text
stringlengths
1
93.6k
self.optimizer.zero_grad()
def clear_buffered_stats(self):
self._buffered_stats.clear()
def get_num_updates(self):
"""Get the number of parameters updates."""
return self._num_updates
def _prepare_sample(self, sample):
if sample is None or len(sample) == 0:
return None
return utils.move_to_cuda(sample)
def dummy_train_step(self, dummy_batch):
"""Dummy training step for warming caching allocator."""
self.train_step(dummy_batch, update_params=False)
self.zero_grad()
self.clear_buffered_stats()
def compute_score_with_logits(logits, labels):
logits = torch.max(logits, 1)[1].data # argmax
one_hots = torch.zeros(*labels.size()).to(logits.device)
one_hots.scatter_(1, logits.view(-1, 1), 1)
scores = (one_hots * labels)
return scores
# <FILESEP>
"""Contains various utility functions for Dynet models."""
import dynet as dy
import numpy as np
def add_dim(variable, dim=0):
""" Adds a dimension to a vector dy.Expression.
Inputs:
variable (dy.Expression): A vector Dynet expression.
dim (int, optional): The dimension to add.
Returns:
dy.Expression with one more dimension (of size 1).
"""
var_size = variable.dim()[0][0]
if dim == 0:
return dy.reshape(variable, (1, var_size))
else:
return dy.reshape(variable, (var_size, 1))
def is_vector(exp):
""" Returns whether the expression is a vector.
Inputs:
exp (dy.Expression): The expression to check.
Returns:
bool, representing whether the expression is a vector.
"""
return len(exp.dim()[0]) == 1
def split_exp(exp, num=2):
""" Splits an expression into n parts.
Inputs:
exp (dy.Expression): A vector Dynet expression.
num (int, optional): The number of parts to split it into.
Returns:
list of dy.Expression, containing the split expression.
"""
assert is_vector(exp)
size = exp.dim()[0][0]
split_amount = int(size / num)
parts = []
for i in range(num):
parts.append(exp[i * split_amount: (i + 1) * split_amount])
return parts
def linear_transform(exp, params):
""" Multiplies a dy.Expression and a set of parameters.
Inputs:
exp (dy.Expression): A Dynet tensor.
params (dy.Parameters): Dynet parameters.
Returns:
dy.Expression representing exp * params.
"""
if is_vector(exp):
exp = add_dim(exp)