def __init__(self, epochs: int = 10, lr: float = 1, coef: float = -2):
super().__init__()
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.epochs = epochs
self.lr = lr
self.coef = coef
self.model = ChildNet().to(self.device)
self.loss = MSELoss()
self.optimizer = Adam(params=self.model.parameters(), lr=lr)
def __init__(self, sizes, nonlin='tanh', device='cpu'):
super(NNRegressor, self).__init__()
if device == 'gpu' and torch.cuda.is_available():
self.device = torch.device('cuda:0')
else:
self.device = torch.device('cpu')
self.sizes = sizes
nlist = dict(relu=torch.relu,
tanh=torch.tanh,
splus=nn.Softplus,
softmax=torch.log_softmax,
linear=func.linear)
self.nonlin = nlist[nonlin]
self.l1 = nn.Linear(self.sizes[0], self.sizes[1]).to(self.device)
self.l2 = nn.Linear(self.sizes[1], self.sizes[2]).to(self.device)
self.output = nn.Linear(self.sizes[2], self.sizes[3]).to(self.device)
self.criterion = MSELoss().to(self.device)
self.optim = None
self.target_size = self.sizes[-1]
self.input_size = self.sizes[0]
self.input_trans = None
self.target_trans = None
def run_detetmenant(in_features, n_iterations=100, size=100):
coder = DetNet(in_features)
dataset = [FloatTensor(make_spd_matrix(in_features)).unsqueeze(0) for _ in range(size)]
dataset = Variable(torch.cat(dataset, 0))
test_size = 30
train_dataset = dataset[:-test_size]
test_dataset = dataset[-test_size:]
optimizer = optim.Adam(coder.parameters(), lr=0.1)
criterion = MSELoss()
for epoch in tqdm(range(1, n_iterations)):
optimizer.zero_grad()
outputs_train = coder(train_dataset)
outputs_test = coder(test_dataset)
loss_train = criterion(outputs_train, train_dataset)
loss_test = criterion(outputs_test, test_dataset)
loss_train.backward(retain_graph=True)
if epoch % 10 == 0:
print("EPOCH: {0}, TRAIN LOSS: {1}, TEST LOSS".format(epoch, loss_train.data[0]), loss_test.data[0])
optimizer.step()
return coder, dataset
class Translator(BasicFamilyReg):
def __init__(self, epochs: int = 10, lr: float = 1, coef: float = -2):
super().__init__()
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.epochs = epochs
self.lr = lr
self.coef = coef
self.model = ChildNet().to(self.device)
self.loss = MSELoss()
self.optimizer = Adam(params=self.model.parameters(), lr=lr)
def fit(self, X_fathers, X_mothers, y_child):
# X = torch.from_numpy(np.concatenate([X_fathers, X_mothers], axis=-1)).float().to(self.device)
X_fathers = np2torch(X_fathers)
X_mothers = np2torch(X_mothers)
y = np2torch(y_child)
for epoch in range(self.epochs):
self.optimizer.zero_grad()
output = self.model(X_fathers, X_mothers)
loss = self.loss.forward(output, y)
loss.backward()
self.optimizer.step()
print("loss", loss)
def predict(self, X_fathers, X_mothers):
X_fathers = np2torch(X_fathers)
X_mothers = np2torch(X_mothers)
with torch.no_grad():
y_pred = self.model(X_fathers, X_mothers)
# y_pred = y_pred + self.coef * np2torch(config.age_kid_direction)
return self.add_random_gender(y_pred)
def __init__(self, env):
self.env = env
action_shape = env.action_space.n
self.action_embed = ActionEmbedding(env)
self.rand_net = AtariNoveltyNet((210, 160), action_shape)
self.dist_net = AtariNoveltyNet((210, 160), action_shape)
self.optimizer = Adam(lr=1e-3, params=self.dist_net.parameters())
self.criterion = MSELoss()
self.device = 'cpu'
def get_lossfn(params, encoder, data):
if params.loss == "mse":
return MSELoss()
elif params.loss == "cosine":
return CosineLoss()
elif params.loss == "ce":
return CrossEntropyLoss(ignore_index=0) # ignore padding symbol
elif params.loss == "fliploss":
if params.baseloss == "cosine":
baseloss = CosineLoss()
elif params.baseloss == "mse":
baseloss = MSELoss()
else:
raise ValueError("Unknown base loss {params.baseloss}.")
bclf = train_binary_classifier(data['Sx'], data['Sy'], encoder, params)
params.latent_binary_classifier = bclf
return FlipLoss(baseloss, bclf, lambda_clfloss=params.lambda_clfloss)
def ratio_mse_den(log_p_pred,
log_r_pred,
t_pred,
y_true,
r_true,
t_true,
log_r_clip=10.):
r_true = torch.clamp(r_true, np.exp(-log_r_clip), np.exp(log_r_clip))
r_pred = torch.exp(torch.clamp(log_r_pred, -log_r_clip, log_r_clip))
return MSELoss()(y_true * r_pred, y_true * r_true)
def _loss(input_x, output_x, lat_mu, lat_var) -> Variable:
mse_loss = MSELoss()(input_x, output_x)
kl_loss = 0.5 * torch.sum(
torch.exp(lat_var) + lat_mu.pow(2) - 1.0 - lat_var)
print('mse_loss ' + str(mse_loss))
print('kl_loss ' + str(kl_loss))
return mse_loss + kl_loss
def ratio_mse_num(log_p_pred,
log_r_pred,
t_pred,
y_true,
r_true,
t_true,
log_r_clip=10.):
r_true = torch.clamp(r_true, np.exp(-log_r_clip), np.exp(log_r_clip))
r_pred = torch.exp(torch.clamp(log_r_pred, -log_r_clip, log_r_clip))
return MSELoss()((1. - y_true) * (1. / r_pred),
(1. - y_true) * (1. / r_true))
def __init__(self, model):
'''
'''
# initialize the module using super() constructor
super(RegressionTrain, self).__init__()
# assign the architectures
self.model = model
# assign the weights for each task
self.weights = torch.nn.Parameter(torch.ones(model.n_tasks).float())
# loss function
self.mse_loss = MSELoss()
def __init__(self, state_dim, action_dim, learning_rate, tau, gamma):
self.state_dim = state_dim
self.action_dim = action_dim
self.tau = tau
self.gamma = gamma
self.critic_network = self.create_critic_network()
self.target_critic_network = self.create_critic_network()
self.optimizer = optim.Adam(self.critic_network.parameters(),
lr=learning_rate)
self.loss = MSELoss()
def _loss(input_x,
output_x,
lat_mu,
lat_var,
real_one_hot,
gen_one_hot,
show_partial=False) -> Variable:
mse_loss = MSELoss()(input_x, output_x)
kld_loss = torch.mean(
-0.5 * torch.sum(1 + lat_var - lat_mu**2 - lat_var.exp(), dim=1),
dim=0)
# KL divergence policzyć jak nalezy
# dane syntetyczne
deemb_loss = MSELoss()(real_one_hot, gen_one_hot)
# return deemb_loss
if show_partial:
print('\tmse_loss ' + str(mse_loss.item()))
print('\tkld_loss ' + str(kld_loss.item()))
print('\tdeemb_loss ' + str(deemb_loss.item()))
return mse_loss + kld_loss * 0.0001 + deemb_loss
def run_test_mutag(in_features, n_iterations=2000):
coder = MatrixEncoder(in_features)
dataset = dt.build_dataset(dt.read_test_graphs(), nodes_number=in_features)[0]
dataset = [dt.adjacency_tensor(x).unsqueeze(0) for x in dataset]
shuffle(dataset)
dataset = Variable(torch.cat(dataset, 0))
test_size = 30
train_dataset = dataset[:-test_size]
test_dataset = dataset[-test_size:]
optimizer = optim.Adam(coder.parameters(), lr=3e-1)
criterion = MSELoss()
loss_data = []
for epoch in tqdm(range(1, n_iterations)):
optimizer.zero_grad()
outputs_train = coder(train_dataset)
outputs_test = coder(test_dataset)
loss_train = criterion(outputs_train, train_dataset)
loss_test = criterion(outputs_test, test_dataset)
loss_train.backward(retain_graph=True)
if epoch % 10 == 0:
print("EPOCH: {0}, TRAIN LOSS: {1}, TEST LOSS".format(epoch, loss_train.data[0]), float(loss_test.data[0]))
if epoch == 1000:
optimizer.state_dict()['param_groups'][0]['lr'] == optimizer.state_dict()['param_groups'][0]['lr'] * 0.1
loss_data.append((loss_train.data[0], float(loss_test.data[0])))
optimizer.step()
fig = plt.figure()
loss_data = np.array(loss_data)
plt.plot(range(len(loss_data[20:])), loss_data[20:, 0], label='train', c='r')
plt.plot(range(len(loss_data[20:])), loss_data[20:, 1], label='test', c='b')
plt.title("MSELoss per epoch for train/test")
plt.xlabel('epoch')
plt.ylabel('MSELoss')
plt.grid()
plt.legend()
plt.show()
fig.savefig('loss.pdf', format='pdf')
return coder, dataset, loss_data
def train(player_model, player_idx, states, actions, is_winner):
# Get rewards for each state
rewards = [
REWARD_SIZE * (1 - EPSILON_DECAY)**i for i in range(len(states))
][::-1]
if not is_winner:
rewards = [-r for r in rewards]
# print("Rewards:", rewards)
# Get model predictions
cnn_inputs = [grid_to_cnn_input(s + 1, NUM_PLAYERS + 1) for s in states]
cnn_inputs = torch.cat(cnn_inputs)
cnn_outputs = player_model(cnn_inputs)
# Compute pseudo-ground truth
cnn_gt = []
out = cnn_outputs.detach().numpy()
for s, r, a, y in zip(states, rewards, actions, out):
gt = np.where(s == -1, y, 0) # empty space
gt.ravel()[a] = y.ravel()[a] + r # taken action
gt = gt.clip(0, 1, out=gt)
cnn_gt.append(gt)
# Create tensors
cnn_gt = torch.tensor(cnn_gt)
# Compute loss
loss = MSELoss()(cnn_outputs, cnn_gt)
print(loss)
# Backprop gradients
player_model.zero_grad()
loss.backward()
# Update weights
optimizer = Adam(player_model.parameters(), lr=LEARNING_RATE)
optimizer.step()
def __init__(self, name, config):
super().__init__(name)
self.loss: nn.Module
self.mode = config.mode
if config.mode == 'regression':
self.loss = MSELoss()
elif config.mode == 'classification':
self.loss = CrossEntropyLoss()
else:
raise ValueError(
f'config.mode must be in [regression, classification] but was {config.mode}'
)
self.head = FinetuneHead(config)
def __init__(self, log, tb_writer, args):
super(self.__class__, self).__init__()
self.log = log
self.tb_writer = tb_writer
self.args = args
self.loss_fn = MSELoss()
self.net = OmniglotNet(self.loss_fn, args).to(device)
self.fast_net = InnerLoop(self.loss_fn, args).to(device)
self.opt = Adam(self.net.parameters(), lr=args.meta_lr)
self.sampler = BatchSampler(args)
self.memory = ReplayBuffer()
def __init__(self,policy_value_model=None):
# create the model
if(policy_value_model == None):
self.model = resnet_policy_value_model()#torch_policy_value_model()
else:
self.model = policy_value_model
learning_rate = 1e-3
self.optimizer = optim.Adam(self.model.parameters(), lr=learning_rate)
self.policy_loss = nn.PoissonNLLLoss()
# self.policy_loss = MSELoss()
# self.policy_loss = nn.CrossEntropyLoss()
self.value_loss = MSELoss()
def __init__(self, sizes, bandwidth):
super(FourierRegressor, self).__init__()
self.bandwidth = bandwidth
self.sizes = sizes
self.target_size = self.sizes[-1]
self.input_size = self.sizes[0]
self.hidden_size = self.sizes[1]
self.basis = FourierFeatures(self.sizes, self.bandwidth)
self.output = nn.Linear(self.hidden_size, self.target_size)
self.criterion = MSELoss()
self.optim = None
self.input_trans = None
self.target_trans = None
def __init__(
self,
policy_thunk,
value_fn_thunk,
experience_replay,
policy_lr,
value_fn_lr,
gamma,
polyak,
learning_start,
steps_per_train,
batch_size,
l2_coeff,
max_timesteps,
):
self.online_policy = policy_thunk()
self.online_policy.cuda()
self.target_policy = policy_thunk()
self.target_policy.cuda()
self.target_policy.load_state_dict(self.online_policy.state_dict())
self.online_value_fn = value_fn_thunk()
self.online_value_fn.cuda()
self.target_value_fn = value_fn_thunk()
self.target_value_fn.cuda()
self.target_value_fn.load_state_dict(self.online_value_fn.state_dict())
self.experience_replay = experience_replay
self.gamma = gamma
self.polyak = polyak
self.total_steps = 0
self.policy_opt = optim.Adam(self.online_policy.parameters(),
policy_lr,
weight_decay=l2_coeff)
self.value_fn_opt = optim.Adam(self.online_value_fn.parameters(),
value_fn_lr,
weight_decay=l2_coeff)
self.value_fn_criterion = MSELoss()
self.learning_start = learning_start
self.steps_per_train = steps_per_train
self.batch_size = batch_size
self.max_timesteps = max_timesteps
def run_test(in_features, n_iterations=100, size=500):
coder = MatrixEncoder(in_features)
dataset = [FloatTensor(10*make_spd_matrix(in_features)).unsqueeze(0) for _ in range(size)]
dataset = Variable(torch.cat(dataset, 0))
test_size = 10
train_dataset = dataset[:-test_size]
test_dataset = dataset[-test_size:]
optimizer = optim.Adam(coder.parameters(), lr=0.1)
criterion = MSELoss()
loss_data = []
for epoch in tqdm(range(1, n_iterations)):
optimizer.zero_grad()
outputs_train = coder(train_dataset)
outputs_test = coder(test_dataset)
loss_train = criterion(outputs_train, train_dataset)
loss_test = criterion(outputs_test, test_dataset)
loss_train.backward(retain_graph=True)
if epoch % 10 == 0:
print("EPOCH: {0}, TRAIN LOSS: {1}, TEST LOSS".format(epoch, loss_train.data[0]), float(loss_test.data[0]))
loss_data.append((loss_train.data[0], float(loss_test.data[0])))
optimizer.step()
loss_data = np.array(loss_data)
plt.plot(range(len(loss_data[20:])), loss_data[20:, 0], label='train')
plt.plot(range(len(loss_data[20:])), loss_data[20:, 1], label='test')
plt.legend()
plt.show()
return coder, dataset
def train_rae(retrain=False):
# Recurrent Autoencoder
optimizer_parameters = {
"lr": 0.001,
}
criterion = MSELoss(reduction="sum")
optimizer = Adam(word_rae.parameters(), **optimizer_parameters)
mt = CriterionTrainer(
criterion=criterion,
model=word_rae,
optimizer=optimizer,
batch_size=batch_size,
max_epochs=max_epochs,
training_data=train_data,
validation_data=validation_data,
clip_max_norm=0.15,
)
mt.model_name = "WordRAE"
if not retrain:
mt.restore_checkpoint()
mt.train(progress_bar='epoch')
def __init__(self,configuration,learning_rate=1e-3,load_model=None):
# load any saved models
# Number of Columns on the Board.
self.columns = configuration.columns
# Number of Rows on the Board.
self.rows = configuration.rows
# Number of Checkers "in a row" needed to win.
self.inarow = configuration.inarow
if(load_model == None):
self.model = ResNetModel(self.columns,self.rows)
else:
self.model = load_model
self.optimizer = optim.Adam(self.model.parameters(), lr=learning_rate)
self.policy_loss = nn.KLDivLoss()
#self.policy_loss = nn.PoissonNLLLoss()
# self.policy_loss = MSELoss()
# self.policy_loss = nn.CrossEntropyLoss()
self.value_loss = MSELoss()
def fit(self, points, features, scores, n_epoch=2000, lr=5e-3):
datasets = []
optimizer = optim.Adam(params=self.net.parameters(), lr=lr)
mseLoss = MSELoss()
for p, f, s in zip(points, features, scores):
one_set = TensorDataset(to_tensor(p), to_tensor(f), to_tensor(s))
datasets.append(one_set)
train_loader = DataLoader(ConcatDataset(datasets[:-1]),
batch_size=1,
shuffle=True)
test_loader = DataLoader(ConcatDataset(datasets[-1:]),
batch_size=1,
shuffle=False)
for e in range(n_epoch):
print(f"Epoch: {e}")
for i, (p, f, s) in enumerate(train_loader):
out = self.net(p, f)
loss = mseLoss(out, s)
loss.backward()
optimizer.step()
optimizer.zero_grad()
sum_loss = 0
with torch.no_grad():
for i, (p, f, s) in enumerate(train_loader):
out = self.net(p, f)
loss = mseLoss(out, s)
sum_loss += loss
sum_loss /= len(train_loader)
print(f"Train Loss: {sum_loss.item()}")
sum_loss = 0
with torch.no_grad():
for i, (p, f, s) in enumerate(test_loader):
out = self.net(p, f)
loss = mseLoss(out, s)
sum_loss += loss
sum_loss /= len(test_loader)
print(f"Test Loss: {sum_loss.item()}")
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
):
outputs = self.albert(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output)
logits = self.classifier(pooled_output)
outputs = (logits,) + outputs[2:] # add hidden states and attention if they are here
if labels is not None:
if self.num_labels == 1:
# We are doing regression
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), labels.view(-1))
else:
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
outputs = (loss,) + outputs
return outputs # (loss), logits, (hidden_states), (attentions)
def create_setting_pretraindcae(model=None):
"""
This part is fixed for pretrain DCAE for mnist from paper Deep one-class classification setting.
adam are used in paper.
"""
setting = {}
if cfg.pretrain_solver == 'adam':
optimizer = torch.optim.Adam(params=model.parameters(), lr=cfg.pretrain_lr)
elif cfg.pretrain_solver == 'sgd':
optimizer = torch.optim.SGD(lr=cfg.pretrain_lr, momentum=cfg.pretrain_momentum, nesterov=True, params=model.parameters())
else:
raise ValueError('invalid pretrain solver for using: {}'.format(cfg.pretrain_solver))
setting['optim'] = optimizer
if cfg.ae_loss == 'l2':
#from loss import MSEReconstructionError
#loss = MSEReconstructionError()
print('using MSE')
loss = MSELoss(reduction='none')
if cfg.ae_loss == 'ce':
loss = BCELoss()
setting['criterion'] = loss
return setting
def __init__(self,
gamma,
epsilon_start,
epsilon_end,
epsilon_decay,
alpha,
target_update,
max_iter,
tau,
batch_size=16,
dropout_ratio=0.25):
self.gamma = gamma
self.epsilon_start = epsilon_start
self.epsilon_end = epsilon_end
self.epsilon_decay = epsilon_decay
self.alpha = alpha
self.target_update = target_update
self.max_iter = max_iter
self.batch_size = batch_size
self.dropout_ratio = dropout_ratio
self.tau = tau
self.tag = "g" + str(self.gamma) + "e" + str(self.epsilon_decay) + "lr" + str(self.alpha) + "t" \
+ str(self.target_update) + "b" + str(self.batch_size) + "d" + str(self.dropout_ratio) + "tau" + str(self.tau)
self.memory = ReplayMemory(5000, self.batch_size)
self.env = gym.make("LunarLander-v2")
self.n_actions = self.env.action_space.n
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.policy_net = DQN(self.dropout_ratio)
self.target_net = DQN(self.dropout_ratio)
self.policy_net = self.policy_net.float()
self.target_net = self.target_net.float()
self.target_net.load_state_dict(self.policy_net.state_dict())
self.optimizer = torch.optim.Adam(self.policy_net.parameters(),
lr=self.alpha)
self.loss = MSELoss()
def __init__(
self,
model_thunk,
experience_replay,
lr,
gamma,
learning_start,
steps_per_train,
batch_size,
max_timesteps,
final_eps,
exploration_fraction,
steps_per_target_update,
adam_eps,
):
self.online_network = model_thunk()
self.online_network.cuda()
self.target_network = model_thunk()
self.target_network.cuda()
self.target_network.load_state_dict(self.online_network.state_dict())
self.experience_replay = experience_replay
self.eps = 1.0
self.gamma = gamma
self.total_steps = 0
self.opt = optim.Adam(self.online_network.parameters(),
lr,
eps=adam_eps)
self.criterion = MSELoss()
self.learning_start = learning_start
self.steps_per_train = steps_per_train
self.batch_size = batch_size
self.max_timesteps = max_timesteps
self.final_eps = final_eps
self.exploration_fraction = exploration_fraction
self.steps_per_target_update = steps_per_target_update
valid_sampler = SubsetRandomSampler(val_indices)
train_loader = DataLoader(dataset,
batch_size=256,
num_workers=0,
sampler=train_sampler)
valid_loader = DataLoader(dataset,
batch_size=256,
num_workers=0,
sampler=valid_sampler)
model = LSTMModel(no_outputs=1)
model = model.cuda()
optimizer = optim.RMSprop(model.parameters(), lr=1e-3)
scheduler = StepLR(optimizer, 1, 0.999)
criterion = MSELoss()
criterion = criterion.cuda()
print_each = 20
def train(epoch):
running_loss = 0.0
model.train()
for i, data in enumerate(train_loader):
inputs = data['images'].cuda()
targets = data['commands'].cuda()
optimizer.zero_grad()
weights_a, weights_b, weights_s = find_weights(targets[:, 0], bins)
import torch.nn as nn
from torch.nn.modules.loss import MSELoss
from torch.optim.adam import Adam
from ..pipeline import Pipeline
from ..learner import LearningAlgorithm
# noinspection PyAbstractClass
class DummyNet(nn.Module):
def __init__(self) -> None:
super().__init__()
self.fc = nn.Linear(2, 2)
net = DummyNet()
learner = LearningAlgorithm(MSELoss(), Adam(net.parameters()))
class TestPipeline:
def test_constructor_rejects_invalid_elements(self) -> None:
# Invalid pre-processor(s)
with pytest.raises(AssertionError):
# noinspection PyTypeChecker
Pipeline(1, net, learner, None)
# Invalid net
with pytest.raises(AssertionError):
# noinspection PyTypeChecker
Pipeline(None, 1, learner, None)
with pytest.raises(AssertionError):
# noinspection PyTypeChecker
Pipeline(None, None, learner, None)
parser.add_argument('--reduce_lr_patience',
required=False,
type=int,
default=3)
parser.add_argument('--denoising_factor',
required=False,
type=float,
default=0.5)
parser.add_argument('--epochs_no_improve',
required=False,
type=float,
default=5)
parser.add_argument('--loss',
required=False,
type=_Loss,
default=MSELoss(reduction='mean'))
parser.add_argument('--auto_encoder_embedding',
required=False,
default='tmp/reversal_auto_encoder_5000_500.pt')
parser.add_argument('--load_attnn_model',
required=False,
type=bool,
default=False)
parser.add_argument('--pseudo_label_iterations',
required=False,
type=int,
default=10)
# Models parameters
parser.add_argument('--autoencoder_shape',
required=False,