fitting_straight_line_pytorch___09___train_model.py

losses = []         # to keep track of the epoch lossese 
slope_list = []     # to keep track of the slope learnt by the model
intercept_list = [] # to keep track of the intercept learnt by the model

EPOCHS = 2500
print('\nTRAINING...')
for epoch in range(EPOCHS):
    # We need to clear the gradients of the optimizer before running the back-propagation in PyTorch
    optimizer.zero_grad() 
    
    # Feeding the input data in the model and getting out the predictions
    pred_y = model(data_x)

    # Calculating the loss using the model's predictions and the real y values
    loss = criterion(pred_y, data_y) 

    # Back-Propagation
    loss.backward() 
    
    # Updating all the trainable parameters
    optimizer.step()
    
    # Appending the loss.item() (a scalar value)
    losses.append(loss.item())
    
    # Appending the learnt slope and intercept   
    slope_list.append(model.linear.weight.item())
    intercept_list.append(model.linear.bias.item())
    
    # We print out the losses after every 2000 epochs
    if (epoch)%100 == 0:
        print('loss: ', loss.item())

utput.py

TRAINING...
loss:  290066.9375
loss:  233668.875
loss:  186075.625
loss:  146441.828125
loss:  113859.40625
loss:  87455.7265625
loss:  66397.1875
loss:  49895.44140625
loss:  37213.265625
loss:  27672.890625
loss:  20661.083984375
loss:  15636.8427734375
loss:  12134.056640625
loss:  9762.8232421875
loss:  8207.25390625
loss:  7220.21630859375
loss:  6615.59326171875
loss:  6258.6826171875
loss:  6055.97216796875
loss:  5945.3681640625
loss:  5887.44677734375
loss:  5858.3232421875
loss:  5844.21875
loss:  5837.568359375
loss:  5834.42822265625

Thanks ! It was very interesting article. But in practice while this approach works, there are faster functions for fitting than DNN training.
My favorite for now is np.polyfit() which gives you the best polynomial coefficients for a poly of given degree.