Skip to content
Snippets Groups Projects
Commit 7a79b40c authored by Rolf Sievert's avatar Rolf Sievert
Browse files

Added validation

parent 2b45fcf5
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
python/main.py
+ 87
− 45
View file @ 7a79b40c
import os import os
import glob import glob
from math import sqrt from math import sqrt
...@@ -13,6 +12,10 @@ from train import train ...@@ -13,6 +12,10 @@ from train import train
from test import test from test import test
import numpy as np import numpy as np
from constants import MEAN_MAP, STD_MAP from constants import MEAN_MAP, STD_MAP
from torch.utils.data.sampler import SubsetRandomSampler
# Set local working directory
os.chdir(os.path.dirname(os.path.realpath(__file__)))
################################## ##################################
## RESULTS ## ## RESULTS ##
...@@ -59,6 +62,8 @@ LEARNING_RATE = 0.001 # DEAFULT IS 0.001 ...@@ -59,6 +62,8 @@ LEARNING_RATE = 0.001 # DEAFULT IS 0.001
## PREPARE DATA ## ## PREPARE DATA ##
################################## ##################################
NUM_EPOCHS = 30
TRAIN_LOGGING = False
CHANNELS = 1 if SPACE == "L" else 3 CHANNELS = 1 if SPACE == "L" else 3
MEAN = MEAN_MAP[SPACE] MEAN = MEAN_MAP[SPACE]
STD = MEAN_MAP[SPACE] STD = MEAN_MAP[SPACE]
...@@ -72,61 +77,98 @@ transform_comp = transforms.Compose([ ...@@ -72,61 +77,98 @@ transform_comp = transforms.Compose([
transforms.Normalize(MEAN, STD), transforms.Normalize(MEAN, STD),
]) ])
trainset = torchvision.datasets.CIFAR10(root='./data', NUM_WORKERS = 2
train=True, # If running on windows, change to 0 workers (does not work otherwise)
if os.name == 'nt':
NUM_WORKERS = 0
def sampleSplit(dataset, split):
"""
Splits indexes of dataset.
"""
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(split * dataset_size))
# Shuffle indices
np.random.seed(123)
np.random.shuffle(indices)
train_indices, val_indices = indices[:split], indices[split:]
# Creating PT data samplers and loaders:
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
return train_sampler, valid_sampler
if __name__ == '__main__':
trainvalset = torchvision.datasets.CIFAR10(root='./data',
train=True,
download=True,
transform=transform_comp)
# Get random indices for train and validation data
train_sampler, val_sampler = sampleSplit(trainvalset, 0.75)
trainloader = torch.utils.data.DataLoader(trainvalset,
batch_size=256,
sampler=train_sampler,
num_workers=NUM_WORKERS)
valloader = torch.utils.data.DataLoader(trainvalset,
batch_size=256,
sampler=val_sampler,
num_workers=NUM_WORKERS)
testset = torchvision.datasets.CIFAR10(root='./data',
train=False,
download=True, download=True,
transform=transform_comp) transform=transform_comp)
trainloader = torch.utils.data.DataLoader(trainset, testloader = torch.utils.data.DataLoader(testset,
batch_size=256, batch_size=128,
shuffle=True, shuffle=False,
num_workers=2) num_workers=NUM_WORKERS)
testset = torchvision.datasets.CIFAR10(root='./data', ##################################
train=False, ## PREPARE NETWORK ##
download=True, ##################################
transform=transform_comp)
testloader = torch.utils.data.DataLoader(testset, use_cuda = torch.cuda.is_available()
batch_size=128, if use_cuda:
shuffle=False, print("Cuda support enabled.")
num_workers=2) else:
################################## print("Cuda support DISABLED.")
## PREPARE NETWORK ##
##################################
use_cuda = torch.cuda.is_available() classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', #model = cvlNet()
'ship', 'truck') model = goalNet(CHANNELS)
#model = cvlNet() if use_cuda:
model = goalNet(CHANNELS) model.cuda()
cudnn.benchmark = True
if use_cuda: objective = nn.CrossEntropyLoss()
model.cuda()
cudnn.benchmark = True
objective = nn.CrossEntropyLoss() optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE) ##################################
## TRAIN NETWORK ##
##################################
################################## model, loss_log, acc_log = train(model,
## TRAIN NETWORK ## trainloader,
################################## valloader,
optimizer,
start_epoch = 1 objective,
num_epochs = 50 use_cuda,
model, loss_log, acc_log = train(model, start_epoch=1,
trainloader, num_epochs=NUM_EPOCHS,
optimizer, log=TRAIN_LOGGING)
objective,
use_cuda,
start_epoch,
num_epochs=num_epochs)
################################## ##################################
## TEST NETWORK ## ## TEST NETWORK ##
################################## ##################################
test_acc = test(model, testloader, use_cuda) test_acc = test(model, testloader, use_cuda)
\ No newline at end of file
python/test.py
+ 2
− 2
View file @ 7a79b40c
...@@ -5,7 +5,6 @@ from torch.autograd import Variable ...@@ -5,7 +5,6 @@ from torch.autograd import Variable
def test(model, testloader, use_cuda): def test(model, testloader, use_cuda):
correct = 0 correct = 0
total = 0 total = 0
for idx, (inputs, targets) in enumerate(testloader): for idx, (inputs, targets) in enumerate(testloader):
...@@ -14,7 +13,8 @@ def test(model, testloader, use_cuda): ...@@ -14,7 +13,8 @@ def test(model, testloader, use_cuda):
if use_cuda: if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda() inputs, targets = inputs.cuda(), targets.cuda()
inputs = torch.tensor(inputs, requires_grad=False) #inputs = torch.tensor(inputs, requires_grad=False)
inputs = inputs.clone().detach().requires_grad_(False)
# Feed-forward the network # Feed-forward the network
outputs = model(inputs) outputs = model(inputs)
......
python/train.py
+ 29
− 3
View file @ 7a79b40c
...@@ -11,7 +11,28 @@ import os ...@@ -11,7 +11,28 @@ import os
#- Backpropagate the loss until the first layer. #- Backpropagate the loss until the first layer.
#- Update the weights for different convolution layers based on the gradients. #- Update the weights for different convolution layers based on the gradients.
def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_epochs): def validate(model, valloader, use_cuda, epoch):
correct = 0
total = 0
for idx, (inputs, targets) in enumerate(valloader):
# Move data to GPU if CUDA is available
if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda()
#inputs = torch.tensor(inputs, requires_grad=False)
inputs = inputs.clone().detach().requires_grad_(False)
# Feed-forward the network
outputs = model(inputs)
_, predicted = torch.max(outputs.data, 1)
total += targets.size(0)
correct += (predicted == targets).sum().item()
print('Validation accuracy after epoch %d: %d %%' % (epoch, 100 * correct / total))
def train(model, trainloader, valloader, optimizer, objective, use_cuda, start_epoch, num_epochs, log=True):
loss_log = [] loss_log = []
acc_log = [] acc_log = []
...@@ -33,7 +54,8 @@ def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_e ...@@ -33,7 +54,8 @@ def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_e
if use_cuda: if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda() inputs, targets = inputs.cuda(), targets.cuda()
inputs, targets = torch.tensor(inputs, requires_grad=True), torch.tensor(targets).long() #inputs, targets = torch.tensor(inputs, requires_grad=True), torch.tensor(targets).long()
inputs, targets = inputs.clone().detach().requires_grad_(True), targets.clone().detach()
# Clear the gradients of all variables # Clear the gradients of all variables
optimizer.zero_grad() optimizer.zero_grad()
...@@ -60,7 +82,8 @@ def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_e ...@@ -60,7 +82,8 @@ def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_e
total += targets.size(0) total += targets.size(0)
correct += (predicted == targets).sum().item() correct += (predicted == targets).sum().item()
print('Loss: {:.8f} | Acc: {:.2f}% ({}/{})'.format((train_loss/(batch_idx+1)), 100.*correct/total, correct, total)) if log:
print('Loss: {:.8f} | Acc: {:.2f}% ({}/{})'.format((train_loss/(batch_idx+1)), 100.*correct/total, correct, total))
loss_log.append((train_loss/(batch_idx+1))) loss_log.append((train_loss/(batch_idx+1)))
acc_log.append(100.*correct/total) acc_log.append(100.*correct/total)
...@@ -76,6 +99,9 @@ def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_e ...@@ -76,6 +99,9 @@ def train(model, trainloader, optimizer, objective, use_cuda, start_epoch, num_e
file_path = './checkpoints/checkpoint_{}.ckpt'.format(epoch) file_path = './checkpoints/checkpoint_{}.ckpt'.format(epoch)
torch.save(state, file_path) torch.save(state, file_path)
# Run and print validation
validate(model, valloader, use_cuda, epoch)
# Save the final model # Save the final model
torch.save(model.state_dict(), './models/cvlnet_trained_cifar10_final') torch.save(model.state_dict(), './models/cvlnet_trained_cifar10_final')
print('Training Finished!') print('Training Finished!')
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment