In this Google Colab, I set up a shallow neural net and train it on a supervised data set of BBC News Texts with category tags. For the full description of this project, see my post Supervised Machine Learning & BBC Text Classification on my portfolio website.
- Libraries: Pandas, Numpy, PyTorch, Sklearn, Plotly
- Data: www.kaggle.com/datasets/hgultekin/bbcnewsarchive
The data set has the following distribution:
The neural net is set up like this:
class ClassificationNet(nn.Module):
def __init__(self):
super(ClassificationNet, self).__init__()
self.layers = nn.Sequential(
nn.Linear(23699, 512), #feature number, first layer size
nn.Hardtanh(),
nn.Linear(512,5) #we have 5 categories!
)
def forward(self, x): #forward pass, input shape(data points, feature number)
sm = nn.Softmax(dim=1) #activation function
x = sm(self.layers(x))
return x
net = ClassificationNet()
# learning rate
lrt=0.01
# optimizer
optimizer = optim.Adam(net.parameters(), lr=lrt)
# loss/criterion
criterion = nn.CrossEntropyLoss()
# epochs / number of training iterations
epochs = 50
After training, we get the following graph for our training and validation loss:
Our evaluation on the test set produces the following confusion matrix:
