Notebooks

Name	Description	Category	Link
Training pix2pix	This notebook shows a simple pipeline for training pix2pix on a simple dataset. Most of the code is based on this implementation.	GAN
One Place	This notebook shows how to train, test then deploy models in the browser directly from one notebook. We use a simple XOR example to prove this simple concept.	Deployment
TPU vs GPU	Google recently allowed training on TPUs for free on colab. This notebook explains how to enable TPU training. Also, it reports some benchmarks using mnist dataset by comparing TPU and GPU performance.	TPU
Keras Custom Data Generator	This notebook shows to create a custom data genertor in keras.	Data Generatation
Eager Execution (1)	As we know that TenosrFlow works with static graphs. So, first you have to create the graph then execute it later. This makes debugging a bit complicated. With Eager Execution you can now evalute operations directly without creating a session.	Dynamic Graphs
Eager Execution (2)	In this notebook I explain different concepts in eager execution. I go over variables, ops, gradients, custom gradients, callbacks, metrics and creating models with tf.keras and saving/restoring them.	Dynamic Graphs
Sketcher	Create a simple app to recognize 100 drawings from the quickdraw dataset. A simple CNN model is created and served to deoploy in the browser to create a sketch recognizer app.	Deployment
QuickDraw10	In this notebook we provide QuickDraw10 as an alternative for MNIST. A script is provided to download and load a preprocessed dataset for 10 classes with training and testing split. Also, a simple CNN model is implemented for training and testing.	Data Preperation
Autoencoders	Autoencoders consists of two structures: the encoder and the decoder. The encoder network downsamples the data into lower dimensions and the decoder network reconstructs the original data from the lower dimension representation. The lower dimension representation is usually called latent space representation.	Auto-encoder
Weight Transfer	In this tutorial we explain how to transfer weights from a static graph model built with TensorFlow to a dynamic graph built with Keras. We will first train a model using Tensorflow then we will create the same model in keras and transfer the trained weights between the two models.	Weights Save and Load
BigGan (1)	Create some cool gifs by interpolation in the latent space of the BigGan model. The model is imported from tensorflow hub.	GAN
BigGan (2)	In this notebook I give a basic introduction to bigGans. I also, how to interpolate between z-vector values. Moreover, I show the results of multiple experiments I made in the latent space of BigGans.	GAN
Mask R-CNN	In this notebook a pretrained Mask R-CNN model is used to predict the bounding box and the segmentation mask of objects. I used this notebook to create the dataset for training the pix2pix model.	Segmentation
QuickDraw Strokes	A notebook exploring the drawing data of quickdraw. I also illustrate how to make a cool animation of the drawing process in colab.	Data Preperation
U-Net	The U-Net model is a simple fully convolutional neural network that is used for binary segmentation i.e foreground and background pixel-wise classification. In this notebook we use it to segment cats and dogs from arbitrary images.	Segmentation
Localizer	A simple CNN with a regression branch to predict bounding box parameters. The model is trained on a dataset of dogs and cats with bounding box annotations around the head of the pets.	Object Localization
Classification and Localization	We create a simple CNN with two branches for classification and locazliation of cats and dogs.	Classification, Localization
Transfer Learning	A notebook about using Mobilenet for transfer learning in TensorFlow. The model is very fast and achieves 97% validation accuracy on a binary classification dataset.	Transfer Learning
Hand Detection	In this task we want to localize the right and left hands for each person that exists in a single frame. It acheives around 0.85 IoU.	Detection
Face Detection	In this task we used a simple version of SSD for face detection. The model was trained on less than 3K images using TensorFlow with eager execution	Detection
TensorFlow 2.0	In this task we use the brand new TF 2.0 with default eager execution. We explore, tensors, gradients, dataset and many more.	Platform
SC-FEGAN	In this notebook, you can play directly with the SC-FEGAN for face-editting directly in the browser.	GAN
Swift for TensorFlow	Swift for TensorFlow is a next-generation platform for machine learning that incorporates differentiable programming. In this notebook a go over its basics and also how to create a simple NN and CNN.	Platform
GCN	Ever asked yourself how to use convolution networks for non Euclidean data for instance graphs ? GCNs are becoming increasingly popular to solve such problems. I used Deep GCNs to classify spammers & non-spammers.	Platform