ONNX model and run it.

• TorchONNXModel wraps an ONNXModel and transforms inputs and outputs from numpy to torch.

Testing

• classification.py tests bones for classifiers on MNIST or CIFAR10

• conditional.py tests class conditional layers with a conditional classification task argmin L(f(x, z), y) where x is a MNIST sample, z a class label, and y = 1 if z is the correct label for x, 0 otherwise.