- how do you get your editor to show code hints from python docstrings?
- where can we find a graph of the softmax function?
- from dataaspirant.com
- what is the purpose and function of the "moment" parameters (beta1 and beta2) in the Adam optimization algorithm?
- what tools are available for visualizing the layers in a neural network?
- TensorBoard's Graph Visualization Tool draws diagrams for tensorflow models (not sure how to get this to work with Keras)
- what tools are available for visualizing saliency maps for a neural network?
- keras-vis is a popular library for creating saliency maps
- what tools are available for visualizing connections between neural network states and input data with a "brain-like" diagram?
- what is the meaning of the "interpolation" parameter to the ImageDataGenerator class in Keras?
- Keras's flow_from_directory docs seem to indicate that this interpolation parameter is used when trying to adjust an image to fit the required input size
- It looks like these resizing options are coming from the PIL Image.resize routine
- There's more explanation of these options in these videos
- Keras's flow_from_directory docs seem to indicate that this interpolation parameter is used when trying to adjust an image to fit the required input size
##Introduction
This repo contains python notebooks and info for the Fall 2019 IUSB Deep Learning Class.
The data is from Kaggle's Cats and Dogs challenge.
https://www.kaggle.com/c/dogs-vs-cats
- Reiterating the importance of training data Dogs/cats
- Keras's functional vs. sequential APIs
- Optimizers (ADAM, SGD)
- Learning rate
- Activation functions
- final dense layer (sigmoid for binary classification vs. softmax for choosing from multiple classes)
- hidden layers (usually relu or some relu variant)
Discussing them what an HDF5 "file" is would be useful. (It's a zip file, but instead of files, it holds data structures)
- data set naming (train, test, holdout VS train, valid, holdout VS train, valid, test VS ...)
- emphasize that mnist and dogs-cats don't have a holdout set
- emphasize that different names are used in different contexts
- emphasize that you have to look it up each time through
Discuss Overfitting and Local Minima/Maxima
Local Minima/Maxima https://youtu.be/IHZwWFHWa-w?t=409
Overfitting discussion https://machinelearningmastery.com/introduction-to-regularization-to-reduce-overfitting-and-improve-generalization-error/