Posts

It’s been around 5 months since I released my last blog on DCGAN Review and Implementation using PyTorch C++ API and I’ve missed writing blogs badly! Straight the to the point, I’m back! But before we start, the PyTorch C++ Frontend has gone through several changes and thanks to the awesome contributors around the world, it resembles the Python API more than it ever did! Since a lot of things have changed, I have also updated my previous blogs (tested on 1.4 Stable build). ...

I’m pleased to start a series of blogs on GANs and their implementation with PyTorch C++ API. We’ll be starting with one of the initial GANs - DCGANs (Deep Convolutional Generative Adversarial Networks). The authors (Soumith Chintala, Radford and Luke Metz) in this Seminal Paper on DCGANs introduced DCGANs to the world like this: We introduce a class of CNNs called deep convolutional generative adversarial networks (DCGANs), that have certain architectural constraints, and demonstrate that they are a strong candidate for unsupervised learning. Training on various image datasets, we show convincing evidence that our deep convolutional adversarial pair learns a hierarchy of representations from object parts to scenes in both the generator and discriminator. Additionally, we use the learned features for novel tasks - demonstrating their applicability as general image representations. ...

Introduction to Xeus Cling Today, we are going to run our C++ codes in the Jupyter Notebook. Sounds ambitious? Not much. Let’s see how we do it using Xeus Cling. I’ll quote the definition of Xeus Cling on the official documentation website. xeus-cling is a Jupyter kernel for C++ based on the C++ interpreter cling and the native implementation of the Jupyter protocol xeus. Just like we use Python Kernel in the Jupyter Notebook, we can also use a C++ based interpreter cling combined with a Jupyter protocol called Xeus to reach closer to implementing C++ code in the notebook. ...

Transfer Learning – Before we go ahead and discuss the Why question of Transfer Learning, let’s have a look at What is Transfer Learning? Let’s have a look at the Notes from CS231n on Transfer Learning: In practice, very few people train an entire Convolutional Network from scratch (with random initialization), because it is relatively rare to have a dataset of sufficient size. Instead, it is common to pretrain a ConvNet on a very large dataset (e.g. ImageNet, which contains 1.2 million images with 1000 categories), and then use the ConvNet either as an initialization or a fixed feature extractor for the task of interest. ...

In the last blog, we had discussed all but training and results of our custom CNN network on Dogs vs Cats dataset. Today, we’ll be making some small changes in the network and discussing training and results of the task. I’ll start with the network overview again, where we used a network similar to VGG-16 (with one extra Fully Connected Layer in the end). While there are absolutely no problems with that network, but since the dataset contains a lot of images (25000 in training dataset) and we were using (200x200x3) input shape to the network (which is 120,000 floating point numbers), this leads to high memory consumption. In short, I was out of RAM to store these many images during program execution. ...

Hi Everyone! So excited to be back with another blog in the series of PyTorch C++ Blogs. Today, we are going to see a practical example of applying a CNN to a Custom Dataset - Dogs vs Cats. This is going to be a short post of showing results and discussion about hyperparameters and loss functions for the task, as code snippets and explanation has been provided here, here and here. ...

Recap of the last blog Before we move on, it’s important what we covered in the last blog. We’ll be going forward from loading Custom Dataset to now using the dataset to train our VGG-16 Network. Previously, we were able to load our custom dataset using the following template: ...

I’m happy to announce a Series of Blog Posts on PyTorch C++ API. Check out the blogs in the series here. Happy Reading!

Overview: How C++ API loads data? In the last blog, we discussed application of a VGG-16 Network on MNIST Data. For those, who are reading this blog for the first time, here is how we had loaded MNIST data: auto data_loader = torch::data::make_data_loader<torch::data::samplers::SequentialSampler>( std::move(torch::data::datasets::MNIST("../../data").map(torch::data::transforms::Normalize<>(0.13707, 0.3081))).map( torch::data::transforms::Stack<>()), 64); Let’s break this piece by piece, as for beginners, this may be unclear. First, we ask the C++ API to load data (images and labels) into tensors. ...

Environment Setup [Ubuntu 16.04, 18.04] Note: If you have already finished installing PyTorch C++ API, please skip this section. Download libtorch: CPU Version: wget https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-latest.zip -O libtorch.zip GPU Version (CUDA 9.0): wget https://download.pytorch.org/libtorch/cu90/libtorch-shared-with-deps-latest.zip -O libtorch.zip GPU Version (CUDA 10.0): wget https://download.pytorch.org/libtorch/cu100/libtorch-shared-with-deps-latest.zip Unzip libtorch.zip: unzip libtorch.zip We’ll use the absolute path of extracted directory (libtorch) later on. Implementation The VGG-16 Network is shown in the Figure below. We’ll start of by first including libtorch header file. ...