- Update code to work with TF 1.0+ and Keras 2.0+
- Reduce the size of the repo by compressing png better and reducing the resolution of very large images.
- (Semi)-automate the rendering of the notebooks somehow...
- Systematically include architecture diagrams in notebooks, possibly using using the name of Keras classes in nodes.
-
Highlight the importance of time-based cross-validation splits and other cross-validation splits: to measure the ability of the model to generalize either to the future, to new users or to new items.
-
Lecture: do not tie explicit feedback to regression metrics, it would be possible to use ranking metrics for explicit feedback.
-
Embedding diagrams should include the one-hot vector of the data points that is multiplied with the embedding matrices to emphasize the fact that embedding matrices holds model parameters and not training data (the training sample is the one-hot vector representation of the user / item).
-
Introduce convolution using the Keras API only. This can be done by using using the following test construct:
Sequential([Convolution2D(**conv_params)]).predict(test_img) -
Add exercises that ask to estimate the parameter size and shape of the output of a stack of Conv + Pooling layers and then ask to use keras to write a program to empirically check the results.
-
Factorize out the PASCAL VOC annotation extraction in a helper module to hide the complexity of setting up the learning task so as to focus the students attention on the model architecture and less on the complexity of annotation preprocessing.
-
Introduce the matplotlib utility to display bounding box earlier so as to display samples from the training set before introducing the ground truth representation of the labels, the IoU think and the models themselves.
-
Use the matplotlib utility to display pairs of overlapping boxes and their IoU value on a random image. On the of the box can be the true annotation from the sample and the other an arbitrary overlapping bbox.
-
Do a cleaner train / validation split, once and for all at the beginning of the notebook.