why my model does not learn? Same on every epoch (Pytorch)

from sklearn import datasets
import pandas as pd
import numpy as np
import torch
from torch import nn

#loading the dataset
(data, target) = datasets.load_diabetes(as_frame=True,return_X_y=True) #with the as_frame=True data: pd.DataFrame


# converting data,target to tensors
data = torch.tensor(data.values,dtype=torch.float)
target = torch.tensor(target.values,dtype=torch.float)


#split the data 80% train 20% testing
a = 0.8
train_data , train_target = data[:int(a*len(data))] , data[:int(a*len(data))]
test_data , test_target = data[int(a*len(data)):] , data[int(a*len(data)):]


#constructing the model
# for this dataset dimentionality is 10 so the in_features will be 10
model = nn.Sequential(
    nn.Linear(in_features=10, out_features=128),
    nn.Linear(in_features=128, out_features=128),
    nn.Linear(in_features=128, out_features=1)
)

#loss fn , optimizer
loss_fn = nn.L1Loss() #binary cross entropy
optimizer = torch.optim.SGD(params = model.parameters(),lr=0.001) #stochastic gradient descent

#training loop
epochs = 1000

for epoch in range(epochs):
    #1. make prediction
    model.train()
    train_pred = model(train_data)
    
    loss = loss_fn(train_pred, train_target)

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    model.eval()
    with torch.inference_mode():
        test_pred = model(test_data)
        loss_test = loss_fn(test_pred, test_target)

    if epoch%(epochs//min(10,epochs))==0: print(f"{epoch} - training loss: {round(float(loss),4)} | test loss: {round(float(loss_test),4)}")- training loss: {loss} | test loss: {loss_test}")

Output

0 – training loss: 0.0837 | test loss: 0.0806

100 – training loss: 0.0433 | test loss: 0.0431

200 – training loss: 0.0426 | test loss: 0.0425

300 – training loss: 0.042 | test loss: 0.0419

400 – training loss: 0.0414 | test loss: 0.0414

500 – training loss: 0.0408 | test loss: 0.0408

600 – training loss: 0.0403 | test loss: 0.0403

700 – training loss: 0.0398 | test loss: 0.0398

800 – training loss: 0.0393 | test loss: 0.0394

900 – training loss: 0.0388 | test loss: 0.0389

Answer

First, as it was mentioned in the comments, you probably meant:

train_data, train_target = data[:int(a*len(data))] , target[:int(a*len(data))]
test_data, test_target = data[int(a*len(data)):] , target[int(a*len(data)):]

Next, your target size is not consistent with the output size (this should give a warning). Using

loss = loss_fn(train_pred, train_target.unsqueeze(1))

and

loss_test = loss_fn(test_pred, test_target.unsqueeze(1))

should give you some traction.

Advertisement

Answer