def train(model, crit, optim, dataset, embedding_sz, device, epoch):
for i, data in enumerate(dataset):
img, lbl = data
img = img.to(device)
z = torch.empty(img.shape[0], embedding_sz, device=device).normal_()
# Real images
optim['D'].zero_grad()
score_real = Discriminator(img)
l_real = optim(score_real, torch.ones(img.shape[0]))
# Fake images
G_z = Generator(z)
score_fake = Discriminator(G_z.detach())
l_fake = optim(score_fake, torch.zeros(img.shape[0]))
l_disc = 0.5*(l_fake + l_real)
l_disc.backward()
optim['D'].step()
# Generator Training
optim['G'].zero_grad()
score_fake = Discriminator(G_z)
l_gen = optim(score_fake, torch.ones(img.shape[0]))
l_gen.backward()
optim['G'].step()
def build(self, model_params):
cfg_as_dict = self.cfg.asdict()
name = cfg_as_dict.pop("name", self.defaults.name)
optim = self.optimizers[name]
return optim(model_params, **cfg_as_dict)
def loss_on_random_task(initial_model, K, num_steps, optim=torch.optim.SGD):
"""
trains the model on a random sine task and measures the loss curve.
for each n in num_steps_measured, records the model function after n gradient updates.
"""
# copy MAML model into a new object to preserve MAML weights during training
model = nn.Sequential(
OrderedDict([('l1', nn.Linear(600, 300)), ('relu1', nn.ReLU()),
('l2', nn.Linear(300, 150)), ('relu2', nn.ReLU()),
('l3', nn.Linear(150, 70)), ('relu3', nn.ReLU()),
('l4', nn.Linear(70, 20)), ('relu4', nn.ReLU()),
('l5', nn.Linear(20, 1))]))
model.load_state_dict(initial_model.state_dict())
criterion = nn.MSELoss()
optimiser = optim(model.parameters(), 0.01)
# train model on a random task
X, y = tasks.sample_data(K)
losses = []
for step in range(1, num_steps + 1):
loss = criterion(model(X), y) / K
losses.append(loss.item())
# compute grad and update inner loop weights
model.zero_grad()
loss.backward()
optimiser.step()
return losses
def train_net(net, train_loader, test_loader, only_fc=True,
optim=optim.Adam,
loss_fn=nn.CrossEntropyLoss(),
n_iter=10, device='cpu'):
net.to(device)
train_losses = []
train_acc, val_acc = [], []
if only_fc:
# only optimize the fc layer
optimizer = optim(net.fc.parameters())
else:
optimizer = optim(net.parameters())
for epoch in range(n_iter):
net.train()
n = 0
n_acc = 0
running_loss = 0.0
for i, (xx, yy) in tqdm.tqdm(enumerate(train_loader),
total=len(train_loader)):
xx = xx.to(device)
yy = yy.to(device)
out = net(xx)
loss = loss_fn(out, yy)
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_loss += loss.item()
n += len(xx)
_, y_pred = out.max(1)
n_acc += (y_pred == yy).float().sum().item()
train_losses.append(running_loss/i)
train_acc.append(n_acc / n)
# test the model
val_acc.append(eval_net(net, test_loader, device))
# print the result
print("epoch : {} \n\t train_losses : {} \n\t train_acc : {} \n\t val_acc: {}".format(
epoch, train_losses[-1], train_acc[-1], val_acc[-1]), flush=True)
def train(self, batches=10, steps=100, optim=optim.SGD, **kwargs):
"""Training the model."""
optimizer = optim([self._pos], **kwargs)
for b_idx in range(batches):
optimizer.zero_grad()
for _ in range(steps):
self.loss.backward(retain_graph=True)
optimizer.step()
log.info(f"batch: {b_idx + 1}/{batches}\tloss: {self.loss}")
def __init__(self,
observation_shape,
actions_shape,
q_lr=.001,
p_lr=.001,
optim=optim.Adam):
self.Q_stable = DQN(observation_shape, actions_shape)
self.Q_unstable = DQN(observation_shape, actions_shape)
self.Q_stable.load_state_dict(self.Q_unstable.state_dict())
self.Q_stable.eval()
self.Policy_stable = PolicyNet(observation_shape, actions_shape)
self.Policy_unstable = PolicyNet(observation_shape, actions_shape)
self.Policy_stable.load_state_dict(self.Policy_unstable.state_dict())
self.Policy_stable.eval()
self.Q_optim = optim(self.Q_unstable.parameters(), lr=q_lr)
self.Policy_optim = optim(self.Policy_unstable.parameters(), lr=p_lr)
def model_functions_at_training(initial_model,
X,
y,
sampled_steps,
x_axis,
optim=torch.optim.SGD,
lr=0.01):
"""
trains the model on X, y and measures the loss curve.
for each n in sampled_steps, records model(x_axis) after n gradient updates.
"""
# copy MAML model into a new object to preserve MAML weights during training
model = nn.Sequential(
OrderedDict([('l1', nn.Linear(600, 300)), ('relu1', nn.ReLU()),
('l2', nn.Linear(300, 150)), ('relu2', nn.ReLU()),
('l3', nn.Linear(150, 70)), ('relu3', nn.ReLU()),
('l4', nn.Linear(70, 20)), ('relu4', nn.ReLU()),
('l5', nn.Linear(20, 1))]))
model.load_state_dict(initial_model.state_dict())
criterion = nn.MSELoss()
optimiser = optim(model.parameters(), lr)
# train model on a random task
num_steps = max(sampled_steps)
K = X.shape[0]
losses = []
outputs = {}
for step in range(1, num_steps + 1):
loss = criterion(model(X), y) / K
losses.append(loss.item())
# compute grad and update inner loop weights
model.zero_grad()
loss.backward()
optimiser.step()
# plot the model function
if step in sampled_steps:
outputs[step] = model(
torch.tensor(x_axis,
dtype=torch.float).view(-1, 1)).detach().numpy()
outputs['initial'] = initial_model(
torch.tensor(x_axis, dtype=torch.float).view(-1, 1)).detach().numpy()
return outputs, losses
def maxloss_perturbs(
model,
criterion,
images,
labels,
epsilon=0.3,
optim=optim.SGD,
optim_params={"lr": 0.03},
n_epoches=100,
verbose=True,
device=None,
):
model.eval()
if device is not None:
model = model.to(device)
images = images.to(device)
labels = labels.to(device)
original_images = images
perturbs = torch.rand(images.shape, device=images.device) / 1000
optimizer = optim([perturbs], **optim_params)
for e in range(n_epoches + 1):
images = images.detach()
images.requires_grad = True
optimizer.zero_grad()
output = model(images)
loss = -1 * criterion(output, labels)
loss.backward()
optimizer.step()
if e == n_epoches:
return torch.clamp(perturbs, -epsilon, epsilon)
images = images + perturbs
images = torch.max(
torch.min(images, original_images + epsilon),
original_images - epsilon,
)
def continuous_optim(tensor_list,
train_data,
loss_fun,
epochs=10,
val_data=None,
other_args=dict()):
"""
Train a tensor network using gradient descent on input dataset
Args:
tensor_list: List of tensors encoding the network being trained
train_data: The data used to train the network
loss_fun: Scalar-valued loss function of the type
tens_list, data -> scalar_loss
(This depends on the task being learned)
epochs: Number of epochs to train for. When val_data is given,
setting epochs=None implements early stopping
val_data: The data used for validation
other_args: Dictionary of other arguments for the optimization,
with some options below (feel free to add more)
optim: Choice of Pytorch optimizer (default='SGD')
lr: Learning rate for optimizer (default=1e-3)
bsize: Minibatch size for training (default=100)
reps: Number of times to repeat
training data per epoch (default=1)
print: Whether to print info (default=True)
dyn_print: use dynamic printing (default=False)
hist: Whether to return losses
from train and val sets (default=False)
momentum: Momentum value for
continuous optimization (default=0)
cvg_threshold: threshold to test convergence of
optimization (optimization is stopped if
|(prev_loss - cur_loss)/prev_loss| < cvg_threshold
If None, convergence is not checked. If epochs is
set as well, then optimziation is stopped either when
convergence criteria is met or when epochs is reached
(default:None)
lr_scheduler: a function taking an optimizer as input
and returning a learning rate scheduler for this optimizer
(default:None)
save_optimizer_state: if True, other_args should have an empty
dict for the key optimizer_state. This dict will contain
{optimizer_state: optimizer state_dict,
lr_scheduler_state: scheduler state_dict (if any)}
after the function returns. (default:False)
load_optimzer_state: a dictionnary that will be used to
initialize the optimizer (and scheduler if any) from a
previously saved optimizer state.
(default: None)
grad_masking_function: a function taking the list of tensor
parameters between the backward pass and the optimizer step
(can be used to e.g. zero out parts of the gradient)
(default: None)
stop_condition: a function taking the training and validation loss
as input after each epoch and returning True if optimization
should be stopped (default: None)
Returns:
better_list: List of tensors with same shape as tensor_list, but
having been optimized using the appropriate optimizer.
When validation data is given, the model with the
lowest validation loss is output, otherwise the model
with lowest training loss
first_loss: Initial loss of the model on the validation set,
before any training. If no val set is provided, the
first training loss is instead returned
best_loss: The value of the validation/training loss for the
model output as better_list
best_epoch: epoch at which best_model was found
loss_record: If hist=True in other_args, history of all validation
and training losses is returned as a tuple of Pytorch
vectors (train_loss, val_loss), with each vector
having length equal to number of epochs of training.
When no validation loss is provided, the second item
(val_loss) is an empty tensor.
"""
# Check input and initialize local record variables
early_stop = epochs is None
has_val = val_data is not None
optim = other_args['optim'] if 'optim' in other_args else 'SGD'
lr = other_args['lr'] if 'lr' in other_args else 1e-3
bsize = other_args['bsize'] if 'bsize' in other_args else 100
reps = other_args['reps'] if 'reps' in other_args else 1
prnt = other_args['print'] if 'print' in other_args else True
hist = other_args['hist'] if 'hist' in other_args else False
dyn_print = other_args['dyn_print'] if 'dyn_print' in other_args else False
lr_scheduler = other_args[
'lr_scheduler'] if 'lr_scheduler' in other_args else None
cvg_threshold = other_args[
'cvg_threshold'] if 'cvg_threshold' in other_args else None
save_optimizer_state = other_args[
'save_optimizer_state'] if 'save_optimizer_state' in other_args else None
load_optimizer_state = other_args[
'load_optimizer_state'] if 'load_optimizer_state' in other_args else None
grad_masking_function = other_args[
'grad_masking_function'] if 'grad_masking_function' in other_args else None
momentum = other_args['momentum'] if 'momentum' in other_args else 0
stop_condition = other_args[
'stop_condition'] if 'stop_condition' in other_args else None
if save_optimizer_state and (not 'optimizer_state' in other_args):
raise ValueError(
"an empty dictionnary should be passed as the optimizer_state argument to store the"
" optimizer state.")
if early_stop and not has_val:
raise ValueError("Early stopping (epochs=None) requires val_data "
"to be input")
loss_rec, first_loss, best_loss, best_network, best_epoch = [], None, np.infty, tensor_list, 0
if hist: loss_record = ([], []) # (train_record, val_record)
# Function to maybe print, conditioned on `prnt`
m_print = lambda s: print(s, end='\r'
if dyn_print else '\n') if prnt else None
# Function to record loss information and return whether to stop
def record_loss(new_loss, new_network, epoch_num):
# Load record variables from outer scope
nonlocal loss_rec, first_loss, best_loss, best_network, best_epoch
# Check for first and best loss
if best_loss is None or new_loss < best_loss:
best_loss, best_network, best_epoch = new_loss, new_network, epoch_num
if first_loss is None:
first_loss = new_loss
# Update loss record and check for early stopping. If you want to
# change early stopping criteria, this is the place to do it.
window = 2 # Number of epochs kept for checking early stopping
warmup = 1 # Number of epochs before early stopping is checked
if len(loss_rec) < window:
stop, loss_rec = False, loss_rec + [new_loss]
else:
# stop = new_loss > sum(loss_rec)/len(loss_rec)
stop = (new_loss > max(loss_rec)) and (epoch_num >= warmup)
loss_rec = loss_rec[1:] + [new_loss]
return stop
# Another loss logging function, but for recording *all* loss history
@torch.no_grad()
def loss_history(new_loss, is_val):
if not hist: return
nonlocal loss_record
loss_record[int(is_val)].append(new_loss)
# Function to run TN on validation data
@torch.no_grad()
def run_val(t_list):
val_loss = []
# Note that `batchify` uses different logic for different types
# of input, so update batchify when you work on tensor completion
for batch in batchify(val_data):
val_loss.append(loss_fun(t_list, batch))
if has_val:
val_loss = torch.mean(torch.tensor(val_loss))
return val_loss
# Copy tensor_list so the original is unchanged
tensor_list = copy_network(tensor_list)
# Record the initial validation loss (if we validation dataset)
if has_val: record_loss(run_val(tensor_list), tensor_list, 0)
# Initialize optimizer, using only the keyword args in the
optim = getattr(torch.optim, optim)
opt_args = signature(optim).parameters.keys()
kwargs = {'lr': lr, 'momentum': momentum} # <- Add new options here
kwargs = {k: v for (k, v) in kwargs.items() if k in opt_args}
optim = optim(tensor_list, **kwargs) # Initialize the optimizer
if lr_scheduler: # instantiate learning rate scheduler
scheduler = lr_scheduler(optim)
if load_optimizer_state:
optim.load_state_dict(
other_args["load_optimizer_state"]["optimizer_state"])
if lr_scheduler:
scheduler.load_state_dict(
other_args["load_optimizer_state"]["lr_scheduler_state"])
# Loop over validation and training for given number of epochs
ep = 1
prev_loss = np.infty
while epochs is None or ep <= epochs:
# Train network on all the training data
#from copy import deepcopy
prev_tensor_list = copy_network(tensor_list)
#prev_tensor_list = tensor_list
train_loss, num_train = 0., 0
for batch in batchify(train_data, batch_size=bsize, reps=reps):
loss = loss_fun(tensor_list, batch)
optim.zero_grad()
loss.backward()
if grad_masking_function:
grad_masking_function(tensor_list)
optim.step()
with torch.no_grad():
num_train += 1
train_loss += loss
train_loss /= num_train
if lr_scheduler:
scheduler.step(train_loss)
loss_history(train_loss, is_val=False)
val_loss = run_val(tensor_list) if has_val else None
val_loss_str = f"Val. loss: {val_loss.data:.10f}" if has_val else ""
m_print(
f"EPOCH {ep} {'('+str(reps)+' reps)' if reps > 1 else ''}\t\t{val_loss_str}\t\t Train loss: {train_loss.data:.10f}\t\t Convergence: {np.abs(train_loss-prev_loss)/prev_loss:.10f}"
)
# Get validation loss if we have it, otherwise record training loss
if has_val:
# Get and record validation loss, check early stopping condition
loss_history(val_loss, is_val=True)
if record_loss(
val_loss,
copy_network(tensor_list) if has_val else prev_tensor_list,
ep) and early_stop:
print(f"\nEarly stopping condition reached")
break
else:
record_loss(
train_loss,
copy_network(tensor_list) if has_val else prev_tensor_list, ep)
if cvg_threshold and np.abs(train_loss -
prev_loss) / prev_loss < cvg_threshold:
print(f"\nConvergence criteria reached")
break
if stop_condition and stop_condition(train_loss=train_loss,
val_loss=val_loss):
print(f"\nStopping condition reached")
break
prev_loss = train_loss
ep += 1
m_print("")
# Save the optimizer state if needed
if save_optimizer_state:
other_args["optimizer_state"]["optimizer_state"] = optim.state_dict()
if lr_scheduler:
other_args["optimizer_state"][
"lr_scheduler_state"] = scheduler.state_dict()
if hist:
loss_record = tuple(torch.tensor(fr) for fr in loss_record)
return best_network, first_loss, best_loss, best_epoch, loss_record
else:
return best_network, first_loss, best_loss
def __init__(self):
super(MAML_coat, self).__init__()
self.model = nn.Sequential(OrderedDict([
('l1', nn.Linear(600,300)),
('relu1', nn.ReLU()),
('l2', nn.Linear(300,1500)),
('relu2', nn.ReLU()),
('l3', nn.Linear(150,70)),
('relu3', nn.ReLU()),
('l4', nn.Linear(70,20),
('relu4', nn.ReLU()),
('l5', nn.Linear(20,1))
]))
def forward(self, x):
return self.model(x)
def parameterised(self, x, weights):
# like forward, but uses ``weights`` instead of ``model.parameters()``
# it'd be nice if this could be generated automatically for any nn.Module...
x = nn.functional.linear(x, weights[0], weights[1])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[2], weights[3])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[4], weights[5])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[6], weights[7])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[8], weights[9])
return x
# %%
class MAMLModel(nn.Module):
def __init__(self):
super(MAMLModel, self).__init__()
self.model = nn.Sequential(OrderedDict([
('l1', nn.Linear(1200,600)),
('relu1', nn.ReLU()),
('l2', nn.Linear(600,200)),
('relu2', nn.ReLU()),
('l3', nn.Linear(200,100)),
('relu3', nn.ReLU()),
('l4', nn.Linear(100,50)),
('relu4', nn.ReLU()),
('l5', nn.Linear(50,25)),
('relu5', nn.ReLU()),
('l6', nn.Linear(25,1))
]))
def forward(self, x):
return self.model(x)
def parameterised(self, x, weights):
# like forward, but uses ``weights`` instead of ``model.parameters()``
# it'd be nice if this could be generated automatically for any nn.Module...
x = nn.functional.linear(x, weights[0], weights[1])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[2], weights[3])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[4], weights[5])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[6], weights[7])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[8], weights[9])
x = nn.functional.relu(x)
x = nn.functional.linear(x, weights[10], weights[11])
return x
# %%
class MAML():
def __init__(self, model, inner_lr, meta_lr, K=10, inner_steps=1, tasks_per_meta_batch=1000):
# important objects
# self.tasks = tasks
self.model = model
self.weights = list(model.parameters()) # the maml weights we will be meta-optimising
self.criterion = nn.MSELoss()
self.meta_optimiser = torch.optim.Adam(self.weights, meta_lr)
# hyperparameters
self.inner_lr = inner_lr
self.meta_lr = meta_lr
self.K = K
self.inner_steps = inner_steps # with the current design of MAML, >1 is unlikely to work well
self.tasks_per_meta_batch = tasks_per_meta_batch
# metrics
self.plot_every = 10
self.print_every = 500
self.meta_losses = []
def inner_loop(self):
# reset inner model to current maml weights
temp_weights = [w.clone() for w in self.weights]
# perform training on data sampled from task
X, y = data_loader_rand(self.K)
for step in range(self.inner_steps):
loss = self.criterion(self.model.parameterised(X, temp_weights), y) / self.K
# compute grad and update inner loop weights
grad = torch.autograd.grad(loss, temp_weights)
temp_weights = [w - self.inner_lr * g for w, g in zip(temp_weights, grad)]
# sample new data for meta-update and compute loss
X, y = data_loader_rand(self.K)
loss = self.criterion(self.model.parameterised(X, temp_weights), y) / self.K
return loss
def main_loop(self, num_iterations):
epoch_loss = 0
for iteration in range(1, num_iterations+1):
# compute meta loss
meta_loss = 0
for i in range(self.tasks_per_meta_batch):
#task = self.tasks.sample_task()
#meta_loss += self.inner_loop(task)
meta_loss += self.inner_loop()
# compute meta gradient of loss with respect to maml weights
meta_grads = torch.autograd.grad(meta_loss, self.weights)
# assign meta gradient to weights and take optimisation step
for w, g in zip(self.weights, meta_grads):
w.grad = g
self.meta_optimiser.step()
# log metrics
epoch_loss += meta_loss.item() / self.tasks_per_meta_batch
if iteration % self.print_every == 0:
print("{}/{}. loss: {}".format(iteration, num_iterations, epoch_loss / self.plot_every))
if iteration % self.plot_every == 0:
self.meta_losses.append(epoch_loss / self.plot_every)
epoch_loss = 0
# %%
tasks = Sine_Task_Distribution(0.1, 5, 0, np.pi, -5, 5)
maml = MAML(MAMLModel(), tasks, inner_lr=0.01, meta_lr=0.001)
# %%
maml.main_loop(num_iterations=10000)
#%%
plt.plot(maml.meta_losses)
#%%
def loss_on_random_task(initial_model, K, num_steps, optim=torch.optim.SGD):
"""
trains the model on a random sine task and measures the loss curve.
for each n in num_steps_measured, records the model function after n gradient updates.
"""
# copy MAML model into a new object to preserve MAML weights during training
model = nn.Sequential(OrderedDict([
('l1', nn.Linear(1200,600)),
('relu1', nn.ReLU()),
('l2', nn.Linear(600,200)),
('relu2', nn.ReLU()),
('l3', nn.Linear(200,100)),
('relu3', nn.ReLU()),
('l4', nn.Linear(100,50)),
('relu4', nn.ReLU()),
('l5', nn.Linear(50,25)),
('relu5', nn.ReLU()),
('l6', nn.Linear(25,1))
]))
model.load_state_dict(initial_model.state_dict())
criterion = nn.MSELoss()
optimiser = optim(model.parameters(), 0.01)
# train model on a random task
task = tasks.sample_task()
X, y = task.sample_data(K)
losses = []
for step in range(1, num_steps+1):
loss = criterion(model(X), y) / K
losses.append(loss.item())
# compute grad and update inner loop weights
model.zero_grad()
loss.backward()
optimiser.step()
return losses
#%%
def average_losses(initial_model, n_samples, K=10, n_steps=10, optim=torch.optim.SGD):
"""
returns the average learning trajectory of the model trained for ``n_iterations`` over ``n_samples`` tasks
"""
x = np.linspace(-5, 5, 2) # dummy input for test_on_new_task
avg_losses = [0] * K
for i in range(n_samples):
losses = loss_on_random_task(initial_model, K, n_steps, optim)
avg_losses = [l + l_new for l, l_new in zip(avg_losses, losses)]
avg_losses = [l / n_samples for l in avg_losses]
return avg_losses
#%%
def mixed_pretrained(iterations=500):
"""
returns a model pretrained on a selection of ``iterations`` random tasks.
"""
# set up model
model = nn.Sequential(OrderedDict([
('l1', nn.Linear(1200,600)),
('relu1', nn.ReLU()),
('l2', nn.Linear(600,200)),
('relu2', nn.ReLU()),
('l3', nn.Linear(200,100)),
('relu3', nn.ReLU()),
('l4', nn.Linear(100,50)),
('relu4', nn.ReLU()),
('l5', nn.Linear(50,25)),
('relu5', nn.ReLU()),
('l6', nn.Linear(25,1))
]))
optimiser = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = nn.MSELoss()
# fit the model
for i in range(iterations):
model.zero_grad()
x, y = tasks.sample_task().sample_data(10)
loss = criterion(model(x), y)
loss.backward()
optimiser.step()
return model
#%%
pretrained = mixed_pretrained(10000)
plt.plot(average_losses(maml.model.model, n_samples=5000, K=10), label='maml')
plt.plot(average_losses(pretrained, n_samples=5000, K=10), label='pretrained')
plt.legend()
plt.title("Average learning trajectory for K=10, starting from initial weights")
plt.xlabel("gradient steps taken with SGD")
plt.show()
#%%
plt.plot(average_losses(maml.model.model, n_samples=5000, K=10, optim=torch.optim.Adam), label='maml')
plt.plot(average_losses(pretrained, n_samples=5000, K=10, optim=torch.optim.Adam), label='pretrained')
plt.legend()
plt.title("Average learning trajectory for K=10, starting from initial weights")
plt.xlabel("gradient steps taken with Adam")
plt.show()
#%%
def model_functions_at_training(initial_model, X, y, sampled_steps, x_axis, optim=torch.optim.SGD, lr=0.01):
"""
trains the model on X, y and measures the loss curve.
for each n in sampled_steps, records model(x_axis) after n gradient updates.
"""
# copy MAML model into a new object to preserve MAML weights during training
model = nn.Sequential(OrderedDict([
('l1', nn.Linear(1200,600)),
('relu1', nn.ReLU()),
('l2', nn.Linear(600,200)),
('relu2', nn.ReLU()),
('l3', nn.Linear(200,100)),
('relu3', nn.ReLU()),
('l4', nn.Linear(100,50)),
('relu4', nn.ReLU()),
('l5', nn.Linear(50,25)),
('relu5', nn.ReLU()),
('l6', nn.Linear(25,1))
]))
model.load_state_dict(initial_model.state_dict())
criterion = nn.MSELoss()
optimiser = optim(model.parameters(), lr)
# train model on a random task
num_steps = max(sampled_steps)
K = X.shape[0]
losses = []
outputs = {}
for step in range(1, num_steps+1):
loss = criterion(model(X), y) / K
losses.append(loss.item())
# compute grad and update inner loop weights
model.zero_grad()
loss.backward()
optimiser.step()
# plot the model function
if step in sampled_steps:
outputs[step] = model(torch.tensor(x_axis, dtype=torch.float).view(-1, 1)).detach().numpy()
outputs['initial'] = initial_model(torch.tensor(x_axis, dtype=torch.float).view(-1, 1)).detach().numpy()
return outputs, losses
#%%
def plot_sampled_performance(initial_model, model_name, task, X, y, optim=torch.optim.SGD, lr=0.01):
x_axis = np.linspace(-5, 5, 1000)
sampled_steps=[1,10]
outputs, losses = model_functions_at_training(initial_model,
X, y,
sampled_steps=sampled_steps,
x_axis=x_axis,
optim=optim, lr=lr)
plt.figure(figsize=(15,5))
# plot the model functions
plt.subplot(1, 2, 1)
plt.plot(x_axis, task.true_function(x_axis), '-', color=(0, 0, 1, 0.5), label='true function')
plt.scatter(X, y, label='data')
plt.plot(x_axis, outputs['initial'], ':', color=(0.7, 0, 0, 1), label='initial weights')
for step in sampled_steps:
plt.plot(x_axis, outputs[step],
'-.' if step == 1 else '-', color=(0.5, 0, 0, 1),
label='model after {} steps'.format(step))
plt.legend(loc='lower right')
plt.title("Model fit: {}".format(model_name))
# plot losses
plt.subplot(1, 2, 2)
plt.plot(losses)
plt.title("Loss over time")
plt.xlabel("gradient steps taken")
plt.show()
#%%
K = 10
task = tasks.sample_task()
X, y = task.sample_data(K)
plot_sampled_performance(maml.model.model, 'MAML', task, X, y)
#%%
plot_sampled_performance(pretrained, 'pretrained at lr=0.02', task, X, y, lr=0.02)
#%%
K = 5
task = tasks.sample_task()
X, y = task.sample_data(K)
plot_sampled_performance(maml.model.model, 'MAML', task, X, y)
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('mode', choices=['train', 'predict'])
arg('run_root')
arg('trained_weight')
arg('--model', default='densenet169')
arg('--optimizer', default='adamw')
arg('--pretrained', type=int, default=1)
arg('--warmingup', type=int, default=1)
arg('--batch-size', type=int, default=32)
arg('--step', type=int, default=1)
arg('--workers', type=int, default=2 if ON_KAGGLE else 4)
arg('--lr', type=float, default=1e-4)
arg('--patience', type=int, default=3)
arg('--epochs', type=int, default=100)
arg('--epoch-size', type=int)
arg('--clean', action='store_true')
arg('--tta', type=int, default=4)
arg('--debug', action='store_true')
args = parser.parse_args()
run_root = Path(args.run_root)
trained_weight = Path(args.trained_weight)
seed_everything(seed=2333)
model_conv = getattr(models, args.model)(num_classes=N_CLASSES,
pretrained=args.pretrained)
model_conv.cuda()
optim = getattr(optimizers, args.optimizer)(params=model_conv.parameters())
if run_root.exists() and args.clean:
shutil.rmtree(run_root)
run_root.mkdir(exist_ok=True, parents=True)
(run_root / 'params.json').write_text(
json.dumps(vars(args), indent=4, sort_keys=True))
batch_size = args.batch_size,
num_workers = args.workers
if trained_weight.exists():
if (os.path.exists(trained_weight)):
model_conv.load_state_dict(torch.load(trained_weight),
strict=False)
if args.mode == 'train':
train_loader, valid_loader = read_data(run_root, batch_size,
num_workers)
if args.warmingup:
model_conv.freeze_basemodel()
n_epochs = 1
p = 0
valid_loss_min = float("inf")
optimizer = optim(lr=1e-2)
for i in range(n_epochs):
start_time = time.time()
avg_loss, avg_corrects, avg_auc = train(i)
avg_val_loss, avg_val_corrects = test()
elapsed_time = time.time() - start_time
print('Epoch {}/{} \t loss={:.4f} acc={:.4f} auc={:.4f} \t val_loss={:.4f} val_acc={:.4f} \t time={:.2f}s'.format(\
i + 1, n_epochs, avg_loss, avg_corrects, avg_auc, avg_val_loss, avg_val_corrects, elapsed_time))
optimizer = optim(lr=args.lr)
model_conv.unfreeze_model()
n_epochs = args.epochs
patience = args.patience
for i in range(n_epochs):
start_time = time.time()
avg_loss, avg_corrects, avg_auc = train(model_conv, train_loader,
i)
avg_val_loss, avg_val_corrects = test(model_conv, valid_loader)
elapsed_time = time.time() - start_time
print('Epoch {}/{} \t loss={:.4f} acc={:.4f} auc={:.4f} \t val_loss={:.4f} val_acc={:.4f} \t time={:.2f}s'.format(\
i + 1, n_epochs, avg_loss, avg_corrects, avg_auc, avg_val_loss, avg_val_corrects, elapsed_time))
if avg_val_loss <= valid_loss_min:
print('Validation loss decreased ({:.6f} --> {:.6f}). Saving model ...'.format(\
valid_loss_min,avg_val_loss))
torch.save(model_conv.state_dict(), 'model.pt')
valid_loss_min = avg_val_loss
p = 0
if avg_val_loss > valid_loss_min:
p += 1
print(f'{p} epochs of increasing val loss')
if p >= 1:
print('Decrease learning rate')
optimizer = optim(lr=args.lr / 10)
if p > patience:
print('Stopping training')
stop = True
break
else:
num_tta = args.tta
test_loader = read_test(root_path, batch_size, num_workers)
model_conv.eval().cuda()
for tta in range(num_tta):
preds = []
for batch_i, (data, target) in enumerate(test_loader):
data, target = data.cuda(), target.cuda()
output = model_conv(data).detach()
pr = output[:, 0].cpu().numpy()
for i in pr:
preds.append(i)
test_preds = pd.DataFrame({
'imgs': test_set.image_files_list,
'preds': preds
})
test_preds['imgs'] = test_preds['imgs'].apply(
lambda x: x.split('.')[0])
sub = pd.read_csv(f'{run_root}/data/sample_submission.csv')
sub = pd.merge(sub, test_preds, left_on='id', right_on='imgs')
sub = sub[['id', 'preds']]
sub.columns = ['id', 'label']
sub.head()
sub.to_csv('single_model_' + str(tta) + '.csv', index=False)
del model_conv
gc.collect()
torch.cuda.empty_cache()
def train(network,
train_dataloader,
test_dataloader,
trainDataset,
testDataset,
vocab,
epochs,
learning_range=0.001,
load_checkpoint=False,
checkpoint_file=CHECKPOINT_FILE,
criterion=nn.CrossEntropyLoss(),
optim=torch.optim.Adam):
print("TRAIN STARTED!")
log_file = open(TRAIN_LOG_FILE, 'w')
train_loss_epochs = []
test_loss_epochs = []
optimizer = optim(network.parameters(), lr=learning_range)
best_test_score = 10**6
if load_checkpoint:
open_checkpoint(network,
optimizer,
is_best=False,
filename=checkpoint_file)
try:
for epoch in range(epochs):
train_loss_sum = 0.0
train_loss_count = 0
sample_id = 0
print("Epoch: {} Training".format(epoch + 1))
for sample in tqdm(train_dataloader):
sample_id += 1
torch.cuda.empty_cache()
features = sample['image']
ann_ids = sample['anns']
batch_size = features.shape[0]
lengths = sample['ann_len']
max_len = lengths.max()
captions = load_anns(trainDataset,
ann_ids,
max_len,
prepare=lambda w: vocab(w))
captions = captions.long()
lengths, perm_index = lengths.sort(0, descending=True)
lengths = lengths.numpy()
captions = captions[perm_index]
features = features[perm_index]
captions = Variable(captions)
features = Variable(features)
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
outputs = network.forward(features, captions, lengths)
targets = targets.cuda()
# outputs = outputs.cpu()
# targets = targets.cpu()
loss_batch = criterion(outputs, targets)
train_loss_sum += loss_batch.data[0]
train_loss_count += 1.0
loss_batch.backward()
optimizer.step()
optimizer.zero_grad()
del features, captions, loss_batch, sample, outputs, targets, lengths
if sample_id % 200 == 0:
gc.collect()
gc.collect()
train_loss_epochs.append(train_loss_sum / train_loss_count)
test_loss_sum = 0.0
test_loss_count = 0
sample_id = 0
torch.cuda.empty_cache()
print("Epoch: {} Testing".format(epoch + 1))
for sample in tqdm(test_dataloader):
sample_id += 1
torch.cuda.empty_cache()
features = sample['image']
ann_ids = sample['anns']
batch_size = features.shape[0]
lengths = sample['ann_len']
max_len = lengths.max()
captions = load_anns(testDataset,
ann_ids,
max_len,
prepare=lambda w: vocab(w))
captions = captions.long()
lengths, perm_index = lengths.sort(0, descending=True)
lengths = lengths.numpy()
captions = captions[perm_index]
features = features[perm_index]
captions = Variable(captions).cuda()
features = Variable(features).cuda()
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
outputs = network.forward(features, captions, lengths)
targets = targets.cuda()
# outputs = outputs.cpu()
# targets = targets.cpu()
loss_batch = criterion(outputs, targets)
test_loss_sum += loss_batch.data[0]
test_loss_count += 1.0
del features, captions, loss_batch, sample, outputs, targets, lengths
if sample_id % 200 == 0:
gc.collect()
test_loss_epochs.append(test_loss_sum / (test_loss_count))
test_network(network, testDataset, vocab, epoch)
is_best = test_loss_epochs[-1] < best_test_score
best_test_score = min(test_loss_epochs[-1], best_test_score)
save_checkpoint(
{
'net': network,
'epoch': epoch + 1,
'state_dict': network.state_dict(),
'best_test_score': best_test_score,
'optimizer': optimizer.state_dict(),
},
is_best,
filename='checkpoints/checkpoint_{}.pth.tar'.format(epoch + 1))
log_file.write(
'\rEpoch {0}... (Train/Test) Loss: {1:.3f}/{2:.3f}\n'.format(
epoch, train_loss_epochs[-1], test_loss_epochs[-1]))
sys.stdout.write(
'\rEpoch {0}... (Train/Test) Loss: {1:.3f}/{2:.3f}\n'.format(
epoch, train_loss_epochs[-1], test_loss_epochs[-1]))
gc.collect()
except KeyboardInterrupt:
pass
# plt.figure(figsize=(12, 5))
# plt.plot(train_loss_epochs[1:], label='Train')
# plt.plot(test_loss_epochs[1:], label='Test')
# plt.xlabel('Epochs', fontsize=16)
# plt.ylabel('Loss', fontsize=16)
# plt.legend(loc=0, fontsize=16)
# plt.grid('on')
# plt.savefig(TRAIN_PLT_FILE)
#
gc.collect()
print("TRAIN ENDED!")
def main(_run, _config, _log):
settings = Settings()
common_params, data_params, net_params, train_params, eval_params = settings['COMMON'], settings['DATA'], settings[
'NETWORK'], settings['TRAINING'], settings['EVAL']
if _run.observers:
os.makedirs(f'{_run.observers[0].dir}/snapshots', exist_ok=True)
for source_file, _ in _run.experiment_info['sources']:
os.makedirs(os.path.dirname(f'{_run.observers[0].dir}/source/{source_file}'),
exist_ok=True)
_run.observers[0].save_file(source_file, f'source/{source_file}')
shutil.rmtree(f'{_run.observers[0].basedir}/_sources')
set_seed(_config['seed'])
cudnn.enabled = True
cudnn.benchmark = True
torch.cuda.set_device(device=_config['gpu_id'])
torch.set_num_threads(1)
_log.info('###### Load data ######')
data_name = _config['dataset']
if data_name == 'BCV':
make_data = meta_data
else:
print(f"data name : {data_name}")
raise ValueError('Wrong config for dataset!')
tr_dataset, val_dataset, ts_dataset = make_data(_config)
trainloader = DataLoader(
dataset=tr_dataset,
batch_size=_config['batch_size'],
shuffle=True,
num_workers=_config['n_work'],
pin_memory=False, #True load data while training gpu
drop_last=True
)
_log.info('###### Create model ######')
model = fs.FewShotSegmentorDoubleSDnet(net_params).cuda()
model.train()
_log.info('###### Set optimizer ######')
optim = torch.optim.Adam
optim_args = {"lr": train_params['learning_rate'],
"weight_decay": train_params['optim_weight_decay'],}
# "momentum": train_params['momentum']}
optim_c = optim(list(model.conditioner.parameters()), **optim_args)
optim_s = optim(list(model.segmentor.parameters()), **optim_args)
scheduler_s = lr_scheduler.StepLR(optim_s, step_size=100, gamma=0.1)
scheduler_c = lr_scheduler.StepLR(optim_c, step_size=100, gamma=0.1)
criterion = losses.DiceLoss()
if _config['record']: ## tensorboard visualization
_log.info('###### define tensorboard writer #####')
_log.info(f'##### board/train_{_config["board"]}_{date()}')
writer = SummaryWriter(f'board/train_{_config["board"]}_{date()}')
iter_print = _config["iter_print"]
iter_n_train = len(trainloader)
_log.info('###### Training ######')
for i_epoch in range(_config['n_steps']):
epoch_loss = 0
for i_iter, sample_batched in enumerate(trainloader):
# Prepare input
s_x = sample_batched['s_x'].cuda() # [B, Support, slice_num=1, 1, 256, 256]
X = s_x.squeeze(2) # [B, Support, 1, 256, 256]
s_y = sample_batched['s_y'].cuda() # [B, Support, slice_num, 1, 256, 256]
Y = s_y.squeeze(2) # [B, Support, 1, 256, 256]
Y = Y.squeeze(2) # [B, Support, 256, 256]
q_x = sample_batched['q_x'].cuda() # [B, slice_num, 1, 256, 256]
query_input = q_x.squeeze(1) # [B, 1, 256, 256]
q_y = sample_batched['q_y'].cuda() # [B, slice_num, 1, 256, 256]
y2 = q_y.squeeze(1) # [B, 1, 256, 256]
y2 = y2.squeeze(1) # [B, 256, 256]
y2 = y2.type(torch.LongTensor).cuda()
entire_weights = []
for shot_id in range(_config["n_shot"]):
input1 = X[:, shot_id, ...] # use 1 shot at first
y1 = Y[:, shot_id, ...] # use 1 shot at first
condition_input = torch.cat((input1, y1.unsqueeze(1)), dim=1)
weights = model.conditioner(condition_input) # 2, 10, [B, channel=1, w, h]
entire_weights.append(weights)
# pdb.set_trace()
avg_weights=[[],[None, None, None, None]]
for i in range(9):
weight_cat = torch.cat([weights[0][i] for weights in entire_weights],dim=1)
avg_weight = torch.mean(weight_cat,dim=1,keepdim=True)
avg_weights[0].append(avg_weight)
avg_weights[0].append(None)
output = model.segmentor(query_input, avg_weights)
loss = criterion(F.softmax(output, dim=1), y2)
optim_s.zero_grad()
optim_c.zero_grad()
loss.backward()
optim_s.step()
optim_c.step()
epoch_loss += loss
if iter_print:
print(f"train, iter:{i_iter}/{iter_n_train}, iter_loss:{loss}", end='\r')
scheduler_c.step()
scheduler_s.step()
print(f'step {i_epoch+1}: loss: {epoch_loss} ')
if _config['record']:
batch_i = 0
frames = []
query_pred = output.argmax(dim=1)
query_pred = query_pred.unsqueeze(1)
frames += overlay_color(q_x[batch_i,0], query_pred[batch_i].float(), q_y[batch_i,0])
# frames += overlay_color(s_xi[batch_i], blank, s_yi[batch_i], scale=_config['scale'])
visual = make_grid(frames, normalize=True, nrow=2)
writer.add_image("train/visual", visual, i_epoch)
save_fname = f'{_run.observers[0].dir}/snapshots/last.pth'
torch.save(model.state_dict(),save_fname)
def main():
# Get arguments
args = parse_args()
# Set random seed
torch.manual_seed(args.seed)
# Cuda
use_cuda = False
if torch.cuda.is_available():
if not args.cuda:
print("WARNING: You have a CUDA device, so you \
should probably run with --cuda")
else:
use_cuda = True
torch.cuda.manual_seed(args.seed)
# Load data + text fields
print('=' * 89)
train_iter, val_iter, test_iter, SRC, TRG = utils.load_dataset(
batch_size=args.batch_size,
use_pretrained_emb=args.pretrained_emb,
save_dir=SAVE_DIR
)
print('=' * 89)
# Intialize model
enc = models.EncoderRNN(
input_size=len(SRC.vocab),
emb_size=(SRC.vocab.vectors.size(1)
if args.pretrained_emb == 'fastText'
else args.emb_size),
embeddings=(SRC.vocab.vectors
if args.pretrained_emb == 'fastText'
else None),
max_norm=args.emb_maxnorm,
padding_idx=SRC.vocab.stoi['<pad>'],
hidden_size=args.hidden_size,
num_layers=args.num_layers,
dropout=args.dropout,
bidirectional=args.bidirectional
)
decoder = models.AttnDecoderRNN if args.attention else models.DecoderRNN
dec = decoder(
enc_num_directions=enc.num_directions,
enc_hidden_size=args.hidden_size,
use_context=args.use_context,
input_size=len(TRG.vocab),
emb_size=(TRG.vocab.vectors.size(1)
if args.pretrained_emb
else args.emb_size),
embeddings=(TRG.vocab.vectors
if args.pretrained_emb
else None),
max_norm=args.emb_maxnorm,
padding_idx=TRG.vocab.stoi['<pad>'],
hidden_size=args.hidden_size,
num_layers=args.num_layers,
dropout=args.dropout,
bidirectional=False # args.bidirectional
)
model = models.Seq2Seq(enc, dec, use_cuda=use_cuda)
if use_cuda:
model.cuda()
print(model)
# Intialize loss
criterion = torch.nn.CrossEntropyLoss(
ignore_index=TRG.vocab.stoi["<pad>"])
# Create optimizer
if args.optimizer == 'Adam':
optim = torch.optim.Adam
elif args.optimizer == 'Adadelta':
optim = torch.optim.Adadelta
elif args.optimizer == 'Adagrad':
optim = torch.optim.Adagrad
else:
optim = torch.optim.SGD
optimizer = optim(model.parameters(), lr=args.lr)
# Create scheduler
lambda_lr = lambda epoch: 0.5 if epoch > 8 else 1
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lambda_lr)
# Train
best_val_loss = None
fname = './{}/{}.pt'.format(SAVE_DIR, args.save)
print('=' * 89)
try:
for epoch in range(1, args.epochs+1):
epoch_start_time = time.time()
attns = train(epoch, model, train_iter, criterion, optimizer,
use_cuda, args, SRC, TRG)
val_loss = evaluate(model, val_iter, criterion, use_cuda)
# Log results
print('-' * 89)
print('| end of epoch {:3d} | time: {:5.2f}s '
'| valid loss {:5.2f} | valid ppl {:8.2f}'.format(
epoch, (time.time() - epoch_start_time),
val_loss, math.exp(val_loss)))
print('-' * 89)
# Save the model if validation loss is best we've seen so far
if not best_val_loss or val_loss < best_val_loss:
if not os.path.isdir(SAVE_DIR):
os.makedirs(SAVE_DIR)
torch.save(model, fname)
best_val_loss = val_loss
# Anneal learning rate
scheduler.step()
except KeyboardInterrupt:
print('-' * 89)
print('Exiting from training early')
# Load the best saved model
with open(fname, 'rb') as f:
model = torch.load(f)
# Run on test data
test_loss = evaluate(model, test_iter, criterion, use_cuda)
# Log results
print('=' * 89)
print('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(
test_loss, math.exp(test_loss)))
print('=' * 89)