def check_ldp(model_name, dataset, perc) :
if dataset == 'imagenette' :
path = untar_data(URLs.IMAGENETTE)
elif dataset == 'cifar10' :
path = untar_data(URLs.CIFAR)
elif dataset == 'imagewoof' :
path = untar_data(URLs.IMAGEWOOF)
new_path = path/('new' + str(perc))
val = 'val'
sz = 224
stats = imagenet_stats
tfms = get_transforms(do_flip=False)
load_name = dataset
if dataset == 'cifar10' :
val = 'test'
sz = 32
stats = cifar_stats
load_name = dataset[ : -2]
data = ImageDataBunch.from_folder(new_path, train = 'train', valid = 'val', test = 'test', bs = 64, size = sz, ds_tfms = tfms).normalize(stats)
if model_name == 'resnet10' :
net = resnet10(pretrained = False, progress = False)
elif model_name == 'resnet14' :
net = resnet14(pretrained = False, progress = False)
elif model_name == 'resnet18' :
net = resnet18(pretrained = False, progress = False)
elif model_name == 'resnet20' :
net = resnet20(pretrained = False, progress = False)
elif model_name == 'resnet26' :
net = resnet26(pretrained = False, progress = False)
savename = '../saved_models/' + dataset + '/less_data' + str(perc) + '/' + model_name + '_classifier/model0.pt'
net.load_state_dict(torch.load(savename, map_location = 'cpu'))
net.cuda()
ld_stagewise_acc = _get_accuracy(data.valid_dl, net)
return ld_stagewise_acc
for i, (images, labels) in enumerate(data.valid_dl):
if torch.cuda.is_available():
images = torch.autograd.Variable(images).cuda().float()
labels = torch.autograd.Variable(labels).cuda()
else:
images = torch.autograd.Variable(images).float()
labels = torch.autograd.Variable(labels)
# Forward pass
y_pred = net(images)
loss = F.cross_entropy(y_pred, labels)
val.append(loss.item())
val_loss = sum(val) / len(val)
val_loss_list.append(val_loss)
val_acc = _get_accuracy(data.valid_dl, net)
experiment.log_metric("train_loss", train_loss)
experiment.log_metric("val_loss", val_loss)
experiment.log_metric("val_acc", val_acc * 100)
print('epoch : ', epoch + 1,
' / ', hyper_params["num_epochs"], ' | TL : ',
round(train_loss, 6), ' | VL : ', round(val_loss, 6), ' | VA : ',
round(val_acc * 100, 6))
if (val_acc * 100) > min_val:
print('saving model')
min_val = val_acc * 100
torch.save(net.state_dict(), savename)
for i, (images, labels) in enumerate(data.valid_dl) :
if torch.cuda.is_available():
images = torch.autograd.Variable(images).cuda().float()
labels = torch.autograd.Variable(labels).cuda()
else :
images = torch.autograd.Variable(images).float()
labels = torch.autograd.Variable(labels)
# Forward pass
outputs = net(images)
loss = F.cross_entropy(outputs, labels)
val.append(loss)
val_loss = (sum(val) / len(val)).item()
val_loss_list.append(val_loss)
val_acc = _get_accuracy(data.valid_dl, net)
val_acc_list.append(val_acc)
print('epoch : ', epoch + 1, ' / ', hyper_params['num_epochs'], ' | TL : ', round(train_loss, 4), ' | VL : ', round(val_loss, 4), ' | VA : ', round(val_acc * 100, 6))
if (val_acc * 100) > min_val :
print('saving model')
min_val = val_acc * 100
torch.save(net.state_dict(), savename)
# plt.plot(range(hyper_params['num_epochs']), train_loss_list, 'r', label = 'training_loss')
# plt.plot(range(hyper_params['num_epochs']), val_loss_list, 'b', label = 'validation_loss')
# plt.legend()
# plt.savefig('../figures/' + str(hyper_params['dataset']) + '/resnet14_no_teacher/training_losses' + str(repeated) + '.jpg')
# plt.close()
# plt.plot(range(hyper_params['num_epochs']), val_acc_list, 'r', label = 'validation_accuracy')