def forward(self, x):
x = self.decoder_unpool0(x)
p = self.decoder_block0(x)
x = F.leaky_relu(x + p)
x = self.decoder_unpool1(x)
p = self.decoder_block1(x)
x = F.leaky_relu(x + p)
x = self.decoder_unpool2(x)
p1 = self.decoder_block2(x)
p2 = self.decoder_block2_shortcut(x)
x = F.leaky_relu(p1 + p2)
p1 = self.decoder_block3(x)
p2 = self.decoder_block3_shortcut(x)
x = F.leaky_relu(p1 + p2)
x = self.decoder_block4(x)
# x = x[:,0,:,:,:]
# x = F.softmax(x, axis=1)
x = paddle.sum(x, axis=1)
x = paddle.clip(x, min=0, max=1)
return x
def add_input(self, x, condition=None):
"""Compute the output distribution (represented by its parameters) for
a step. It works similarily with the ``forward`` method but in a
``step-in-step-out`` fashion.
Parameters
-----------
x : Tensor [shape=(B,)]
A step of the input waveform.
condition : Tensor, optional [shape=(B, C_cond)]
A step of the upsampled condition. Defaults to None.
Returns
--------
Tensor: [shape=(B, C_output)]
A step of the parameters of the output distributions.
"""
# Causal Conv
if self.loss_type == "softmax":
x = paddle.clip(x, min=-1., max=0.99999)
x = quantize(x, self.output_dim)
x = self.embed(x) # (B, C)
else:
x = paddle.unsqueeze(x, -1) # (B, 1)
x = self.embed(x) # (B, C)
# Residual & Skip-conenection & linears
z = self.resnet.add_input(x, condition)
z = F.relu(self.proj2(F.relu(self.proj1(z)))) # (B, C)
# Output
y = self.proj3(z)
return y
def generate_relative_positions_embeddings(self,
length,
depth,
max_relative_position=127):
vocab_size = max_relative_position * 2 + 1
range_vec = paddle.arange(length)
range_mat = paddle.tile(range_vec, repeat_times=[length]).reshape(
(length, length))
distance_mat = range_mat - paddle.t(range_mat)
distance_mat_clipped = paddle.clip(distance_mat.astype('float32'),
-max_relative_position,
max_relative_position)
final_mat = distance_mat_clipped + max_relative_position
embeddings_table = np.zeros([vocab_size, depth])
for pos in range(vocab_size):
for i in range(depth // 2):
embeddings_table[pos, 2 * i] = np.sin(
pos / np.power(10000, 2 * i / depth))
embeddings_table[pos, 2 * i + 1] = np.cos(
pos / np.power(10000, 2 * i / depth))
embeddings_table_tensor = paddle.to_tensor(embeddings_table,
dtype='float32')
flat_relative_positions_matrix = final_mat.reshape((-1, ))
one_hot_relative_positions_matrix = paddle.nn.functional.one_hot(
flat_relative_positions_matrix.astype('int64'),
num_classes=vocab_size)
embeddings = paddle.matmul(one_hot_relative_positions_matrix,
embeddings_table_tensor)
my_shape = final_mat.shape
my_shape.append(depth)
embeddings = embeddings.reshape(my_shape)
return embeddings
def forward(self, raw_features, coarse_volumes):
n_views_rendering = coarse_volumes.shape[1]
raw_features = paddle.split(raw_features,
num_or_sections=raw_features.shape[1],
axis=1)
volume_weights = []
for i in range(n_views_rendering):
raw_feature = paddle.squeeze(raw_features[i], axis=1)
# print("torch.Size([batch_size, 9, 32, 32, 32]) ---",raw_feature.shape)
volume_weight = self.layer1(raw_feature)
# print("torch.Size([batch_size, 16, 32, 32, 32]) ---",volume_weight.shape) #
volume_weight = self.layer2(volume_weight)
# print("torch.Size([batch_size, 8, 32, 32, 32]) ---",volume_weight.shape) #
volume_weight = self.layer3(volume_weight)
# print("torch.Size([batch_size, 4, 32, 32, 32]) ---",volume_weight.shape) #
volume_weight = self.layer4(volume_weight)
# print("torch.Size([batch_size, 2, 32, 32, 32]) ---",volume_weight.shape) #
volume_weight = self.layer5(volume_weight)
# print("torch.Size([batch_size, 1, 32, 32, 32]) ---",volume_weight.shape) #
volume_weight = paddle.squeeze(volume_weight, axis=1)
# print("torch.Size([batch_size, 32, 32, 32]) ---",volume_weight.shape) #
volume_weights.append(volume_weight)
volume_weights = paddle.transpose(paddle.stack(volume_weights),
perm=[1, 0, 2, 3, 4])
volume_weights = paddle.nn.functional.softmax(volume_weights, axis=1)
# print("torch.Size([batch_size, n_views, 32, 32, 32]) ---",volume_weights.shape) #
# print("torch.Size([batch_size, n_views, 32, 32, 32]) ---",coarse_volumes.shape) #
coarse_volumes = coarse_volumes * volume_weights
coarse_volumes = paddle.sum(coarse_volumes, axis=1)
return paddle.clip(coarse_volumes, min=0, max=1)
def forward(self, x, condition=None):
"""Forward pass of ``WaveNet``.
Parameters
-----------
x : Tensor [shape=(B, T)]
The input waveform.
condition : Tensor, optional [shape=(B, C_cond, T)]
the upsampled condition. Defaults to None.
Returns
-------
Tensor: [shape=(B, T, C_output)]
The parameters of the output distributions.
"""
# Causal Conv
if self.loss_type == "softmax":
x = paddle.clip(x, min=-1., max=0.99999)
x = quantize(x, self.output_dim)
x = self.embed(x) # (B, T, C)
else:
x = paddle.unsqueeze(x, -1) # (B, T, 1)
x = self.embed(x) # (B, T, C)
x = paddle.transpose(x, perm=[0, 2, 1]) # (B, C, T)
# Residual & Skip-conenection & linears
z = self.resnet(x, condition)
z = paddle.transpose(z, [0, 2, 1])
z = F.relu(self.proj2(F.relu(self.proj1(z))))
y = self.proj3(z)
return y
def mu_law_encode(x: Tensor, mu: int = 256, quantized: bool = True) -> Tensor:
"""Mu-law encoding.
Compute the mu-law decoding given an input code.
When quantized is True, the result will be converted to
integer in range [0,mu-1]. Otherwise, the resulting signal
is in range [-1,1]
Parameters:
x(Tensor): the input tensor of arbitrary shape to be encoded.
mu(int): the maximum value (depth) of encoded signal. The signal will be
clip to be in range [0,mu-1].
quantized(bool): indicate whether the signal will quantized to integers.
Examples:
.. code-block:: python
import paddle
import paddleaudio.functional as F
F.mu_law_encode(paddle.randn((2, 8)))
>> Tensor(shape=[2, 8], dtype=int32, place=CUDAPlace(0), stop_gradient=True,
[[0, 5, 30, 255, 255, 255, 12, 13],
[0, 241, 8, 243, 7, 35, 84, 228]])
Reference:
https://en.wikipedia.org/wiki/%CE%9C-law_algorithm
"""
mu = mu - 1
y = paddle.sign(x) * paddle.log1p(mu * paddle.abs(x)) / math.log1p(mu)
if quantized:
y = (y + 1) / 2 * mu + 0.5 # convert to [0 , mu-1]
y = paddle.clip(y, min=0, max=mu).astype('int32')
return y
def dynamic_k_matching(self, cost_matrix, pairwise_ious, num_gt):
match_matrix = np.zeros_like(cost_matrix.numpy())
# select candidate topk ious for dynamic-k calculation
topk_ious, _ = paddle.topk(pairwise_ious, self.candidate_topk, axis=0)
# calculate dynamic k for each gt
dynamic_ks = paddle.clip(topk_ious.sum(0).cast('int'), min=1)
for gt_idx in range(num_gt):
_, pos_idx = paddle.topk(cost_matrix[:, gt_idx],
k=dynamic_ks[gt_idx],
largest=False)
match_matrix[:, gt_idx][pos_idx.numpy()] = 1.0
del topk_ious, dynamic_ks, pos_idx
# match points more than two gts
extra_match_gts_mask = match_matrix.sum(1) > 1
if extra_match_gts_mask.sum() > 0:
cost_matrix = cost_matrix.numpy()
cost_argmin = np.argmin(cost_matrix[extra_match_gts_mask, :],
axis=1)
match_matrix[extra_match_gts_mask, :] *= 0.0
match_matrix[extra_match_gts_mask, cost_argmin] = 1.0
# get foreground mask
match_fg_mask_inmatrix = match_matrix.sum(1) > 0
match_gt_inds_to_fg = match_matrix[match_fg_mask_inmatrix, :].argmax(1)
return match_gt_inds_to_fg, match_fg_mask_inmatrix
def compute_softmax_loss(self, y, t):
"""Compute the loss when output distributions are categorial
distributions.
Parameters
----------
y : Tensor [shape=(B, T, C_output)]
The logits of the output distributions.
t : Tensor [shape=(B, T)]
The target audio. The audio is first quantized then used as the
target.
Notes
-------
Output distributions whose input contains padding is neglected in
loss computation. So the first ``context_size`` steps does not
contribute to the loss.
Returns
--------
Tensor: [shape=(1,)]
The loss.
"""
# context size is not taken into account
y = y[:, self.context_size:, :]
t = t[:, self.context_size:]
t = paddle.clip(t, min=-1.0, max=0.99999)
quantized = quantize(t, n_bands=self.output_dim)
label = paddle.unsqueeze(quantized, -1)
loss = F.softmax_with_cross_entropy(y, label)
reduced_loss = paddle.mean(loss)
return reduced_loss
def degree_norm(graph, mode="indegree"):
"""Calculate the degree normalization of a graph
Args:
graph: the graph object from (:code:`Graph`)
mode: which degree to be normalized ("indegree" or "outdegree")
return:
A tensor with shape (num_nodes, 1).
"""
assert mode in [
'indegree', 'outdegree'
], "The degree_norm mode should be in ['indegree', 'outdegree']. But recieve mode=%s" % mode
if mode == "indegree":
degree = graph.indegree()
elif mode == "outdegree":
degree = graph.outdegree()
norm = paddle.cast(degree, dtype=paddle.get_default_dtype())
norm = paddle.clip(norm, min=1.0)
norm = paddle.pow(norm, -0.5)
norm = paddle.reshape(norm, [-1, 1])
return norm
def get_loss(self, pred_hm, pred_wh, target_hm, box_target, target_weight):
pred_hm = paddle.clip(F.sigmoid(pred_hm), 1e-4, 1 - 1e-4)
hm_loss = self.hm_loss(pred_hm, target_hm)
H, W = target_hm.shape[2:]
mask = paddle.reshape(target_weight, [-1, H, W])
avg_factor = paddle.sum(mask) + 1e-4
base_step = self.down_ratio
shifts_x = paddle.arange(0, W * base_step, base_step, dtype='int32')
shifts_y = paddle.arange(0, H * base_step, base_step, dtype='int32')
shift_y, shift_x = paddle.tensor.meshgrid([shifts_y, shifts_x])
base_loc = paddle.stack([shift_x, shift_y], axis=0)
base_loc.stop_gradient = True
pred_boxes = paddle.concat(
[0 - pred_wh[:, 0:2, :, :] + base_loc, pred_wh[:, 2:4] + base_loc],
axis=1)
pred_boxes = paddle.transpose(pred_boxes, [0, 2, 3, 1])
boxes = paddle.transpose(box_target, [0, 2, 3, 1])
boxes.stop_gradient = True
pred_boxes, boxes, mask = self.filter_box_by_weight(
pred_boxes, boxes, mask)
mask.stop_gradient = True
wh_loss = self.wh_loss(pred_boxes, boxes, iou_weight=mask.unsqueeze(1))
wh_loss = wh_loss / avg_factor
ttf_loss = {'hm_loss': hm_loss, 'wh_loss': wh_loss}
return ttf_loss
def forward(self, x):
x = x.unsqueeze(1)
x = self.conv1(x)
x = self.relu(x)
x = self.bn1(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = x.reshape((x.shape[0], -1, x.shape[-1]))
w = self.attention(x)
if self.encoder_type == "SAP":
x = paddle.sum(x * w, axis=2)
elif self.encoder_type == "ASP":
mu = paddle.sum(x * w, axis=2)
sg = paddle.sum((x**2) * w, axis=2) - mu**2
sg = paddle.clip(sg, min=1e-5)
sg = paddle.sqrt(sg)
x = paddle.concat((mu, sg), 1)
x = x.reshape((x.shape[0], -1))
x = self.fc(x)
return x
def forward(self, dist_feat):
dist = paddle.clip(dist_feat.squeeze(), 1.0,
self.cut_dist - 1e-6).astype('int64') - 1
eh_emb = self.dist_embedding_layer(dist)
eh_emb = self.dist_input_layer(eh_emb)
# eh_emb = paddle.cast(eh_emb, 'float64')
return eh_emb
def forward(self, input_data):
expert_outputs = []
for i in range(0, self.expert_num):
linear_out = self._param_expert[i](input_data)
expert_output = F.relu(linear_out)
expert_outputs.append(expert_output)
expert_concat = paddle.concat(x=expert_outputs, axis=1)
expert_concat = paddle.reshape(expert_concat,
[-1, self.expert_num, self.expert_size])
output_layers = []
for i in range(0, self.gate_num):
cur_gate_linear = self._param_gate[i](input_data)
cur_gate = F.softmax(cur_gate_linear)
cur_gate = paddle.reshape(cur_gate, [-1, self.expert_num, 1])
cur_gate_expert = paddle.multiply(x=expert_concat, y=cur_gate)
cur_gate_expert = paddle.sum(x=cur_gate_expert, axis=1)
cur_tower = self._param_tower[i](cur_gate_expert)
cur_tower = F.relu(cur_tower)
out = self._param_tower_out[i](cur_tower)
out = F.softmax(out)
out = paddle.clip(out, min=1e-15, max=1.0 - 1e-15)
output_layers.append(out)
return output_layers
def forward(self):
fpn_rois = self.input('FpnRois', 0)
areas = self.bbox_area(fpn_rois)
scale = paddle.sqrt(areas)
num_level = self.max_level - self.min_level + 1
target_level = paddle.log(scale / self.refer_scale + 1e-06) / np.log(2)
target_level = paddle.floor(self.refer_level + target_level)
target_level = paddle.clip(target_level,
min=self.min_level,
max=self.max_level)
rois = list()
rois_idx_order = list()
for level in range(self.min_level, self.max_level + 1):
level_tensor = paddle.full_like(target_level, fill_value=level)
res = paddle.equal(target_level, level_tensor)
res = paddle.squeeze(res, axis=1)
res = paddle.cast(res, dtype='int32')
index = paddle.nonzero(res)
roi = paddle.gather(fpn_rois, index, axis=0)
rois.append(roi)
rois_idx_order.append(index)
rois_idx_order = paddle.concat(rois_idx_order, axis=0)
size = paddle.shape(rois_idx_order)[0]
_, rois_idx_restore = paddle.topk(rois_idx_order,
axis=0,
sorted=True,
largest=False,
k=size)
#rois_idx_restore = paddle.cast(rois_idx_restore, dtype='int32')
return {'MultiFpnRois': rois, 'RestoreIndex': [rois_idx_restore]}
def cdist(self, a, b):
a_s = paddle.norm(a, p=2, axis=-1).pow(2)
b_s = paddle.norm(b, p=2, axis=-1).pow(2)
dist_score = -2 * paddle.bmm(a, b.transpose(
[0, 2, 1])) + b_s.unsqueeze(-2) + a_s.unsqueeze(-1)
dist_score = paddle.sqrt(paddle.clip(dist_score, min=1e-30))
return dist_score
def forward(self, outputs, targets):
""" This performs the loss computation.
Parameters:
outputs: dict of tensors, see the output specification of the model for the format
targets: list of dicts, such that len(targets) == batch_size.
The expected keys in each dict depends on the losses applied, see each loss' doc
"""
outputs_without_aux = {k: v for k, v in outputs.items() if k !=\
'aux_outputs'}
indices = self.matcher(outputs_without_aux, targets)
num_boxes = sum(len(t['labels']) for t in targets)
num_boxes = paddle.to_tensor([num_boxes], dtype=torch.float, device
=next(iter(outputs.values())).device)
if is_dist_avail_and_initialized():
torch2paddle.all_reduce(num_boxes)
num_boxes = paddle.clip(num_boxes / get_world_size(), min=1).item()
losses = {}
for loss in self.losses:
losses.update(self.get_loss(loss, outputs, targets, indices,
num_boxes))
if 'aux_outputs' in outputs:
for i, aux_outputs in enumerate(outputs['aux_outputs']):
indices = self.matcher(aux_outputs, targets)
for loss in self.losses:
if loss == 'masks':
continue
kwargs = {}
if loss == 'labels':
kwargs = {'log': False}
l_dict = self.get_loss(loss, aux_outputs, targets,
indices, num_boxes, **kwargs)
l_dict = {(k + f'_{i}'): v for k, v in l_dict.items()}
losses.update(l_dict)
return losses
def degree_norm(self, g):
degree = g.indegree() + 1 # self loop
norm = paddle.cast(degree, dtype=paddle.get_default_dtype())
norm = paddle.clip(norm, min=1.0)
norm = paddle.pow(norm, -0.5)
norm = paddle.reshape(norm, [-1, 1])
return norm
def forward(self, sender_receiver, is_training):
# Construct permutation input
sender_emb = self.feature_emb_edge(
paddle.cast(sender_receiver[0, :], 'int32'))
receiver_emb = self.feature_emb_edge(
paddle.cast(sender_receiver[1, :], 'int32'))
_input = paddle.multiply(sender_emb, receiver_emb)
h_relu = self.dropout(self.relu(self.linear1(_input)))
loc = self.linear2(h_relu)
if is_training:
u = paddle.rand(loc.shape, dtype=loc.dtype)
u.stop_gradient = False
logu = paddle.log2(u)
logmu = paddle.log2(1 - u)
sum_log = loc + logu - logmu
s = F.sigmoid(sum_log / self.temp)
s = s * (self.inter_max - self.inter_min) + self.inter_min
else:
s = F.sigmoid(loc) * (self.inter_max -
self.inter_min) + self.inter_min
s = paddle.clip(s, min=0, max=1)
l0_penaty = F.sigmoid(
loc -
self.temp * np.log2(-self.inter_min / self.inter_max)).mean()
return s, l0_penaty
def forward(self, input, target):
"""
Args:
inputs: feature matrix with shape (batch_size, feat_dim)
target: ground truth labels with shape (num_classes)
"""
inputs = input["features"]
if self.normalize_feature:
inputs = 1. * inputs / (paddle.expand_as(
paddle.norm(inputs, p=2, axis=-1, keepdim=True), inputs) +
1e-12)
bs = inputs.shape[0]
# compute distance
dist = paddle.pow(inputs, 2).sum(axis=1, keepdim=True).expand([bs, bs])
dist = dist + dist.t()
dist = paddle.addmm(input=dist,
x=inputs,
y=inputs.t(),
alpha=-2.0,
beta=1.0)
dist = paddle.clip(dist, min=1e-12).sqrt()
# hard negative mining
is_pos = paddle.expand(target, (bs, bs)).equal(
paddle.expand(target, (bs, bs)).t())
is_neg = paddle.expand(target, (bs, bs)).not_equal(
paddle.expand(target, (bs, bs)).t())
# `dist_ap` means distance(anchor, positive)
## both `dist_ap` and `relative_p_inds` with shape [N, 1]
'''
dist_ap, relative_p_inds = paddle.max(
paddle.reshape(dist[is_pos], (bs, -1)), axis=1, keepdim=True)
# `dist_an` means distance(anchor, negative)
# both `dist_an` and `relative_n_inds` with shape [N, 1]
dist_an, relative_n_inds = paddle.min(
paddle.reshape(dist[is_neg], (bs, -1)), axis=1, keepdim=True)
'''
dist_ap = paddle.max(paddle.reshape(paddle.masked_select(dist, is_pos),
(bs, -1)),
axis=1,
keepdim=True)
# `dist_an` means distance(anchor, negative)
# both `dist_an` and `relative_n_inds` with shape [N, 1]
dist_an = paddle.min(paddle.reshape(paddle.masked_select(dist, is_neg),
(bs, -1)),
axis=1,
keepdim=True)
# shape [N]
dist_ap = paddle.squeeze(dist_ap, axis=1)
dist_an = paddle.squeeze(dist_an, axis=1)
# Compute ranking hinge loss
y = paddle.ones_like(dist_an)
loss = self.ranking_loss(dist_an, dist_ap, y)
return {"TripletLossV2": loss}
def forward(self, state):
x = F.relu(self.l1(state))
x = F.relu(self.l2(x))
mean = self.mean(x)
log_std = F.relu(self.std(x))
log_std = paddle.clip(log_std, self.log_min_std, self.log_max_std)
return mean, log_std
def forward(self, obs):
x = F.relu(self.l1(obs))
x = F.relu(self.l2(x))
act_mean = self.mean_linear(x)
act_std = self.std_linear(x)
act_log_std = paddle.clip(act_std, min=LOG_SIG_MIN, max=LOG_SIG_MAX)
return act_mean, act_log_std
def test_clip_dygraph(self):
place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda(
) else fluid.CPUPlace()
paddle.disable_static(place)
data_shape = [1, 9, 9, 4]
data = np.random.random(data_shape).astype('float32')
images = paddle.to_variable(data, dtype='float32')
v_min = paddle.to_variable(np.array([0.2], dtype=np.float32))
v_max = paddle.to_variable(np.array([0.8], dtype=np.float32))
out_1 = paddle.clip(images, min=0.2, max=0.8)
out_2 = paddle.clip(images, min=0.2, max=0.9)
out_3 = paddle.clip(images, min=v_min, max=v_max)
self.assertTrue(np.allclose(out_1.numpy(), data.clip(0.2, 0.8)))
self.assertTrue(np.allclose(out_2.numpy(), data.clip(0.2, 0.9)))
self.assertTrue(np.allclose(out_3.numpy(), data.clip(0.2, 0.8)))
def clip_grad_value_(parameters, clip_value):
r"""Clips gradient of an iterable of parameters at specified value.
Gradients are modified in-place.
Arguments:
parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a
single Tensor that will have gradients normalized
clip_value (float or int): maximum allowed value of the gradients.
The gradients are clipped in the range
:math:`\left[\text{-clip\_value}, \text{clip\_value}\right]`
"""
if isinstance(parameters, paddle.Tensor):
parameters = [parameters]
clip_value = float(clip_value)
for p in filter(lambda p: p.grad is not None, parameters):
paddle.clip(p.grad, min=-clip_value, max=clip_value)
def update_target_network_clip(self):
for param_q, param_k in zip(self.towers[0].parameters(),
self.towers[1].parameters()):
# paddle.assign((param_k * self.m + param_q * (1. - self.m)), param_k)
paddle.assign(
param_k - (1 - self.m) * paddle.clip(
(param_k - param_q), min=-1.0, max=1.0), param_k)
param_k.stop_gradient = True
def backward(R_p):
Z = []
for _ in range(self.num):
Z.append(self.X)
Spp = []
Spn = []
for z, rp, rn in zip(Z, R_p):
Spp.append(safe_divide(paddle.clip(rp, min=0), z))
Spn.append(safe_divide(paddle.clip(rp, max=0), z))
Cpp = self.gradprop(Z, self.X, Spp)[0]
Cpn = self.gradprop(Z, self.X, Spn)[0]
Rp = self.X * (Cpp * Cpn)
return Rp
def forward(self, distance, label):
label = -1 * (2 * label - 1)
# print(label, distance)
pos_num = paddle.sum((label == 1).astype('float32')) + 0.0001
neg_num = paddle.sum((label == -1).astype('float32')) + 0.0001
loss_1 = paddle.sum((1 + label) / 2 * paddle.pow(distance, 2)) / pos_num
loss_2 = paddle.sum((1 - label) / 2 * paddle.pow(paddle.clip(self.margin - distance, min=0.0), 2)) / neg_num
loss = loss_1 + loss_2
return loss
def _get_Adc_loss(self, feed_dict, node_repr):
node_i_repr = paddle.gather(node_repr, feed_dict['Ad_node_i'])
node_j_repr = paddle.gather(node_repr, feed_dict['Ad_node_j'])
node_ij_repr = paddle.concat([node_i_repr, node_j_repr], 1)
logits = self.Adc_mlp.forward(node_ij_repr)
atom_dist = paddle.clip(feed_dict['Ad_atom_dist'], 0.0, 20.0)
atom_dist_id = paddle.cast(atom_dist / 20.0 * self.Adc_vocab, 'int64')
loss = self.Adc_loss(logits, atom_dist_id)
return loss
def get_loss(self, heatmap, size, offset, weights, inputs):
heatmap_target = inputs['heatmap']
size_target = inputs['size']
offset_target = inputs['offset']
index = inputs['index']
mask = inputs['index_mask']
heatmap = paddle.clip(F.sigmoid(heatmap), 1e-4, 1 - 1e-4)
heatmap_loss = self.focal_loss(heatmap, heatmap_target)
size = paddle.transpose(size, perm=[0, 2, 3, 1])
size_n, size_h, size_w, size_c = size.shape
size = paddle.reshape(size, shape=[size_n, -1, size_c])
index = paddle.unsqueeze(index, 2)
batch_inds = list()
for i in range(size_n):
batch_ind = paddle.full(shape=[1, index.shape[1], 1],
fill_value=i,
dtype='int64')
batch_inds.append(batch_ind)
batch_inds = paddle.concat(batch_inds, axis=0)
index = paddle.concat(x=[batch_inds, index], axis=2)
pos_size = paddle.gather_nd(size, index=index)
mask = paddle.unsqueeze(mask, axis=2)
size_mask = paddle.expand_as(mask, pos_size)
size_mask = paddle.cast(size_mask, dtype=pos_size.dtype)
pos_num = size_mask.sum()
size_mask.stop_gradient = True
size_target.stop_gradient = True
size_loss = F.l1_loss(pos_size * size_mask,
size_target * size_mask,
reduction='sum')
size_loss = size_loss / (pos_num + 1e-4)
offset = paddle.transpose(offset, perm=[0, 2, 3, 1])
offset_n, offset_h, offset_w, offset_c = offset.shape
offset = paddle.reshape(offset, shape=[offset_n, -1, offset_c])
pos_offset = paddle.gather_nd(offset, index=index)
offset_mask = paddle.expand_as(mask, pos_offset)
offset_mask = paddle.cast(offset_mask, dtype=pos_offset.dtype)
pos_num = offset_mask.sum()
offset_mask.stop_gradient = True
offset_target.stop_gradient = True
offset_loss = F.l1_loss(pos_offset * offset_mask,
offset_target * offset_mask,
reduction='sum')
offset_loss = offset_loss / (pos_num + 1e-4)
det_loss = weights['heatmap'] * heatmap_loss + weights[
'size'] * size_loss + weights['offset'] * offset_loss
return {
'det_loss': det_loss,
'heatmap_loss': heatmap_loss,
'size_loss': size_loss,
'offset_loss': offset_loss
}
def compute_mog_loss(self, y, t):
"""Compute the loss where output distributions is a mixture of
Gaussians distributions.
Parameters
-----------
y : Tensor [shape=(B, T, C_output)]
The parameterd of the output distribution. It is the concatenation
of 3 parts, the logits of every distribution, the mean of each
distribution and the log standard deviation of each distribution.
Each part's shape is (B, T, n_mixture), where ``n_mixture`` means
the number of Gaussians in the mixture.
t : Tensor [shape=(B, T)]
The target audio.
Notes
-------
Output distributions whose input contains padding is neglected in
loss computation. So the first ``context_size`` steps does not
contribute to the loss.
Returns
--------
Tensor: [shape=(1,)]
The loss.
"""
n_mixture = self.output_dim // 3
# context size is not taken in to account
y = y[:, self.context_size:, :]
t = t[:, self.context_size:]
w, mu, log_std = paddle.split(y, 3, axis=2)
# 100.0 is just a large float
log_std = paddle.clip(log_std, min=self.log_scale_min, max=100.)
inv_std = paddle.exp(-log_std)
p_mixture = F.softmax(w, -1)
t = paddle.unsqueeze(t, -1)
if n_mixture > 1:
# t = F.expand_as(t, log_std)
t = paddle.expand(t, [-1, -1, n_mixture])
x_std = inv_std * (t - mu)
exponent = paddle.exp(-0.5 * x_std * x_std)
pdf_x = 1.0 / math.sqrt(2.0 * math.pi) * inv_std * exponent
pdf_x = p_mixture * pdf_x
# pdf_x: [bs, len]
pdf_x = paddle.sum(pdf_x, -1)
per_sample_loss = -paddle.log(pdf_x + 1e-9)
loss = paddle.mean(per_sample_loss)
return loss
def forward(self, inputs):
pool = self.pool2d_gap(inputs)
pool = paddle.squeeze(pool, axis=[2, 3])
squeeze = self.squeeze(pool)
squeeze = F.relu(squeeze)
excitation = self.excitation(squeeze)
excitation = paddle.clip(x=excitation, min=0, max=1)
excitation = paddle.unsqueeze(excitation, axis=[2, 3])
out = paddle.multiply(inputs, excitation)
return out
