def __call__(self, x, pad):
pad = paddle.reshape(pad, shape=[2, -1])
pad = paddle.transpose(pad, perm=[1, 0])
pad = paddle.reverse(pad, axis=[0])
pad = paddle.flatten(pad)
pad = paddle.cast(pad, dtype="int32")
pad1, pad2 = paddle.split(pad, num_or_sections=2, axis=0)
x = paddle.nn.functional.pad(x=x, pad=pad1, **self.layer_attrs)
x = paddle.transpose(x, perm=[2, 3, 0, 1])
x = paddle.nn.functional.pad(x=x, pad=pad2, **self.layer_attrs)
out = paddle.transpose(x, perm=[2, 3, 0, 1])
return out
def iou_single(a, b, mask, n_class):
valid = mask == 1
valid_flatten = paddle.reshape(valid, (-1, ))
valid_flatten = paddle.cast(valid_flatten, dtype="int32")
index = where(valid_flatten == 1)
if index.shape[0] == 0:
return paddle.zeros((1, ))
index = paddle.reshape(index, (1, -1))
a_flatten = paddle.reshape(a, (1, -1))
a = paddle.index_sample(a_flatten, index)
a = paddle.reshape(a, (-1, ))
b_flatten = paddle.reshape(b, (1, -1))
b = paddle.index_sample(b_flatten, index)
b = paddle.reshape(b, (-1, ))
miou = []
for i in range(n_class):
inter = paddle.logical_and(a == i, b == i)
inter = paddle.cast(inter, dtype='float32')
union = paddle.logical_or(a == i, b == i)
union = paddle.cast(union, dtype='float32')
miou.append(paddle.sum(inter) / (paddle.sum(union) + EPS))
miou = sum(miou) / len(miou)
return miou
def gen_acrostic(model, start_words, prefix_words=None):
result = []
start_words_len = len(start_words)
input = paddle.to_tensor([word2ix['<START>']])
input = paddle.reshape(input, [1, 1])
# 指示已经生成了几句藏头诗
index = 0
pre_word = '<START>'
hidden = None
# 存在风格前缀，则生成hidden
if prefix_words:
for word in prefix_words:
output, hidden = model(input, hidden)
input = paddle.to_tensor([word2ix[word]])
input = paddle.reshape(input, [1, 1])
# 开始生成诗句
for i in range(Config.max_gen_len):
output, hidden = model(input, hidden)
_, top_index = paddle.fluid.layers.topk(output[0], k=1)
top_index = top_index.numpy()[0]
w = ix2word[top_index]
# 说明上个字是句末
if pre_word in {'。', '，', '?', '！', '<START>'}:
if index == start_words_len:
break
else:
w = start_words[index]
index += 1
# print(w,word2ix[w])
input = paddle.to_tensor([word2ix[w]])
input = paddle.reshape(input, [1, 1])
else:
input = paddle.to_tensor([top_index])
input = paddle.reshape(input, [1, 1])
result.append(w)
pre_word = w
result = ''.join(result)
return result
def delta2rbox(self, rrois, deltas, means, stds, wh_ratio_clip=1e-6):
"""
:param rrois: (cx, cy, w, h, theta)
:param deltas: (dx, dy, dw, dh, dtheta)
:param means: means of anchor
:param stds: stds of anchor
:param wh_ratio_clip: clip threshold of wh_ratio
:return:
"""
deltas = paddle.reshape(deltas, [-1, 5])
rrois = paddle.reshape(rrois, [-1, 5])
pd_means = paddle.ones(shape=[5]) * means
pd_stds = paddle.ones(shape=[5]) * stds
denorm_deltas = deltas * pd_stds + pd_means
dx = denorm_deltas[:, 0]
dy = denorm_deltas[:, 1]
dw = denorm_deltas[:, 2]
dh = denorm_deltas[:, 3]
dangle = denorm_deltas[:, 4]
max_ratio = np.abs(np.log(wh_ratio_clip))
dw = paddle.clip(dw, min=-max_ratio, max=max_ratio)
dh = paddle.clip(dh, min=-max_ratio, max=max_ratio)
rroi_x = rrois[:, 0]
rroi_y = rrois[:, 1]
rroi_w = rrois[:, 2]
rroi_h = rrois[:, 3]
rroi_angle = rrois[:, 4]
gx = dx * rroi_w * paddle.cos(rroi_angle) - dy * rroi_h * paddle.sin(
rroi_angle) + rroi_x
gy = dx * rroi_w * paddle.sin(rroi_angle) + dy * rroi_h * paddle.cos(
rroi_angle) + rroi_y
gw = rroi_w * dw.exp()
gh = rroi_h * dh.exp()
ga = np.pi * dangle + rroi_angle
ga = (ga + np.pi / 4) % np.pi - np.pi / 4
bboxes = paddle.stack([gx, gy, gw, gh, ga], axis=-1)
return bboxes
def forward(self, generator_prediction_scores,
discriminator_prediction_scores, generator_labels,
discriminator_labels, attention_mask):
# generator loss
gen_loss = self.gen_loss_fct(
paddle.reshape(generator_prediction_scores, [-1, self.vocab_size]),
paddle.reshape(generator_labels, [-1]))
# todo: we can remove 4 lines after when CrossEntropyLoss(reduction='mean') improved
umask_positions = paddle.zeros_like(generator_labels).astype(
paddle.get_default_dtype())
mask_positions = paddle.ones_like(generator_labels).astype(
paddle.get_default_dtype())
mask_positions = paddle.where(generator_labels == -100,
umask_positions, mask_positions)
if mask_positions.sum() == 0:
gen_loss = paddle.to_tensor([0.0])
else:
gen_loss = gen_loss.sum() / mask_positions.sum()
# discriminator loss
seq_length = discriminator_labels.shape[1]
disc_loss = self.disc_loss_fct(
paddle.reshape(discriminator_prediction_scores, [-1, seq_length]),
discriminator_labels.astype(paddle.get_default_dtype()))
if attention_mask is not None:
umask_positions = paddle.ones_like(discriminator_labels).astype(
paddle.get_default_dtype())
mask_positions = paddle.zeros_like(discriminator_labels).astype(
paddle.get_default_dtype())
use_disc_loss = paddle.where(attention_mask, disc_loss,
mask_positions)
umask_positions = paddle.where(attention_mask, umask_positions,
mask_positions)
disc_loss = use_disc_loss.sum() / umask_positions.sum()
else:
total_positions = paddle.ones_like(discriminator_labels).astype(
paddle.get_default_dtype())
disc_loss = disc_loss.sum() / total_positions.sum()
return self.gen_weight * gen_loss + self.disc_weight * disc_loss
def forward(self, inputs, labels=None):
"""
inputs: [btc, time_steps]
labels: [btc, time_steps]
"""
batch_size, seq_length = inputs.shape[0], inputs.shape[1]
init_hidden_data = np.zeros(
(self.num_layers, batch_size, self.hidden_size), dtype='float32')
init_hidden = paddle.to_tensor(data=init_hidden_data,
dtype=None,
place=None,
stop_gradient=True)
x_emb = self.embedding(inputs)
# 添加dropout 随机丢失
if self.dropout_layer is not None:
x_emb = self.dropout_layer(x_emb)
rnn_out, last_hidden = self.rnn(x_emb, init_hidden)
output = self.classifier(rnn_out)
if labels is not None:
logits = paddle.reshape(output, shape=[-1, self.vocab_size])
labels = paddle.reshape(labels, shape=[-1, 1])
loss = nn.functional.softmax_with_cross_entropy(logits=logits,
label=labels,
soft_label=False,
ignore_index=1)
# 计算recall@20 指标
acc = paddle.metric.accuracy(input=logits, label=labels, k=20)
loss = paddle.reshape(loss, shape=[-1, seq_length])
# 综合所有batch和序列长度的loss
loss = paddle.mean(loss, axis=[0])
loss = paddle.sum(loss)
# loss = paddle.mean(loss)
return loss, acc
return output
def forward(self, inputs, gsrm_word_pos, gsrm_slf_attn_bias1,
gsrm_slf_attn_bias2):
# ===== GSRM Visual-to-semantic embedding block =====
b, t, c = inputs.shape
pvam_features = paddle.reshape(inputs, [-1, c])
word_out = self.fc0(pvam_features)
word_ids = paddle.argmax(F.softmax(word_out), axis=1)
word_ids = paddle.reshape(x=word_ids, shape=[-1, t, 1])
#===== GSRM Semantic reasoning block =====
"""
This module is achieved through bi-transformers,
ngram_feature1 is the froward one, ngram_fetaure2 is the backward one
"""
pad_idx = self.char_num
word1 = paddle.cast(word_ids, "float32")
word1 = F.pad(word1, [1, 0], value=1.0 * pad_idx, data_format="NLC")
word1 = paddle.cast(word1, "int64")
word1 = word1[:, :-1, :]
word2 = word_ids
enc_inputs_1 = [word1, gsrm_word_pos, gsrm_slf_attn_bias1]
enc_inputs_2 = [word2, gsrm_word_pos, gsrm_slf_attn_bias2]
gsrm_feature1 = self.wrap_encoder0(enc_inputs_1)
gsrm_feature2 = self.wrap_encoder1(enc_inputs_2)
gsrm_feature2 = F.pad(gsrm_feature2, [0, 1],
value=0.,
data_format="NLC")
gsrm_feature2 = gsrm_feature2[:, 1:, ]
gsrm_features = gsrm_feature1 + gsrm_feature2
gsrm_out = self.mul(gsrm_features)
b, t, c = gsrm_out.shape
gsrm_out = paddle.reshape(gsrm_out, [-1, c])
return gsrm_features, word_out, gsrm_out
def _postprocessing_by_level(self, locations, box_cls, box_reg, box_ctn,
scale_factor):
"""
Args:
locations (Variables): anchor points for current layer, [H*W, 2]
box_cls (Variables): categories prediction, [N, C, H, W], C is the number of classes
box_reg (Variables): bounding box prediction, [N, 4, H, W]
box_ctn (Variables): centerness prediction, [N, 1, H, W]
scale_factor (Variables): [h_scale, w_scale] for input images
Return:
box_cls_ch_last (Variables): score for each category, in [N, C, M]
C is the number of classes and M is the number of anchor points
box_reg_decoding (Variables): decoded bounding box, in [N, M, 4]
last dimension is [x1, y1, x2, y2]
"""
act_shape_cls = self._merge_hw(box_cls)
box_cls_ch_last = paddle.reshape(x=box_cls, shape=act_shape_cls)
box_cls_ch_last = F.sigmoid(box_cls_ch_last)
act_shape_reg = self._merge_hw(box_reg)
box_reg_ch_last = paddle.reshape(x=box_reg, shape=act_shape_reg)
box_reg_ch_last = paddle.transpose(box_reg_ch_last, perm=[0, 2, 1])
box_reg_decoding = paddle.stack([
locations[:, 0] - box_reg_ch_last[:, :, 0],
locations[:, 1] - box_reg_ch_last[:, :, 1],
locations[:, 0] + box_reg_ch_last[:, :, 2],
locations[:, 1] + box_reg_ch_last[:, :, 3]
],
axis=1)
box_reg_decoding = paddle.transpose(box_reg_decoding, perm=[0, 2, 1])
act_shape_ctn = self._merge_hw(box_ctn)
box_ctn_ch_last = paddle.reshape(x=box_ctn, shape=act_shape_ctn)
box_ctn_ch_last = F.sigmoid(box_ctn_ch_last)
# recover the location to original image
im_scale = paddle.concat([scale_factor, scale_factor], axis=1)
box_reg_decoding = box_reg_decoding / im_scale
box_cls_ch_last = box_cls_ch_last * box_ctn_ch_last
return box_cls_ch_last, box_reg_decoding
def apply_rotary_position_embeddings(sinusoidal_pos,
query_layer,
key_layer,
value_layer=None):
# https://kexue.fm/archives/8265
# sin [batch_size, num_heads, sequence_length, embed_size_per_head//2]
# cos [batch_size, num_heads, sequence_length, embed_size_per_head//2]
sin, cos = paddle.chunk(sinusoidal_pos, 2, axis=-1)
# sin [θ0,θ1,θ2......θd/2-1] -> sin_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
sin_pos = paddle.reshape(paddle.stack([sin, sin], axis=-1),
sinusoidal_pos.shape)
# cos [θ0,θ1,θ2......θd/2-1] -> cos_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
cos_pos = paddle.reshape(paddle.stack([cos, cos], axis=-1),
sinusoidal_pos.shape)
# rotate_half_query_layer [-q1,q0,-q3,q2......,-qd-1,qd-2]
rotate_half_query_layer = paddle.reshape(
paddle.stack(
[-query_layer[:, :, :, 1::2], query_layer[:, :, :, 0::2]],
axis=-1),
query_layer.shape,
)
query_layer = query_layer * cos_pos + rotate_half_query_layer * sin_pos
# rotate_half_key_layer [-k1,k0,-k3,k2......,-kd-1,kd-2]
rotate_half_key_layer = paddle.reshape(
paddle.stack([-key_layer[:, :, :, 1::2], key_layer[:, :, :, 0::2]],
axis=-1),
key_layer.shape,
)
key_layer = key_layer * cos_pos + rotate_half_key_layer * sin_pos
if value_layer is not None:
# rotate_half_value_layer [-v1,v0,-v3,v2......,-vd-1,vd-2]
rotate_half_value_layer = paddle.reshape(
paddle.stack(
[-value_layer[:, :, :, 1::2], value_layer[:, :, :, 0::2]],
axis=-1),
value_layer.shape,
)
value_layer = value_layer * cos_pos + rotate_half_value_layer * sin_pos
return query_layer, key_layer, value_layer
return query_layer, key_layer
def _weight_norm(v, g, dim):
shape = v.shape
ndims = len(shape)
if dim == -1:
v_normalized = v / (paddle.sqrt(paddle.sum(paddle.square(v))) + 1e-12)
elif dim == 0:
p_matrix = paddle.reshape(v, (shape[0], -1))
v_normalized = F.l2_normalize(p_matrix, axis=1)
v_normalized = paddle.reshape(v_normalized, shape)
elif dim == ndims - 1:
p_matrix = paddle.reshape(v, (-1, shape[-1]))
v_normalized = F.l2_normalize(p_matrix, axis=0)
v_normalized = paddle.reshape(v_normalized, shape)
else:
perm = list(range(ndims))
perm[0] = dim
perm[dim] = 0
p_transposed = paddle.transpose(v, perm)
transposed_shape = p_transposed.shape
p_matrix = paddle.reshape(p_transposed, (p_transposed.shape[0], -1))
v_normalized = F.l2_normalize(p_matrix, axis=1)
v_normalized = paddle.reshape(v_normalized, transposed_shape)
v_normalized = paddle.transpose(v_normalized, perm)
weight = F.elementwise_mul(v_normalized,
g,
axis=dim if dim is not None else -1)
return weight
def get_feature_by_coordinate(self, x, coord, offset_h, offset_w,
padded_x_w):
x = paddle.reshape(x, [0, 0, -1])
index = paddle.cast(
coord[:, :, :, :self.N] * padded_x_w,
dtype='int64') + coord[:, :, :, self.N:] # offset_x*w + offset_y
index = paddle.unsqueeze(index, 1)
index = paddle.tile(index, [1, self.in_channel, 1, 1, 1])
index = paddle.reshape(index, (0, 0, -1))
x_range = list(range(3))
dim = 2
x_range[0] = dim
x_range[dim] = 0
x_swaped = paddle.transpose(x, perm=x_range)
index_range = list(range(3))
index_range[0] = dim
index_range[dim] = 0
index_swaped = paddle.transpose(index, perm=index_range)
x_shape = layers.shape(x_swaped)
index_shape = layers.shape(index_swaped)
prod = paddle.prod(x_shape[1:], keepdim=True)
x_swaped_flattend = paddle.reshape(x_swaped, [-1])
index_swaped_flattend = paddle.reshape(index_swaped, [-1])
index_swaped_flattend *= prod
bias = paddle.arange(start=0, end=prod, step=1, dtype='float32')
bias = paddle.tile(bias, index_shape[0])
index_swaped_flattend += bias
gathered = paddle.gather(x_swaped_flattend, index_swaped_flattend)
gathered = paddle.reshape(gathered, layers.shape(index_swaped))
x_offset = paddle.transpose(gathered, perm=x_range)
x_offset = paddle.reshape(
x_offset, (-1, self.in_channel, offset_h, offset_w, self.N))
return x_offset
def forward(self, x):
out = self.conv1(x)
rp = F.adaptive_max_pool2d(out, (self.s, 1))
cp = F.adaptive_max_pool2d(out, (1, self.s))
p = paddle.reshape(self.conv_p(rp), (x.shape[0], self.k, self.s, self.s))
q = paddle.reshape(self.conv_q(cp), (x.shape[0], self.k, self.s, self.s))
p = F.sigmoid(p)
q = F.sigmoid(q)
p = p / paddle.sum(p, axis=3, keepdim=True)
q = q / paddle.sum(q, axis=2, keepdim=True)
p = paddle.reshape(p, (x.shape[0], self.k, 1, self.s, self.s))
p = paddle.expand(p, (x.shape[0], self.k, x.shape[1] // self.k, self.s, self.s))
p = paddle.reshape(p, (x.shape[0], x.shape[1], self.s, self.s))
q = paddle.reshape(q, (x.shape[0], self.k, 1, self.s, self.s))
q = paddle.expand(q, (x.shape[0], self.k, x.shape[1] // self.k, self.s, self.s))
q = paddle.reshape(q, (x.shape[0], x.shape[1], self.s, self.s))
p = self.resize_mat(p, x.shape[2] // self.s)
q = self.resize_mat(q, x.shape[2] // self.s)
y = paddle.matmul(p, x)
y = paddle.matmul(y, q)
y = self.conv2(y)
return y
def forward(self, keys, querys, mask):
"""Calculate forward propagation of tacotron2 decoder.
Parameters
----------
keys: Tensor[shape=(B, T_key, C)]
Batch of the sequences of padded output from encoder.
querys: Tensor[shape(B, T_query, C)]
Batch of the sequences of padded mel spectrogram.
mask: Tensor
Mask generated with text length. Shape should be (B, T_key, T_query) or broadcastable shape.
Returns
-------
mel_output: Tensor [shape=(B, T_query, C)]
Output sequence of features.
stop_logits: Tensor [shape=(B, T_query)]
Output sequence of stop logits.
alignments: Tensor [shape=(B, T_query, T_key)]
Attention weights.
"""
querys = paddle.reshape(
querys,
[querys.shape[0], querys.shape[1] // self.reduction_factor, -1])
querys = paddle.concat([
paddle.zeros(shape=[querys.shape[0], 1, querys.shape[-1]],
dtype=querys.dtype), querys
],
axis=1)
querys = self.prenet(querys)
self._initialize_decoder_states(keys)
self.mask = mask
mel_outputs, stop_logits, alignments = [], [], []
while len(mel_outputs
) < querys.shape[1] - 1: # Ignore the last time step
query = querys[:, len(mel_outputs), :]
mel_output, stop_logit, attention_weights = self._decode(query)
mel_outputs += [mel_output]
stop_logits += [stop_logit]
alignments += [attention_weights]
alignments = paddle.stack(alignments, axis=1)
stop_logits = paddle.concat(stop_logits, axis=1)
mel_outputs = paddle.stack(mel_outputs, axis=1)
return mel_outputs, stop_logits, alignments
def _network(self, hidden, cell, init_actions=None, is_inference=False):
actions = []
entropies = []
sample_log_probs = []
with fluid.unique_name.guard('Controller'):
self._create_parameter()
inputs = self.g_emb
for idx in range(len(self.range_tables)):
logits, output, states = self._lstm(inputs,
hidden,
cell,
token_idx=idx)
hidden, cell = np.squeeze(states)
probs = paddle.nn.functional.softmax(logits, axis=1)
if is_inference:
action = paddle.argmax(probs, axis=1)
else:
if init_actions:
action = paddle.slice(init_actions,
axes=[1],
starts=[idx],
ends=[idx + 1])
action = paddle.squeeze(action, axis=[1])
action.stop_gradient = True
else:
action = fluid.layers.sampling_id(probs)
actions.append(action)
log_prob = paddle.nn.functional.softmax_with_cross_entropy(
logits,
paddle.reshape(action, shape=[paddle.shape(action), 1]),
axis=1)
sample_log_probs.append(log_prob)
entropy = log_prob * paddle.exp(-1 * log_prob)
entropy.stop_gradient = True
entropies.append(entropy)
action_emb = paddle.cast(action, dtype=np.int64)
inputs = paddle.static.nn.embedding(
action_emb,
size=(self.max_range_table, self.hidden_size),
param_attr=paddle.ParamAttr(
name='emb_w', initializer=uniform_initializer(1.0)))
self.sample_log_probs = paddle.concat(sample_log_probs, axis=0)
entropies = paddle.stack(entropies)
self.sample_entropies = paddle.sum(entropies)
return actions
def forward(self, x):
n, c, h, w = x.shape
# query: n, c, h * w
query = paddle.reshape(x, (n, c, h * w))
# key: n, h * w, c
key = paddle.reshape(x, (n, c, h * w))
key = paddle.transpose(key, (0, 2, 1))
# sim: n, c, c
sim = paddle.bmm(query, key)
# The danet author claims that this can avoid gradient divergence
sim = paddle.max(sim, axis=-1, keepdim=True).expand_as(sim) - sim
sim = F.softmax(sim, axis=-1)
# feat: from (n, c, h * w) to (n, c, h, w)
value = paddle.reshape(x, (n, c, h * w))
feat = paddle.bmm(sim, value)
feat = paddle.reshape(feat, (n, c, h, w))
out = self.gamma * feat + x
return out
def forward(self, n2n_g, n_feats, edge_feats, edge_feat_dist):
feature = []
for att_l in self.att_layers:
feature.append(att_l(n2n_g, n_feats, edge_feats, edge_feat_dist))
feature = paddle.stack(feature, axis=1)
if self.merge == "cat":
feature = paddle.reshape(feature, [-1, self.num_heads * self.hidden_dim])
if self.merge == "mean":
feature = paddle.mean(feature, axis=1)
return feature
def rzz_gate_matrix(params):
"""
RZZ gate
:return:
"""
theta = params
re_a = paddle.cos(theta / 2)
re_b = paddle.zeros([1], 'float64')
im_a = paddle.sin(theta / 2)
im_b = paddle.zeros([1], 'float64')
re = paddle.reshape(
paddle.concat([
re_a, re_b, re_b, re_b, re_b, re_a, re_b, re_b, re_b, re_b, re_a,
re_b, re_b, re_b, re_b, re_a
]), [4, 4])
im = paddle.reshape(
paddle.concat([
-im_a, im_b, im_b, im_b, im_b, im_a, im_b, im_b, im_b, im_b, im_a,
im_b, im_b, im_b, im_b, -im_a
]), [4, 4])
return re + im * paddle.to_tensor([1j], 'complex128')
def forward(self, words, wp):
word_embed = self.word_embed(words)
mask = words != self.pad_index
seq_lens = paddle.sum(paddle.cast(mask, "int32"), axis=-1)
x, _ = self.lstm(word_embed, sequence_length=seq_lens)
x = paddle.reshape(
index_sample(x, wp),
shape=[wp.shape[0], wp.shape[1], x.shape[2]],
)
words = paddle.index_sample(words, wp)
return words, x
def forward(self, input, seq_lens):
embedded = self.embedding(input)
self.embedded = embedded
output, hidden = self.lstm(embedded,
sequence_length=paddle.to_tensor(
seq_lens, dtype='int32'))
encoder_feature = paddle.reshape(
output, [-1, 2 * config.hidden_dim]) # B * t_k x 2*hidden_dim
encoder_feature = self.W_h(encoder_feature)
return output, encoder_feature, hidden
def test_shape_omit_dims(self):
for dtype in self._dtypes:
x_np = np.random.randn(
2, 3, 4).astype(dtype) + 1j * np.random.randn(2, 3,
4).astype(dtype)
shape = (0, -1)
shape_ = (2, 12)
for place in self._places:
with dg.guard(place):
x_var = dg.to_variable(x_np)
y_var = paddle.reshape(x_var, shape)
y_np = y_var.numpy()
self.assertTrue(np.allclose(np.reshape(x_np, shape_), y_np))
def valid_flags(self, featmap_size, valid_size):
feat_h, feat_w = featmap_size
valid_h, valid_w = valid_size
assert valid_h <= feat_h and valid_w <= feat_w
valid_x = paddle.zeros([feat_w], dtype='int32')
valid_y = paddle.zeros([feat_h], dtype='int32')
valid_x[:valid_w] = 1
valid_y[:valid_h] = 1
valid_xx, valid_yy = self._meshgrid(valid_x, valid_y)
valid = valid_xx & valid_yy
valid = paddle.reshape(valid, [-1, 1])
valid = paddle.expand(valid, [-1, self.num_base_anchors]).reshape([-1])
return valid
def forward(self, inputs):
left_emb = self.emb(inputs[0])
right_emb = self.emb(inputs[1])
cross = paddle.matmul(left_emb, right_emb, transpose_y=True)
cross = paddle.reshape(cross, [-1, 1, cross.shape[1], cross.shape[2]])
conv = self.conv(cross)
if self.conv_act == "relu":
conv = F.relu(conv)
if self.pool_type == "max":
pool = F.max_pool2d(conv,
kernel_size=self.pool_size,
stride=self.pool_stride,
padding=self.pool_padding)
reshape = paddle.reshape(pool, [
-1,
list(pool.shape)[1] * list(pool.shape)[2] * list(pool.shape)[3]
])
hid = self.fc1(reshape)
if self.hidden_act == "relu":
relu_hid = F.relu(hid)
prediction = self.fc2(relu_hid)
return prediction
def __split_heads_qkv(queries, keys, values, n_head, d_key, d_value):
"""
Reshape input tensors at the last dimension to split multi-heads
and then transpose. Specifically, transform the input tensor with shape
[bs, max_sequence_length, n_head * hidden_dim] to the output tensor
with shape [bs, n_head, max_sequence_length, hidden_dim].
"""
# The value 0 in shape attr means copying the corresponding dimension
# size of the input as the output dimension size.
reshaped_q = paddle.reshape(x=queries, shape=[0, 0, n_head, d_key])
# permuate the dimensions into:
# [batch_size, n_head, max_sequence_len, hidden_size_per_head]
q = paddle.transpose(x=reshaped_q, perm=[0, 2, 1, 3])
# For encoder-decoder attention in inference, insert the ops and vars
# into global block to use as cache among beam search.
reshaped_k = paddle.reshape(x=keys, shape=[0, 0, n_head, d_key])
k = paddle.transpose(x=reshaped_k, perm=[0, 2, 1, 3])
reshaped_v = paddle.reshape(x=values,
shape=[0, 0, n_head, d_value])
v = paddle.transpose(x=reshaped_v, perm=[0, 2, 1, 3])
return q, k, v
def test(model):
model.eval()
avg_acc = [[], []]
for batch_id, data in enumerate(test_reader):
img = paddle.to_tensor(data[0])
img = paddle.reshape(img, [-1, 1, 28, 28])
label = paddle.to_tensor(data[1])
label = paddle.reshape(label, [-1, 1])
out = model(img)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
avg_acc[0].append(acc_top1.numpy())
avg_acc[1].append(acc_top5.numpy())
if batch_id % 100 == 0:
_logger.info(
"Test | step {}: acc1 = {:}, acc5 = {:}".format(
batch_id, acc_top1.numpy(), acc_top5.numpy()))
_logger.info("Test | Average: acc_top1 {}, acc_top5 {}".format(
np.mean(avg_acc[0]), np.mean(avg_acc[1])))
return np.mean(avg_acc[0]), np.mean(avg_acc[1])
def forward(self, hidden_states, masked_positions=None):
if masked_positions is not None:
hidden_states = paddle.reshape(hidden_states,
[-1, hidden_states.shape[-1]])
hidden_states = paddle.tensor.gather(hidden_states,
masked_positions)
hidden_states = self.transform(hidden_states)
hidden_states = self.activation(hidden_states)
hidden_states = self.layer_norm(hidden_states)
logits = paddle.tensor.matmul(hidden_states,
self.decoder_weight,
transpose_y=True) + self.decoder_bias
return logits
def __combine_heads(x):
"""
Transpose and then reshape the last two dimensions of inpunt tensor x
so that it becomes one dimension, which is reverse to __split_heads.
"""
if len(x.shape) != 4:
raise ValueError("Input(x) should be a 4-D Tensor.")
trans_x = paddle.transpose(x, perm=[0, 2, 1, 3])
# The value 0 in shape attr means copying the corresponding dimension
# size of the input as the output dimension size.
return paddle.reshape(
x=trans_x, shape=[0, 0, trans_x.shape[2] * trans_x.shape[3]])
def forward(self, inputs):
with paddle.static.amp.fp16_guard():
if self.data_format == "NHWC":
inputs = paddle.tensor.transpose(inputs, [0, 2, 3, 1])
inputs.stop_gradient = True
y = self.conv(inputs)
y = self.pool2d_max(y)
for block in self.block_list:
y = block(y)
y = self.pool2d_avg(y)
y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y)
return y
def test():
model.eval()
accuracies = []
losses = []
for (x, y) in val_loader:
with paddle.no_grad():
logits = model(x)
y = paddle.reshape(y, (-1, 1))
loss = loss_fn(logits, y)
acc = acc_fn(logits, y)
accuracies.append(np.mean(acc.numpy()))
losses.append(np.mean(loss.numpy()))
return np.mean(accuracies) * 100, np.mean(losses)
def forward(self, inputs):
y = self.conv1_1(inputs)
y = self.conv1_2(y)
y = self.conv1_3(y)
y = self.pool2d_max(y)
att = None
y = {0: y, 1: att}
for block in self.block_list:
y = block(y)
y = self.pool2d_avg(y[0])
y = paddle.reshape(y, shape=[-1, self.pool2d_avg_channels])
y = self.out(y)
return y
def forward(self, x):
x = self.inception_stem(x)
for inception_block in self.inception_block_list:
x = inception_block(x)
if self.with_pool:
x = self.avg_pool(x)
if self.num_classes > 0:
x = paddle.reshape(x, shape=[-1, 2048])
x = self.dropout(x)
x = self.fc(x)
return x
