def forward(self, x):
"""Inputs of forward function
Args:
x: the sequence fed to the positional encoder model (required).
Shape:
x: [sequence length, batch size, embed dim]
output: [sequence length, batch size, embed dim]
Examples:
>>> output = pos_encoder(x)
"""
w_pe = self.pe[:paddle.shape(x)[-1], :]
w1 = self.linear1(self.avg_pool_1(x).squeeze()).unsqueeze(0)
w_pe = w_pe * w1
w_pe = paddle.transpose(w_pe, [1, 2, 0])
w_pe = paddle.unsqueeze(w_pe, 2)
h_pe = self.pe[:paddle.shape(x).shape[-2], :]
w2 = self.linear2(self.avg_pool_2(x).squeeze()).unsqueeze(0)
h_pe = h_pe * w2
h_pe = paddle.transpose(h_pe, [1, 2, 0])
h_pe = paddle.unsqueeze(h_pe, 3)
x = x + w_pe + h_pe
x = paddle.transpose(
paddle.reshape(x,
[x.shape[0], x.shape[1], x.shape[2] * x.shape[3]]),
[2, 0, 1])
return self.dropout(x)
def scatter_paddle(self, refined_seg_logits, point_indices, point_logits):
"""
paddle version scatter : equal to pytorch version scatter(-1,point_indices,point_logits).
Args:
refined_seg_logits(Tensor): shape=[batch_size, channels, height * width]
point_indices(Tensor): shape=[batch_size, channels, height * width]
point_logits(Tensor): shape[batch_size, channels, height * width]
Returns:
scattered refined_seg_logits(Tensor).
"""
original_shape = paddle.shape(
refined_seg_logits) # [batch_size, channels, height * width]
new_refined_seg_logits = refined_seg_logits.flatten(0, 1) # [N*C,H*W]
offsets = (paddle.arange(paddle.shape(new_refined_seg_logits)[0]) *
paddle.shape(new_refined_seg_logits)[1]).unsqueeze(
-1) # [N*C,1]
point_indices = point_indices.flatten(0, 1) # [N*C,H*W]
new_point_indices = (point_indices + offsets).flatten()
point_logits = point_logits.flatten() # [N*C*H*W]
refined_seg_logits = paddle.scatter(refined_seg_logits.flatten(),
new_point_indices,
point_logits,
overwrite=True)
return refined_seg_logits.reshape(shape=original_shape)
def paste_mask(self, masks, boxes, im_h, im_w):
"""
Paste the mask prediction to the original image.
"""
x0_int, y0_int = 0, 0
x1_int, y1_int = im_w, im_h
x0, y0, x1, y1 = paddle.split(boxes, 4, axis=1)
N = masks.shape[0]
img_y = paddle.arange(y0_int, y1_int) + 0.5
img_x = paddle.arange(x0_int, x1_int) + 0.5
img_y = (img_y - y0) / (y1 - y0) * 2 - 1
img_x = (img_x - x0) / (x1 - x0) * 2 - 1
# img_x, img_y have shapes (N, w), (N, h)
if self.assign_on_cpu:
paddle.set_device('cpu')
gx = img_x[:, None, :].expand(
[N, paddle.shape(img_y)[1],
paddle.shape(img_x)[1]])
gy = img_y[:, :, None].expand(
[N, paddle.shape(img_y)[1],
paddle.shape(img_x)[1]])
grid = paddle.stack([gx, gy], axis=3)
img_masks = F.grid_sample(masks, grid, align_corners=False)
return img_masks[:, 0]
def forward(self, x):
feats = self.backbone(x)
feats = [feats[i] for i in self.backbone_indices]
fpn_feats = self.neck(feats) # [n,256,64,128]*3 & [n,256,128,256]
pfn_logits = self.fpnhead(
fpn_feats
) # segmainoutput decode_head[0] 512*1024->[n, 19, 64, 128]
point_logits = self.pointhead(
fpn_feats, pfn_logits) # segpointoutput decode_head[1]
if self.training:
logit_list = [
F.interpolate(logit,
paddle.shape(x)[2:],
mode='bilinear',
align_corners=self.align_corners)
for logit in pfn_logits
]
logit_list.append(point_logits)
else:
logit_list = [
F.interpolate(logit,
paddle.shape(x)[2:],
mode='bilinear',
align_corners=self.align_corners)
for logit in point_logits
]
return logit_list
def forward(self, input, mask=None):
"""
Args:
input (obj: `paddle.Tensor`) of shape (batch, seq_len, input_size): Tensor containing the features of the input sequence.
mask (obj: `paddle.Tensor`, optional, defaults to `None`) of shape (batch, seq_len) :
Tensor is a bool tensor, whose each element identifies whether the input word id is pad token or not.
"""
weight = self.input_weight.tile(
repeat_times=(paddle.shape(input)[0], 1, 1))
bias = self.bias.tile(repeat_times=(paddle.shape(input)[0], 1, 1))
# Shape: (batch_size, max_seq_len, hidden_size)
word_squish = paddle.bmm(input, weight) + bias
att_context_vector = self.att_context_vector.tile(
repeat_times=(paddle.shape(input)[0], 1, 1))
# Shape: (batch_size, max_seq_len, 1)
att_score = paddle.bmm(word_squish, att_context_vector)
if mask is not None:
# mask, remove the effect of 'PAD'
mask = paddle.cast(mask, dtype='float32')
mask = mask.unsqueeze(axis=-1)
inf_tensor = paddle.full(
shape=paddle.shape(mask), dtype='float32', fill_value=-INF)
att_score = paddle.multiply(att_score, mask) + paddle.multiply(
inf_tensor, (1 - mask))
att_weight = F.softmax(att_score, axis=1)
# Shape: (batch_size, hidden_size)
reps = paddle.bmm(input.transpose(perm=(0, 2, 1)),
att_weight).squeeze(-1)
return reps, att_weight
def get_target_tensor(self, prediction, target_is_real):
"""Create label tensors with the same size as the input.
Parameters:
prediction (tensor) - - tpyically the prediction from a discriminator
target_is_real (bool) - - if the ground truth label is for real images or fake images
Returns:
A label tensor filled with ground truth label, and with the size of the input
"""
if target_is_real:
if not hasattr(self, 'target_real_tensor'):
self.target_real_tensor = paddle.full(
shape=paddle.shape(prediction),
fill_value=self.target_real_label,
dtype='float32')
target_tensor = self.target_real_tensor
else:
if not hasattr(self, 'target_fake_tensor'):
self.target_fake_tensor = paddle.full(
shape=paddle.shape(prediction),
fill_value=self.target_fake_label,
dtype='float32')
target_tensor = self.target_fake_tensor
# target_tensor.stop_gradient = True
return target_tensor
def forward(self, inputs):
"""
Get SOLOv2MaskHead output.
Args:
inputs(list[Tensor]): feature map from each necks with shape of [N, C, H, W]
Returns:
ins_pred(Tensor): Output of SOLOv2MaskHead head
"""
feat_all_level = F.relu(self.convs_all_levels[0](inputs[0]))
for i in range(1, self.range_level):
input_p = inputs[i]
if i == (self.range_level - 1):
input_feat = input_p
x_range = paddle.linspace(
-1, 1, paddle.shape(input_feat)[-1], dtype='float32')
y_range = paddle.linspace(
-1, 1, paddle.shape(input_feat)[-2], dtype='float32')
y, x = paddle.meshgrid([y_range, x_range])
x = paddle.unsqueeze(x, [0, 1])
y = paddle.unsqueeze(y, [0, 1])
y = paddle.expand(
y, shape=[paddle.shape(input_feat)[0], 1, -1, -1])
x = paddle.expand(
x, shape=[paddle.shape(input_feat)[0], 1, -1, -1])
coord_feat = paddle.concat([x, y], axis=1)
input_p = paddle.concat([input_p, coord_feat], axis=1)
feat_all_level = paddle.add(feat_all_level,
self.convs_all_levels[i](input_p))
ins_pred = F.relu(self.conv_pred(feat_all_level))
return ins_pred
def forward(self, x):
# Encoder
input_shape = paddle.shape(x)[2:]
x = self.conv_bn0(x) # 1/2
shortcut = self.conv_bn1(x) # shortcut
x = F.max_pool2d(x, kernel_size=3, stride=2, padding=1) # 1/4
x = self.block1(x) # 1/8
x = self.block2(x) # 1/16
# Decoder
x = self.depthwise_separable0(x)
shortcut_shape = paddle.shape(shortcut)[2:]
x = F.interpolate(
x,
shortcut_shape,
mode='bilinear',
align_corners=self.align_corners)
x = paddle.concat(x=[shortcut, x], axis=1)
x = self.depthwise_separable1(x)
logit = self.depthwise_separable2(x)
logit = F.interpolate(
logit,
input_shape,
mode='bilinear',
align_corners=self.align_corners)
return [logit]
def forward_test(self, src):
bs = paddle.shape(src)[0]
if self.encoder is not None:
src = self.positional_encoding(paddle.transpose(src, [1, 0, 2]))
memory = self.encoder(src)
else:
memory = paddle.transpose(paddle.squeeze(src, 2), [2, 0, 1])
dec_seq = paddle.full((bs, 1), 2, dtype=paddle.int64)
dec_prob = paddle.full((bs, 1), 1., dtype=paddle.float32)
for len_dec_seq in range(1, 25):
dec_seq_embed = paddle.transpose(self.embedding(dec_seq),
[1, 0, 2])
dec_seq_embed = self.positional_encoding(dec_seq_embed)
tgt_mask = self.generate_square_subsequent_mask(
paddle.shape(dec_seq_embed)[0])
output = self.decoder(dec_seq_embed,
memory,
tgt_mask=tgt_mask,
memory_mask=None,
tgt_key_padding_mask=None,
memory_key_padding_mask=None)
dec_output = paddle.transpose(output, [1, 0, 2])
dec_output = dec_output[:, -1, :]
word_prob = F.softmax(self.tgt_word_prj(dec_output), axis=1)
preds_idx = paddle.argmax(word_prob, axis=1)
if paddle.equal_all(
preds_idx,
paddle.full(paddle.shape(preds_idx), 3, dtype='int64')):
break
preds_prob = paddle.max(word_prob, axis=1)
dec_seq = paddle.concat(
[dec_seq, paddle.reshape(preds_idx, [-1, 1])], axis=1)
dec_prob = paddle.concat(
[dec_prob, paddle.reshape(preds_prob, [-1, 1])], axis=1)
return [dec_seq, dec_prob]
def forward(self, conv_out):
last_out = self.lateral_convs[-1](conv_out[-1])
f = last_out
fpn_feature_list = [last_out]
for i in reversed(range(len(conv_out) - 1)):
conv_x = conv_out[i]
conv_x = self.lateral_convs[i](conv_x)
prev_shape = paddle.shape(conv_x)[2:]
f = conv_x + F.interpolate(
f, prev_shape, mode='bilinear', align_corners=True)
fpn_feature_list.append(self.fpn_out[i](f))
output_size = paddle.shape(fpn_feature_list[-1])[2:]
x = self.scale_heads[0](fpn_feature_list[-1])
for index in range(len(self.scale_heads) - 2, 0, -1):
x = x + F.interpolate(self.scale_heads[index](
fpn_feature_list[index]),
size=output_size,
mode='bilinear',
align_corners=self.align_corners)
x = self.cls_seg(x)
if self.enable_auxiliary_loss:
dsn = self.dsn(conv_out[2])
return [x, dsn]
else:
return [x]
def _topk(self, scores):
k = self.max_per_img
shape_fm = paddle.shape(scores)
shape_fm.stop_gradient = True
cat, height, width = shape_fm[1], shape_fm[2], shape_fm[3]
# batch size is 1
scores_r = paddle.reshape(scores, [cat, -1])
topk_scores, topk_inds = paddle.topk(scores_r, k)
topk_scores, topk_inds = paddle.topk(scores_r, k)
topk_ys = topk_inds // width
topk_xs = topk_inds % width
topk_score_r = paddle.reshape(topk_scores, [-1])
topk_score, topk_ind = paddle.topk(topk_score_r, k)
k_t = paddle.full(paddle.shape(topk_ind), k, dtype='int64')
topk_clses = paddle.cast(paddle.floor_divide(topk_ind, k_t), 'float32')
topk_inds = paddle.reshape(topk_inds, [-1])
topk_ys = paddle.reshape(topk_ys, [-1, 1])
topk_xs = paddle.reshape(topk_xs, [-1, 1])
topk_inds = paddle.gather(topk_inds, topk_ind)
topk_ys = paddle.gather(topk_ys, topk_ind)
topk_xs = paddle.gather(topk_xs, topk_ind)
return topk_score, topk_inds, topk_clses, topk_ys, topk_xs
def __call__(self, x, index):
if self.dim < 0:
self.dim += len(x.shape)
x_range = list(range(len(x.shape)))
x_range[0] = self.dim
x_range[self.dim] = 0
x_swaped = paddle.transpose(x, perm=x_range)
index_range = list(range(len(index.shape)))
index_range[0] = self.dim
index_range[self.dim] = 0
index_swaped = paddle.transpose(index, perm=index_range)
dtype = index.dtype
x_shape = paddle.shape(x_swaped)
index_shape = paddle.shape(index_swaped)
prod = paddle.cast(paddle.prod(x_shape), dtype=dtype) / x_shape[0]
x_swaped_flattend = paddle.flatten(x_swaped)
index_swaped_flattend = paddle.flatten(index_swaped)
index_swaped_flattend *= prod
bias = paddle.arange(start=0, end=prod, dtype=dtype)
bias = paddle.reshape(bias, x_shape[1:])
bias = paddle.crop(bias, index_shape[1:])
bias = paddle.flatten(bias)
bias = paddle.tile(bias, [index_shape[0]])
index_swaped_flattend += bias
gathered = paddle.index_select(x_swaped_flattend, index_swaped_flattend)
gathered = paddle.reshape(gathered, index_swaped.shape)
out = paddle.transpose(gathered, perm=x_range)
return out
def forward(self,
input_ids,
position_ids=None,
attention_mask=None,
use_cache=False,
cache=None):
self.checkpoints = []
if position_ids is None:
past_length = 0
if cache is not None:
past_length = paddle.shape(cache[0].k)[-2]
position_ids = paddle.arange(past_length,
paddle.shape(input_ids)[-1] +
past_length,
dtype='int64')
position_ids = position_ids.unsqueeze(0)
# .expand_as(input_ids)
position_ids = paddle.fluid.layers.expand_as(
position_ids, input_ids)
embedding_output = self.embeddings(input_ids=input_ids,
position_ids=position_ids)
encoder_outputs = self.decoder(embedding_output,
memory=None,
tgt_mask=None,
use_cache=use_cache,
cache=cache)
self.checkpoints.extend(self.decoder.checkpoints)
return encoder_outputs
def forward(self, x):
outputs = []
if self.data_format == 'NCHW':
interpolate_shape = paddle.shape(x)[2:]
axis = 1
else:
interpolate_shape = paddle.shape(x)[1:3]
axis = -1
for block in self.aspp_blocks:
y = block(x)
outputs.append(y)
if self.image_pooling:
img_avg = self.global_avg_pool(x)
img_avg = F.interpolate(
img_avg,
interpolate_shape,
mode='bilinear',
align_corners=self.align_corners,
data_format=self.data_format)
outputs.append(img_avg)
x = paddle.concat(outputs, axis=axis)
x = self.conv_bn_relu(x)
x = self.dropout(x)
return x
def net(self, inputs, is_infer=False):
word2vec_model = Word2VecLayer(self.sparse_feature_number,
self.sparse_feature_dim,
self.neg_num,
emb_name="emb",
emb_w_name="emb_w",
emb_b_name="emb_b")
true_logits, neg_logits = word2vec_model.forward(inputs)
label_ones = paddle.full(shape=[paddle.shape(true_logits)[0], 1],
fill_value=1.0)
label_zeros = paddle.full(
shape=[paddle.shape(true_logits)[0], self.neg_num], fill_value=0.0)
true_logits = paddle.nn.functional.sigmoid(true_logits)
true_xent = paddle.nn.functional.binary_cross_entropy(
true_logits, label_ones)
neg_logits = paddle.nn.functional.sigmoid(neg_logits)
neg_xent = paddle.nn.functional.binary_cross_entropy(
neg_logits, label_zeros)
cost = paddle.add(true_xent, neg_xent)
avg_cost = paddle.mean(x=cost)
self._cost = avg_cost
fetch_dict = {'loss': avg_cost}
return fetch_dict
def forward(self,
decoder_input_ids=None,
decoder_attention_mask=None,
encoder_output=None,
memory_mask=None,
cache=None):
if decoder_attention_mask is None:
decoder_length = paddle.shape(decoder_input_ids)[-1]
decoder_attention_mask = paddle.tensor.triu(
(paddle.full(
(decoder_length, decoder_length),
-np.inf,
dtype=paddle.get_default_dtype())),
1)
decoder_inputs_embeds = self.embed_tokens(decoder_input_ids)
past_key_values_length = paddle.shape(cache[0][0].k)[
2] if cache is not None else 0
decoder_inputs_embed_pos = self.decoder_embed_positions(
decoder_input_ids.shape, past_key_values_length)
hidden_states = decoder_inputs_embeds + decoder_inputs_embed_pos
hidden_states = self.decoder_layernorm_embedding(hidden_states)
decoder_input = self.decoder_dropout(hidden_states)
decoder_output = self.decoder(
tgt=decoder_input,
memory=encoder_output,
tgt_mask=decoder_attention_mask,
memory_mask=memory_mask,
cache=cache)
return decoder_output
def forward(self, x, patch_embed_size):
""" Forward function.
Args:
x (Tensor): Input tensor of decoder.
patch_embed_size (Tensor): The height and width of the patch embed tensor.
Returns:
list[Tensor]: Segmentation results.
"""
x = self.proj_input(x)
cls_token = self.cls_token.expand((paddle.shape(x)[0], -1, -1))
x = paddle.concat([x, cls_token], axis=1)
for block in self.blocks:
x = block(x)
x = self.decoder_norm(x)
patches, masks = x[:, :-self.num_classes], x[:, -self.num_classes:]
patches = self.proj_patch(patches)
masks = self.proj_class(masks)
patches = patches / paddle.norm(patches, axis=-1, keepdim=True)
masks = masks / paddle.norm(masks, axis=-1, keepdim=True)
masks = patches @ masks.transpose((0, 2, 1))
masks = masks.reshape(
(0, 0, self.num_classes)) # For export inference model
masks = self.mask_norm(masks)
#[b, (h w), c] -> [b, c, h, w]
h, w = patch_embed_size[0], patch_embed_size[1]
masks = masks.reshape((0, h, w, paddle.shape(masks)[-1]))
masks = masks.transpose((0, 3, 1, 2))
return [masks]
def row_column_shuffle(embedding):
embedding = paddle.transpose(embedding, perm=[1, 0])
corrupted_embedding = paddle.transpose(embedding[paddle.randperm(
paddle.shape(embedding)[0])],
perm=[1, 0])
return corrupted_embedding[paddle.randperm(
paddle.shape(corrupted_embedding)[0])]
def __call__(self, hm, wh, reg, im_shape, scale_factor):
heat = self._simple_nms(hm)
scores, inds, clses, ys, xs = self._topk(heat)
scores = paddle.tensor.unsqueeze(scores, [1])
clses = paddle.tensor.unsqueeze(clses, [1])
reg_t = paddle.transpose(reg, [0, 2, 3, 1])
# Like TTFBox, batch size is 1.
# TODO: support batch size > 1
reg = paddle.reshape(reg_t, [-1, paddle.shape(reg_t)[-1]])
reg = paddle.gather(reg, inds)
xs = paddle.cast(xs, 'float32')
ys = paddle.cast(ys, 'float32')
xs = xs + reg[:, 0:1]
ys = ys + reg[:, 1:2]
wh_t = paddle.transpose(wh, [0, 2, 3, 1])
wh = paddle.reshape(wh_t, [-1, paddle.shape(wh_t)[-1]])
wh = paddle.gather(wh, inds)
if self.regress_ltrb:
x1 = xs - wh[:, 0:1]
y1 = ys - wh[:, 1:2]
x2 = xs + wh[:, 2:3]
y2 = ys + wh[:, 3:4]
else:
x1 = xs - wh[:, 0:1] / 2
y1 = ys - wh[:, 1:2] / 2
x2 = xs + wh[:, 0:1] / 2
y2 = ys + wh[:, 1:2] / 2
n, c, feat_h, feat_w = hm.shape[:]
padw = (feat_w * self.down_ratio - im_shape[0, 1]) / 2
padh = (feat_h * self.down_ratio - im_shape[0, 0]) / 2
x1 = x1 * self.down_ratio
y1 = y1 * self.down_ratio
x2 = x2 * self.down_ratio
y2 = y2 * self.down_ratio
x1 = x1 - padw
y1 = y1 - padh
x2 = x2 - padw
y2 = y2 - padh
bboxes = paddle.concat([x1, y1, x2, y2], axis=1)
scale_y = scale_factor[:, 0:1]
scale_x = scale_factor[:, 1:2]
scale_expand = paddle.concat([scale_x, scale_y, scale_x, scale_y],
axis=1)
boxes_shape = paddle.shape(bboxes)
boxes_shape.stop_gradient = True
scale_expand = paddle.expand(scale_expand, shape=boxes_shape)
bboxes = paddle.divide(bboxes, scale_expand)
if self.for_mot:
results = paddle.concat([bboxes, scores, clses], axis=1)
return results, inds
else:
results = paddle.concat([clses, scores, bboxes], axis=1)
return results, paddle.shape(results)[0:1]
def _dice_loss(self, input, target):
input = paddle.reshape(input, shape=(paddle.shape(input)[0], -1))
target = paddle.reshape(target, shape=(paddle.shape(target)[0], -1))
a = paddle.sum(input * target, axis=1)
b = paddle.sum(input * input, axis=1) + 0.001
c = paddle.sum(target * target, axis=1) + 0.001
d = (2 * a) / (b + c)
return 1 - d
def test_positive_attr_shape(self):
x = paddle.zeros([1, 3, 5, 7])
self.assertEqual(
paddle.jit.dy2static.choose_shape_attr_or_api(
x.shape, paddle.shape(x)), x.shape)
self.assertEqual(
paddle.jit.dy2static.choose_shape_attr_or_api(
x.shape, paddle.shape(x), 3), x.shape[3])
def forward(self, src_word, trg_word):
src_max_len = paddle.shape(src_word)[-1]
trg_max_len = paddle.shape(trg_word)[-1]
base_attn_bias = paddle.cast(
src_word == self.bos_id,
dtype=paddle.get_default_dtype()).unsqueeze([1, 2]) * -1e9
src_slf_attn_bias = base_attn_bias
src_slf_attn_bias.stop_gradient = True
trg_slf_attn_bias = paddle.tensor.triu(
(paddle.ones(
(trg_max_len, trg_max_len),
dtype=paddle.get_default_dtype()) * -np.inf),
1)
trg_slf_attn_bias.stop_gradient = True
trg_src_attn_bias = paddle.tile(base_attn_bias, [1, 1, trg_max_len, 1])
src_pos = paddle.cast(
src_word != self.bos_id, dtype="int64") * paddle.arange(
start=0, end=src_max_len)
trg_pos = paddle.cast(
trg_word != self.bos_id, dtype="int64") * paddle.arange(
start=0, end=trg_max_len)
src_emb = self.src_word_embedding(src_word)
src_pos_emb = self.src_pos_embedding(src_pos)
src_emb = src_emb + src_pos_emb
enc_input = F.dropout(
src_emb, p=self.dropout,
training=self.training) if self.dropout else src_emb
with paddle.static.amp.fp16_guard():
if self.waitk >= src_max_len or self.waitk == -1:
# Full sentence
enc_outputs = [
self.encoder(
enc_input, src_mask=src_slf_attn_bias)
]
else:
# Wait-k policy
enc_outputs = []
for i in range(self.waitk, src_max_len + 1):
enc_output = self.encoder(
enc_input[:, :i, :],
src_mask=src_slf_attn_bias[:, :, :, :i])
enc_outputs.append(enc_output)
trg_emb = self.trg_word_embedding(trg_word)
trg_pos_emb = self.trg_pos_embedding(trg_pos)
trg_emb = trg_emb + trg_pos_emb
dec_input = F.dropout(
trg_emb, p=self.dropout,
training=self.training) if self.dropout else trg_emb
dec_output = self.decoder(
dec_input,
enc_outputs,
tgt_mask=trg_slf_attn_bias,
memory_mask=trg_src_attn_bias)
predict = self.linear(dec_output)
return predict
def test_shape_with_var(self):
with program_guard(Program(), Program()):
x = paddle.static.data(shape=[-1, 1, 3], name='x')
fake_var = paddle.randn([2, 3])
target_shape = [
-1, paddle.shape(fake_var)[0], paddle.shape(fake_var)[1]
]
out = paddle.expand(x, shape=target_shape)
self.assertListEqual(list(out.shape), [-1, -1, -1])
def _gen_proposal(self, scores, bbox_deltas, anchors, inputs):
"""
scores (list[Tensor]): Multi-level scores prediction
bbox_deltas (list[Tensor]): Multi-level deltas prediction
anchors (list[Tensor]): Multi-level anchors
inputs (dict): ground truth info
"""
prop_gen = self.train_proposal if self.training else self.test_proposal
im_shape = inputs['im_shape']
# Collect multi-level proposals for each batch
# Get 'topk' of them as final output
bs_rois_collect = []
bs_rois_num_collect = []
batch_size = paddle.slice(paddle.shape(im_shape), [0], [0], [1])
# Generate proposals for each level and each batch.
# Discard batch-computing to avoid sorting bbox cross different batches.
for i in range(batch_size):
rpn_rois_list = []
rpn_prob_list = []
rpn_rois_num_list = []
for rpn_score, rpn_delta, anchor in zip(scores, bbox_deltas,
anchors):
rpn_rois, rpn_rois_prob, rpn_rois_num, post_nms_top_n = prop_gen(
scores=rpn_score[i:i + 1],
bbox_deltas=rpn_delta[i:i + 1],
anchors=anchor,
im_shape=im_shape[i:i + 1])
if rpn_rois.shape[0] > 0:
rpn_rois_list.append(rpn_rois)
rpn_prob_list.append(rpn_rois_prob)
rpn_rois_num_list.append(rpn_rois_num)
if len(scores) > 1:
rpn_rois = paddle.concat(rpn_rois_list)
rpn_prob = paddle.concat(rpn_prob_list).flatten()
if rpn_prob.shape[0] > post_nms_top_n:
topk_prob, topk_inds = paddle.topk(rpn_prob,
post_nms_top_n)
topk_rois = paddle.gather(rpn_rois, topk_inds)
else:
topk_rois = rpn_rois
topk_prob = rpn_prob
else:
topk_rois = rpn_rois_list[0]
topk_prob = rpn_prob_list[0].flatten()
bs_rois_collect.append(topk_rois)
bs_rois_num_collect.append(paddle.shape(topk_rois)[0])
bs_rois_num_collect = paddle.concat(bs_rois_num_collect)
return bs_rois_collect, bs_rois_num_collect
def test_negative_attr_shape(self):
x = paddle.zeros([7])
self.assertEqual(
paddle.jit.dy2static.choose_shape_attr_or_api([-1],
paddle.shape(x), 0),
paddle.shape(x)[0])
self.assertEqual(
paddle.jit.dy2static.choose_shape_attr_or_api([-1],
paddle.shape(x)),
paddle.shape(x))
def forward(self, x):
y = paddle.rand(shape=[1, 3, 4, 4])
w1 = paddle.shape(y)[0]
w2 = paddle.shape(x)[0]
while w2 != w1:
x = F.avg_pool2d(x, kernel_size=3, padding=1, stride=2)
w2 = paddle.shape(x)[0]
return x + y
def forward(self,
input_ids,
position_ids=None,
attention_mask=None,
use_cache=False,
cache=None):
self.checkpoints = []
if position_ids is None:
past_length = 0
if cache is not None:
past_length = paddle.shape(cache[0].k)[-2]
position_ids = paddle.arange(past_length,
paddle.shape(input_ids)[-1] +
past_length,
dtype='int64')
position_ids = position_ids.unsqueeze(0)
position_ids = paddle.fluid.layers.expand_as(
position_ids, input_ids)
embedding_output = self.embeddings(input_ids=input_ids,
position_ids=position_ids)
if _global_parallel_strategy == "pp":
auto.shard_tensor(input_ids,
dist_attr={
"process_mesh":
PP_MESH_LIST[0],
"dims_mapping":
[-1 for i in range(len(input_ids.shape))]
})
if _global_parallel_strategy == "dp_pp":
auto.shard_tensor(
input_ids,
dist_attr={
"process_mesh":
DPPP_MESH_LIST[0],
"dims_mapping":
[0] + [-1 for i in range(len(input_ids.shape) - 1)]
})
if _global_parallel_strategy == "dp_mp_pp":
auto.shard_tensor(
input_ids,
dist_attr={
"process_mesh":
DPMPPP_MESH_LIST[0],
"dims_mapping":
[0] + [-1 for i in range(len(input_ids.shape) - 1)]
})
encoder_outputs = self.decoder(embedding_output,
memory=None,
tgt_mask=attention_mask,
use_cache=use_cache,
cache=cache)
self.checkpoints.extend(self.decoder.checkpoints)
return encoder_outputs
def dmr_fcn_attention(item_eb,
item_his_eb,
context_his_eb,
mask,
mode='SUM'):
mask = paddle.equal(mask, paddle.ones_like(mask))
item_eb_tile = paddle.tile(item_eb,
[1, paddle.shape(mask)[1]]) # B, T*E
item_eb_tile = paddle.reshape(
item_eb_tile,
[-1, paddle.shape(mask)[1], item_eb.shape[-1]]) # B, T, E
if context_his_eb is None:
query = item_eb_tile
else:
query = paddle.concat([item_eb_tile, context_his_eb], axis=-1)
query = self.query_layer2(query)
query = self.query_prelu2(query)
dmr_all = paddle.concat(
[
query, item_his_eb, query - item_his_eb,
query * item_his_eb
],
axis=-1)
att_layer_1 = self.att_layer1_layer2(dmr_all)
att_layer_1 = F.sigmoid(att_layer_1)
att_layer_2 = self.att_layer2_layer2(att_layer_1)
att_layer_2 = F.sigmoid(att_layer_2)
att_layer_3 = self.att_layer3_layer2(att_layer_2) # B, T, 1
att_layer_3 = paddle.reshape(
att_layer_3, [-1, 1, paddle.shape(item_his_eb)[1]]) # B,1,T
scores = att_layer_3
scores = scores.reshape([-1, 1, self.history_length]) ##
# Mask
key_masks = paddle.unsqueeze(mask, 1) # B,1,T
paddings = paddle.ones_like(scores) * (-2**32 + 1)
paddings_no_softmax = paddle.zeros_like(scores)
scores = paddle.where(key_masks, scores, paddings) # [B, 1, T]
scores_no_softmax = paddle.where(key_masks, scores,
paddings_no_softmax)
scores = F.softmax(scores)
if mode == 'SUM':
output = paddle.matmul(scores, item_his_eb) # [B, 1, H]
output = paddle.sum(output, axis=1) # B,E
else:
scores = paddle.reshape(scores,
[-1, paddle.shape(item_his_eb)[1]])
output = item_his_eb * paddle.unsqueeze(scores, -1)
output = paddle.reshape(output, paddle.shape(item_his_eb))
return output, scores, scores_no_softmax
def select_topk(heat_map, K=100):
"""
Args:
heat_map: heat_map in [N, C, H, W]
K: top k samples to be selected
score: detection threshold
Returns:
"""
#batch, c, height, width = paddle.shape(heat_map)
batch, c = heat_map.shape[:2]
height = paddle.shape(heat_map)[2]
width = paddle.shape(heat_map)[3]
# First select topk scores in all classes and batchs
# [N, C, H, W] -----> [N, C, H*W]
heat_map = paddle.reshape(heat_map, (batch, c, -1))
# Both in [N, C, K]
topk_scores_all, topk_inds_all = paddle.topk(heat_map, K)
# topk_inds_all = topk_inds_all % (height * width) # todo: this seems redudant
topk_ys = (topk_inds_all // width).astype("float32")
topk_xs = (topk_inds_all % width).astype("float32")
# Select topK examples across channel
# [N, C, K] -----> [N, C*K]
topk_scores_all = paddle.reshape(topk_scores_all, (batch, -1))
# Both in [N, K]
topk_scores, topk_inds = paddle.topk(topk_scores_all, K)
topk_clses = (topk_inds // K).astype("float32")
# First expand it as 3 dimension
topk_inds_all = paddle.reshape(
_gather_feat(paddle.reshape(topk_inds_all, (batch, -1, 1)), topk_inds),
(batch, K))
topk_ys = paddle.reshape(
_gather_feat(paddle.reshape(topk_ys, (batch, -1, 1)), topk_inds),
(batch, K))
topk_xs = paddle.reshape(
_gather_feat(paddle.reshape(topk_xs, (batch, -1, 1)), topk_inds),
(batch, K))
return dict({
"topk_score": topk_scores,
"topk_inds_all": topk_inds_all,
"topk_clses": topk_clses,
"topk_ys": topk_ys,
"topk_xs": topk_xs
})
def drop_path(x, drop_prob=0., training=False):
if drop_prob == 0. or not training:
return x
keep_prob = 1 - drop_prob
B = paddle.shape(x)[0]
ndim = len(paddle.shape(x))
shape = (B,) + (1,) * (ndim - 1) # work with diff dim tensors, not just 2D ConvNets
random_tensor = keep_prob + paddle.rand(shape, dtype=x.dtype)
random_tensor = random_tensor.floor() # binarize
output = x / keep_prob * random_tensor
return output
