xrsrke/tinypytorch

A deep learning framework from scratch

tinypytorch

import torch
from tinypytorch.data import get_local_data
from tinypytorch.model import initialize_parameters, Model
from tinypytorch.metrics import accuracy

Dependence libraries - nbdev - torch - matplotlib - pytest

This file will become your README and also the index of your documentation.

Install

pip install tinypytorch

Train a neural network

Data

x_train, y_train, x_valid, y_valid = get_local_data()

x_train.shape, y_train.shape

(torch.Size([50000, 784]), torch.Size([50000]))

y_train[1]

tensor(0)

y_train[1:5]

tensor([0, 4, 1, 9])

Initialize hyperparameters

n, m = x_train.shape # num rows and columns

c = y_train.max() + 1

n, m, c

(50000, 784, tensor(10))

nh = 50 # num hidden

w1, b1, w2, b2 = initialize_parameters(m, nh)

w1.shape, b1.shape

(torch.Size([784, 50]), torch.Size([50]))

w2.shape, b2.shape

(torch.Size([50, 1]), torch.Size([1]))

• Training set’s shape: (50000, 784)
• Weight’s shape: (784, 50)
• Bias’s shape: (50)

The first layer (Lin): (50000, 784) x (784, 50) + (50)

model = Model(w1, b1, w2, b2)

loss = model(x_train, y_train)

Model.__call__
l=<tinypytorch.model.Lin object>
Lin.forward
inp=torch.Size([50000, 784])
w=torch.Size([784, 50])
b=torch.Size([50])
output.shape=torch.Size([50000, 50])
x.shape=torch.Size([50000, 50])
Model.__call__
l=<tinypytorch.model.ReLU object>
x.shape=torch.Size([50000, 50])
Model.__call__
l=<tinypytorch.model.Lin object>
Lin.forward
inp=torch.Size([50000, 50])
w=torch.Size([50, 1])
b=torch.Size([1])
output.shape=torch.Size([50000, 1])
x.shape=torch.Size([50000, 1])

loss

tensor(26.1652)

model.backward()

Example 2

bs = 64

xb = x_train[0:64]