def read_site_level_correlation(model, X, y_true,
lambda_1=1, lambda_2=1, lambda_3=1,
mode='prod_pooling'):
read_level_probability, site_level_probability, read_representation = model.get_read_site_probability(X)
# Calculating site probability through pooling of read level probability and classifier
if mode == 'prod_pooling':
pooled_read_probability = 1 - torch.prod(1 - read_level_probability, axis=1)
elif mode == 'mean_pooling':
pooled_read_probability = torch.mean(read_level_probability, axis=1)
elif mode == 'max_pooling':
pooled_read_probability = torch.max(read_level_probability, axis=1).values
site_level_loss = BCELoss()(site_level_probability.flatten(), y_true.float())
read_level_loss = BCELoss()(pooled_read_probability, y_true.float())
# Calculate correlation between site level probabiltiy and read level
device = pooled_read_probability.device
vx = site_level_probability - torch.mean(site_level_probability.flatten())
vy = pooled_read_probability - torch.mean(pooled_read_probability)
eps = torch.Tensor([1e-6]).to(device)
pcc = torch.sum(vx * vy) / (torch.sqrt(torch.sum(vx ** 2) + eps) * torch.sqrt(torch.sum(vy ** 2) + eps))
pcc_loss = 1 - pcc
loss = lambda_1 * site_level_loss + lambda_2 * read_level_loss + lambda_3 * pcc_loss
return {'y_pred': site_level_probability, 'loss': loss, 'y_pred_read': pooled_read_probability,
'site_loss': site_level_loss, 'read_loss': read_level_loss, 'pcc_loss': pcc_loss}
def __init__(self, hparams, fold):
super().__init__()
# Save hparams for later
self.hparams = hparams
self.fold = fold
if self.hparams.dataset == "US8K":
self.dataset_folder = config.path_to_UrbanSound8K
self.nb_classes = 10
self.best_scores = [0] * 5
elif self.hparams.dataset == "ESC50":
self.dataset_folder = config.path_to_ESC50
self.nb_classes = 50
self.best_scores = [0] * 5
elif self.hparams.dataset == "SONYCUST":
self.dataset_folder = config.path_to_SONYCUST
self.nb_classes = 31
self.best_scores = [0] * 10
else:
None
#
# Settings for the SED models
model_param = {"classes_num": self.nb_classes}
self.model = Cnn10(**model_param)
if self.hparams.dataset != "SONYCUST":
self.loss = BCELoss(reduction="none")
else:
self.loss_c = BCELoss(reduction="none")
self.loss_f = Masked_loss(BCELoss(reduction="none"))
self.mixup_augmenter = Mixup(self.hparams.alpha, self.hparams.seed)
def __init__(self, hparams):
super().__init__()
# Save hparams for later
self.hparams = hparams
self.path_to_talnet = config.audioset
self.num_classes_dict = {'coarse': 8, 'fine': 23, 'both': 31}
#hack
self.data_prepared = False
self.prepare_data()
num_meta = len(self.dataset[0]['metadata'])
# Model parameters
model_param = dict(
(k, self.hparams.__dict__[k])
for k in ("dropout", "pooling", "n_conv_layers", "kernel_size",
"n_pool_layers", "embedding_size", "batch_norm"))
self.model = TALNetV3(
self.hparams,
num_mels=64,
num_meta=num_meta,
num_classes=self.num_classes_dict[self.hparams.output_mode])
# Load every pretrained layers matching our layers
pretrained_dict = torch.load(self.path_to_talnet,
map_location='cpu')['model']
model_dict = self.model.state_dict()
pretrained_dict = {
k: v
for k, v in pretrained_dict.items()
if k in model_dict and v.size() == model_dict[k].size()
}
model_dict.update(pretrained_dict)
self.model.load_state_dict(model_dict)
self.loss_c = BCELoss(reduction='none')
self.loss_f = Masked_loss(BCELoss(reduction='none'))
def _define_loss(self):
loss_functions = {
"Generator": BCELoss(),
"Adversary": BCELoss(),
"L1": L1Loss()
}
return loss_functions
def __init__(self, hparams):
super().__init__()
# Save hparams for later
self.hparams = hparams
self.num_classes_dict = {'coarse': 8, 'fine': 23, 'both': 31}
# To avoid calling prepare_data 2 times
self.data_prepared = False
self.prepare_data()
num_meta = len(self.dataset[0]['metadata'])
# Model parameters
model_param = {
'n_conv_layers': hparams.n_conv_layers,
'kernel_size': (3, 3),
'n_pool_layers': hparams.n_pool_layers,
'embedding_size': hparams.embedding_size,
'norm': 'GN',
'conv_pool_strat': 'max',
'conv_activation': hparams.conv_activation,
'pooling': hparams.pooling,
'dropout': hparams.dropout,
'n_head': hparams.n_head,
'd_kv': hparams.d_kv,
'dropout_transfo': hparams.dropout_transfo,
'num_meta': num_meta,
'meta_emb': hparams.meta_embedding_size,
'num_mels': hparams.num_mels,
'num_classes': self.num_classes_dict[self.hparams.output_mode]
}
self.model = TALNetV2_meta(**model_param)
self.loss_c = BCELoss(reduction='none')
self.loss_f = Masked_loss(BCELoss(reduction='none'))
def weighted_binary_cross_entropy_loss(model, X, y_true):
y_pred = model(X)
_, counts = torch.unique(y_true, return_counts=True)
pos_weight, neg_weight = counts
sample_weights = torch.where(y_true == 0, neg_weight, pos_weight)
loss = BCELoss(reduce=False)(y_pred, y_true.float()) * sample_weights
loss = loss.mean()
return create_train_results(y_pred, loss)
def _define_loss(self):
loss_functions = {
"Generator": BCELoss(),
"Adversary": BCELoss(),
"Discrete": CrossEntropyLoss(),
"Continuous": NormalNegativeLogLikelihood()
}
return loss_functions
def _define_loss(self):
if self.adv_type == "Discriminator":
loss_functions = {"Generator": BCELoss(), "Adversary": BCELoss()}
elif self.adv_type == "Critic":
loss_functions = {"Generator": WassersteinLoss(), "Adversary": WassersteinLoss()}
else:
raise NotImplementedError("'adv_type' must be one of Discriminator or Critic.")
return loss_functions
def main():
args = parse_training_args("SRGAN")
epochs = args.epochs
load_path = args.load
init_path = args.init
out_path = args.out
cuda = args.cuda
device = torch.device(
'cuda' if torch.cuda.is_available() and cuda else 'cpu')
g_net = Generator().to(device)
g_criterion = PerceptualLoss(feature_extractor=TruncatedVgg(),
content_criterion=MSELoss(),
adversarial_criterion=BCELoss()).to(device)
g_optimizer = Adam(params=filter(lambda p: p.requires_grad,
g_net.parameters()),
lr=1e-4)
g_scheduler = ReduceLROnPlateau(optimizer=g_optimizer,
factor=0.5,
patience=3,
verbose=True)
d_net = Discriminator().to(device)
d_criterion = DiscriminatorLoss(criterion=BCELoss()).to(device)
d_optimizer = Adam(params=filter(lambda p: p.requires_grad,
d_net.parameters()),
lr=1e-4)
d_scheduler = ReduceLROnPlateau(optimizer=d_optimizer,
factor=0.5,
patience=3,
verbose=True)
converter = Converter()
dataset = ImageNetDataset(json_path='data/train.json', converter=converter)
data_loader = DataLoader(dataset=dataset,
batch_size=8,
num_workers=4,
pin_memory=True,
shuffle=True)
trainer = GANTrainer(g_net=g_net,
g_criterion=g_criterion,
g_optimizer=g_optimizer,
g_scheduler=g_scheduler,
d_net=d_net,
d_criterion=d_criterion,
d_optimizer=d_optimizer,
d_scheduler=d_scheduler,
data_loader=data_loader,
device=device)
if init_path:
trainer.load_pretrained_generator(init_path)
if load_path:
trainer.load(load_path)
trainer.train(max_epochs=epochs, save_path=out_path)
def __init__(self, anchors, num_classes, img_dim):
super(YOLOLayer, self).__init__()
self.anchors = anchors
self.num_anchors = len(anchors)
self.num_classes = num_classes
self.bbox_attrs = 5 + num_classes
self.img_dim = img_dim
self.ignore_thres = 0.5
self.lambda_coord = 1
self.mse_loss = MSELoss()
self.bce_loss = BCELoss()
def _get_loss(F, model, batch, forward_barch_fun, **kws):
device = torch.device(F.device)
y = batch['label'].to(device)
y_ = forward_barch_fun(F, model, batch, **kws)
loss = BCELoss()(y_, y)
sc = {}
sc['loss'] = loss.item()
sc['auc'] = np.mean([
calc_auc(y_.cpu().detach().numpy()[:, i],
y.cpu().detach().numpy()[:, i]) for i in range(y.shape[1])
])
sc['score'] = scoring(sc['loss'])
return loss, sc
def loss(self, input_x, output_x, lat_mean: torch.Tensor,
lat_logvar: torch.Tensor) -> (Variable, Variable, Variable):
real_cred_id, \
real_debt_id, \
real_log_balance, \
real_perc_withdrawed, \
real_trans_types, \
real_trans_ops, \
real_trans_descs, \
real_bank_code_debtor = self.cut_emb_data(input_x)
gen_cred_id, \
gen_debt_id, \
gen_log_balance, \
gen_perc_withdrawed, \
gen_trans_types, \
gen_trans_ops, \
gen_trans_descs, \
gen_bank_code_debtor = output_x
loss_cred_id = MSELoss()(real_cred_id, gen_cred_id)
loss_debt_id = MSELoss()(real_debt_id, gen_debt_id)
loss_log_balance = MSELoss()(real_log_balance, gen_log_balance)
loss_perc_withdrawed = MSELoss()(real_perc_withdrawed,
gen_perc_withdrawed)
loss_trans_types = BCELoss()(gen_trans_types,
real_trans_types.detach())
loss_trans_ops = BCELoss()(gen_trans_ops, real_trans_ops.detach())
loss_trans_descs = BCELoss()(gen_trans_descs,
real_trans_descs.detach())
loss_bank_code_debtor = BCELoss()(gen_bank_code_debtor,
real_bank_code_debtor.detach())
recon_loss = 0*loss_cred_id + \
0*loss_debt_id + \
loss_log_balance + \
loss_perc_withdrawed + \
loss_trans_types + \
loss_trans_ops + \
loss_trans_descs + \
loss_bank_code_debtor
kld_loss = torch.mean(-0.5 * torch.sum(
1 + lat_logvar - lat_mean.pow(2) - lat_logvar.exp(), dim=1),
dim=0)
return recon_loss + kld_loss * 5e-3, recon_loss, kld_loss
def __init__(self, hparams, fold):
super().__init__()
# Save hparams for later
self.hparams = hparams
self.fold = fold
if self.hparams.dataset == "US8K":
self.dataset_folder = config.path_to_UrbanSound8K
self.nb_classes = 10
self.input_size = (162, 64)
self.best_scores = [0] * 5
elif self.hparams.dataset == "ESC50":
self.dataset_folder = config.path_to_ESC50
self.nb_classes = 50
self.input_size = (200, 64)
self.best_scores = [0] * 5
elif self.hparams.dataset == "SONYCUST":
self.dataset_folder = config.path_to_SONYCUST
self.nb_classes = 31
self.input_size = (400, 64)
self.best_scores = [0] * 10
else:
None
model_param = {
"num_mels": hparams.num_mels,
"num_classes": self.nb_classes,
}
self.model = TALNetV3NoMeta(hparams.__dict__, **model_param)
# Load every pretrained layers matching our layers
pretrained_dict = torch.load(config.audioset,
map_location="cpu")["model"]
model_dict = self.model.state_dict()
pretrained_dict = {
k: v
for k, v in pretrained_dict.items()
if k in model_dict and v.size() == model_dict[k].size()
}
model_dict.update(pretrained_dict)
self.model.load_state_dict(model_dict)
if self.hparams.dataset != "SONYCUST":
self.loss = BCELoss(reduction="none")
else:
self.loss_c = BCELoss(reduction="none")
self.loss_f = Masked_loss(BCELoss(reduction="none"))
def deep_learning_setup(self):
"""
Sets up the optimizer and loss criterion
"""
if self.optimizer_type == 'Adam':
self.optimizer = Adam(filter(lambda p: p.requires_grad,
self.detector.parameters()),
lr=self.lr)
elif self.optimizer_type == 'RMSprop':
self.optimizer = RMSprop(filter(lambda p: p.requires_grad,
self.detector.parameters()),
lr=self.lr)
elif self.optimizer_type == 'SGD':
self.optimizer = SGD(filter(lambda p: p.requires_grad,
self.detector.parameters()),
lr=self.lr)
else:
raise NotImplementedError("No such optimizer implemented!!!")
if self.loss_type == 'BCE':
self.criterion = BCELoss().to(device)
elif self.loss_type == 'CrossEntropy':
self.criterion = CrossEntropyLoss().to(device)
elif self.loss_type == 'MSE':
self.criterion = MSELoss().to(device)
else:
raise NotImplementedError("No such loss function implemented!!!")
def __init__(self, model, lr, weight_decay):
super(trainDeepPix, self).__init__()
self.model = model
self.lossC = BCELoss()
self.lossS = BCELoss()
if CUDA:
self.model = self.model.cuda()
self.lossC = self.lossC.cuda()
self.lossS = self.lossS.cuda()
self.optimizer = Adam(self.model.parameters(),
lr=lr,
weight_decay=weight_decay)
def ordinalLoss(y_pred, y_true, n, padded_value_indicator=PADDED_Y_VALUE):
"""
Ordinal loss.
:param y_pred: predictions from the model, shape [batch_size, slate_length, n]
:param y_true: ground truth labels, shape [batch_size, slate_length]
:param n: number of ordinal values, int
:param padded_value_indicator: an indicator of the y_true index containing a padded item, e.g. -1
:return: loss value, a torch.Tensor
"""
# device = get_torch_device()
y_pred = y_pred.clone()
y_true = with_ordinals(y_true.clone(), n)
mask = y_true == padded_value_indicator
valid_mask = y_true != padded_value_indicator
ls = BCELoss(reduction='none')(y_pred, y_true)
ls[mask] = 0.0
document_loss = torch.sum(ls, dim=2)
# sum_valid = torch.sum(valid_mask, dim=2).type(torch.float32) > torch.tensor(0.0, dtype=torch.float32, device=device)
sum_valid = torch.sum(valid_mask, dim=2).type(
torch.float32) > torch.tensor(0.0, dtype=torch.float32)
loss_output = torch.sum(document_loss) / torch.sum(sum_valid)
return loss_output
def forward_sub(self, sequence_output, attention_mask, sub_labels,
batch_size):
logits = self.classifier(sequence_output)
outputs = (logits,
) # add hidden states and attention if they are here
# logger.warning(f"logits.shape: {logits.shape}")
loss_fct = BCELoss(reduction='none')
# loss_fct = BCEWithLogitsLoss(reduction='none')
loss_sig = nn.Sigmoid()
# Only keep active parts of the loss
# active_logits = logits.view(-1, self.num_labels)
active_logits = logits.view(-1, 2)
active_logits = loss_sig(active_logits)
active_logits = active_logits**2
# logger.warning(f"active_logits.shape: {active_logits.shape}")
if sub_labels is not None:
# active_labels = labels.view(-1, self.num_labels).float()
active_labels = sub_labels.view(-1, 2).float()
# logger.warning(
# f"active_logits.shape: {active_logits.shape}, active_labels.shape: {active_labels.shape}"
# )
loss = loss_fct(active_logits, active_labels)
# loss = loss.view(-1,sequence_output.size()[1],2)
loss = loss.view(batch_size, -1, 2)
loss = torch.mean(loss, 2)
loss = torch.sum(attention_mask * loss) / torch.sum(attention_mask)
outputs = (loss, ) + outputs
else:
outputs = (torch.tensor(0.0), active_logits)
# outputs = (torch.tensor(0.0), logits)
return outputs
def site_level_entropy_cluster(model, X, y_true,
lambda_1=1, lambda_2=1,
mode='prod_pooling'):
read_level_probability, site_level_probability, read_representation = model.get_read_site_probability(X)
# Calculating site probability through pooling of read level probability and classifier
site_level_loss = BCELoss()(site_level_probability, y_true.float())
# Calculate loss based on the compactness of reads with similar probability score
with torch.no_grad():
prob = read_level_probability.detach().cpu().numpy().reshape(-1, 1)
kmeans = KMeans(2).fit(prob)
labels = kmeans.labels_
read_representation = read_representation.reshape(-1, read_representation.shape[-1])
probability_mask = (labels == 1)
pos_reps = read_representation[probability_mask]
neg_reps = read_representation[~probability_mask]
device = read_representation.device
eps = torch.Tensor([1e-08]).to(device)
rep_loss = 0
for reps in [pos_reps, neg_reps]:
if len(reps) > 1:
reps = reps.div(torch.max(reps.pow(2).sum(1, keepdim=True).pow(0.5), eps))
rep_similarities = reps @ reps.t()
norm_constants = rep_similarities.numel() - rep_similarities.shape[0]
avg_similarities = (torch.sum(rep_similarities) - torch.sum(torch.diagonal(rep_similarities))) / norm_constants # zeroing the diagonal
sim_loss = -torch.log(avg_similarities)
rep_loss += sim_loss
loss = lambda_1 * site_level_loss + lambda_2 * rep_loss
return {'y_pred': site_level_probability, 'loss': loss,
'site_loss': site_level_loss, 'sim_loss': sim_loss}
def forward(self, input_ids, pos_tag_ids, senti_word_ids, polarity_ids, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None,
labels=None):
outputs = self.roberta(input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
pos_ids=pos_tag_ids,
senti_word_ids=senti_word_ids,
polarity_ids=polarity_ids
)
sequence_output = outputs[0]
logits = self.classifier(sequence_output)
logits = self.sigmoid_layer(logits)
outputs = (logits,) + outputs[2:]
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 = BCELoss()
loss = loss_fct(logits.view(-1), labels.view(-1))
outputs = (loss,) + outputs
return outputs # (loss), logits, (hidden_states), (attentions)
def __init__(self,
se_loss=False,
se_weight=0.2,
nclass=-1,
aux=False,
aux_weight=0.4,
weight=None,
size_average=True,
ignore_index=-1,
reduction='mean'):
super(SegmentationLosses_oc, self).__init__(weight,
ignore_index=ignore_index,
reduction=reduction)
self.ignore_index = ignore_index
self.se_loss = se_loss
self.aux = aux
self.nclass = nclass
self.se_weight = se_weight
self.aux_weight = aux_weight
self.bceloss = BCELoss(weight, reduction=reduction)
self.bce = BCEWithLogitsLoss(weight=None,
size_average=None,
reduce=None,
reduction='mean',
pos_weight=None)
def bce(y_pred, y_true, indices=None, padded_value_indicator=PADDED_Y_VALUE):
"""
Binary Cross-Entropy loss.
:param y_pred: predictions from the model, shape [batch_size, slate_length]
:param y_true: ground truth labels, shape [batch_size, slate_length]
:param padded_value_indicator: an indicator of the y_true index containing a padded item, e.g. -1
:return: loss value, a torch.Tensor
"""
device = get_torch_device()
y_pred = y_pred.clone()
y_true = y_true.clone()
mask = y_true == padded_value_indicator
valid_mask = y_true != padded_value_indicator
ls = BCELoss(reduction='none')(y_pred, y_true)
ls[mask] = 0.0
document_loss = torch.sum(ls, dim=-1)
sum_valid = torch.sum(valid_mask, dim=-1).type(
torch.float32) > torch.tensor(0.0, dtype=torch.float32, device=device)
loss_output = torch.sum(document_loss) / torch.sum(sum_valid)
return loss_output
def __init__(self):
super().__init__()
self.criterion = BCELoss()
self.old_classes = []
self.old_model = None
self.old_logits = None
def __init__(self,
se_loss=False,
se_weight=0.2,
nclass=-1,
aux=False,
aux_weight=0.4,
weight=None,
size_average=True,
ignore_index=-1,
reduction='mean'):
super(SegmentationLosses_contour_BoundaryRelax,
self).__init__(weight,
ignore_index=ignore_index,
reduction=reduction)
self.se_loss = se_loss
self.aux = aux
self.nclass = nclass
self.se_weight = se_weight
self.aux_weight = aux_weight
self.bceloss = BCELoss(weight, reduction=reduction)
self.gamma = 2.0
self.alpha = 1.0
self.label_relax = RelaxedBoundaryLossToTensor(ignore_id=ignore_index,
num_classes=nclass)
self.label_relax_loss = ImgWtLossSoftNLL(classes=nclass,
ignore_index=ignore_index,
weights=None,
upper_bound=1.0,
norm=False)
def __init__(self, config, special_slot_label_number, slot_embedding_type='with_BIO', matching_similarity_type='xy1', sentence_encoder_shared=True, slot_embedding_dependent_query_encoding='none', matching_similarity_y='ctx', matching_similarity_function='dot', task_adaptive_projection='none', pretrained_tf_model=None, device=None):
'''
slot_embedding_type: with_BIO, with_BI, without_BIO
matching_similarity_type: xy, x1y, xy1, x1y1, rxy
slot_embedding_dependent_query_encoding: none, ctx, desc
matching_similarity_y: ctx, desc, ctx_desc
task_adaptive_projection: none, LSM, adaptive_finetune, LEN
'''
super().__init__()
self.config = config
self.device = device
self.special_slot_label_number = special_slot_label_number
self.slot_embedding_type = slot_embedding_type
self.matching_similarity_type = matching_similarity_type
self.matching_similarity_y = matching_similarity_y
#sentence_encoder_shared = True #default 'True'; False
bilinear = False #default 'False'; True, False
crf_trainable_balance_weight = False # default 'False'; False
if self.matching_similarity_type == 'rxy':
crf_trainable_balance_weight = True
#slot_embedding_dependent_query_encoding = True # True, False
self.slot_embedding_dependent_query_encoding = slot_embedding_dependent_query_encoding
# token embedding layer (shared between query and support sets)
self.query_sentence_encoder = SequenceEncoder_with_pure_bert(config, pretrained_tf_model, device=self.device)
if sentence_encoder_shared:
self.support_sentence_encoder = self.query_sentence_encoder
else:
self.support_sentence_encoder = SequenceEncoder_with_pure_bert(config, pretrained_tf_model, device=self.device)
encoder_output_dim = self.query_sentence_encoder.get_output_dim()
self.special_slot_label_embeddings = nn.Embedding(special_slot_label_number, encoder_output_dim)
self.slot_tagger_projection_x = Projection(config, encoder_output_dim, device=self.device)
self.slot_tagger_projection_y = self.slot_tagger_projection_x
self.slot_tagger = MatchingClassifier(config, matching_similarity_function=matching_similarity_function)
projected_tag_embedding_dim = self.slot_tagger_projection_y.get_output_dim()
if config.task_st == "slot_tagger":
self.slot_tag_loss_fct = CrossEntropyLoss(ignore_index=config.output_tag_pad_id, size_average=False)
elif config.task_st == "slot_tagger_with_adaptive_crf":
self.crf_transition_layer = TransitionLayer(projected_tag_embedding_dim)
self.crf = CRFLoss(trainable_balance_weight=crf_trainable_balance_weight, device=self.device)
self.slot_tag_loss_fct = self.crf.neg_log_likelihood_loss
elif config.task_st == "slot_tagger_with_abstract_crf":
self.crf_transition_layer = AbstractTransitionLayer(device=self.device)
self.crf = CRFLoss(trainable_balance_weight=crf_trainable_balance_weight, device=self.device)
self.slot_tag_loss_fct = self.crf.neg_log_likelihood_loss
else:
exit()
self.intent_multi_class = (config.task_sc_type == "multi_cls_BCE")
self.intent_classifier = MatchingClassifier(config, matching_similarity_function=matching_similarity_function)
if self.intent_multi_class:
self.intent_loss_fct = BCELoss(size_average=False)
else:
self.intent_loss_fct = CrossEntropyLoss(size_average=False)
self.init_weights()
def main():
params = vars(
train_parse_args(hyperparameters_default=HYPERPARAMETERS,
paths_default=PATHS))
if not params['disable_cuda'] and torch.cuda.is_available():
params['device'] = torch.device('cuda:0')
else:
params['device'] = torch.device('cpu')
print(torch.cuda.is_available(), params['device'])
loaders = get_data_loaders(imgs_dir=params['imgs_dir'],
labels_filename=params['labels_filename'],
batch_size=params['batch_size'],
n_imgs=params['n_imgs'])
model = ResNet()
#model = SimpleClassifier()
model.cuda()
optimizer = Adam(model.parameters(), lr=3e-4)
criterion = BCELoss()
trainer = Trainer(params, model, optimizer, criterion)
trainer.run(loaders)
def __init__(self, args: argparse.Namespace):
super().__init__()
self.args = args
self.model = ConvolutionClassifier()
self.loss_fn = BCELoss()
self.predict = [[] for i in range(919)]
self.label = [[] for i in range(919)]
def __init__(self, vocab_size_1=16, vocab_size_2=4, word_dim=300,
hidden_dim=256, batch_size=64, num_layers=2, use_dropout=True,
dropout=0.5, use_cuda=False):
super(AttentionSeq2Seq, self).__init__()
self.hidden_dim = hidden_dim
self.num_layers = num_layers
self.use_cuda = use_cuda
self.W1 = nn.Linear(hidden_dim*2, hidden_dim*2)
self.W2 = nn.Linear(hidden_dim*2, hidden_dim*2)
self.W3 = nn.Linear(hidden_dim*2, vocab_size_2)
self.W4 = nn.Linear(hidden_dim*2, vocab_size_2)
self.v = nn.Linear(hidden_dim*2, 1)
if use_dropout:
self.encoder = nn.LSTM(word_dim, hidden_dim, dropout=dropout,
num_layers=self.num_layers, batch_first=True,
bidirectional=True)
self.decoder = nn.LSTM(word_dim, hidden_dim, dropout=dropout,
num_layers=self.num_layers, batch_first=True,
bidirectional=True)
else:
self.encoder = nn.LSTM(word_dim, hidden_dim,
num_layers=self.num_layers, batch_first=True,
bidirectional=True)
self.decoder = nn.LSTM(word_dim, hidden_dim,
num_layers=self.num_layers, batch_first=True,
bidirectional=True)
self.encembedding = nn.Embedding(vocab_size_1, word_dim)
self.decembedding = nn.Embedding(vocab_size_2, word_dim)
self.batch_size = batch_size
self.loss = BCELoss()
def __init__(self, args, word_embeddings: TextFieldEmbedder,
vocab: Vocabulary) -> None:
super().__init__(vocab)
# parameters
self.args = args
self.word_embeddings = word_embeddings
# layers
self.event_embedding = EventEmbedding(self.args, self.word_embeddings)
self.lstm = DynamicLSTM(self.args.embedding_size * 2,
self.args.hidden_size,
num_layers=1,
batch_first=True)
self.attention = NoQueryAttention(self.args.hidden_size +
self.args.embedding_size * 2,
score_function='bi_linear')
self.score = Score(self.args.hidden_size,
self.args.embedding_size,
threshold=self.args.threshold)
# metrics
self.accuracy = BooleanAccuracy()
self.f1_score = F1Measure(positive_label=1)
self.loss_function = BCELoss()
def eval_genome(g, c):
cppn = create_cppn(
g, c,
['k', 'C_in', 'C_out', 'z'],
['w', 'b'])
cnn = AudioCNN(cppn)
criterion = BCELoss()
optimizer = torch.optim.SGD(cnn.parameters(), lr=0.001)
y_ = getattr(c, "y")
loss = None
for i in range(20):
out = cnn(c.data.view(c.data.shape[0], 1, -1))
optimizer.zero_grad()
out = torch.sigmoid(cnn.l1(out.view(-1, 72)))
loss = criterion(out, y_)
loss.backward()
optimizer.step()
return 1 / (1 + loss.detach().item())
def __init__(self, dev, dis):
from torch.nn import BCELoss
super().__init__(dev, dis)
# define the criterion object
self.criterion = BCELoss()
