Torchélie is a set of tools for PyTorch. It includes losses, optimizers, algorithms, utils, layers, models and training loops.

Feedback is absolutely welcome.

You may want to read the detailed docs

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

It depends on Pytorch (obvi), and has an optional dependency on OpenCV for some transforms (Canny, as of today). It also depends on Visdom for realtime visualizations, plotting, etc.

To install visdom: pip install visdom. Then, you need to run a Visdom server with python -m visdom.server, direct your browser to http://localhost:8097. Now you're ready to use VisdomLogger and enjoy realtime tracking of your experiments.

Torchelie API is beta and can be a bit unstable. Minor breaking changes can happen.

Code, README, docs and tests might be out of sync in general. Please tell me if you notice anything wrong.

Let's say you want to do the hello-world of deep learning: MNIST handwritten digits classification. Let's also assume that you already have your training and testing datasets organised properly, e.g. coming from the Kaggle archive:

$ tree mnist_png

mnist_png
├── testing
│   ├── 0
│   ├── 1
│   ├── 2
│   ├── 3
│   ├── 4
│   ├── 5
│   ├── 6
│   ├── 7
│   ├── 8
│   └── 9
└── training
    ├── 0
    ├── 1
    ├── 2
    │   ├── 10009.png
    │   ├── 10016.png
    │   └── [...]
    ├── 3
    ├── 4
    ├── 5
    ├── 6
    ├── 7
    ├── 8
    └── 9

Torchelie comes with a classification "recipe" out-of-the-box, which can be used directly to train your a model straight from the command line:

$ python3 -m torchelie.recipes.classification --trainset mnist_png/training --testset mnist_png/testing

[...]
 | Ep. 0 It 1 | {'lr_0': '0.0100', 'acc': '0.0938', 'loss': '3.1385'}
 | Ep. 0 It 11 | {'lr_0': '0.0100', 'acc': '0.2017', 'loss': '2.4109'}
 | Ep. 0 It 21 | {'lr_0': '0.0100', 'acc': '0.3185', 'loss': '2.0410'}
 | Ep. 0 It 31 | {'lr_0': '0.0100', 'acc': '0.3831', 'loss': '1.8387'}
 | Ep. 0 It 41 | {'lr_0': '0.0100', 'acc': '0.4451', 'loss': '1.6513'}
[...]
Test | Ep. 1 It 526 | [...] 'acc': '0.9799', 'loss': '0.0797' [...]
 | Ep. 1 It 556 | {'lr_0': '0.0100', 'acc': '0.9588', 'loss': '0.1362'}
 | Ep. 1 It 566 | {'lr_0': '0.0100', 'acc': '0.9606', 'loss': '0.1341'}

Want to run it on your laptop which doesnt have a GPU? Simply add the --device cpu option!

With a simple use case and a properly organized dataset, we already saw how Torchelie can help experiment quickly. But what just happened?

The classification recipe is a whole ready-to-use training loop which:

• handles all the image loading
• uses the ResNet18 model from PyTorch's Torchvision to classify images from the training dataset
• computes a cross entropy loss on the predicted outputs
• uses RAdamW to optimize the model along the way
• periodically (default every 1k iterations) assess the accuracy of the trained model using the test dataset
• gives as much insights as possible during the training through:
  • stdout (as shown above)
  • visdom (TODO)

The cool thing is that all these building blocks are available!

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.

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.

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:

• CondSeq 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
• SpadeResBlock instead uses Spade2d
• AutoGANGenBlock is a block for AutoGAN
• SNResidualDiscrBlock is a residual block with spectral normalization

• 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
• attention56 from Residual Attention Networks

Debug models:

• VggDebug
• ResNetDebug
• PreactResNetDebug

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)

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

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.
• ResizedCrop: deterministic version of torchvision.transforms.RandomResizedCrop

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

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.
• OneCycle implements 1cycle policy

• 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.
• WithIndexDataset returns the sample's index as well. Useful if you want to retrieve the sample or associate something to it.
• CachedDataset lazily caches a dataset so that next iterations won't access the original storage or recompute the initial dataset's transforms

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

• 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

• 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

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.

• 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.

Since OpenCV is an optional dependency, you might want to run tests in such a setup (therefore not testing Canny). You can do so by excluding the require_opencv pytest custom marker like so:

pytest -m "not require_opencv"

Code is formatted using YAPF.

For now, the CI doesn't check for code format, and the config files for yapf isn't there, but do your best to format your code using YAPF (or at least comply with PEP8 🙂)

Code is linted using Flake8. Do your best to send code that don't make it scream too loud 😉

You can run it like this:

flake8 torchelie

Despite typing being optional in Python, type hints can save a lot of time on a project such as Torchélie. This project is type-checked using mypy. Make sure it passes successfully, and consider adding type hints where it makes sense to do so when contributing code!

You can run it like this:

mypy torchelie

Common widespread naming best practices apply.

That being said, please specifically try to avoid using l as a variable name, even for iterators. First, because of E741 (see PEP8 "names to avoid"), second because in the context of Torchélie it might mean layer, label, loss, length, line, or other words that are spread among the codebase. Therefore, using l would make it considerably harder to understand code when reading it.