The Python environment for deep learning has a new golden boy in the form of the PyTorch package. It seems faster and less convoluted than Tensorflow, other than more scalable. At this stage, it works for Linux and MacOs only, so I am going to need a double boot on my Win 10 laptop, again, as for Salome Meca 2017 for nuclear engineering design. Good!

GitHub not

I am reconsidering my commitment to GitHub for showing my data science & deep learning code. Actually, I will not put anything into my account there for the time being. I am neither a public persona living off conferences nor a developer looking for a job, so GitHub is worth very little to me. Fundamentally, I am hearing more and more stories about GitHub accounts being raided by unscrupolous firms who steal the code and resell it under their copyright without a problem. I prefer to keep my things private, then, only releasing reproducible results in a formal manner.

Kaggle Planet

My first image classification, deep learning competition on Kaggle ended last night. It was Kaggle Planet and I finished 172/938 , unimpressive but pretty solid for a noob. It was a multilabel classification problem with 100k images from the Amazon Forest, for which I implemented a few state of the art architectures in Keras and ensembled their results. Help came from public kernels and forum topics, so a great learning experience on top of the MOOC by Jeremy Howard and the latest Deep Learning with Python book by Francois Chollet. Now putting all together for a post-competition Jupyter notebook I can reuse for similar tasks in the future. Thanks, Kaggle!


I am becoming pretty used to Keras for image classification deep learning but still need control of the size and the time required to train and test my networks. One procedure I absolutely need to master in a short time is fit_generator() , which makes training / testing computations batch by batch instead of storing all the images in memory, 16GB RAM going away pretty fast. This would help me process very big datasets and bigger image sizes than the 64×64 pixels I am using now. DenseNet-121 is ok with several tricks and 5 k-folds already, but ResNet-50 and Inception-4 are not yet, both requiring 224×224 images.

DL Python

A new book by Francois Chollet, the inventor of, is going to be released in Oct 2017 under the title Deep Learning with Python . It will present the state of art neural networks architectures and implementations, starting from Theano and TensorFlow libraries on top of his own excellent Keras wrapper. The ebook format is available in pre-release already at a 42% discount, coupon code: deeplearning. The ebook is going to save me a lot of time, just purchased and planning to study extensively this week.

DL pretrained

I am using one current competition from to make experiments with deep learning. They employ pretrained / weighted neural network models for image classification distributed by Keras for the Amazon Planet competition. Pretraining is a way to transfer learning, that is to say to use models trained on standard image datasets to make predictions on new image datasets, only changing the last layer or two in the model. My new laptop with a GTX 1070 8 GB graphics card comes just handy for working locally with the VGG-16 and ResNet-50 models, my best compromise between hardware affordability and results accuracy. Better and smarter approaches are described in the forums here, food for thought!