Strange results measuring accuracy of NN

1

I’m trying to make a model to predict prices using the Boston dataset from sklearn.datasets

# -*- coding: utf-8 -*-
from sklearn.datasets import load_boston
from torch.utils.data.dataset import Dataset
from torch.utils.data import DataLoader
import torch.nn.functional as F
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(13, 64)
        self.fc2 = nn.Linear(64, 32)
        self.fc3 = nn.Linear(32, 16)
        self.fc4 = nn.Linear(16,1)

    def forward(self, x):
        x = self.fc1(x)
        x = F.relu(x)
        x = self.fc2(x)
        x = F.relu(x)
        x = self.fc3(x)
        x = F.relu(x)
        return self.fc4(x)

net = Net()

class BostonData(Dataset):
    __xs = np.array(0)
    __ys = np.array(0)

    def __init__(self, train = True):
        df = load_boston()
        index = int(len(df["data"]) * 0.7)
        if train:
            self.__xs = np.array(df["data"][0:index])
            self.__ys = np.array(df["target"][0:index])
        else:
            self.__xs = np.array(df["data"][index:])
            self.__ys = np.array(df["target"][index:])

        self.__xs = (self.__xs - self.__xs.mean()) / self.__xs.std()

    def __getitem__(self, index):
        return self.__xs[index], self.__ys[index]

    def __len__(self):
        return len(self.__xs)


dset_train = BostonData(train = True)
dset_test = BostonData(train = False)
train_loader = DataLoader(dset_train, batch_size=10, shuffle=True)
test_loader = DataLoader(dset_train, batch_size=10, shuffle=True)

After training the model, the strange behavior that I’m talking is on the following part. I want to measure an error average between the true values and the predictions using the test set as follows

res = 0
with torch.no_grad():
    for testbatch in test_loader:
        X,y = testbatch
        output = net(X.float())
        for i,elem in enumerate(output):
            res = res + np.abs(y[i] - elem)
print(res/len(dset_train))

This gives me no error whatsoever, but I was curious about if the model was oferfitting, so I changed the test_loader for the train_loader

res = 0
with torch.no_grad():
    for testbatch in train_loader:
        X,y = testbatch
        output = net(X.float())
        for i,elem in enumerate(output):
            res = res + np.abs(y[i] - elem)
print(res/len(dset_train))

Executing only these lines of code (I’m using Spyder as IDE) brings the same result for res/len(dset_train) as doing it for test_loader , which seems really strange. Is there any error that I didn’t notice?

2

As a side note, with torch.no_grad(): should be for validation only. While training, you should not have the with torch.no_grad(): block. This may result in problems you defined as strange results.

3

This is the code of the training process

optimizer = optim.Adam(net.parameters(), lr = 0.001)
criterion = torch.nn.MSELoss() 
EPOCHS = 500000

lloss = []

for epoch in range(EPOCHS):
    avg_train_loss = 0
    for trainbatch in train_loader:
        X,y = trainbatch
        net.zero_grad()
        output = net(X.float())
        loss = criterion(output[:, -1], y)
        loss.backward()
        optimizer.step()
        avg_train_loss += loss.item() / len(train_loader)
    lloss.append(avg_train_loss)

There is no no_grad() there, and I think it has nothing to do with the problem I presented

4

You could try printing several “intermediate” variables, like X and y, output, res, to make sure that you get different values between the train loader and test loader.

The most important thing is checking y[i] against elem, this way you will see if your model makes mistakes.

One thing I’ve noticed in your code (but it might be due to copy/pasting and not be present in your actual code) is in the first error measurement (on the test set, i.e. the 2nd code block in your message): you print res/len(dset_train). Shouldn’t it be res/len(dset_test) ?

5

I saw the problem

train_loader = DataLoader(dset_train, batch_size=10, shuffle=True) test_loader = DataLoader(dset_train, batch_size=10, shuffle=True)

Something strange when I call DataLoader :wink:
Thanks for replying

1 Like