CODEWITHSUNDEEP
pytorch
Moth
Moth

Reputation: 96

Applying any optimization causes values to be NaN

I am working on a classifier model for this data set: https://archive.ics.uci.edu/ml/datasets/ILPD+%28Indian+Liver+Patient+Dataset%29 and I have come up with this code in pytorch:

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from sklearn import preprocessing
from sklearn.model_selection import train_test_split
import pandas as pd
import numpy as np


ILPDataset = pd.read_csv('preprocessed-ilpd.csv')
ILPDataset["Age"] = pd.to_numeric(ILPDataset["Age"])
LabelEncoder = preprocessing.LabelEncoder()
LabelEncoder.fit(ILPDataset["Gender"])
ILPDataset["Gender"] = LabelEncoder.transform(ILPDataset["Gender"])
print(ILPDataset["Gender"].describe())
ILPDataset["AAP"] = preprocessing.scale(ILPDataset["AAP"])
ILPDataset["SgAlAm"] = preprocessing.scale(ILPDataset["SgAlAm"])
ILPDataset["SgApAm"] = preprocessing.scale(ILPDataset["SgApAm"])
Features = ["Age","Gender","TB","DB","AAP","SgAlAm","SgApAm","TP","ALB","A/G"]
ILPDFeatures = ILPDataset[Features]
ILPDTarget = ILPDataset["Selector"]

X_Train, X_Test, Y_Train, Y_Test = train_test_split(ILPDFeatures,ILPDTarget,test_size=0.2,random_state=0)

print(X_Train.shape)
print(Y_Train.shape)
print(X_Test.shape)
print(Y_Test.shape)
torch.set_default_tensor_type(torch.DoubleTensor)
TrainX = torch.from_numpy(X_Train.values).double()
TestX = torch.from_numpy(X_Test.values).double()


TrainY = torch.from_numpy(Y_Train.values).long().view(1,-1)[0]
TestY = torch.from_numpy(Y_Test.values).long().view(1,-1)[0]
TrainY.reshape(TrainY.shape[0],1)
TestY.reshape(TestY.shape[0],1)
print(X_Train.shape[1])
input_layers = X_Train.shape[1]
output_layers = 2
hidden = 100

class Net(nn.Module):
    def __init__(self):
        super(Net,self).__init__()
        self.fc1 = nn.Linear(input_layers,hidden)
        self.fc2 = nn.Linear(hidden,hidden)
        self.fc3 = nn.Linear(hidden,output_layers)

    def forward(self,x):
        x = torch.sigmoid(self.fc1(x))
        x = torch.sigmoid(self.fc2(x))
        x = self.fc3(x)
        return F.softmax(x,dim=-1)

model = Net()
optimizer = torch.optim.SGD(model.parameters(),lr = 0.0001,momentum = 0.9)
loss_fn = nn.NLLLoss()
epochs_data = []
epochs = 601
print(TrainX.shape)
print(TestY.shape)
for epoch in range(1,3):
    optimizer.zero_grad()
    Ypred = model(TrainX)
    loss = loss_fn(Ypred,TrainY)
    loss.backward()
    optimizer.step()
    print(Ypred)
    if epoch%100==0:
        print("Epoch - %d, (%d%%) "%(epoch,epoch/epochs*100))

YPred_Test = model(TestX)
Loss_Test = loss_fn(YPred_Test,TestY)
print(Loss_Test.item())
print(YPred_Test)
print(YPred_Test)
print(TestY)

I have used different optimizers with different LRs and Momentum but on applying each one of them after loss.backward(), it turns the values in the tensor to NaN. I looked up the different answers on SO but even after trying them out I couldn't solve the bug.

Upvotes: 0

Views: 2234

Answers (2)

Moth
Moth

Reputation: 96

The problem is caused by not normalizing every column in the dataset. I have no idea why but normalizing column will solve the problem.

Upvotes: 1

Harshit Kumar
Harshit Kumar

Reputation: 12837

The problem is that you're using softmax with NLLLoss, which doesn't work together. The documentation on softmax states:

This module doesn’t work directly with NLLLoss, which expects the Log to be computed between the Softmax and itself. Use LogSoftmax instead (it’s faster and has better numerical properties).

Use log_softmax in the last layer with NLLLoss as follows:

def forward(self,x):
    ...
    return F.log_softmax(x,dim=-1)
...
loss_fn = nn.NLLLoss()

Upvotes: 2

Related Questions