We Keep Coding

Nested Cross Validation in PyTorch

I am working on a multiclass image classification problem that has 4k labeled images. Currently, I am using cross-validation.However, I want to try nested-cross-validation for the problem in pytorch but I couldn't find it.
Is it possible to use nested-cross-validation in pytorch? if so, how?
I built a simple pipeline below. Is it correct to implement nested-cross-validation like this.
import torch
from torch.utils.data import DataLoader,SubsetRandomSampler
from sklearn.model_selection import KFold
from torchvision import datasets
input_size = (256,3,224,244)
target_size = (256,)
class Dataset(datasets.VisionDataset):
def __init__(self):
super().__init__(self)
self.images = torch.rand(input_size).float()
self.targets = torch.randint(0,3,target_size)
def __getitem__(self, index: int) -> any:
return self.images[index],self.targets[index]
def __len__(self) -> int:
return len(self.images)
class BasicModel(torch.nn.Module):
def __init__(self) -> None:
super(BasicModel,self).__init__()
self.conv = torch.nn.Conv2d(3,16,kernel_size=(5,5))
self.adp = torch.nn.AdaptiveAvgPool2d(1)
self.linear = torch.nn.Linear(16,3)
def forward(self,x):
x = self.conv(x)
x = self.adp(x)
x = x.view(x.size(0),-1)
x = self.linear(x)
return x
data_ids = [*range(input_size[0])]
data = Dataset()
model = BasicModel()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),lr=0.01)
k_fold = 5
kfold_test = KFold(n_splits=k_fold, shuffle=True)
num_epochs = 2
for test_fold,(remain_ids,test_ids) in enumerate(kfold_test.split(data_ids)):
test_sampler = SubsetRandomSampler(test_ids)
testLoader = DataLoader(data,sampler=test_sampler,batch_size=2)
kfold_val = KFold(n_splits=k_fold-1, shuffle=True)
for epoch in range(num_epochs):
for val_fold,(train_ids,val_ids) in enumerate(kfold_val.split(remain_ids)):
train_sampler = SubsetRandomSampler(train_ids)
trainLoader = DataLoader(data,sampler=train_sampler,batch_size=2)
val_sampler = SubsetRandomSampler(val_ids)
valLoader = DataLoader(data,sampler=val_sampler,batch_size=2)
model.train()
for image,target in trainLoader:
with torch.cuda.amp.autocast():
output = model(image)
loss = criterion(output,target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f"train loss:{loss.item()}")
model.eval()
with torch.no_grad():
for image,target in valLoader:
output = model(image)
loss = criterion(output,target)
print(f"test loss:{loss.item()}")
model.eval()
with torch.no_grad():
for image,target in testLoader:
output = model(image)
loss = criterion(output,target)
print(f"test loss:{loss.item()}")

Continue training with torch.save and torch.load - key error messages

I am new to Torch and using a code template for a masked-cnn model. In order to be prepared if the training is interrupted, I have used torch.save and torch.load in my code, but I think I cannot use this alone for continuing training sessions? I start training by:
model = train_mask_net(64)
This calls the function train_mask_net where I have included torch.save in the epoch loop. I wanted to load one of the saved models and continue training with torch.load in front of the loop, but I got "key error" messages for the optimizer, loss and epoch call. Should I have made a specific checkpoint function as I have seen in some tutorials or is there a possibility that I can continue training with the files saved by the torch.saved command?
def train_mask_net(num_epochs=1):
data = MaskDataset(list(data_mask.keys()))
data_loader = torch.utils.data.DataLoader(data, batch_size=8, shuffle=True, num_workers=4)
model = XceptionHourglass(max_clz+2)
model.cuda()
dp = torch.nn.DataParallel(model)
loss = nn.CrossEntropyLoss()
params = [p for p in dp.parameters() if p.requires_grad]
optimizer = torch.optim.RMSprop(params, lr=2.5e-4, momentum=0.9)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=6,
gamma=0.9)
checkpoint = torch.load('imaterialist2020-pretrain-models/maskmodel_160.model_ep17')
#print(checkpoint)
model.load_state_dict(checkpoint)
#optimizer.load_state_dict(checkpoint)
#epoch = checkpoint['epoch']
#loss = checkpoint['loss']
for epoch in range(num_epochs):
print(epoch)
total_loss = []
prog = tqdm(data_loader, total=len(data_loader))
for i, (imag, mask) in enumerate(prog):
X = imag.cuda()
y = mask.cuda()
xx = dp(X)
# to 1D-array
y = y.reshape((y.size(0),-1)) # batch, flatten-img
y = y.reshape((y.size(0) * y.size(1),)) # flatten-all
xx = xx.reshape((xx.size(0), xx.size(1), -1)) # batch, channel, flatten-img
xx = torch.transpose(xx, 2, 1) # batch, flatten-img, channel
xx = xx.reshape((xx.size(0) * xx.size(1),-1)) # flatten-all, channel
losses = loss(xx, y)
prog.set_description("loss:%05f"%losses)
optimizer.zero_grad()
losses.backward()
optimizer.step()
total_loss.append(losses.detach().cpu().numpy())
torch.save(model.state_dict(), MODEL_FILE_DIR+"maskmodel_%d.model"%attr_image_size[0]+'_ep'+str(epoch)+'_tsave')
prog, X, xx, y, losses = None, None, None, None, None,
torch.cuda.empty_cache()
gc.collect()
return model
I don't think its necessary, but the xceptionhour class looks like this:
class XceptionHourglass(nn.Module):
def __init__(self, num_classes):
super(XceptionHourglass, self).__init__()
self.num_classes = num_classes
self.conv1 = nn.Conv2d(3, 128, 3, 2, 1, bias=True)
self.bn1 = nn.BatchNorm2d(128)
self.mish = Mish()
self.conv2 = nn.Conv2d(128, 256, 3, 1, 1, bias=True)
self.bn2 = nn.BatchNorm2d(256)
self.block1 = HourglassNet(4, 256)
self.bn3 = nn.BatchNorm2d(256)
self.block2 = HourglassNet(4, 256)
...

torch.save(model.state_dict(), PATH) only saves the model weights.
To also save optimizer, loss, epoch, etc., change it to:
torch.save({'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'loss': loss,
'epoch': epoch,
# ...
}, PATH)
To load them:
checkpoint = torch.load(PATH)
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
epoch = checkpoint['epoch']
loss = checkpoint['loss']
More on it here.

Implementing Early Stopping in Pytorch without Torchsample

As a Pytorch newbie (coming from tensorflow), I am unsure of how to implement Early Stopping. My research has led me discover that pytorch does not have a native way to so this. I have also discovered torchsample, but am unable to install it in my conda environment for whatever reason. Is there a simple way to go about applying early stopping without it? Here is my current setup:
class RegressionDataset(Dataset):
def __init__(self, X_data, y_data):
self.X_data = X_data
self.y_data = y_data
def __getitem__(self, index):
return self.X_data[index], self.y_data[index]
def __len__(self):
return len(self.X_data)
train_dataset = RegressionDataset(torch.from_numpy(X_train).float(), torch.from_numpy(y_train).float())
val_dataset = RegressionDataset(torch.from_numpy(X_val).float(), torch.from_numpy(y_val).float())
test_dataset = RegressionDataset(torch.from_numpy(X_test).float(), torch.from_numpy(y_test).float())
# Model Params
EPOCHS = 100
BATCH_SIZE = 1000
LEARNING_RATE = 0.001
NUM_FEATURES = np.shape(X_test)[1]
# Initialize Dataloader
train_loader = DataLoader(dataset = train_dataset, batch_size=BATCH_SIZE, shuffle = True)
val_loader = DataLoader(dataset = val_dataset, batch_size=BATCH_SIZE)
test_loader = DataLoader(dataset = test_dataset, batch_size=BATCH_SIZE)
# Define Neural Network Architecture
class MultipleRegression(nn.Module):
def __init__(self, num_features):
super(MultipleRegression, self).__init__()
# Define architecture
self.layer_1 = nn.Linear(num_features, 16)
self.layer_2 = nn.Linear(16, 32)
self.layer_3 = nn.Linear(32, 25)
self.layer_4 = nn.Linear(25, 20)
self.layer_5 = nn.Linear(20, 16)
self.layer_out = nn.Linear(16, 1)
self.relu = nn.ReLU() # ReLU applied to all layers
# Initialize weights and biases
nn.init.xavier_uniform_(self.layer_1.weight)
nn.init.zeros_(self.layer_1.bias)
nn.init.xavier_uniform_(self.layer_2.weight)
nn.init.zeros_(self.layer_2.bias)
nn.init.xavier_uniform_(self.layer_3.weight)
nn.init.zeros_(self.layer_3.bias)
nn.init.xavier_uniform_(self.layer_4.weight)
nn.init.zeros_(self.layer_4.bias)
nn.init.xavier_uniform_(self.layer_5.weight)
nn.init.zeros_(self.layer_5.bias)
nn.init.xavier_uniform_(self.layer_out.weight)
nn.init.zeros_(self.layer_out.bias)
def forward(self, inputs):
x = self.relu(self.layer_1(inputs))
x = self.relu(self.layer_2(x))
x = self.relu(self.layer_3(x))
x = self.relu(self.layer_4(x))
x = self.relu(self.layer_5(x))
x = self.layer_out(x)
return(x)
def predict(self, test_inputs):
x = self.relu(self.layer_1(test_inputs))
x = self.relu(self.layer_2(x))
x = self.relu(self.layer_3(x))
x = self.relu(self.layer_4(x))
x = self.relu(self.layer_5(x))
x = self.layer_out(x)
return(x)
# Check for GPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
model = MultipleRegression(NUM_FEATURES)
model.to(device)
print(model)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr = LEARNING_RATE)
# define dictionary to store loss/epochs for training and validation
loss_stats = {
"train": [],
"val": []
}
# begin training
print("Begin Training")
for e in tqdm(range(1, EPOCHS+1)):
# Training
train_epoch_loss = 0
model.train()
for X_train_batch, y_train_batch in train_loader:
X_train_batch, y_train_batch = X_train_batch.to(device), y_train_batch.to(device)
optimizer.zero_grad()
y_train_pred = model(X_train_batch)
train_loss = criterion(y_train_pred, y_train_batch.unsqueeze(1))
train_loss.backward()
optimizer.step()
train_epoch_loss += train_loss.item()
# validation
with torch.no_grad():
val_epoch_loss = 0
model.eval()
for X_val_batch, y_val_batch in val_loader:
X_val_batch, y_val_batch = X_val_batch.to(device), y_val_batch.to(device)
y_val_pred = model(X_val_batch)
val_loss = criterion(y_val_pred, y_val_batch.unsqueeze(1))
val_epoch_loss += val_loss.item()
loss_stats["train"].append(train_epoch_loss/len(train_loader))
loss_stats["val"].append(val_epoch_loss/len(val_loader))
print(f"Epoch {e}: \ Train loss: {train_epoch_loss/len(train_loader):.5f} \ Val loss: {val_epoch_loss/len(val_loader):.5f}")
# Visualize loss and accuracy
train_val_loss_df = pd.DataFrame.from_dict(loss_stats).reset_index().melt(id_vars=["index"]).rename(columns = {"index":"epochs"})
plt.figure()
sns.lineplot(data = train_val_loss_df, x = "epochs", y = "value", hue = "variable").set_title("Train-Val Loss/Epoch")
# Test model
y_pred_list = []
with torch.no_grad():
model.eval()
for X_batch, _ in test_loader:
X_batch = X_batch.to(device)
y_test_pred = model(X_batch)
y_pred_list.append(y_test_pred.cpu().numpy())
y_pred_list = [a.squeeze().tolist() for a in y_pred_list]
y_pred_list = [item for sublist in y_pred_list for item in sublist]
y_pred_list = np.array(y_pred_list)
mse = mean_squared_error(y_test, y_pred_list)
r_square = r2_score(y_test, y_pred_list)
print("Mean Squared Error :", mse)
print("R^2 :", r_square)

A basic way to do this is to keep track of the best validation loss obtained so far.
You can have a variable best_loss = 0 initialized before your loop over epochs (or you could do other things like best loss per epoch, etc.).
After each validation pass then do:
if val_loss > best_loss:
best_loss = val_loss
# At this point also save a snapshot of the current model
torch.save(model, 'my_model_best_loss.pth')
Then, if the best_loss does not improve significantly after some number of training steps, or by the end of the epoch, or if it val_loss gets worse, break out of the loop and terminate the training there.
For implementing algorithms like early stopping (and your training loop in general) you may find it easier to give PyTorch Lightning a try (no affiliation, but it's much easier than trying to roll everything by hand).

How to make prediction on pytorch emotion detection model

I made a CNN model for emotion recognition on 5 emotions. I wanted to test it on an single image to get the individual class predictions for each emotion.
Evaluating the model works, but I can't seem to find how to make a prediction with a single image. How can I do that?
The Model
def conv_block(in_channels, out_channels, pool=False):
layers = [nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
nn.BatchNorm2d(out_channels),
nn.ELU(inplace=True)]
if pool: layers.append(nn.MaxPool2d(2))
return nn.Sequential(*layers)
class ResNet(ImageClassificationBase):
def __init__(self, in_channels, num_classes):
super().__init__()
self.conv1 = conv_block(in_channels, 128)
self.conv2 = conv_block(128, 128, pool=True)
self.res1 = nn.Sequential(conv_block(128, 128), conv_block(128, 128))
self.drop1 = nn.Dropout(0.5)
self.conv3 = conv_block(128, 256)
self.conv4 = conv_block(256, 256, pool=True)
self.res2 = nn.Sequential(conv_block(256, 256), conv_block(256, 256))
self.drop2 = nn.Dropout(0.5)
self.conv5 = conv_block(256, 512)
self.conv6 = conv_block(512, 512, pool=True)
self.res3 = nn.Sequential(conv_block(512, 512), conv_block(512, 512))
self.drop3 = nn.Dropout(0.5)
self.classifier = nn.Sequential(nn.MaxPool2d(6),
nn.Flatten(),
nn.Linear(512, num_classes))
def forward(self, xb):
out = self.conv1(xb)
out = self.conv2(out)
out = self.res1(out) + out
out = self.drop1(out)
out = self.conv3(out)
out = self.conv4(out)
out = self.res2(out) + out
out = self.drop2(out)
out = self.conv5(out)
out = self.conv6(out)
out = self.res3(out) + out
out = self.drop3(out)
out = self.classifier(out)
return out
The fit_one_cycle function is called to train the model
#torch.no_grad()
def evaluate(model, val_loader):
model.eval()
outputs = [model.validation_step(batch) for batch in val_loader]
return model.validation_epoch_end(outputs)
def get_lr(optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
def fit_one_cycle(epochs, max_lr, model, train_loader, val_loader,
weight_decay=0, grad_clip=None, opt_func=torch.optim.SGD):
torch.cuda.empty_cache()
history = []
# Set up custom optimizer with weight decay
optimizer = opt_func(model.parameters(), max_lr, weight_decay=weight_decay)
# Set up one-cycle learning rate scheduler
sched = torch.optim.lr_scheduler.OneCycleLR(optimizer, max_lr, epochs=epochs,
steps_per_epoch=len(train_loader))
for epoch in range(epochs):
# Training Phase
model.train()
train_losses = []
lrs = []
for batch in train_loader:
loss = model.training_step(batch)
train_losses.append(loss)
loss.backward()
# Gradient clipping
if grad_clip:
nn.utils.clip_grad_value_(model.parameters(), grad_clip)
optimizer.step()
optimizer.zero_grad()
# Record & update learning rate
lrs.append(get_lr(optimizer))
sched.step()
# Validation phase
result = evaluate(model, val_loader)
result['train_loss'] = torch.stack(train_losses).mean().item()
result['lrs'] = lrs
model.epoch_end(epoch, result)
history.append(result)
return history
This returns the accuracy and loss, i want to change this so it returns prediction percentages for each class.
def accuracy(outputs, labels):
_, preds = torch.max(outputs, dim=1)
return torch.tensor(torch.sum(preds == labels).item() / len(preds))
class ImageClassificationBase(nn.Module):
def training_step(self, batch):
images, labels = batch
out = self(images)
loss = F.cross_entropy(out, labels)
return loss
def validation_step(self, batch):
images, labels = batch
out = self(images)
loss = F.cross_entropy(out, labels)
acc = accuracy(out, labels)
return {'val_loss': loss, 'val_acc': acc}
def validation_epoch_end(self, outputs):
batch_losses = [x['val_loss'] for x in outputs]
epoch_loss = torch.stack(batch_losses).mean()
batch_accs = [x['val_acc'] for x in outputs]
epoch_acc = torch.stack(batch_accs).mean()
return {'val_loss': epoch_loss.item(), 'val_acc': epoch_acc.item()}
def epoch_end(self, epoch, result):
print("Epoch [{}], last_lr: {:.5f}, train_loss: {:.4f}, val_loss: {:.4f}, val_acc: {:.4f}".format(
epoch, result['lrs'][-1], result['train_loss'], result['val_loss'], result['val_acc']))

Evaluating the model works, but I can't seem to find how to make a
prediction with a single image. How can I do that?
Simply, if you have a single image make sure to:
use additional 1 dimension at the beginning
make sure to use CHW format instead of HWC (or specify that within pytorch, check out how to do that here)
For example:
my_model = CNN(...)
random_image = torch.randn(1, 3, 100, 100) # 3 channels, 100x100 img
BTW. Your accuracy could be written a little simpler like this:
def accuracy(outputs, labels):
preds = torch.argmax(outputs, dim=1)
return torch.sum(preds == labels) / len(preds)
Getting class probability
Similar to argmax you can use softmax which transforms logits (unnormalized probability outputted by your network) into probabilities:
def probability(outputs):
return torch.nn.functional.softmax(outputs, dim=1)

How to solve size mismatch error in pytorch?

I am trying to create a logistic model by using CIFAR10 data in PyTorch. After running the model for evaluation I run into an error :
RuntimeError: size mismatch, m1: [750 x 4096], m2: [1024 x 10] at C:\w\1\s\tmp_conda_3.7_100118\conda\conda-bld\pytorch_1579082551706\work\aten\src\TH/generic/THTensorMath.cpp:136
It seems like input_size is creating a problem, I dont know I am new to this. Please let me know what changes should I make in order to overcome this error.
These are the hyperparameters:
batch_size = 100
learning_rate = 0.001
# Other constants
input_size = 4*4*64
num_classes = 10
This is the cell that downloads and splits the dataset into train, validation and test.
transform = torchvision.transforms.Compose(
[torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5))])
testset = torchvision.datasets.CIFAR10(root='D:\PyTorch\cifar-10-python', train=False,download=False, transform=transform)
trainvalset = torchvision.datasets.CIFAR10(root='D:\PyTorch\cifar-10-python', train=True,download=False, transform=transform)
trainset, valset = torch.utils.data.random_split(trainvalset, [45000, 5000]) # 10% for validation
train_loader = torch.utils.data.DataLoader(trainset, batch_size=50, shuffle=True)
test_loader = torch.utils.data.DataLoader(testset, batch_size=1000, shuffle=False)
val_loader = torch.utils.data.DataLoader(valset, batch_size=1000, shuffle=False)
This is the architecture of my model.
class CifarModel(nn.Module):
def __init__(self):
super().__init__()
self.linear = nn.Linear(input_size, num_classes)
def forward(self, xb):
xb = xb.view(-1, 64*8*8)
#xb = xb.reshape(-1, 784)
print(xb.shape)
out = self.linear(xb)
return out
def training_step(self, batch):
images, labels = batch
out = self(images) # Generate predictions
loss = F.cross_entropy(out, labels) # Calculate loss
return loss
def validation_step(self, batch):
images, labels = batch
out = self(images) # Generate predictions
loss = F.cross_entropy(out, labels) # Calculate loss
acc = accuracy(out, labels) # Calculate accuracy
return {'val_loss': loss.detach(), 'val_acc': acc.detach()}
def validation_epoch_end(self, outputs):
batch_losses = [x['val_loss'] for x in outputs]
epoch_loss = torch.stack(batch_losses).mean() # Combine losses
batch_accs = [x['val_acc'] for x in outputs]
epoch_acc = torch.stack(batch_accs).mean() # Combine accuracies
return {'val_loss': epoch_loss.item(), 'val_acc': epoch_acc.item()}
def epoch_end(self, epoch, result):
print("Epoch [{}], val_loss: {:.4f}, val_acc: {:.4f}".format(epoch, result['val_loss'], result['val_acc']))
model = CifarModel()
def accuracy(outputs, labels):
_, preds = torch.max(outputs, dim=1)
return torch.tensor(torch.sum(preds == labels).item() / len(preds))
def evaluate(model, val_loader):
outputs = [model.validation_step(batch) for batch in val_loader]
return model.validation_epoch_end(outputs)
def fit(epochs, lr, model, train_loader, val_loader, opt_func=torch.optim.SGD):
history = []
optimizer = opt_func(model.parameters(), lr)
for epoch in range(epochs):
# Training Phase
for batch in train_loader:
loss = model.training_step(batch)
loss.backward()
optimizer.step()
optimizer.zero_grad()
# Validation phase
result = evaluate(model, val_loader)
model.epoch_end(epoch, result)
history.append(result)
return history
evaluate(model, val_loader)

Here you are specifying that the number of output classes should be 10:
num_classes = 10
Your forward function does not reflect this:
xb = xb.view(-1, 64*8*8) # you get 750x4096
out = self.linear(xb) # here an input of
# input_size to linear layer = 4*4*64 # 1024
# num_classes = 10
Modify it like this:
xb = xb.view(-1, 64*4*4) # you get 750x1024
out = self.linear(xb) # M1 750x1024 M2 1024x10:
# input_size = 4*4*64 # 1024
# num_classes = 10