import torch
import torch.nn as nn
class conv_block(nn.Module):
def init(self, in_c, out_c):
super().init()
self.conv1 = nn.Conv2d(in_c, out_c, kernel_size=3, padding=1)
self.bn1 = nn.BatchNorm2d(out_c)
self.conv2 = nn.Conv2d(out_c, out_c, kernel_size=3, padding=1)
self.bn2 = nn.BatchNorm2d(out_c)
self.conv3=nn.Conv2d(out_c, out_c, kernel_size=3, padding=1)
self.bn3= nn.BatchNorm2d(out_c)
self.relu=nn.ReLU()
self.conv4=nn.Conv2d(out_c, out_c, kernel_size=3, padding=1)
self.bn4= nn.BatchNorm2d(out_c)
self.relu=nn.ReLU()
self.conv5=nn.Conv2d(out_c, out_c, kernel_size=3, padding=1)
self.bn5= nn.BatchNorm2d(out_c)
self.relu=nn.ReLU()
self.conv6=nn.Conv2d(out_c, out_c, kernel_size=3, padding=1)
self.bn6= nn.BatchNorm2d(out_c)
self.relu=nn.ReLU()
self.conv7=nn.Conv2d(out_c, out_c, kernel_size=3, padding=1)
self.bn7= nn.BatchNorm2d(out_c)
self.relu=nn.ReLU()
self.relu = nn.ReLU()
def forward(self, inputs):
x = self.conv1(inputs)
x = self.bn1(x)
x = self.relu(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.relu(x)
x=self.conv3(x)
x=self.bn3(x)
x=self.relu(x)
x=self.conv4(x)
x=self.bn4(x)
x=self.relu(x)
x=self.conv5(x)
x=self.bn5(x)
x=self.relu(x)
x=self.conv6(x)
x=self.bn6(x)
x=self.relu(x)
x=self.conv7(x)
x=self.bn7(x)
x=self.relu(x)
return x
class encoder_block(nn.Module):
def init(self, in_c, out_c):
super().init()
self.conv = conv_block(in_c, out_c)
self.pool = nn.MaxPool2d((2, 2))
def forward(self, inputs):
x = self.conv(inputs)
p = self.pool(x)
return x, p
class decoder_block(nn.Module):
def init(self, in_c, out_c):
super().init()
self.up = nn.ConvTranspose2d(in_c, out_c, kernel_size=2, stride=2, padding=0)
self.conv = conv_block(out_c+out_c, out_c)
def forward(self, inputs, skip):
x = self.up(inputs)
if x.shape != skip.shape:
x = TF.resize(x, size=skip.shape[2:])
x = torch.cat([x, skip], axis=1)
x = self.conv(x)
return x
class build_unet(nn.Module):
def init(self):
super().init()
""" Encoder """
self.e1 = encoder_block(3, 64)
self.e2 = encoder_block(64, 128)
self.e3 = encoder_block(128, 256)
self.e4 = encoder_block(256, 512)
""" Bottleneck """
self.b = conv_block(512, 1024)
""" Decoder """
self.d1 = decoder_block(1024, 512)
self.d2 = decoder_block(512, 256)
self.d3 = decoder_block(256, 128)
self.d4 = decoder_block(128, 64)
""" Classifier """
self.outputs = nn.Conv2d(64, 1, kernel_size=1, padding=0)
def forward(self, inputs):
""" Encoder """
s1, p1 = self.e1(inputs)
s2, p2 = self.e2(p1)
s3, p3 = self.e3(p2)
s4, p4 = self.e4(p3)
""" Bottleneck """
b = self.b(p4)
""" Decoder """
d1 = self.d1(b, s4)
d2 = self.d2(d1, s3)
d3 = self.d3(d2, s2)
d4 = self.d4(d3, s1)
outputs = self.outputs(d4)
return outputs
if name == “main”:
x = torch.randn((2, 3, 512, 512))
f = build_unet()
y = f(x)
print(y.shape)
This is the code for the model input images greyscale 512X512 with mask labels
The loss fund is:
import torch.nn.functional as F
class DiceLoss(nn.Module):
def init(self, weight=None, size_average=True):
super(DiceLoss, self).init()
def forward(self, inputs, targets, smooth=1):
#comment out if your model contains a sigmoid or equivalent activation layer
inputs = torch.sigmoid(inputs)
#flatten label and prediction tensors
inputs = inputs.view(-1)
targets = targets.view(-1)
intersection = (inputs * targets).sum()
dice = (2.*intersection + smooth)/(inputs.sum() + targets.sum() + smooth)
return 1 - dice
class DiceBCELoss(nn.Module):
def init(self, weight=None, size_average=True):
super(DiceBCELoss, self).init()
def forward(self, inputs, targets, smooth=1):
#comment out if your model contains a sigmoid or equivalent activation layer
inputs = torch.sigmoid(inputs)
#flatten label and prediction tensors
inputs = inputs.view(-1)
targets = targets.view(-1)
intersection = (inputs * targets).sum()
dice_loss = 1 - (2.*intersection + smooth)/(inputs.sum() + targets.sum() + smooth)
BCE = F.binary_cross_entropy(inputs, targets, reduction='mean')
Dice_BCE = BCE + dice_loss
return Dice_BCE
The code for training : from torch.utils.data import DataLoader
from glob import glob
def train(model, loader, optimizer, loss_fn, device):
epoch_loss = 0.0
model.train()
for x, y in loader:
x = x.to(device, dtype=torch.float32)
y = y.to(device, dtype=torch.float32)
optimizer.zero_grad()
y_pred = model(x)
loss = loss_fn(y_pred, y)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
epoch_loss = epoch_loss/len(loader)
return epoch_loss
def evaluate(model, loader, loss_fn, device):
epoch_loss = 0.0
model.eval()
with torch.no_grad():
for x, y in loader:
x = x.to(device, dtype=torch.float32)
y = y.to(device, dtype=torch.float32)
y_pred = model(x)
loss = loss_fn(y_pred, y)
epoch_loss += loss.item()
epoch_loss = epoch_loss/len(loader)
return epoch_loss
if name == “main”:
“”" Seeding “”"
seeding(42)
""" Directories """
create_dir("files")
""" Load dataset """
train_x = sorted(glob("/content/drive/MyDrive/Data_brain/train/image/*")[:10])
train_y = sorted(glob("/content/drive/MyDrive/Data_brain/train/mask/*")[:10])
valid_x = sorted(glob("/content/drive/MyDrive/Data_brain/test/image/*")[:5])
valid_y = sorted(glob("/content/drive/MyDrive/Data_brain/test/mask/*")[:5])
data_str = f"Dataset Size:\nTrain: {len(train_x)} - Valid: {len(valid_x)}\n"
print(data_str)
""" Hyperparameters """
size = (512, 512)
batch_size = 2
num_epochs = 30
lr = 0.05
checkpoint_path = "files/checkpoint.pth"
train_dataset = DriveDataset(train_x, train_y)
valid_dataset = DriveDataset(valid_x, valid_y)
train_loader = DataLoader(
dataset=train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=0
)
valid_loader = DataLoader(
dataset=valid_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=0
)
device = torch.device('cuda')
model = build_unet()
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=5, verbose=True)
loss_fn = DiceBCELoss()
best_valid_loss = float("inf")
val_loss=[]
trai_loss=[]
t=[]
for epoch in range(num_epochs):
start_time = time.time()
train_loss = train(model, train_loader, optimizer, loss_fn, device)
valid_loss = evaluate(model, valid_loader, loss_fn, device)
val_loss.append(valid_loss)
trai_loss.append(train_loss)
t.append(epoch)
""" Saving the model """
if valid_loss < best_valid_loss:
data_str = f"Valid loss improved from {best_valid_loss:2.4f} to {valid_loss:2.4f}. Saving checkpoint: {checkpoint_path}"
print(data_str)
best_valid_loss = valid_loss
torch.save(model.state_dict(), checkpoint_path)
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
data_str = f'Epoch: {epoch+1:02} | Epoch Time: {epoch_mins}m {epoch_secs}s\n'
data_str += f'\tTrain Loss: {train_loss:.3f}\n'
data_str += f'\t Val. Loss: {valid_loss:.3f}\n'
print(data_str)
The model is not even overfitting training data