A deep neural network that predicts whether a tumor is malignant or benign.
Based on the Wisconsin breast cancer diagnosis dataset.
- Data preprocessing
- Flexible neural network architecture
- Feedforward
- Backpropagation
- Gradient descent
- Sigmoid & Softmax activation functions
- Binary cross-entropy loss
- Matrix multiplication
See the subject for more details.
Final Score 125/100
First you need to have your golang workspace set up on your machine.
Then clone this repo into your go-workspace/src/ folder.
git clone https://github.com/dfinnis/Multilayer_Perceptron.git; cd Multilayer_Perceptron
Download dependencies.
go get -d ./...; pip install -r requirements.txt
To run.
go run main.go
Alternatively, build & run the binary.
go build; ./Multilayer_Perceptron
Default behaviour is to split the data into training & test sets, train a model on the training set, show metrics for training & test sets, then save model as model.json.
go run main.go -t -s 4242
While training prints metrics & confusion matrix for training & test sets. After training shows a line graph of loss over training period.
without --predict or --early or --seed, --train uses the entire dataset to train.
go run main.go -p -s 4242
Only predict, don't train. Uses the entire dataset to predict unless seeded with -s.
The default model path is "model.json", but with additional argument FILEPATH model is loaded from given filepath.
Predict prints metrics & confusion matrix for predictions on the test set.
go run main.go -e -s 4242
Provide addtional argument SEED integer for randomization.
This seeds the pseudo-randomization of weights & shuffling of data. Thus a data split, model & loss can be replicated exactly with a given seed. The default seed is the current time.
go run main.go -e
Early stopping. Stop training when test set loss starts to increase. This avoids overfitting & minimizes test loss.
go run main.go -ep 14000
Provide addtional argument EPOCHS to determine length of training. Must be an integer between 0 & 100000.
The default number of epochs is 15000, which is usually around when test loss reaches minimum on default settings.
go run main.go -e -l 0.1
Provide addtional argument LEARNING rate. Must be a float between 0 & 1. The default learning rate is 0.01.
go run main.go -e -a "30 30 2"
Provide addtional argument ARCHITECTURE as string.
The default architecture is "16 16 2". 2 hidden layers with 16 neurons, & an output layer with 2 neurons.
go run main.go -mse
While training print loss metric: Mean squared error. Default loss metric: Binary cross entropy log loss.
go run main.go -rmse
While training print loss metric: Root mean squared error.
go run main.go -q
Don't print architecture or seed or additional metrics while training.
go run main.go data.csv
Any non-flag argument will be read as data path. The default data path is data.csv.
The model is saved every training epoch in json format as a 2d array of neurons with bias & weights. Below is the first 42 lines of a model starting to train. We see the first neuron in the first hidden layer, it has 30 weights corresponding to the 30 neurons in the input layer.
42 provides evaluation.py which splits data.csv randomly into data_training.csv & data_test.csv.
test.sh runs evaluation.py then trains a model with data_training.csv, & predicts with data_test.csv.
./test.sh
However, this is the same as simply running go run main.go which by default splits the data into training & test sets.
The cursor is hidden while training for clean output. If Multilayer_Perceptron is killed, you may need to reset your cursor.
./reset_cursor.sh