def _forward_train(self, encodings: torch.tensor,
context_masks: torch.tensor, entity_masks: torch.tensor,
entity_sizes: torch.tensor, relations: torch.tensor,
rel_masks: torch.tensor):
# get contextualized token embeddings from last transformer layer
context_masks = context_masks.float()
h = self.bert(input_ids=encodings, attention_mask=context_masks)[0]
entity_masks = entity_masks.float()
batch_size = encodings.shape[0]
# classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_clf, entity_spans_pool = self._classify_entities(
encodings, h, entity_masks, size_embeddings)
# classify relations
rel_masks = rel_masks.float().unsqueeze(-1)
h_large = h.unsqueeze(1).repeat(
1, max(min(relations.shape[1], self._max_pairs), 1), 1, 1)
rel_clf = torch.zeros(
[batch_size, relations.shape[1],
self._relation_types]).to(self.rel_classifier.weight.device)
# obtain relation logits
# chunk processing to reduce memory usage
for i in range(0, relations.shape[1], self._max_pairs):
# classify relation candidates
chunk_rel_logits = self._classify_relations(
entity_spans_pool, size_embeddings, relations, rel_masks,
h_large, i)
rel_clf[:, i:i + self._max_pairs, :] = chunk_rel_logits
return entity_clf, rel_clf
def _forward_train_common(self,
encodings: torch.tensor,
context_masks: torch.tensor,
mention_masks: torch.tensor,
mention_sizes: torch.tensor,
entities: torch.tensor,
entity_masks: torch.tensor,
coref_mention_pairs: torch.tensor,
coref_eds,
max_spans=None,
max_coref_pairs=None,
**kwargs):
context_masks = context_masks.float()
mention_masks = mention_masks.float()
entity_masks = entity_masks.float()
h = self.bert(input_ids=encodings,
attention_mask=context_masks)['last_hidden_state']
mention_reprs = self.mention_representation(h,
mention_masks,
max_spans=max_spans)
entity_reprs = self.entity_representation(mention_reprs, entities,
entity_masks)
mention_clf = self.mention_localization(mention_reprs, mention_sizes)
entity_clf = self.entity_classification(entity_reprs)
coref_clf = self.coreference_resolution(mention_reprs,
coref_mention_pairs,
coref_eds,
max_pairs=max_coref_pairs)
return h, mention_reprs, entity_reprs, mention_clf, entity_clf, coref_clf
def forward(self, encodings: torch.tensor, context_masks: torch.tensor, mention_masks: torch.tensor,
mention_sample_masks: torch.tensor, coref_mention_pairs: torch.tensor,
coref_eds: torch.tensor, coref_sample_masks: torch.tensor,
max_spans=None, max_coref_pairs=None, valid_mentions=None, inference=False, **kwargs):
context_masks = context_masks.float()
mention_masks = mention_masks.float()
coref_sample_masks = coref_sample_masks.float()
h = self.bert(input_ids=encodings, attention_mask=context_masks)['last_hidden_state']
mention_reprs = self.mention_representation(h, mention_masks, max_spans=max_spans)
coref_clf = self.coreference_resolution(mention_reprs, coref_mention_pairs, coref_eds, max_pairs=max_coref_pairs)
if inference:
if valid_mentions is None:
valid_mentions = torch.ones(mention_sample_masks.shape, dtype=torch.float).to(context_masks.device)
valid_mentions *= mention_sample_masks
clusters, clusters_sample_masks = misc.create_clusters(coref_clf, coref_mention_pairs, coref_sample_masks,
valid_mentions, self._coref_threshold)
coref_clf = torch.sigmoid(coref_clf)
coref_clf[coref_clf < self._coref_threshold] = 0
coref_clf *= coref_sample_masks
return dict(coref_clf=coref_clf, clusters=clusters, clusters_sample_masks=clusters_sample_masks)
return dict(coref_clf=coref_clf)
def focal_loss(
y_pred: torch.tensor,
y_true: torch.tensor,
alpha: float = 1,
gamma: float = 2,
reduce: bool = True
) -> torch.tensor:
"""Focal loss for training a classifier.
Args:
y_pred (torch.tensor): tensor of predicted probabilities
y_true (torch.tensor): tensor of binary integer targets
alpha (float, optional): balancing term. Defaults to 1.
gamma (float, optional): focusing parameter (larger --> downweights easier samples).
Defaults to 2.
reduce (bool, optional): whether to mean-reduce the loss. Defaults to True.
Returns:
torch.tensor: focal loss (scalar)
"""
bce_loss_term = F.binary_cross_entropy(y_pred.float(), y_true.float(), reduction='none')
p_t = torch.exp(-bce_loss_term)
loss = (alpha * ((1 - p_t) ** gamma) * bce_loss_term)
if reduce:
loss = loss.mean()
return loss
def _forward_encode(self, encodings: torch.tensor,
context_mask: torch.tensor, entity_masks: torch.tensor,
entity_sizes: torch.tensor, relations: torch.tensor,
rel_masks: torch.tensor):
# get contextualized token embeddings from last transformer layer
context_mask = context_mask.float()
h = self.bert(input_ids=encodings, attention_mask=context_mask)[0]
entity_masks = entity_masks.float()
batch_size = encodings.shape[0]
device = self.entity_encoder.weight.device
# encode and classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_encoding, entity_spans_pool = self._encode_entities(
encodings, h, entity_masks, size_embeddings)
# prepare relation encoding
rel_masks = rel_masks.float().unsqueeze(-1)
h_large = h.unsqueeze(1).repeat(
1, max(min(relations.shape[1], self._max_pairs), 1), 1, 1)
rel_encoding = torch.zeros(
[batch_size, relations.shape[1], self.encoding_size]).to(device)
# obtain relation encodings
# chunk processing to reduce memory usage
for i in range(0, relations.shape[1], self._max_pairs):
# classify relation candidates
rel_encoding_chunk = self._encode_relations(
entity_spans_pool, size_embeddings, relations, rel_masks,
h_large, i)
rel_encoding[:, i:i + self._max_pairs, :] = rel_encoding_chunk
return entity_encoding, rel_encoding
def forward(self, encodings: torch.tensor, context_masks: torch.tensor, mention_masks: torch.tensor,
entities: torch.tensor, entity_masks: torch.tensor,
rel_entity_pairs: torch.tensor, rel_sample_masks: torch.tensor,
rel_entity_pair_mp: torch.tensor, rel_mention_pair_ep: torch.tensor,
rel_mention_pairs: torch.tensor, rel_ctx_masks: torch.tensor, rel_pair_masks: torch.tensor,
rel_token_distances: torch.tensor, rel_sentence_distances: torch.tensor, entity_types: torch.tensor,
max_spans: bool = None, max_rel_pairs: bool = None, inference: bool = False, *args, **kwargs):
context_masks = context_masks.float()
mention_masks = mention_masks.float()
entity_masks = entity_masks.float()
h = self.bert(input_ids=encodings, attention_mask=context_masks)['last_hidden_state']
mention_reprs = self.mention_representation(h, mention_masks, max_spans=max_spans)
entity_reprs = self.entity_representation(mention_reprs, entities, entity_masks)
entity_pair_reprs = self.entity_pair_representation(entity_reprs, rel_entity_pairs)
rel_entity_types = util.batch_index(entity_types, rel_entity_pairs)
rel_clf = self.relation_classification(entity_pair_reprs, h, mention_reprs,
rel_entity_pair_mp, rel_mention_pair_ep,
rel_mention_pairs, rel_ctx_masks, rel_pair_masks,
rel_token_distances, rel_sentence_distances, rel_entity_types,
max_pairs=max_rel_pairs)
if inference:
rel_clf = torch.sigmoid(rel_clf)
rel_clf[rel_clf < self._rel_threshold] = 0
rel_clf *= rel_sample_masks.unsqueeze(-1)
return dict(rel_clf=rel_clf)
def _forward_eval(self,
encodings: torch.tensor,
context_mask: torch.tensor,
entity_masks: torch.tensor,
entity_sizes: torch.tensor,
entity_spans: torch.tensor = None,
entity_sample_mask: torch.tensor = None):
# get contextualized token embeddings from last transformer layer
context_mask = context_mask.float()
h = self.bert(input_ids=encodings, attention_mask=context_mask)[0]
# enhance hidden features
orig_shape = h.shape
h = self.feature_enhancer.prepare_input(h, context_mask)
h = self.feature_enhancer(h)
h = self.feature_enhancer.prepare_output(h, orig_shape)
entity_masks = entity_masks.float()
batch_size = encodings.shape[0]
ctx_size = context_mask.shape[-1]
# classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_clf, entity_spans_pool = self._classify_entities(
encodings, h, entity_masks, size_embeddings)
# apply softmax
entity_clf = torch.softmax(entity_clf, dim=2)
return entity_clf
def _forward_inference_common(self,
encodings: torch.tensor,
context_masks: torch.tensor,
mention_masks: torch.tensor,
mention_sizes: torch.tensor,
mention_spans: torch.tensor,
mention_sample_masks: torch.tensor,
max_spans=None,
max_coref_pairs=None):
# get contextualized token embeddings from last transformer layer
context_masks = context_masks.float()
mention_masks = mention_masks.float()
mention_sample_masks = mention_sample_masks.float()
# embed documents
h = self.bert(input_ids=encodings,
attention_mask=context_masks)['last_hidden_state']
# get mention representations
mention_reprs = self.mention_representation(h,
mention_masks,
max_spans=max_spans)
# classify mentions
mention_clf = self.mention_localization(mention_reprs, mention_sizes)
valid_mentions = (
(torch.sigmoid(mention_clf) >= self._mention_threshold).float() *
mention_sample_masks)
# create mention pairs
coref_mention_pairs, coref_mention_eds, coref_sample_masks = misc.create_coref_mention_pairs(
valid_mentions, mention_spans, encodings, self._tokenizer)
coref_sample_masks = coref_sample_masks.float()
# classify coreferences
coref_clf = self.coreference_resolution(mention_reprs,
coref_mention_pairs,
coref_mention_eds,
max_pairs=max_coref_pairs)
# create clusters
clusters, clusters_sample_masks = misc.create_clusters(
coref_clf, coref_mention_pairs, coref_sample_masks, valid_mentions,
self._coref_threshold)
entity_sample_masks = clusters_sample_masks.any(-1).float()
# create entity representations
entity_reprs = self.entity_representation(
mention_reprs, clusters, clusters_sample_masks.float())
# classify entities
entity_clf = self.entity_classification(entity_reprs)
return (h, mention_reprs, entity_reprs, clusters, entity_sample_masks,
coref_sample_masks, clusters_sample_masks, mention_clf,
entity_clf, coref_clf)
def _forward_eval(self,
encodings: torch.tensor,
context_mask: torch.tensor,
entity_masks: torch.tensor,
entity_sizes: torch.tensor,
entity_spans: torch.tensor = None,
entity_sample_mask: torch.tensor = None):
# get contextualized token embeddings from last transformer layer
context_mask = context_mask.float()
h = self.bert(input_ids=encodings, attention_mask=context_mask)[0]
# enhance hidden features
orig_shape = h.shape
h = self.feature_enhancer.prepare_input(h, context_mask)
h = self.feature_enhancer(h)
h = self.feature_enhancer.prepare_output(h, orig_shape)
entity_masks = entity_masks.float()
batch_size = encodings.shape[0]
ctx_size = context_mask.shape[-1]
# classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_clf, entity_spans_pool = self._classify_entities(
encodings, h, entity_masks, size_embeddings)
# ignore entity candidates that do not constitute an actual entity for relations (based on classifier)
relations, rel_masks, rel_sample_masks = self._filter_spans(
entity_clf, entity_spans, entity_sample_mask, ctx_size)
rel_masks = rel_masks.float()
rel_sample_masks = rel_sample_masks.float()
h_large = h.unsqueeze(1).repeat(
1, max(min(relations.shape[1], self._max_pairs), 1), 1, 1)
rel_clf = torch.zeros(
[batch_size, relations.shape[1],
self._relation_types]).to(self.rel_classifier.weight.device)
# obtain relation logits
# chunk processing to reduce memory usage
for i in range(0, relations.shape[1], self._max_pairs):
# classify relation candidates
chunk_rel_logits = self._classify_relations(
entity_spans_pool, size_embeddings, relations, rel_masks,
h_large, i)
# apply sigmoid
chunk_rel_clf = torch.sigmoid(chunk_rel_logits)
rel_clf[:, i:i + self._max_pairs, :] = chunk_rel_clf
rel_clf = rel_clf * rel_sample_masks # mask
# apply softmax
entity_clf = torch.softmax(entity_clf, dim=2)
return entity_clf, rel_clf, relations
def dice(self,
input: torch.tensor,
target: torch.tensor,
weight: float,
epsilon=1e-6) -> float:
"""
Computes DiceCoefficient as defined in https://arxiv.org/abs/1606.04797 given a multi channel input and target.
Assumes the input is a normalized probability, e.g. a result of Sigmoid or Softmax function.
:param input: NxCxSpatial input tensor
:param target: NxCxSpatial target tensor
:param weight: Cx1 tensor of weight per channel. Channels represent the class
:param epsilon: prevents division by zero
:return: dice loss, dice score
"""
assert input.size() == target.size(
), "'input' and 'target' must have the same shape"
input = self._flatten(input)
target = self._flatten(target)
target = target.float()
# Compute per channel Dice Coefficient
intersect = (input * target).sum(-1)
if weight is not None:
intersect = weight * intersect
union = (input * input).sum(-1) + (target * target).sum(-1)
return 2 * (intersect / union.clamp(min=epsilon))
def masked_cross_entropy(logits:torch.tensor, target:torch.tensor, length:torch.tensor):
"""
Args:
logits: A tensor containing a FloatTensor of size
(batch, max_len, num_classes) which contains the
unnormalized probability for each class.
target: A tensor containing a LongTensor of size
(batch, max_len) which contains the index of the true
class for each corresponding step.
length: A tensor containing a LongTensor of size (batch,)
which contains the length of each data in a batch.
Returns:
loss: An average loss value masked by the length.
"""
# logits_flat: (batch * max_len, num_classes)
logits_flat = logits.view(-1, logits.size(-1))
# log_probs_flat: (batch * max_len, num_classes)
log_probs_flat = functional.log_softmax(logits_flat, dim=-1)
# target_flat: (batch * max_len, 1)
target_flat = target.view(-1, 1)
# losses_flat: (batch * max_len, 1)
losses_flat = -torch.gather(log_probs_flat, dim=1, index=target_flat)
# losses: (batch, max_len)
losses = losses_flat.view(*target.size())
# mask: (batch, max_len)
mask = sequence_mask(sequence_length=length, max_len=target.size(1))
losses = losses * mask.float()
loss = losses.sum() / length.float().sum()
return loss
def __init__(
self,
model: torch.nn.Module,
device: torch.device,
optimizer: torch.optim.Adam,
file_path: str,
train_loader: DataLoader,
valid_loader: DataLoader,
test_loader: DataLoader,
unique_cells: Sequence[int],
num_epochs: int,
cells_tensor: torch.tensor,
label_to_cellid: Dict[int, int],
is_distance_distribution: bool
):
self.model = model
self.device = device
self.optimizer = optimizer
self.file_path = file_path
self.train_loader = train_loader
self.valid_loader = valid_loader
self.test_loader = test_loader
self.unique_cells = unique_cells
self.num_epochs = num_epochs
self.cells_tensor = cells_tensor.float().to(self.device)
self.label_to_cellid = label_to_cellid
self.cos = nn.CosineSimilarity(dim=2)
self.best_valid_loss = float("Inf")
if not os.path.exists(self.file_path):
os.mkdir(self.file_path)
self.model_path = os.path.join(self.file_path, 'model.pt')
self.metrics_path = os.path.join(self.file_path, 'metrics.tsv')
self.is_distance_distribution = is_distance_distribution
def print_tensor(x: torch.tensor, pt=True, val=True, shp=True):
"""Print the mean, min, median, std, and size of a tensor tensor
Args:
x:
val: if print the values of the tensor
shp: if print the shape of the tensor
Returns: None
"""
x = x.float()
message = ''
if shp:
message = str(x.shape) + '\n'
if val:
x = x.flatten()
if len(x) != 1:
message += (
'mean = %3.3f, min = %3.3f, max = %3.3f, median = %3.3f, std=%3.3f'
% (x.mean(), x.min(), x.max(), x.median(), x.std()))
else:
message += (f'one element {x[0]}')
if pt:
logging.debug(message)
return message
def forward(self, encodings: torch.tensor, context_masks: torch.tensor, mention_masks: torch.tensor,
mention_sizes: torch.tensor, mention_sample_masks: torch.tensor,
max_spans=None, inference=False, **kwargs):
context_masks = context_masks.float()
mention_masks = mention_masks.float()
h = self.bert(input_ids=encodings, attention_mask=context_masks)['last_hidden_state']
mention_reprs = self.mention_representation(h, mention_masks, max_spans=max_spans)
mention_clf = self.mention_localization(mention_reprs, mention_sizes)
if inference:
mention_clf = torch.sigmoid(mention_clf)
mention_clf[mention_clf < self._mention_threshold] = 0
mention_clf *= mention_sample_masks
return dict(mention_clf=mention_clf)
def _forward_train(self, encodings: torch.tensor,
context_mask: torch.tensor, entity_masks: torch.tensor,
entity_sizes: torch.tensor, relations: torch.tensor,
rel_masks: torch.tensor):
# get contextualized token embeddings from last transformer layer
context_mask = context_mask.float()
h_bert = self.bert(input_ids=encodings, attention_mask=context_mask)[0]
# print("h_bert", h_bert.shape, h_bert[1, :, 1])
# print("mask bert", context_mask.shape, context_mask[1, :])
# lengths = context_mask.sum(dim=1).int().tolist()
# print("lengths", lengths)
# enhance hidden features
orig_shape = h_bert.shape
h = self.feature_enhancer.prepare_input(h_bert, context_mask)
h = self.feature_enhancer(h)
h = self.feature_enhancer.prepare_output(h, orig_shape)
# print("h_fe_prepped", h[1, :, 1])
entity_masks = entity_masks.float()
batch_size = encodings.shape[0]
# classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_clf, entity_spans_pool = self._classify_entities(
encodings, h, entity_masks, size_embeddings)
# classify relations
rel_masks = rel_masks.float().unsqueeze(-1)
h_large = h.unsqueeze(1).repeat(
1, max(min(relations.shape[1], self._max_pairs), 1), 1, 1)
rel_clf = torch.zeros(
[batch_size, relations.shape[1],
self._relation_types]).to(self.rel_classifier.weight.device)
# obtain relation logits
# chunk processing to reduce memory usage
for i in range(0, relations.shape[1], self._max_pairs):
# classify relation candidates
chunk_rel_logits = self._classify_relations(
entity_spans_pool, size_embeddings, relations, rel_masks,
h_large, i)
rel_clf[:, i:i + self._max_pairs, :] = chunk_rel_logits
return entity_clf, rel_clf
def forward(self, encodings: torch.tensor, context_masks: torch.tensor, mention_masks: torch.tensor,
mention_sizes: torch.tensor, entity_sample_masks: torch.tensor,
entities: torch.tensor, entity_masks: torch.tensor, max_spans=None, inference=False, **kwargs):
context_masks = context_masks.float()
mention_masks = mention_masks.float()
h = self.bert(input_ids=encodings, attention_mask=context_masks)['last_hidden_state']
mention_reprs = self.mention_representation(h, mention_masks, max_spans=max_spans)
entity_reprs = self.entity_representation(mention_reprs, entities, entity_masks)
entity_clf = self.entity_classification(entity_reprs)
if inference:
entity_clf = torch.softmax(entity_clf, dim=-1)
entity_clf *= entity_sample_masks.float().unsqueeze(-1)
return dict(entity_clf=entity_clf)
def forward(self,
wav_mask: torch.tensor,
mask_val: float = 0.0) -> torch.tensor:
# make mask
with torch.no_grad():
mel_mask = self.conv(wav_mask.float().unsqueeze(1)).squeeze(1)
mel_mask = (mel_mask != mask_val).float()
return mel_mask
def forward(self, wav_mask: torch.tensor) -> torch.tensor:
# make mask
with torch.no_grad():
wav_mask = F.pad(wav_mask, [0, self.win_length // 2], value=0.)
wav_mask = F.pad(wav_mask, [self.win_length // 2, 0], value=1.)
mel_mask = self.conv(wav_mask.float().unsqueeze(1)).squeeze(1)
mel_mask = torch.ceil(mel_mask)
return mel_mask
def __call__(self, prediction: torch.tensor, target: torch.tensor, **kwargs) -> torch.tensor:
prediction = prediction.transpose(self.axis, -1).contiguous()
target = target.transpose(self.axis, -1).contiguous()
if self.to_float:
target = target.float()
prediction = prediction.view(-1, prediction.shape[-1]) if self.is_2d else prediction.view(-1)
return self.func.__call__(prediction, target.view(-1), **kwargs)
def bce_loss(
y_pred: torch.tensor,
y_true: torch.tensor,
reduce: bool = True,
label_weight_mapping: dict = None
) -> torch.tensor:
weight = None
if label_weight_mapping is not None:
weight = torch.tensor([
[label_weight_mapping[col][int(t[col].item())] for col in range(len(t))]
for t in y_true
])
if torch.cuda.is_available():
weight = weight.cuda()
loss = F.binary_cross_entropy(y_pred.float(), y_true.float(), weight=weight)
if reduce:
loss = loss.mean()
return loss
def forward(self, embedding: torch.tensor, action: torch.tensor = None):
if action is None:
action = torch.zeros((embedding.shape[0], self.action_dim), dtype=torch.float32)
else:
action = action.float()
x = torch.cat([embedding, action], dim=-1)
quality = self.transform(x)
return quality
def f2_loss(pred: torch.tensor, target: torch.tensor, epsilon:float=1e-7) -> float:
beta = 2
y_pred = nn.Sigmoid()(pred)
y_true = target.float()
TP = (y_pred * y_true).sum(1)
prec = TP / (y_pred.sum(1) + epsilon)
rec = TP / (y_true.sum(1) + epsilon)
res = (1 + beta ** 2) * prec * rec / ((beta ** 2) + prec + rec + epsilon)
f1 = res
f1 = f1.clamp(min=0)
return 1 - f1.mean()
def intersectionAndUnionGPU(
preds: torch.tensor,
target: torch.tensor,
num_classes: int,
ignore_index=255) -> Tuple[torch.tensor, torch.tensor, torch.tensor]:
"""
inputs:
preds : shape [H, W]
target : shape [H, W]
num_classes : Number of classes
returns :
area_intersection : shape [num_class]
area_union : shape [num_class]
area_target : shape [num_class]
"""
assert (preds.dim() in [1, 2, 3])
assert preds.shape == target.shape
preds = preds.view(-1)
target = target.view(-1)
preds[target == ignore_index] = ignore_index
intersection = preds[preds == target]
# Addind .float() becausue histc not working with long() on CPU
area_intersection = torch.histc(intersection.float(),
bins=num_classes,
min=0,
max=num_classes - 1)
area_output = torch.histc(preds.float(),
bins=num_classes,
min=0,
max=num_classes - 1)
area_target = torch.histc(target.float(),
bins=num_classes,
min=0,
max=num_classes - 1)
area_union = area_output + area_target - area_intersection
# print(torch.unique(intersection))
return area_intersection, area_union, area_target
def _forward_train(self, encodings: torch.tensor,
context_mask: torch.tensor, entity_masks: torch.tensor,
entity_sizes: torch.tensor):
# get contextualized token embeddings from last transformer layer
context_mask = context_mask.float()
h_bert = self.bert(input_ids=encodings, attention_mask=context_mask)[0]
# enhance hidden features
orig_shape = h_bert.shape
h = self.feature_enhancer.prepare_input(h_bert, context_mask)
h = self.feature_enhancer(h)
h = self.feature_enhancer.prepare_output(h, orig_shape)
entity_masks = entity_masks.float()
batch_size = encodings.shape[0]
# classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_clf, entity_spans_pool = self._classify_entities(
encodings, h, entity_masks, size_embeddings)
return entity_clf
def lr_one_cuml(x: torch.tensor, y: torch.tensor, **kwargs):
from cuml.linear_model import LogisticRegression
initial_device = x.device
x = x.detach()
y = y.float()
model = LogisticRegression(**kwargs)
model = model.fit(x, y)
weight = model.coef_
bias = model.intercept_
weight = torch.as_tensor(weight, device=initial_device).float().t()
bias = torch.as_tensor(bias, device=initial_device).float()
n_iter = model.solver_model.num_iters
return weight, bias, n_iter
def _forward_train(self, encodings: torch.tensor,
context_masks: torch.tensor, entity_masks: torch.tensor,
entity_sizes: torch.tensor): # noqa
# get contextualized token embeddings from last transformer layer
context_masks = context_masks.float()
h = self.bert(input_ids=encodings, attention_mask=context_masks)[0]
# classify entities
size_embeddings = self.size_embeddings(
entity_sizes) # embed entity candidate sizes
entity_clf = self._classify_entities(encodings, h, entity_masks,
size_embeddings)
return entity_clf
def mask_fill(
fill_value: float,
tokens: torch.tensor,
embeddings: torch.tensor,
padding_index: int,
) -> torch.tensor:
"""
Function that masks embeddings representing padded elements.
:param fill_value: the value to fill the embeddings belonging to padded tokens.
:param tokens: The input sequences [bsz x seq_len].
:param embeddings: word embeddings [bsz x seq_len x hiddens].
:param padding_index: Index of the padding token.
"""
padding_mask = tokens.eq(padding_index).unsqueeze(-1)
return embeddings.float().masked_fill_(padding_mask, fill_value).type_as(embeddings)
def forward(self, preds: torch.tensor, labels: torch.tensor):
smooth = 1.
preds = torch.argmax(preds, dim=-1).float()
preds.requires_grad = True
labels = labels.float()
pred_flat = preds.view(-1)
label_flat = labels.view(-1)
intersection = (pred_flat * label_flat).sum()
score = (2. * intersection + smooth) / (pred_flat.sum() +
label_flat.sum() + smooth)
score = 1. - score.sum() / labels.size(0)
return score
def _input_transform(img: torch.tensor):
## bgr2rgb
new_img = torch.zeros(img.shape).to(device)
new_img[:, :, 0] = new_img[:, :, 0] + img[:, :, 2]
new_img[:, :, 1] = new_img[:, :, 1] + img[:, :, 1]
new_img[:, :, 2] = new_img[:, :, 2] + img[:, :, 0]
img = new_img
# reisze
img = resize(img)
img = img.permute(2, 0, 1).contiguous()
#img = img.float().div(255.0)
img = img.float() / 255.0
img = img.unsqueeze(0)
return img
def compute_FB_param(self, features_q: torch.tensor,
gt_q: torch.tensor) -> torch.tensor:
"""
inputs:
features_q : shape [n_tasks, shot, c, h, w]
gt_q : shape [n_tasks, shot, h, w]
updates :
self.FB_param : shape [n_tasks, num_classes]
"""
ds_gt_q = F.interpolate(gt_q.float(),
size=features_q.size()[-2:],
mode='nearest').long()
valid_pixels = (ds_gt_q != 255).unsqueeze(
2) # [n_tasks, shot, num_classes, h, w]
assert (valid_pixels.sum(
dim=(1, 2, 3, 4)) == 0).sum() == 0, valid_pixels.sum(dim=(1, 2, 3,
4))
one_hot_gt_q = to_one_hot(
ds_gt_q, self.num_classes) # [n_tasks, shot, num_classes, h, w]
oracle_FB_param = (valid_pixels * one_hot_gt_q).sum(
dim=(1, 3, 4)) / valid_pixels.sum(dim=(1, 3, 4))
if self.FB_param_type == 'oracle':
self.FB_param = oracle_FB_param
# Used to assess influence of delta perturbation
if self.FB_param_noise != 0:
perturbed_FB_param = oracle_FB_param
perturbed_FB_param[:,
1] += self.FB_param_noise * perturbed_FB_param[:,
1]
perturbed_FB_param = torch.clamp(perturbed_FB_param, 0, 1)
perturbed_FB_param[:, 0] = 1.0 - perturbed_FB_param[:, 1]
self.FB_param = perturbed_FB_param
else:
logits_q = self.get_logits(features_q)
probas = self.get_probas(logits_q).detach()
self.FB_param = (valid_pixels * probas).sum(dim=(1, 3, 4))
self.FB_param /= valid_pixels.sum(dim=(1, 3, 4))
# Compute the relative error
deltas = self.FB_param[:, 1] / oracle_FB_param[:, 1] - 1
return deltas
