class FurnitureNASNetALarge350Finetunned(Module):
def __init__(self, pretrained):
super(FurnitureNASNetALarge350Finetunned, self).__init__()
self.model = nasnetalarge(num_classes=1000, pretrained=pretrained)
filters = self.model.penultimate_filters // 24
self.model.last_linear = Linear(24 * filters, 1024)
self.final_classifier = Sequential(ReLU(inplace=True),
Linear(1024,
512), ReLU(inplace=True),
Linear(512, 128))
for m in self.model.last_linear.modules():
if isinstance(m, Linear):
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()
for m in self.final_classifier.modules():
if isinstance(m, Linear):
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()
# create aliases:
self.stem = ModuleList([
self.model.conv0,
self.model.cell_stem_0,
self.model.cell_stem_1,
])
self.features = ModuleList([
self.model.cell_0, self.model.cell_1, self.model.cell_2,
self.model.cell_3, self.model.cell_4, self.model.cell_5,
self.model.reduction_cell_0, self.model.cell_6, self.model.cell_7,
self.model.cell_8, self.model.cell_9, self.model.cell_10,
self.model.cell_11, self.model.reduction_cell_1,
self.model.cell_12, self.model.cell_13, self.model.cell_14,
self.model.cell_15, self.model.cell_16, self.model.cell_17
])
self.classifier = self.model.last_linear
# Freeze stem and features
for param in self.stem.parameters():
param.requires_grad = False
for param in self.features.parameters():
param.requires_grad = False
def train(self, mode=True):
self.classifier.train(mode)
self.final_classifier.train(mode)
return self
def forward(self, x):
x = self.model.features(x)
x = self.model.logits(x)
x = self.final_classifier(x)
return x
class SimpleDDPGAgent(Module):
def __init__(self, **kwargs):
super(SimpleDDPGAgent, self).__init__()
hidden_size = kwargs['hidden_size']
# actor
self.actor_linears = ModuleList(
[Linear(kwargs['state_dim'], hidden_size[0])])
self.actor_linears.extend([
Linear(hidden_size[i - 1], hidden_size[i])
for i in range(1, len(hidden_size))
])
self.action = Linear(hidden_size[-1], kwargs['action_dim'])
# critic
self.critic_linears = ModuleList([
Linear(kwargs['state_dim'] + kwargs['action_dim'], hidden_size[0])
])
self.critic_linears.extend([
Linear(hidden_size[i - 1], hidden_size[i])
for i in range(1, len(hidden_size))
])
self.q = Linear(hidden_size[-1], 1)
self.relu = ReLU()
self.sigmoid = Sigmoid()
self.tanh = Tanh()
self.apply(init_weights) # xavier uniform init
def act(self, state):
x = state
for l in self.actor_linears:
x = l(x)
x = self.relu(x)
action = self.tanh(self.action(x))
return action
def Q(self, state, action):
x = torch.cat([state, action], dim=1)
for l in self.critic_linears:
x = l(x)
x = self.relu(x)
q = self.q(x)
return q
def get_actor_parameters(self):
return list(self.actor_linears.parameters()) + list(
self.action.parameters())
def get_critic_parameters(self):
return list(self.critic_linears.parameters()) + list(
self.q.parameters())
class Net(Module):
def __init__(self, input_size, hidden_sizes, output_size, lr):
super(Net, self).__init__()
max_potential = 3.0
self.layers = ModuleList()
linear = Linear(input_size, hidden_sizes[0])
self.layers.append(linear)
for i in range(len(hidden_sizes)):
layer = Integrating(hidden_sizes[i], hidden_sizes[i],
max_potential)
self.layers.append(layer)
# self.layers.append(Integrating(hidden_sizes[-1], output_size, max_potential, self.layers[-1]))
self.layers.append(Linear(hidden_sizes[-1], output_size))
self.loss = CrossEntropyLoss()
self.optimizer = torch.optim.Adam(self.layers.parameters(), lr=lr)
def forward(self, x):
for layer in self.layers:
x = layer(x)
return x
def reset_potentials(self):
for layer in self.layers:
if type(layer).__name__ == "Integrating":
layer.reset_potentials()
def __init__(self, num_classes: int, base_net: nn.ModuleList, source_layer_indexes: List[int],
extras: nn.ModuleList, classification_headers: nn.ModuleList,
regression_headers: nn.ModuleList, is_test=False, config=None, device=None):
"""Compose a SSD model using the given components.
"""
super(SSD, self).__init__()
self.num_classes = num_classes
self.base_net = base_net
self.source_layer_indexes = source_layer_indexes
self.extras = extras
self.classification_headers = classification_headers
self.regression_headers = regression_headers
self.is_test = is_test
self.config = config
# register layers in source_layer_indexes by adding them to a module list
self.source_layer_add_ons = nn.ModuleList([t[1] for t in source_layer_indexes
if isinstance(t, tuple) and not isinstance(t, GraphPath)])
self.device = next(base_net.parameters()).device
if device is not None:
self.device = device
self.base_net.to(device)
if is_test:
self.config = config
self.priors = config.priors.to(self.device)
def initmodel(character_encoder, tag_encoder, embedded_dimension):
"""
:param character_encoder:
:param tag_encoder:
:param embedded_dimension:
:return:
"""
character_encoder = copy(character_encoder)
tag_encoder = copy(tag_encoder)
tag_encoder[BD] = len(tag_encoder)
character_encoder[BD] = len(character_encoder)
character_embedding = Embedding(len(character_encoder), embedded_dimension)
tag_embedding = Embedding(len(tag_encoder), embedded_dimension)
encoder_part = LSTM(input_size=embedded_dimension,
hidden_size=LSTMDIM,
num_layers=1,
bidirectional=1).type(DTYPE)
ench0 = randn(2, 1, LSTMDIM).type(DTYPE)
encc0 = randn(2, 1, LSTMDIM).type(DTYPE)
decoder_part = LSTM(input_size=2 * LSTMDIM + embedded_dimension,
hidden_size=LSTMDIM,
num_layers=1).type(DTYPE)
dech0 = randn(2, 1, 2 * LSTMDIM + embedded_dimension).type(DTYPE)
decc0 = randn(2, 1, 2 * LSTMDIM + embedded_dimension).type(DTYPE)
pred = Linear(LSTMDIM, len(character_encoder)).type(DTYPE)
softmax = LogSoftmax().type(DTYPE)
model = ModuleList([
character_embedding, tag_embedding, encoder_part, decoder_part, pred,
softmax
])
optimizer = Adam(model.parameters(), lr=LEARNINGRATE, betas=BETAS)
return {
'model': model,
'optimizer': optimizer,
'cencoder': character_encoder,
'tencoder': tag_encoder,
'cembedding': character_embedding,
'tembedding': tag_embedding,
'enc': encoder_part,
'ench0': ench0,
'encc0': encc0,
'dec': decoder_part,
'dech0': dech0,
'decc0': decc0,
'pred': pred,
'sm': softmax,
'embdim': embedded_dimension
}
def initmodel(cencoder, tencoder, embdim):
cencoder = copy(cencoder)
tencoder = copy(tencoder)
tencoder[BD] = len(tencoder)
cencoder[BD] = len(cencoder)
cembedding = Embedding(len(cencoder), embdim)
tembedding = Embedding(len(tencoder), embdim)
enc = LSTM(input_size=embdim,
hidden_size=LSTMDIM,
num_layers=1,
bidirectional=1).type(DTYPE)
ench0 = randn(2, 1, LSTMDIM).type(DTYPE)
encc0 = randn(2, 1, LSTMDIM).type(DTYPE)
dec = LSTM(input_size=2 * LSTMDIM + embdim,
hidden_size=LSTMDIM,
num_layers=1).type(DTYPE)
dech0 = randn(2, 1, 2 * LSTMDIM + embdim).type(DTYPE)
decc0 = randn(2, 1, 2 * LSTMDIM + embdim).type(DTYPE)
pred = Linear(LSTMDIM, len(cencoder)).type(DTYPE)
sm = LogSoftmax().type(DTYPE)
model = ModuleList([cembedding, tembedding, enc, dec, pred, sm])
optimizer = Adam(model.parameters(), lr=LEARNINGRATE, betas=BETAS)
return {
'model': model,
'optimizer': optimizer,
'cencoder': cencoder,
'tencoder': tencoder,
'cembedding': cembedding,
'tembedding': tembedding,
'enc': enc,
'ench0': ench0,
'encc0': encc0,
'dec': dec,
'dech0': dech0,
'decc0': decc0,
'pred': pred,
'sm': sm,
'embdim': embdim
}
class FurnitureResNet101_350_finetune(Module):
def __init__(self, pretrained=True):
super(FurnitureResNet101_350_finetune, self).__init__()
self.model = resnet101(num_classes=1000, pretrained=pretrained)
self.model.avgpool = AdaptiveAvgPool2d(1)
# create aliases:
self.stem = ModuleList([
self.model.conv1,
self.model.bn1,
self.model.layer1,
self.model.layer2,
])
self.features = ModuleList([
self.model.layer3,
self.model.layer4,
])
self.classifier = self.model.fc
self.final_classifiers = Sequential(
Linear(1000, 1000),
Linear(1000, 128),
)
for m in self.final_classifiers.modules():
if isinstance(m, Linear):
normal_(m.weight, 0, 0.01)
constant_(m.bias, 0)
# freeze internal layers:
for param in self.stem.parameters():
param.requires_grad = False
def forward(self, x):
x = self.model(x)
return self.final_classifiers(x)
class SimplePPOAgent(Module):
def __init__(self, **kwargs):
super(SimplePPOAgent, self).__init__()
hidden_size = kwargs['hidden_size']
# actor
self.actor_bn = BatchNorm1d(kwargs['state_dim'])
self.actor_linears = ModuleList(
[Linear(kwargs['state_dim'], hidden_size[0])])
self.actor_linears.extend([
Linear(hidden_size[i - 1], hidden_size[i])
for i in range(1, len(hidden_size))
])
self.mu = Linear(hidden_size[-1], kwargs['action_dim'])
self.log_var = Linear(hidden_size[-1], kwargs['action_dim'])
# self.log_var = torch.nn.Parameter(torch.zeros(kwargs['action_dim']))
# critic
self.critic_bn = BatchNorm1d(kwargs['state_dim'])
self.critic_linears = ModuleList(
[Linear(kwargs['state_dim'], hidden_size[0])])
self.critic_linears.extend([
Linear(hidden_size[i - 1], hidden_size[i])
for i in range(1, len(hidden_size))
])
self.v = Linear(hidden_size[-1], 1)
self.relu = ReLU()
self.tanh = Tanh()
self.apply(init_weights) # xavier uniform init
def forward(self, state, action=None):
x = state
x = self.actor_bn(x)
for l in self.actor_linears:
x = l(x)
x = self.relu(x)
mu = self.tanh(self.mu(x))
log_var = -self.relu(self.log_var(x))
sigmas = log_var.exp().sqrt()
dists = Normal(mu, sigmas)
if action is None:
action = dists.sample()
log_prob = dists.log_prob(action).sum(dim=-1, keepdim=True)
x = state
x = self.critic_bn(x)
for l in self.critic_linears:
x = l(x)
x = self.relu(x)
v = self.v(x)
return action, log_prob, dists.entropy(), v
def get_actor_parameters(self):
return [
*self.actor_bn.parameters(), *self.actor_linears.parameters(),
*self.mu.parameters(), *self.log_var.parameters()
]
def get_critic_parameters(self):
return [
*self.critic_bn.parameters(), *self.critic_linears.parameters(),
*self.v.parameters()
]
def main(mode, dataset_name, expid, epochs, lr, batch_size, ntrials,
struc_learn_method):
network_dict = {}
result_dict = {}
losses_dict = {}
bayesmodel_dict = {}
# Get network
network = get_network(dataset_name)
length = len(network)
bayesmodel, layout, encoders, nfeatures, nclass = get_bayesnet_model(
dataset_name)
# Obtain models
generator = Generator(nfeatures=nfeatures, ndomain=ndomain)
discriminator = ModuleList([
Discriminator(local_nfeatures=nfeatures[local].sum(), ndomain=ndomain)
for local in network
])
classifier = Classifier(nfeatures=nfeatures, nclass=nclass)
# Obtain dataloader
data_loader, test_data_loader = get_bayesnet_dataloaders(
dataset_name=dataset_name, length=length, batch_size=batch_size)
# Initialize optimizers
G_optimizer = torch.optim.Adam(generator.parameters(),
lr=lr,
betas=(.9, .99))
D_optimizer = torch.optim.Adam(discriminator.parameters(),
lr=lr,
betas=(.9, .99))
G_scheduler = torch.optim.lr_scheduler.StepLR(G_optimizer,
step_size=1,
gamma=1)
# Set up trainer
expname = '{}'.format(dataset_name.upper())
trainer = Trainer(mode, network, bayesmodel, layout, encoder, nfeatures,
expname, generator, discriminator, G_optimizer,
D_optimizer, G_scheduler)
# Train model
expdesc = "{}-e{}lr{}bs{}pe{}nt{}-{}".format(mode, epochs, int(10000 * lr),
batch_size, pretrain_epochs,
ntrials, expid)
generator_state_dict = trainer.train(data_loader,
test_data_loader,
expdesc,
epochs=epochs)
# Get result
display.clear_output(wait=True)
#generator.load_state_dict(generator_state_dict)
generator.eval()
result_dict = {}
ess, dss, rsss, bmps, spas = [], [], [], [], []
for trial in range(ntrials):
iter_test_data = iter(test_data_loader)
sampled_data = [next(iter_test_data).data.numpy() for i in range(16)]
sampled_data = np.concatenate(sampled_data, axis=0)
fixed_latents = Variable(generator.sample_latent(
sampled_data.shape[0]))
generated = generator(fixed_latents.cuda(),
0.1).detach().cpu().data.numpy()
ess.append(energy_statistics(sampled_data, generated))
dss.append(
discriminative_score(sampled_data, generated, diagnose=False))
rsss.append(
relative_structure_score(sampled_data, generated, encoder,
bayesmodel))
bmps.append(
structure_prediction(generated,
encoder,
bayesmodel,
method=struc_learn_method,
score_type='bic'))
spas.append(structure_prediction_accuracy(bayesmodel, bmps[-1]))
print("Evaluating... Progress {:.2f}%".format(
(trial + 1) / ntrials * 100),
end='\r')
ess, dss, rsss, spas = np.array(ess), np.array(dss), np.array(
rsss), np.array(spas)
result_dict['energy_statistics'] = (ess.mean(),
ess.std() / np.sqrt(ntrials))
result_dict['discriminative_score'] = (dss.mean(),
dss.std() / np.sqrt(ntrials))
result_dict['relative_structure_score'] = (rsss.mean(),
rsss.std() / np.sqrt(ntrials))
bayesmodel_dict = bmps[0]
result_dict['structure_prediction_accuracy'] = (spas.mean(), spas.std() /
np.sqrt(ntrials))
losses_dict = {}
losses_dict['wasserstein_loss'] = trainer.losses['D']
losses_dict['energy_statistics'] = trainer.losses['energy_statistics']
losses_dict['relative_structure_score'] = trainer.losses[
'relative_structure_score']
losses_dict['structure_prediction_iteration'] = trainer.losses[
'structure_prediction_iteration']
losses_dict['structure_prediction_accuracy'] = trainer.losses[
'structure_prediction_accuracy']
losses_dict['bayesmodel_predicted'] = trainer.losses[
'bayesmodel_predicted']
network_dict = network