def main():
# TODO: shift to a config file
parser = argparse.ArgumentParser()
parser.add_argument("--filepath", help="directory containing data")
parser.add_argument("--filename", help="file containing data")
parser.add_argument("--remove_first_column",
help="Remove first column of data")
parser.add_argument("--epochs", help="total number of epochs")
parser.add_argument("--batch_size", help="size of a batch")
parser.add_argument("--learning_rate", help="learning rate if using SGD")
args = parser.parse_args()
train_x, train_y, test_x, test_y = get_data(args.filepath, args.filename)
train_ds = Dataset(train_x, train_y) # a dataset just gives a __getitem__
test_ds = Dataset(test_x, test_y)
train_dl = DataLoader(dataset=train_ds, batch_size=int(args.batch_size))
# a dataloader does 2 things: shuffles the data and convert values to a tensor
test_dl = DataLoader(dataset=test_ds, batch_size=2 * int(args.batch_size))
no_of_classes = 1 + int(test_y.max().item())
optim = SGDOptimizerForModules(float(args.learning_rate))
model = LogisticClassification([len(train_x[0]), 100],
[100, no_of_classes], optim)
epochs = int(args.epochs)
callbacks = [
LossAndAccuracyCallback(),
LrScheduler(cosine_schedule(float(args.learning_rate), 0.001)),
LrRecorder()
]
runner = Runner(model, train_dl, test_dl, callbacks)
runner.fit(epochs)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--filepath", help="directory containing the file")
parser.add_argument("--filename", help="file containing data")
parser.add_argument("--epochs", type=int, help="total number of epochs")
parser.add_argument("--batch_size", type=int, help="size of a batch")
parser.add_argument("--learning_rate",
type=float,
help="learning rate if using SGD")
args = parser.parse_args()
train_x, train_y, test_x, test_y = get_data(args.filepath, args.filename)
train_ds = Dataset(train_x, train_y)
test_ds = Dataset(test_x, test_y)
train_dl = DataLoader(dataset=train_ds, batch_size=args.batch_size)
test_dl = DataLoader(dataset=test_ds, batch_size=2 * args.batch_size)
no_of_classes = 1 + (train_y.max().item())
model = LogisticClassification([len(train_x[0]), 100], [100, 50],
[50, no_of_classes])
optim = Optimizer(model.trainable_params, args.learning_rate)
callbacks = [
LossAndAccuracyCallback(),
LrScheduler(cosine_schedule(args.learning_rate, 0.01))
]
runner = RunnerPytorch(model, optim, train_dl, test_dl, callbacks)
runner.fit(args.epochs)
# This is the problem with individual param hooks, will have to plot them seperately
# Its better to get gradients for entire layers
#TODO: add better plots
# Mean of gradients of W
plt.plot(model.layer1.grads_mean[0], label="Layer_1_W")
plt.plot(model.layer2.grads_mean[0], label="Layer_2_W")
plt.plot(model.layer3.grads_mean[0], label="Layer_3_W")
plt.legend(loc="upper right")
plt.title("Means of Weights")
plt.show()
plt.close()
# Variances of gradients of W
plt.plot(model.layer1.grads_variance[0], label="Layer_1_W")
plt.plot(model.layer2.grads_variance[0], label="Layer_2_W")
plt.plot(model.layer3.grads_variance[0], label="Layer_3_W")
plt.legend(loc="upper right")
plt.title("Variances of Weights")
plt.show()
def main():
# TODO: shift to a config file
parser = argparse.ArgumentParser()
parser.add_argument("--filepath", help="directory containing data")
parser.add_argument("--filename", help="file containing data")
parser.add_argument("--remove_first_column",
help="Remove first column of data")
parser.add_argument("--epochs", help="total number of epochs")
parser.add_argument("--batch_size", help="size of a batch")
parser.add_argument(
"--update_rule",
help=
"Matrix or SGD for normal form update or stochastic gradient descent")
parser.add_argument("--learning_rate", help="learning rate if using SGD")
parser.add_argument(
"--loss_type",
help="MSE or ML for mean squared error or maximum likelihood")
parser.add_argument("--regularization", help="L1 or L2 regularization")
parser.add_argument("--regularization_constant",
help="regularization constant")
args = parser.parse_args()
train_x, train_y, test_x, test_y = get_data(args.filepath, args.filename)
train_ds = Dataset(train_x, train_y) # a dataset just gives a __getitem__
test_ds = Dataset(test_x, test_y)
train_dl = DataLoader(dataset=train_ds, batch_size=int(args.batch_size))
# a dataloader does 2 things: shuffles the data and convert values to a tensor
test_dl = DataLoader(dataset=test_ds, batch_size=2 * int(args.batch_size))
torch.autograd.set_detect_anomaly(True)
no_of_classes = 1 + int(test_y.max().item())
# optim = Optimizer(model.trainable_params,float(args.learning_rate))
optim = SGDOptimizerForModules(float(args.learning_rate))
model = LogisticClassification([len(train_x[0]), 100], [100, 50],
[50, no_of_classes], optim)
epochs = int(args.epochs)
batch_size = int(args.batch_size)
for i in range(epochs):
for j, (x, y) in enumerate(train_dl):
print(" Train error for epoch " + str(i) + " and batch " +
str(j + 1) + " : ")
model.forward(x, y)
model.backward()
for j, (x, y) in enumerate(test_dl):
print(" Test error for epoch " + str(i) + " and batch " +
str(j + 1) + " : ")
with torch.no_grad():
model.forward(x, y)
def loadData():
#Create the dataset
dataset = Dataset(csv_file = 'csv.csv', \
transform = transform.Compose([transform.Resize((100,100)), \
transform.ToTensor(),\
transform.Normalize((0.5, 0.5, 0.5), \
(0.5, 0.5, 0.5))]))
dataset_size = len(dataset)
indices = list(range(dataset_size))
split1 = int(np.floor(.6 * dataset_size))
split2 = int(np.floor(.8 * dataset_size))
train_indices, val_indices, test_indices = indices[:split1], indices[
split1:split2], indices[split2:]
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
test_sampler = SubsetRandomSampler(test_indices)
trainloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=train_sampler)
valloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
testloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=test_sampler)
return trainloader, valloader, testloader
#### Save ####
torch.save(net.state_dict(), PATH_TO_MODEL)
print_time("Training time :", debug_level, "TRAIN", time_start, 3)
return
if __name__ == "__main__":
#### Debug settings ####
PRINT_LEVEL = "TRAIN" # Possible modes : DEBUG, INFO, RUN, TRAIN
print_info("Starting training with debug level : " + PRINT_LEVEL,
PRINT_LEVEL, "TRAIN")
#### Pytorch settings ####
torch.set_default_tensor_type(torch.FloatTensor)
print_info("Working device : " + str(DEVICE), PRINT_LEVEL, "INFO")
#### Net ####
Net = DDSPNet().float()
Net = Net.to(DEVICE)
#### Data ####
Dataset = Dataset()
Dataloader = DataLoader(Dataset,
batch_size=BATCH_SIZE,
shuffle=SHUFFLE_DATALOADER)
#### Train ####
train(Net, Dataloader, NUMBER_EPOCHS, PRINT_LEVEL)
train_states = torch.load(FILEPATH_MODEL_LOAD)
model_ft.load_state_dict(train_states['train_states_latest']['model_state_dict'])
optimizer_ft.load_state_dict(train_states['train_states_latest']['optimizer_state_dict'])
train_states_best = train_states['train_states_best']
loss_valid_min = train_states_best['loss_valid_min']
model_save_criteria = train_states_best['model_save_criteria']
else:
train_states = {}
model_save_criteria = np.inf
nEpochs = 100
batchSize = 10
transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])])
train_dataset = Dataset(train_dir, trainImages, transform=transform)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batchSize, shuffle=True)
valid_dataset = Dataset(valid_dir, validImages, transform=transform)
valid_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=batchSize, shuffle=False)
loss_train = []
loss_valid = []
for epoch in range(nEpochs):
running_loss = 0
# epoch_accuracy = 0
running_time_batch = 0
time_batch_start = time.time()
model_ft.train()
print("training...")
from network import UnetPP
from loss import LosswithLogit
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', '-bs', type=int, default=1)
parser.add_argument('--log_dir', type=str, default='log')
parser.add_argument('-lr', '--learning_rate', type=float, default=1e-4)
args = parser.parse_args()
set_logging_path(
path=join(args.log_dir, 'log{}.txt'.format(datetime.now().strftime(
"%Y-%m-%d-%H-%M-%S"))))
trainloader = data.DataLoader(dataset=Dataset(root="./data/stage1_train_neat"),
batch_size=args.batch_size,
shuffle=True)
testloader = data.DataLoader(dataset=Dataset(root="./data/stage1_test_neat"),
batch_size=args.batch_size,
shuffle=False)
device = torch.device("cuda:0")
model = UnetPP()
model.to(device)
criterion = LosswithLogit()
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
#lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=, gamma=)
label_dir = r'C:\CS8395_DLMIC\data\assignment1_data\labels'
label_txt = glob.glob(os.path.join(label_dir, '*els.txt'))
model = ResNet_PC().to(device)
torch.cuda.manual_seed(1)
model.apply(weights_init)
optimizer = torch.optim.Adam(model.parameters(), lr=0.005)
criterion = nn.MSELoss().cuda()
with open(os.path.join(label_txt[0]), 'r') as f:
filenames = f.readlines()
filenames = [item.strip() for item in filenames]
transform = transforms.ToTensor()
train_dataset = Dataset(train_dir, filenames, transform=transform)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=4,
shuffle=True)
# model = ResNet_PC().to(device)
# model.apply(weights_init)
if __name__ != "__main__":
input = torch.rand([4, 3, 326, 490]).float().to(device)
# print(input)
# print(model)
output = model(input)
print("Output shape: ", output.shape) #Output shape: torch.Size([4, 2])
for tBatchIdx, sample in enumerate(train_loader):
print("training...")
# time_batch_load = time.time() - time_batch_start