jamesquirk / cnn-from-scratch Goto Github PK

Currently activation functions are only tested for post-FC layers - try with convs and pools and implement any required changes

Tests are not clean and not isolated enough.

Split out cnn.py into sub dirs and class files

Check whether the transpose operations in the vectorised calculation of dCdF are required in Conv2D layer backprop. It appears that they may cancel each other out.

Optimiser classes are not as clean as they could be.

Interface between Layers and Optimisers at the param update stage needs simplifying.

The aim of this repo is to act as a useful resource for somebody fairly new to CNNs who are looking to understand the inner workings in more detail.

The scope of this project is understanding how CNNs can be implemented using only basic Python libraries (NumPy). The documentation will therefore, describe how the cnn package works and ultimately how to implement a CNN in a repeatable manner; for an arbitrary number of layers.

Outline:

• Basic intro to CNNs
• Convolutional layers
• Pooling layers
• Flatten layers
• FC layers
• Activation (pseudo) layers
• Cost/ loss
• Optimisation algorithms

The objects would extend np.array objects to include additional information such as 'trainable' to avoid the inconvenience of containing them in dictionaries.

• Implement param objects
• refactor layer parameters to use the new object
• update optimisers (name check is not necessary)