The FastAI course seems to be a GANtastic journey (quote by Jakub Langr and Vladimir Bok). After seeing the LAMB optimizer (You et al. 2019) I was completely astonished by the connection between my favorite subjects: Mathematics and Programming showing how we can translate 10 lines of a Mathematical formula into 10 lines of code.
Math –> Code
Code –> Math
It has been love at first sight. It was similar to the one I had for Visual Programming Languages in 2013 when I realized that I can drive my design with mathematical formulas in a parametric and procedural way with algorithms and now thanks to deep learning I can define a mathematical formula based on my design.
Math –> Design
Design –> Math
This passion is driven by the idea of cross-pollination; applying transfer learning between different disciplines and industries. One discipline that is fascinating to me is
Geometrical Deep Learning (GDL). Working in the Non-euclidean space has been always challenging for me and the representation of non-euclidean geometry produces more often inductive bias. You can find GDL applied with Graph neural networks (GNNs) , Graph Convolutional Network (GCN) and many others, it allows us to take advantage of data with inherent relationships, connections, and shared properties.
I arrived here to Silicon Valley almost 3 years ago with a grant for Artificial Intelligence and Building Information Modeling but unfortunately, I had to realize that the market, companies, and technologies were not ready yet for this leap in the Architecture Engineering and Construction. Companies were complaining regarding their lack of big data, without focusing on the quality of the data and a clever approach to the problem. Finally, I can quote someone who strongly believes that you do NOT need a lot of data to do deep learning. This person is Jeremy Howard: he has proved several times that. He defeated giant companies such as Google, IBM, and others in Deep Learning competitions using smarter techniques such as Transfer Learning and Data Augmentation. These are fundamental in order to lay down the basic foundation for cross-pollination.
“Although many have claimed that you need Google-size data sets to do deep learning, this is false. The power of transfer learning (combined with techniques like data augmentation) makes it possible for people to apply pre-trained models to much smaller datasets.”Jeremy Howard, Fastai
During this blog, I will try to use a transfer learning approach as much as possible.
fine_tuning : Transfer Learning
fit_one_cycle : Training from Scratch
“Transfer Learning: Using a pre-trained model for a task different to what it was originally trained for.“Jeremy Howard, Deep Learning for Coders without a Ph.D.
Total trainable params: 21,813,056
These are the parameters of resnet34 architecture after we changed the Head with the new layer. And in the dls ( dataloader ) we modified the dataset with aug_transforms and then we applied the fine-tuning during in the learning part. I combined Chapter 1 and Chapter 2 from Jeremy in this code.
dls = ImageDataLoaders.from_name_func(
path, get_image_files(path), valid_pct=0.2, seed=42,
label_func=is_cat, item_tfms=RandomResizedCrop(224, min_scale=0.5),
learn = cnn_learner(dls, resnet34, metrics=error_rate)
I can’t wait to show you in what project I am working on right now.