Hey @ everybody,
I´m trying to use a CNN Classifier on clinical data (input shape 39,12, rows for values and columns for time intervals) to predict a categorial statement as good / bad result; so i´m using conv2d with the size (1,3) or (1,5) to find time sensitive patterns (each row == 1 clinical parameter) and a max pool pattern of (1,2).
After running the forward path i´m using a sigmoid function on the output node of the last linear network layer to receive a propability between 0 and 1 for category good; using these propabilties to discriminate the loss between them and the labels [0,1] and backpropagate on them.
I ask myself if the architecture ist correct and if it is whatelse can i do to improve performance, because los isnt really decreasing and accuracy is random at the begining of training and ends up getting worse.
# CNN Model
# Output Size of Convolutional Layer by ((Width - FilterSize + (Padding * 2)) / Stride) + 1
import torch.nn as nn
import torch.nn.functional as F
class ConvNet(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(1, 6, (1,5), padding = (0,1), dilation = 1).double()
# could use avg_pool2d
self.pool = nn.MaxPool2d((1,2), 1)
self.conv2 = nn.Conv2d(6, 16, (1,3), padding = (0,1), dilation = 1).double()
self.fc1 = nn.Linear(16 * 39 * 8, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 1)
def forward(self, x):
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = x.float()
x = x.view(-1,16*39*8)
# or use x = torch.flatten(x, 1) to flatten all dimensions except batch
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
# Propabilities through Sigmoid Function
x = torch.sigmoid(x)
return x
def binary(self,y):
# adding function of sigmoid function and round to classify and calculate accuracy; KODIETZ 221017
y = torch.round(y)
return y
# Define model
net = ConvNet()
import torch.optim as optim
# define optimizer & Loss Function; get model
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
# training
total_datapoints = len(trainloader)
for epoch in range(num_epochs):
running_loss = 0.0
epoch_acc = 0.0
epoch_loss = 0.0
print(f'-----------Start Epoch {epoch + 1} -----------------------')
for i,(input_torch,labels) in enumerate(trainloader):
# origin shape: [batch,channels,height,weight]
# here we got [5,1,12,39] = 5,1,468
input_torch = input_torch.double().to(device)
labels = labels.float().to(device)
# Zero the gradients
optimizer.zero_grad()
# Forward pass
outputs = net(input_torch).reshape(1,len(input_torch))[0]
loss = criterion(outputs, labels)
# Accuray for Batch
y_pred = net.binary(outputs)
correct = sum(labels == y_pred).item()
acc = correct / len(input_torch)
epoch_acc += acc
# Backward Propagation and Optimization
loss.backward()
optimizer.step()
# print statistics
epoch_loss += loss.item()
if i % 5 == 0:
print(f'{epoch + 1}, {i + 1}\t|\t loss: {loss.item():.5f} \t| acc: {acc}')
print(f'-----------End Epoch {epoch + 1}: | Loss: {epoch_loss/len(trainloader):.5f} | Acc: {epoch_acc/len(trainloader):.3f}')
thanks for any help!
best regards
- because of an imabalanced dataset i provided a weigth attribute for the loss; but still not performing properly;