Logistic regression example with PyTorch (Marc Lelarge's class "Deep Learning Do-it-Yourself")

logistic_regression_pytorch.py

import torch
import numpy as np
from scipy.stats import bernoulli
from scipy.special import expit

dtype = torch.FloatTensor

# Model
w_source = np.array([2., -3.])
b_source  = np.array([1.])

# Data generation
x = np.random.random((100,2))
y = bernoulli.rvs(expit(np.dot(x, w_source) + b_source))

# Convert to tensors
x_t = torch.from_numpy(x).type(dtype)
y_t = torch.from_numpy(y).type(dtype).unsqueeze(1) # Add 1D for compatibility with the BCELoss

# Init model with sigmoid output
model = torch.nn.Sequential(
    torch.nn.Linear(2, 1),
    torch.nn.Sigmoid()
)

model.train()

loss_fn = torch.nn.BCELoss() # Binary cross-entropy loss
optimizer = torch.optim.SGD(model.parameters(), lr=0.01) # PyTorch automatically detects weights to optimize

for epoch in range(1000):
    y_pred = model(x_t)
    # The loss function is expecting 1D float tensors for both ground truth and predictions
    loss = loss_fn(y_pred, y_t)
    if epoch % 100 == 0:
        print("progress:", "epoch:", epoch, "loss",loss.item())
    # Zero gradients, perform a backward pass, and update the weights.
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    
print("estimation of the parameters:")
for param in model.parameters():
    print(param)