def test_get_correlated_mask(self):
loss = NTXentLoss()
for bsz in range(1, 1000):
mask = loss._torch_get_correlated_mask(bsz)
# correct number of zeros in mask
self.assertAlmostEqual(mask.sum(), 4 * (bsz * bsz - bsz))
# if mask is correct,
# (1 - mask) * v adds up the first and second half of v
v = torch.randn((2 * bsz))
mv = torch.mv(1. - mask.float(), v)
vv = (v[bsz:] + v[:bsz]).repeat(2)
self.assertAlmostEqual((mv - vv).pow(2).sum(), 0.)
def test_forward_pass_memory_bank_cuda(self):
if not torch.cuda.is_available():
return
loss = NTXentLoss(memory_bank_size=64)
for bsz in range(1, 100):
batch = torch.randn(2 * bsz, 32).cuda()
l = loss(batch)
def __init__(self):
super().__init__()
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
self.projection_head = SimCLRProjectionHead(512, 2048, 2048)
# enable gather_distributed to gather features from all gpus
# before calculating the loss
self.criterion = NTXentLoss(gather_distributed=True)
def __init__(self):
super().__init__()
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
self.projection_head = NNCLRProjectionHead(512, 512, 128)
self.prediction_head = NNCLRPredictionHead(128, 512, 128)
self.memory_bank = NNMemoryBankModule(size=4096)
self.criterion = NTXentLoss()
def test_forward_pass(self):
loss = NTXentLoss(memory_bank_size=0)
for bsz in range(1, 20):
batch_1 = torch.randn((bsz, 32))
batch_2 = torch.randn((bsz, 32))
# symmetry
l1 = loss(batch_1, batch_2)
l2 = loss(batch_2, batch_1)
self.assertAlmostEqual((l1 - l2).pow(2).item(), 0.)
def test_forward_pass_neg_temp(self):
loss = NTXentLoss(temperature=-1., memory_bank_size=0)
for bsz in range(1, 100):
batch_1 = torch.randn((bsz, 32))
batch_2 = torch.randn((bsz, 32))
# symmetry
l1 = loss(torch.cat((batch_1, batch_2), 0))
l2 = loss(torch.cat((batch_2, batch_1), 0))
self.assertAlmostEqual((l1 - l2).pow(2).item(), 0.)
def test_forward_pass_1d(self):
loss = NTXentLoss()
for bsz in range(1, 100):
batch_1 = torch.randn((bsz, 1))
batch_2 = torch.randn((bsz, 1))
# symmetry
l1 = loss(torch.cat((batch_1, batch_2), 0))
l2 = loss(torch.cat((batch_2, batch_1), 0))
self.assertAlmostEqual((l1 - l2).pow(2).item(), 0.)
def __init__(self):
super().__init__()
# create a ResNet backbone and remove the classification head
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
hidden_dim = resnet.fc.in_features
self.projection_head = SimCLRProjectionHead(hidden_dim, hidden_dim,
128)
self.criterion = NTXentLoss()
def __init__(self):
super().__init__()
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
self.projection_head = MoCoProjectionHead(512, 512, 128)
self.backbone_momentum = copy.deepcopy(self.backbone)
self.projection_head_momentum = copy.deepcopy(self.projection_head)
deactivate_requires_grad(self.backbone_momentum)
deactivate_requires_grad(self.projection_head_momentum)
self.criterion = NTXentLoss(memory_bank_size=4096)
def test_forward_pass_cuda(self):
if torch.cuda.is_available():
loss = NTXentLoss(memory_bank_size=0)
for bsz in range(1, 100):
batch_1 = torch.randn((bsz, 32)).cuda()
batch_2 = torch.randn((bsz, 32)).cuda()
# symmetry
l1 = loss(torch.cat((batch_1, batch_2), 0))
l2 = loss(torch.cat((batch_2, batch_1), 0))
self.assertAlmostEqual((l1 - l2).pow(2).item(), 0.)
else:
pass
def test_with_values(self):
for n_samples in [1, 2, 4]:
for dimension in [1, 2, 16, 64]:
for temperature in [0.1, 1, 10]:
out0 = np.random.normal(0, 1, size=(n_samples, dimension))
out1 = np.random.normal(0, 1, size=(n_samples, dimension))
with self.subTest(msg=f"out0.shape={out0.shape}, temperature={temperature}"):
out0 = torch.FloatTensor(out0)
out1 = torch.FloatTensor(out1)
loss_function = NTXentLoss(temperature=temperature)
l1 = float(loss_function(out0, out1))
l2 = float(loss_function(out1, out0))
l1_manual = self.calc_ntxent_loss_manual(out0, out1, temperature=temperature)
l2_manual = self.calc_ntxent_loss_manual(out0, out1, temperature=temperature)
self.assertAlmostEqual(l1, l2, places=5)
self.assertAlmostEqual(l1, l1_manual, places=5)
self.assertAlmostEqual(l2, l2_manual, places=5)
def test_with_correlated_embedding(self):
for n_samples in [1, 2, 8, 16]:
for memory_bank_size in [0, 1, 2, 8, 15, 16, 17]:
for temperature in [0.1, 1, 7]:
out0 = np.random.random((n_samples, 1))
out1 = np.random.random((n_samples, 1))
out0 = np.concatenate([out0, 2 * out0], axis=1)
out1 = np.concatenate([out1, 2 * out1], axis=1)
out0 = torch.FloatTensor(out0)
out1 = torch.FloatTensor(out1)
out0.requires_grad = True
with self.subTest(msg=f"n_samples: {n_samples}, memory_bank_size: {memory_bank_size},"
f"temperature: {temperature}"):
loss_function = NTXentLoss(temperature=temperature, memory_bank_size=memory_bank_size)
if memory_bank_size > 0:
for i in range(int(memory_bank_size / n_samples) + 2):
# fill the memory bank over multiple rounds
loss = float(loss_function(out0, out1))
expected_loss = -1 * np.log(1 / (memory_bank_size + 1))
else:
loss = float(loss_function(out0, out1))
expected_loss = -1 * np.log(1 / (2 * n_samples - 1))
self.assertAlmostEqual(loss, expected_loss, places=5)
def _train_cli(cfg, is_cli_call=True):
input_dir = cfg['input_dir']
if input_dir and is_cli_call:
input_dir = fix_input_path(input_dir)
if 'seed' in cfg.keys():
seed = cfg['seed']
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if cfg["trainer"]["weights_summary"] == "None":
cfg["trainer"]["weights_summary"] = None
if torch.cuda.is_available():
device = 'cuda'
elif cfg['trainer'] and cfg['trainer']['gpus']:
device = 'cpu'
cfg['trainer']['gpus'] = 0
else:
device = 'cpu'
distributed_strategy = None
if cfg['trainer']['gpus'] > 1:
distributed_strategy = 'ddp'
if cfg['loader']['batch_size'] < 64:
msg = 'Training a self-supervised model with a small batch size: {}! '
msg = msg.format(cfg['loader']['batch_size'])
msg += 'Small batch size may harm embedding quality. '
msg += 'You can specify the batch size via the loader key-word: '
msg += 'loader.batch_size=BSZ'
warnings.warn(msg)
# determine the number of available cores
if cfg['loader']['num_workers'] < 0:
cfg['loader']['num_workers'] = cpu_count()
state_dict = None
checkpoint = cfg['checkpoint']
if cfg['pre_trained'] and not checkpoint:
# if checkpoint wasn't specified explicitly and pre_trained is True
# try to load the checkpoint from the model zoo
checkpoint, key = get_ptmodel_from_config(cfg['model'])
if not checkpoint:
msg = 'Cannot download checkpoint for key {} '.format(key)
msg += 'because it does not exist! '
msg += 'Model will be trained from scratch.'
warnings.warn(msg)
elif checkpoint:
checkpoint = fix_input_path(checkpoint) if is_cli_call else checkpoint
if checkpoint:
# load the PyTorch state dictionary and map it to the current device
if is_url(checkpoint):
state_dict = load_state_dict_from_url(
checkpoint, map_location=device)['state_dict']
else:
state_dict = torch.load(checkpoint,
map_location=device)['state_dict']
# load model
resnet = ResNetGenerator(cfg['model']['name'], cfg['model']['width'])
last_conv_channels = list(resnet.children())[-1].in_features
features = nn.Sequential(
get_norm_layer(3, 0),
*list(resnet.children())[:-1],
nn.Conv2d(last_conv_channels, cfg['model']['num_ftrs'], 1),
nn.AdaptiveAvgPool2d(1),
)
model = _SimCLR(features,
num_ftrs=cfg['model']['num_ftrs'],
out_dim=cfg['model']['out_dim'])
if state_dict is not None:
load_from_state_dict(model, state_dict)
criterion = NTXentLoss(**cfg['criterion'])
optimizer = torch.optim.SGD(model.parameters(), **cfg['optimizer'])
dataset = LightlyDataset(input_dir)
cfg['loader']['batch_size'] = min(cfg['loader']['batch_size'],
len(dataset))
collate_fn = ImageCollateFunction(**cfg['collate'])
dataloader = torch.utils.data.DataLoader(dataset,
**cfg['loader'],
collate_fn=collate_fn)
encoder = SelfSupervisedEmbedding(model, criterion, optimizer, dataloader)
encoder.init_checkpoint_callback(**cfg['checkpoint_callback'])
encoder.train_embedding(**cfg['trainer'], strategy=distributed_strategy)
print(
f'Best model is stored at: {bcolors.OKBLUE}{encoder.checkpoint}{bcolors.ENDC}'
)
os.environ[cfg['environment_variable_names']
['lightly_last_checkpoint_path']] = encoder.checkpoint
return encoder.checkpoint
def _train_cli(cfg, is_cli_call=True):
data = cfg['data']
download = cfg['download']
root = cfg['root']
if root and is_cli_call:
root = fix_input_path(root)
input_dir = cfg['input_dir']
if input_dir and is_cli_call:
input_dir = fix_input_path(input_dir)
if 'seed' in cfg.keys():
seed = cfg['seed']
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if torch.cuda.is_available():
device = 'cuda'
elif cfg['trainer'] and cfg['trainer']['gpus']:
device = 'cpu'
cfg['trainer']['gpus'] = 0
if cfg['loader']['batch_size'] < 64:
msg = 'Training a self-supervised model with a small batch size: {}! '
msg = msg.format(cfg['loader']['batch_size'])
msg += 'Small batch size may harm embedding quality. '
msg += 'You can specify the batch size via the loader key-word: '
msg += 'loader.batch_size=BSZ'
warnings.warn(msg)
state_dict = None
checkpoint = cfg['checkpoint']
if cfg['pre_trained'] and not checkpoint:
# if checkpoint wasn't specified explicitly and pre_trained is True
# try to load the checkpoint from the model zoo
checkpoint, key = get_ptmodel_from_config(cfg['model'])
if not checkpoint:
msg = 'Cannot download checkpoint for key {} '.format(key)
msg += 'because it does not exist! '
msg += 'Model will be trained from scratch.'
warnings.warn(msg)
elif checkpoint:
checkpoint = fix_input_path(checkpoint) if is_cli_call else checkpoint
if checkpoint:
# load the PyTorch state dictionary and map it to the current device
if is_url(checkpoint):
state_dict = load_state_dict_from_url(
checkpoint, map_location=device
)['state_dict']
else:
state_dict = torch.load(
checkpoint, map_location=device
)['state_dict']
# load model
model = ResNetSimCLR(**cfg['model'])
if state_dict is not None:
model.load_from_state_dict(state_dict)
criterion = NTXentLoss(**cfg['criterion'])
optimizer = torch.optim.SGD(model.parameters(), **cfg['optimizer'])
dataset = LightlyDataset(root,
name=data, train=True, download=download,
from_folder=input_dir)
cfg['loader']['batch_size'] = min(
cfg['loader']['batch_size'],
len(dataset)
)
collate_fn = ImageCollateFunction(**cfg['collate'])
dataloader = torch.utils.data.DataLoader(dataset,
**cfg['loader'],
collate_fn=collate_fn)
encoder = SelfSupervisedEmbedding(model, criterion, optimizer, dataloader)
encoder.init_checkpoint_callback(**cfg['checkpoint_callback'])
encoder = encoder.train_embedding(**cfg['trainer'])
print('Best model is stored at: %s' % (encoder.checkpoint))
return encoder.checkpoint
collate_fn = SimCLRCollateFunction(
input_size=32,
gaussian_blur=0.,
)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=256,
collate_fn=collate_fn,
shuffle=True,
drop_last=True,
num_workers=8,
)
criterion = NTXentLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.06)
print("Starting Training")
for epoch in range(10):
total_loss = 0
for (x0, x1), _, _ in dataloader:
x0 = x0.to(device)
x1 = x1.to(device)
z0 = model(x0)
z1 = model(x1)
loss = criterion(z0, z1)
total_loss += loss.detach()
loss.backward()
optimizer.step()
optimizer.zero_grad()
dataset = LightlyDataset.from_torch_dataset(cifar10)
# or create a dataset from a folder containing images or videos:
# dataset = LightlyDataset("path/to/folder")
collate_fn = MoCoCollateFunction(input_size=32)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=256,
collate_fn=collate_fn,
shuffle=True,
drop_last=True,
num_workers=8,
)
criterion = NTXentLoss(memory_bank_size=4096)
optimizer = torch.optim.SGD(model.parameters(), lr=0.06)
print("Starting Training")
for epoch in range(10):
total_loss = 0
for (x_query, x_key), _, _ in dataloader:
update_momentum(model.backbone, model.backbone_momentum, m=0.99)
update_momentum(model.projection_head, model.projection_head_momentum, m=0.99)
x_query = x_query.to(device)
x_key = x_key.to(device)
query = model(x_query)
key = model.forward_momentum(x_key)
loss = criterion(query, key)
total_loss += loss.detach()
loss.backward()
def __init__(self):
super().__init__()
resnet = torchvision.models.resnet18()
self.backbone = nn.Sequential(*list(resnet.children())[:-1])
self.projection_head = SimCLRProjectionHead(512, 2048, 2048)
self.criterion = NTXentLoss()
def test_forward_pass_memory_bank(self):
loss = NTXentLoss(memory_bank_size=64)
for bsz in range(1, 100):
batch = torch.randn(2 * bsz, 32)
l = loss(batch)
def test_forward_pass_memory_bank(self):
loss = NTXentLoss(memory_bank_size=64)
for bsz in range(1, 20):
batch_1 = torch.randn((bsz, 32))
batch_2 = torch.randn((bsz, 32))
l = loss(batch_1, batch_2)
