Convolutional Neural Network

This is a template implementation of a convolutional neural network for regression or classification tasks. When calling the script it is possible to specify which case we are interested in and the network will adapt accordingly.

Comet ML

When running the script, every (hyper)parameter is synced on Comet ML, and so are the training/validation curves. In order to be able to see the results, all you have to do is insert you API key in comet.config in the space I have provided for you. Also, please add the name project name as this is important when logging the experiments.

Input organisation

Place in Input/Images the images you wish to classify or apply a regression upon. Similarly, in Input/Labels put either the label or the quantitative value corresponding to the image at hand. Notice: image and label must have the same name across directories, e.g. Input/Images/foo.jpg ans Input/Labels/foo.yaml. If this condition is not met, it will not be possible to associate each image with the corresponding label.

Running the script

In order to run the script it is sufficient to type

python main.py

However, there are a number of (hyper) parameters that can be passed to specify a few details.

[-h] # Print Help

[--project [PROJECT]] # String representing the name of the project (for logging purposes)

[--feature [FEATURE]] # String representing the feature to be considered (in the case of labels in yaml file)

[--test_size [TEST_SIZE]] # Float in (0,1) representing the percentage to be kept for testing

[--validation_size [VALIDATION_SIZE]] # Float in (0,1) representing the percentageto be kept for validation

[--seed [SEED]] # Integer representing the seed used in train/test/validation splitting

[--batch [BATCH]] # Integer for batch size

[--loss [{MAE,MSE,MAPE}]] # Loss function to be used at training time

[--optimizer [{Adam,SGD}]] # Optimizer to be used at training time

[--learning_rate [LEARNING_RATE]] # Learning rate for all types of optimizer

[--filters [FILTERS]] # Tuple with dimension of filters to apply in the CNN

[--regress [REGRESS]] # Whether to apply or not regression as last layer

[--beta_1 [BETA_1]] # Beta_1 parameter for Adam

[--beta_2 [BETA_2]] # Beta_1 parameter for Adam

[--epsilon [EPSILON]] # Epsilon parameter for Adam

[--momentum [MOMENTUM]] # Momentum parameter for SGD

[--epochs [EPOCHS]] # Integer for training epochs