def list_examples(batch, device=None, padding=None):
if len(batch) == 0:
raise ValueError('batch is empty')
if padding is not None:
raise NotImplementedError
first_elem = batch[0]
if isinstance(first_elem, tuple):
result = []
if not isinstance(padding, tuple):
padding = [padding] * len(first_elem)
for i in six.moves.range(len(first_elem)):
result.append([to_device(device, example[i]) for example in batch])
return tuple(result)
elif isinstance(first_elem, dict):
result = {}
if not isinstance(padding, dict):
padding = {key: padding for key in first_elem}
for key in first_elem:
result[key] = [to_device(device, example[key])
for example in batch]
return result
else:
raise NotImplementedError
return to_device(device, _concat_arrays(batch, padding))
def convert_hybrid(batch, gpu):
def to_device_batch(batch, gpu):
if gpu is None:
return batch
elif gpu < 0:
return [chainer.dataset.to_device(gpu, x) for x in batch]
else:
xp = cuda.cupy.get_array_module(*batch)
concat = xp.concatenate(batch, axis=0)
sections = np.cumsum([len(x) for x in batch[:-1]], dtype='i')
concat_dev = chainer.dataset.to_device(gpu, concat)
batch_dev = cuda.cupy.split(concat_dev, sections)
return batch_dev
return {
'ids': [x1 for x1, t1, t2, t3 in batch],
'ts_link':
to_device(x=np.array([t1 for _, t1, t2, t3 in batch], dtype='i'),
device=gpu),
'ts_type':
to_device(x=np.array([t2 for _, t1, t2, t3 in batch], dtype='i'),
device=gpu),
'ts_link_type':
to_device(x=np.array([t3 for _, t1, t2, t3 in batch], dtype='i'),
device=gpu)
}
def coupling_converter(batch, device):
list_array = list()
list_dists = list()
list_targets = list()
list_pairs_index = list()
with_target = 'fc' in batch[0]['targets'].columns
for i, d in enumerate(batch):
list_array.append(d['graphs'].input_array)
list_dists.append(d['graphs'].dists)
if with_target:
list_targets.append(
d['targets'][['fc', 'sd', 'pso', 'dso']].values.astype(np.float32))
sample_index = np.full((len(d['targets']), 1), i)
atom_index = d['targets'][['atom_index_0', 'atom_index_1']].values
list_pairs_index.append(np.concatenate([sample_index, atom_index], axis=1))
input_array = to_device(device, np.stack(list_array))
dists = to_device(device, np.stack(list_dists))
pairs_index = np.concatenate(list_pairs_index)
array = {'input_array': input_array, 'dists': dists, 'pairs_index': pairs_index}
if with_target:
array['targets'] = to_device(device, np.concatenate(list_targets))
return array
def concat_yolo(batch, device=None):
images = []
bbox = []
label = []
for b in batch:
images.append(b[0])
bbox.append(to_device(device, np.array(b[1], dtype=np.float32)))
label.append(to_device(device, np.array(b[2], dtype=np.int32)))
images = concat_examples(images, device)
return images, bbox, label
def data_convert(
batch: List[Data],
device=None,
):
if len(batch) == 0:
raise ValueError('batch is empty')
return dict(
input_array=[to_device(device, d.input_array) for d in batch],
target_ids=[to_device(device, d.target_ids) for d in batch],
vec=[to_device(device, d.vec) for d in batch],
)
def converter(self, batch, device=-1):
# alternative to chainer.dataset.concat_examples
DATA_SHAPE = 40 * 3 # 40 log filterbank
xs = [np.load(path).astype(np.float32) for path, _ in batch]
delta_x = [delta(x, 3) for x in xs]
delta_delta_x = [delta(x, 3) for x in delta_x]
Xs = [
to_device(self.device,
np.concatenate((a, b, c), axis=1).astype(np.float32))
for a, (b, c) in zip(xs, zip(delta_x, delta_delta_x))
]
# Xs = [F.concat((X, self.xp.zeros(((self.stacked_frames + self.skip_size) - len(X), DATA_SHAPE), dtype=self.xp.float32)), axis=0) if len(X) < (self.stacked_frames + self.skip_size) else X for X in Xs]
# Xs = [F.pad_sequence([X[i:i + self.stacked_frames] for i in range(0, len(X), self.skip_size)]).reshape(-1, DATA_SHAPE * self.stacked_frames) for X in Xs]
word_label = [
self.xp.asarray(lab[0]).astype(self.xp.int32) for _, lab in batch
]
char_lable = [
self.xp.asarray(lab[1]).astype(self.xp.int32) for _, lab in batch
]
lable_batch = char_lable
return Xs, lable_batch
def animate_ff(model, frames, actions, mean_image, args):
initial_frames = np.concatenate([
normalize_frame(frame) - mean_image
for frame in frames[0:args.skip_frames]
])
initial_action = actions[args.skip_frames - 1].astype(np.float32)
input_frames, input_action = concat_examples(batch=[(initial_frames,
initial_action)],
device=args.gpu)
print('input_frames shape: {}'.format(input_frames.shape))
predicted_frames = []
ground_truths = []
with chainer.no_backprop_mode():
for next_index in range(args.skip_frames, len(frames)):
print('next_index: {}'.format(next_index))
predicted_frame = model((input_frames, input_action))
next_frames = F.concat((input_frames, predicted_frame),
axis=1)[:, 3:, :, :]
next_action = actions[next_index].astype(np.float32)
input_frames = next_frames
input_action = to_device(device=args.gpu,
x=next_action.reshape((-1, 3, 1)))
# print('next_frames shape: {}'.format(input_frames.shape))
# print('next_action shape: {}'.format(input_action.shape))
# Keep predicted image
predicted_frame.to_cpu()
predicted_frames.append(
converter.chw2hwc(predicted_frame.data[0] + mean_image))
ground_truth = frames[next_index]
ground_truths.append(converter.chw2hwc(ground_truth))
return ground_truths, predicted_frames
def concat_examples_batch(batch, device=None):
train_frame = []
train_actions = []
target_frames = []
for ((frame, actions), target) in batch:
train_frame.append(frame)
train_actions.append(actions)
target_frames.append(target)
train_frame = np.array(train_frame)
train_actions = np.array(train_actions)
target_frames = np.array(target_frames)
if not device is None:
train_frame = to_device(device, train_frame)
train_actions = to_device(device, train_actions)
target_frames = to_device(device, target_frames)
# print('train frame shape: {}, train actions shape: {}, target_frames shape: {}'.format(train_frame.shape, train_actions.shape, target_frames.shape))
return dict(input=(train_frame, train_actions), target=target_frames)
def predict(start_image, goal_image, generator, posterior, transition,
subtractor, classifier, args):
start = to_device(device=args.gpu,
x=start_image.reshape((1, ) + start_image.shape))
start = subtract_background(subtractor, start)
goal = to_device(device=args.gpu,
x=goal_image.reshape((1, ) + goal_image.shape))
goal = subtract_background(subtractor, goal)
print('start image shape: ', start.shape)
print('goal image shape: ', goal.shape)
xp = np if args.gpu < 0 else cp
loc = xp.zeros(shape=(1), dtype=np.float32)
scale = xp.ones(shape=(1), dtype=np.float32)
_Normal = Normal(loc=loc, scale=scale)
sequence = []
chainer.config.train = False
start_state, _ = find_closest_latent_state(start, generator, transition,
classifier, args)
goal_state, _ = find_closest_latent_state(goal, generator, transition,
classifier, args)
with chainer.no_backprop_mode():
start_state = start_state.reshape((1, ) + start_state.shape)
goal_state = goal_state.reshape((1, ) + goal_state.shape)
s_current = start_state
print('current_state: ', s_current.shape)
step = (goal_state - start_state) / args.steps
z = _Normal.sample(sample_shape=(1, 4))
z = F.squeeze(z)
z = z.reshape((1, ) + z.shape)
for i in range(1, args.steps):
s_next = start_state + step * i
x = F.concat((z, s_current, s_next), axis=1)
x = F.reshape(x, shape=x.shape + (1, 1))
o = generator(x)
_, o_next = F.split_axis(o, 2, axis=1, force_tuple=True)
o_next.to_cpu()
sequence.append(o_next.data[0][0])
s_current = s_next
return sequence
def convert_sg(batch, device):
if len(batch) == 0: raise ValueError('batch is empty')
result = [
to_device(device,
_concat_arrays([example[0] for example in batch], None)),
[example[1] for example in batch]
]
return tuple(result)
def test_to_device(self):
src_xp = get_xp(self.src_gpu)
dst_xp = get_xp(self.dst_gpu)
x = src_xp.array([1], 'i')
y = dataset.to_device(self.device, x)
self.assertIsInstance(y, dst_xp.ndarray)
self.assertEqual(int(y.device), self.device)
def animate_lstm(model, frames, actions, mean_image, args):
with chainer.no_backprop_mode():
for step in range(args.init_steps - 1):
print('next_index: {}'.format(step))
frame = frames[step]
normalized_frame = normalize_frame(frame)
input_frame = normalized_frame - mean_image
input_frame = to_device(device=args.gpu,
x=input_frame.reshape((-1, ) +
input_frame.shape))
print('input_frame shape: {}'.format(input_frame.shape))
input_action = actions[step + 1].astype(np.float32)
input_action = to_device(device=args.gpu,
x=input_action.reshape((-1, 3, 1)))
model((input_frame, input_action))
predicted_frames = []
ground_truths = []
frame = frames[args.init_steps - 1]
normalized_frame = normalize_frame(frame)
next_frame = normalized_frame - mean_image
next_frame = to_device(device=args.gpu,
x=input_frame.reshape((-1, ) + next_frame.shape))
with chainer.no_backprop_mode():
for next_index in range(args.init_steps - 1, len(frames) - 1):
print('next_index: {}'.format(next_index))
input_frame = next_frame
input_action = actions[next_index].astype(np.float32)
input_action = to_device(device=args.gpu,
x=input_action.reshape((-1, 3, 1)))
predicted_frame = model((input_frame, input_action))
next_frame = chainer.Variable(predicted_frame.array)
# Keep predicted image
predicted_frame.to_cpu()
predicted_frames.append(
converter.chw2hwc(predicted_frame.data[0] + mean_image))
ground_truth = frames[next_index]
ground_truths.append(converter.chw2hwc(ground_truth))
return ground_truths, predicted_frames
def kdnet_converter(batch, device=None, padding=None):
# concat_examples to CPU at first.
result = concat_examples(batch, device=None, padding=padding)
out_list = []
for elem in result:
if elem.dtype != object:
# Send to GPU for int/float dtype array.
out_list.append(to_device(device, elem))
else:
# Do NOT send to GPU for dtype=object array.
out_list.append(elem)
return tuple(out_list)
def test_to_device(self):
src_xp = get_xp(self.src_gpu)
dst_xp = get_xp(self.dst_gpu)
x = src_xp.array([1], 'i')
y = dataset.to_device(self.device, x)
self.assertIsInstance(y, dst_xp.ndarray)
if self.device is not None and self.device >= 0:
self.assertEqual(int(y.device), self.device)
if self.device is None and self.src_gpu:
self.assertEqual(int(x.device), int(y.device))
def forward(self, *args, **kwargs):
if isinstance(self.label_key, int):
if not (-len(args) <= self.label_key < len(args)):
msg = 'Label key %d is out of bounds' % self.label_key
raise ValueError(msg)
t = args[self.label_key]
if self.label_key == -1:
args = args[:-1]
else:
args = args[:self.label_key] + args[self.label_key + 1:]
elif isinstance(self.label_key, str):
if self.label_key not in kwargs:
msg = 'Label key "%s" is not found' % self.label_key
raise ValueError(msg)
t = kwargs[self.label_key]
del kwargs[self.label_key]
self.y = None
self.loss = None
self.accuracy = None
self.y = self.predictor(*args, **kwargs)
self.loss = self.lossfun(self.y, t)
if -1000 < self.loss.data / self.batch_size < 1000:
reporter.report({'loss': self.loss / self.batch_size}, self)
# reporter.report({'char_loss': char_loss}, self)
else:
print('loss f****d up!!!!!!!!!!!!!')
if self.compute_accuracy:
wer = 0
ys = [y.data[:n] for y, n in zip(F.stack(self.y[0], 1), self.y[1])]
target = to_device(-1, t)
print(len(ys[0]), len(target[0]))
out = remove_blank(F.argmax(ys[0], axis=1).data)
out = [int(o) for o in out]
print(out)
print(target[0])
for yy, tt in zip(ys, target):
out = remove_blank(F.argmax(yy, axis=1).data)
out = [int(o) for o in out]
wer += _wer(out, tt)
reporter.report({'accuracy': wer / len(ys)}, self)
return self.loss
def concat_example(batch, device):
if len(batch) == 0:
raise ValueError('batch is empty')
first_elem = batch[0]
assert isinstance(first_elem, tuple)
result = []
for i in range(len(first_elem)):
result.append([example[i] for example in batch])
# labels should be ndarray
result[-1] = dataset.to_device(device, numpy.array(result[-1],
numpy.int32))
return tuple(result)
def convert(batch, gpu):
"""
"""
def to_device_batch(batch, gpu):
if gpu is None:
return batch
elif gpu < 0:
return [chainer.dataset.to_device(gpu, x) for x in batch]
else:
xp = cuda.cupy.get_array_module(*batch)
concat = xp.concatenate(batch, axis=0)
sections = np.cumsum([len(x) for x in batch[:-1]], dtype='i')
concat_dev = chainer.dataset.to_device(gpu, concat)
batch_dev = cuda.cupy.split(concat_dev, sections)
return batch_dev
return {'x1s': to_device_batch([x1 for x1, x2, t in batch], gpu=gpu),
'x2s': to_device_batch([x2 for x1, x2, t in batch], gpu=gpu),
't': to_device(x=np.array([t for _, _, t in batch], dtype='i'), device=gpu)
}
def convert(
minibatch: List[Tuple[
np.ndarray,
np.ndarray, np.ndarray, np.ndarray, np.ndarray,
np.ndarray
]],
device: Optional[int]
) -> Tuple[ndarray, ndarray, ndarray, ndarray, ndarray, ndarray]:
# Append eos to the end of sentence
eos_ = np.array([EOS], 'i')
(
src_batch,
ga_batch, wo_batch, ni_batch, ga2_batch,
tgt_batch
) = zip(*minibatch)
with chainer.no_backprop_mode():
src_sentences = \
[Variable(np.hstack((sentence, eos_))) for sentence in src_batch]
ga_sentences = \
[Variable(np.hstack((sentence, eos_))) for sentence in ga_batch]
wo_sentences = \
[Variable(np.hstack((sentence, eos_))) for sentence in wo_batch]
ni_sentences = \
[Variable(np.hstack((sentence, eos_))) for sentence in ni_batch]
ga2_sentences = \
[Variable(np.hstack((sentence, eos_))) for sentence in ga2_batch]
tgt_sentences = \
[Variable(np.hstack((sentence, eos_))) for sentence in tgt_batch]
src_block = F.pad_sequence(src_sentences, padding=PAD).data
ga_block = F.pad_sequence(ga_sentences, padding=PAD).data
wo_block = F.pad_sequence(wo_sentences, padding=PAD).data
ni_block = F.pad_sequence(ni_sentences, padding=PAD).data
ga2_block = F.pad_sequence(ga2_sentences, padding=PAD).data
tgt_block = F.pad_sequence(tgt_sentences, padding=PAD).data
return (
to_device(device, src_block),
to_device(device, ga_block),
to_device(device, wo_block),
to_device(device, ni_block),
to_device(device, ga2_block),
to_device(device, tgt_block)
)
def converter(self, batch, device=-1):
# alternative to chainer.dataset.concat_examples
DATA_SHAPE = 40 # 40 log filterbank
Xs = [
to_device(self.device,
np.load(path).astype(np.float32)) for path, _ in batch
]
Xs = [
F.concat((X,
self.xp.zeros(
((self.stacked_frames + self.skip_size) - len(X),
DATA_SHAPE),
dtype=self.xp.float32)),
axis=0) if len(X) <
(self.stacked_frames + self.skip_size) else X for X in Xs
]
Xs = [
F.pad_sequence([
X[i:i + self.stacked_frames]
for i in range(0, len(X), self.skip_size)
]).reshape(-1, DATA_SHAPE * self.stacked_frames) for X in Xs
]
word_label = [
self.xp.asarray(lab[0]).astype(self.xp.int32) for _, lab in batch
]
char_lable = [
self.xp.asarray(lab[1]).astype(self.xp.int32) for _, lab in batch
]
lable_batch = (word_label, char_lable)
return Xs, lable_batch