the cobweb is a tiny neural network library that I created to gain deeper understanding of how neural nets are working underneath.
first of all checkout the requirements.txt
and simply use pip install cubweb command
let's create a simple NN using cobweb first you need to import MLP (model)
from cobweb.model import MLP
now we will create our model object.
you should pass your input (your training data) size as the first argument
and a list of tuples for each layer input size and it's activation function name
for instance imagine we want a NN with 4 layers such that the first layer has 12 neurons with tanh activation
the second layer has 8 neurons with tanh activation
the third layer has 4 neurons with the relu activation function
and the last layer has 3 neuron with the softmax activation functin and our inputs are 1*3 arrays.
our model will be created like this:
m = MLP(3,[(12,'tanh'),(8,"tanh"),(4,"relu"),(3,"softmax")])
Tip
the default activation is relu so if you just pass a number or a tuple just with one elemnt like these are valid!
m = MLP(3,[(12,'tanh'),(8,"tanh"),(4,),(3,"softmax")])
m = MLP(3,[(12,'tanh'),(8,"tanh"),4,(3,"softmax")])
you can pass these as activatins
- "relu"
- "tanh"
- "softmax"
Warning
make sure you are always using the softmax activation with the cross_entropy loss and as the last layer!
because of the backpropagation of softmax and cross_entropy loss if you use it in other ways maybe lead to error or your network doesn't learn.
in other activations and losses you are totally free.
now let's train our NN
for this we will use **m.update()**
update method gets these as input:
- training data batch
- target datas
- loss_function (as str)
- number of epochs
- learning rate
by running this line of code our NN will be training:
m.update(x,y,loss_function="cross_entropy",epochs=100,learning_rate=0.01)
Tip
the default valu for loss_function is "mse" which is MSE(mean squared error) loss.
the default value for epochs is 100
the default value for learning rate is 0.01
- "mse" for MSE
- "mae" for MAE
- "cross_entropy" for categorical cross entropy
also even you can use r2 score as loss function by passing "r2"
special tahnks to Andrej Karpathy
I learned the most of these codes by his Micro Grad.