def main():
datasets = create_or_restore(IMDB)
train_data = datasets['train_unsup']
test_data = datasets['test_unsup']
bs = 50
bptt = 70
train = SequenceIterator(to_sequence(train_data), bptt, bs)
valid = SequenceIterator(to_sequence(test_data), bptt, bs)
lm = LanguageModel(
vocab_sz=train_data.vocab.size,
embed_sz=400, n_hidden=1150)
dev = device(force_cpu=True) if args.use_cpu else device(args.cuda)
print('Selected device: %s' % dev)
opt = optim.Adam(
lm.parameters(), lr=1e-3, weight_decay=1e-7, betas=(0.8, 0.99))
cycle_length = len(train_data) // bs
sched = CosineAnnealingLR(opt, t_max=cycle_length, cycle_mult=1, eta_min=1e-5)
loop = Loop(lm, opt, sched, device=dev)
loop.run(train_data=train, valid_data=valid,
loss_fn=F.cross_entropy,
metrics=[accuracy],
callbacks=default_callbacks())
best_model = loop['Checkpoint'].best_model
print('Best model: %s' % best_model)
with open('best', 'w') as file:
file.write(best_model + '\n')
def main():
bs = 64
n_anchors = 4
dataset = VOCDataset(TRAIN_JSON, TRAIN_JPEG, device=DEVICE)
loader = VOCDataLoader(dataset, batch_size=bs, num_workers=0)
# plotter = VOCPlotter(id2cat=dataset.id2cat, figsize=(12, 10))
#
# for images, (boxes, classes) in iter(loader):
# with plotter:
# plotter.plot_boxes(*to_np(images, boxes, classes))
# break # a single batch to verify everything works
n_classes = len(dataset.id2cat)
cycle_len = math.ceil(len(dataset)/bs)
model = SSD(n_classes=n_classes, bias=-3.)
optimizer = optim.Adam(model.parameters(), lr=1e-2)
scheduler = CosineAnnealingLR(optimizer, t_max=cycle_len)
loop = Loop(model, optimizer, scheduler, device=DEVICE)
anchors, grid_sizes = [
x.to(DEVICE) for x in (
t(make_grid(n_anchors), requires_grad=False).float(),
t([1/n_anchors], requires_grad=False).unsqueeze(1))]
bce_loss = BinaryCrossEntropyLoss(n_classes)
loss_fn = lambda x, y: ssd_loss(x, y, anchors, grid_sizes, bce_loss, n_classes)
loop.run(
train_data=loader,
epochs=100,
loss_fn=loss_fn,
callbacks=[Logger()]
)
def main():
bs, bptt = 32, 16
field = Field(lower=True, tokenize=list)
dataset = TextDataset(field, keep_new_lines=True, min_freq=5)
factory = lambda seq: SequenceIterator(seq, bptt, bs)
dataset.build(train=TRAIN_DIR, valid=VALID_DIR, iterator_factory=factory)
model = RNN(dataset.vocab_size,
n_factors=50,
batch_size=128,
n_hidden=256,
n_recurrent=3,
architecture=nn.LSTM)
optimizer = optim.Adam(model.parameters(), lr=1e-2)
cycle_length = dataset['train'].total_iters
scheduler = CosineAnnealingLR(optimizer,
t_max=cycle_length / 2,
eta_min=1e-5)
callbacks = [EarlyStopping(patience=50), Logger(), History(), Checkpoint()]
loop = Loop(Stepper(model, optimizer, scheduler, F.nll_loss))
loop.run(train_data=dataset['train'],
valid_data=dataset['valid'],
callbacks=callbacks,
epochs=500)
best_model = loop['Checkpoint'].best_model
model.load_state_dict(torch.load(best_model))
text = generate_text(model, field, seed='Deep song')
pretty_print(text)
def main():
batch_size = 10000
num_workers = cpu_count()
data_transforms = {
'train':
transforms.Compose([
transforms.RandomRotation(4),
transforms.RandomAffine(degrees=0, translate=(0.05, 0.05)),
transforms.ToTensor(),
transforms.Normalize(*STATS)
]),
'valid':
transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize(*STATS)])
}
datasets = load_dataset(data_transforms,
batch_size=batch_size,
num_workers=num_workers)
n_samples = len(datasets['train']['loader'])
n_batches = math.ceil(n_samples / batch_size)
model = ResNet(10)
opt = optim.Adam(model.parameters(), lr=1e-2)
sched = CosineAnnealingLR(opt, T_max=n_batches / 4, eta_min=1e-5)
loop = Loop(model, opt, sched, device=DEVICE)
loop.run(train_data=datasets['train']['loader'],
valid_data=datasets['valid']['loader'],
loss_fn=F.cross_entropy,
metrics=[accuracy],
callbacks=default_callbacks(),
epochs=3)
best_model = loop['Checkpoint'].best_model
weights = torch.load(best_model)
model.load_state_dict(weights)
x, y = random_sample(datasets['valid']['dataset'])
y_pred = model(x.to(DEVICE))
valid_acc = accuracy(y_pred, y.to(DEVICE))
title = f'Validation accuracy: {valid_acc:2.2%}'
titles = [str(x) for x in to_np(y_pred.argmax(dim=1))]
show_predictions(images=to_np(x.permute(0, 3, 2, 1)),
suptitle=title,
titles=titles)
dummy_input = torch.randn(16, 1, 28, 28, requires_grad=True).cuda()
torch.onnx.export(model, dummy_input, 'trivial.onnx', export_params=True)
core_ml_model = convert('digits.onnx')
core_ml_model.save('digits.mlmodel')
print('CoreML model was saved onto disk')
def main():
path = '/home/ck/data/cifar10/train'
dataset = LabelledImagesDataset(labels_from='folders',
root=path,
batch_size=2048,
one_hot=False,
transforms=[to_xy, as_tensor])
train_data = iter(dataset)
n = len(train_data)
model = SimpleResNet([10, 20, 40, 80, 160], 10).cuda()
optimizer = optim.Adam(model.parameters(), lr=1e-2)
schedule = CosineAnnealingLR(optimizer, t_max=n / 2, eta_min=1e-5)
loop = Loop(model, optimizer, schedule)
loop.run(train_data=train_data, callbacks=[Logger()])
def main():
bs = 64
bptt = 8
n_factors = 50
n_hidden = 512
n_recurrent = 2
n_epochs = 100
field = Field(lower=True, tokenize=list)
dataset = TextDataset(field, min_freq=5)
factory = lambda seq: SequenceIterator(seq, bptt, bs)
dataset.build(train=TRAIN_DIR, valid=VALID_DIR, iterator_factory=factory)
model = RNN(dataset.vocab_size,
n_factors=n_factors,
batch_size=bs,
n_hidden=n_hidden,
n_recurrent=n_recurrent,
architecture=nn.LSTM)
optimizer = optim.Adam(model.parameters(), lr=1e-2)
cycle_length = dataset['train'].total_iters
scheduler = CosineAnnealingLR(optimizer, t_max=cycle_length, eta_min=1e-5)
loop = Loop(model, optimizer, scheduler, device=DEVICE)
loop.run(train_data=dataset['train'],
valid_data=dataset['valid'],
epochs=n_epochs,
callbacks=[
EarlyStopping(patience=50),
Logger(),
History(),
Checkpoint()])
best_model = loop['Checkpoint'].best_model
model.load_state_dict(torch.load(best_model))
text = generate_text(model, field, seed='For thos')
pretty_print(text)
def main():
root = Path.home() / 'data' / 'cifar10'
data_transforms = {
'train':
transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=MEAN, std=STD)
]),
'valid':
transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize(mean=MEAN, std=STD)])
}
datasets, loaders, dataset_sizes = {}, {}, {}
for name in ('train', 'valid'):
dataset = ImageFolder(root / name, data_transforms[name])
training = name == 'train'
datasets[name] = dataset
loaders[name] = DataLoader(dataset=dataset,
batch_size=256,
shuffle=training,
num_workers=cpu_count())
dataset_sizes[name] = len(dataset)
n = len(datasets['train'])
model = CustomResNet()
optimizer = optim.Adam(model.parameters(), lr=1e-2, weight_decay=1e-5)
schedule = CosineAnnealingLR(optimizer,
t_max=n,
eta_min=1e-5,
cycle_mult=2)
loop = Loop(model, optimizer, schedule, device=DEVICE)
callbacks = [
History(),
CSVLogger(),
Logger(),
EarlyStopping(patience=50),
Checkpoint()
]
loop.run(train_data=loaders['train'],
valid_data=loaders['valid'],
callbacks=callbacks,
loss_fn=F.cross_entropy,
metrics=[accuracy],
epochs=150)
dataset = datasets['valid']
loader = DataLoader(dataset=dataset, batch_size=8, shuffle=True)
x, y = next(iter(loader))
state = torch.load(loop['Checkpoint'].best_model)
model.load_state_dict(state)
predictions = model(x.cuda())
labels = predictions.argmax(dim=1)
verbose = [dataset.classes[name] for name in labels]
imshow(utils.make_grid(x), title=verbose)
from settings import *
objects = []
providers = [
] # глобальные обработчики события отрисовки. Такие обработчики не привязаны к объектам
events_handlers = [] # обрботчики событий pygame
if __name__ == '__main__':
from core.loop import Loop
from core.key_bindings import KeyBindings
from objects.interface.menu import Menu
from core.game import Game
pygame.display.init()
pygame.display.set_caption(name)
screen = pygame.display.set_mode((width, height))
pygame.font.init()
game = Game(objects)
objects.append(Menu(screen, game))
loop = Loop(screen, objects)
game.loop = loop
KeyBindings.register(pygame.K_ESCAPE, lambda: loop.stop())
KeyBindings.register(pygame.K_q, lambda: loop.stop())
loop.run()