def forward(self, x):
x = self.conv_1(x)
x = self.BN_1(x)
x = F.relu(x)
x = F.max_pool2d(x, kernel_size=2) # 第1个MAX_POOL层
x = self.conv_2(x)
x = self.BN_2(x)
x = F.relu(x)
x = F.max_pool2d(x, kernel_size=2) # 第2个MAX_POOL层
x = self.conv_3(x)
x = self.BN_3(x)
x = F.relu(x)
x = F.max_pool2d(x, kernel_size=2) # 第3个MAX_POOL层
x = self.conv_4(x)
x = self.BN_4(x)
x = F.relu(x)
x = F.max_pool2d(x, kernel_size=2) # 第4个MAX_POOL层
x = paddle.flatten(x, 1, -1) ## flatten
x = self.linear(x) # linear
output = x
return output
def build_model(self, exec_mode):
feed_shape = list(self.feed_shape[0])
if self.is_ipu_mode(exec_mode):
feed_shape[0] = 1
x = paddle.static.data(
name=self.feed_list[0], shape=feed_shape, dtype='float32')
with paddle.static.ipu_shard_guard(index=0, stage=0):
# using fp32
x = paddle.static.nn.conv2d(input=x, num_filters=3, filter_size=3)
x = paddle.static.nn.batch_norm(x, act='relu')
x = F.max_pool2d(x, kernel_size=2, stride=2)
with paddle.static.ipu_shard_guard(index=1, stage=1):
# using fp16
with paddle.static.amp.fp16_guard():
x = paddle.static.nn.conv2d(
input=x, num_filters=6, filter_size=3)
x = paddle.static.nn.batch_norm(x, act='relu')
x = F.max_pool2d(x, kernel_size=2, stride=2)
with paddle.static.ipu_shard_guard(index=2, stage=2):
# using fp32
x = paddle.static.nn.fc(x, size=10)
loss = paddle.mean(x)
self.fetch_list = [loss.name]
def forward(self, x):
logit_list = []
x = self.enco1(x)
x, ind1 = F.max_pool2d(x, kernel_size=2, stride=2, return_mask=True)
size1 = x.shape
x = self.enco2(x)
x, ind2 = F.max_pool2d(x, kernel_size=2, stride=2, return_mask=True)
size2 = x.shape
x = self.enco3(x)
x, ind3 = F.max_pool2d(x, kernel_size=2, stride=2, return_mask=True)
size3 = x.shape
x = self.enco4(x)
x, ind4 = F.max_pool2d(x, kernel_size=2, stride=2, return_mask=True)
size4 = x.shape
x = self.enco5(x)
x, ind5 = F.max_pool2d(x, kernel_size=2, stride=2, return_mask=True)
size5 = x.shape
x = F.max_unpool2d(x,
indices=ind5,
kernel_size=2,
stride=2,
output_size=size4)
x = self.deco1(x)
x = F.max_unpool2d(x,
indices=ind4,
kernel_size=2,
stride=2,
output_size=size3)
x = self.deco2(x)
x = F.max_unpool2d(x,
indices=ind3,
kernel_size=2,
stride=2,
output_size=size2)
x = self.deco3(x)
x = F.max_unpool2d(x,
indices=ind2,
kernel_size=2,
stride=2,
output_size=size1)
x = self.deco4(x)
x = F.max_unpool2d(x, indices=ind1, kernel_size=2, stride=2)
x = self.deco5(x)
logit_list.append(x)
return logit_list
def forward(self, x, pool_size=(2, 2), pool_type='avg'):
x = self.conv1(x)
x = self.bn1(x)
x = F.relu(x)
if pool_type == 'max':
x = F.max_pool2d(x, kernel_size=pool_size)
elif pool_type == 'avg':
x = F.avg_pool2d(x, kernel_size=pool_size)
elif pool_type == 'avg+max':
x = F.avg_pool2d(x, kernel_size=pool_size) + F.max_pool2d(x, kernel_size=pool_size)
else:
raise Exception(
f'Pooling type of {pool_type} is not supported. It must be one of "max", "avg" and "avg+max".')
return x
def forward(self, x):
channel_att_sum = None
for pool_type in self.pool_types:
if pool_type == 'avg':
avg_pool = F.avg_pool2d(x, (x.shape[2], x.shape[3]),
stride=(x.shape[2], x.shape[3]))
channel_att_raw = self.mlp(avg_pool)
elif pool_type == 'max':
max_pool = F.max_pool2d(x, (x.shape[2], x.shape[3]),
stride=(x.shape[2], x.shape[3]))
channel_att_raw = self.mlp(max_pool)
elif pool_type == 'lp':
lp_pool = F.lp_pool2d(x,
2, (x.shape[2], x.shape[3]),
stride=(x.shape[2], x.shape[3]))
channel_att_raw = self.mlp(lp_pool)
elif pool_type == 'lse':
# LSE pool only
lse_pool = logsumexp_2d(x)
channel_att_raw = self.mlp(lse_pool)
if channel_att_sum is None:
channel_att_sum = channel_att_raw
else:
channel_att_sum = channel_att_sum + channel_att_raw
scale = F.sigmoid(channel_att_sum)
scale = paddle.unsqueeze(scale, 2)
scale = paddle.unsqueeze(scale, 3)
scale = paddle.expand_as(scale, x)
return x * scale
def forward(self, feat, img_metas=None):
if img_metas is not None:
assert len(img_metas[0]) == feat.shape[0]
valid_ratios = None
if img_metas is not None and self.mask:
valid_ratios = img_metas[-1]
h_feat = feat.shape[2] # bsz c h w
feat_v = F.max_pool2d(feat,
kernel_size=(h_feat, 1),
stride=1,
padding=0)
feat_v = feat_v.squeeze(2) # bsz * C * W
feat_v = paddle.transpose(feat_v, perm=[0, 2, 1]) # bsz * W * C
holistic_feat = self.rnn_encoder(feat_v)[0] # bsz * T * C
if valid_ratios is not None:
valid_hf = []
T = holistic_feat.shape[1]
for i, valid_ratio in enumerate(valid_ratios):
valid_step = min(T, math.ceil(T * valid_ratio)) - 1
valid_hf.append(holistic_feat[i, valid_step, :])
valid_hf = paddle.stack(valid_hf, axis=0)
else:
valid_hf = holistic_feat[:, -1, :] # bsz * C
holistic_feat = self.linear(valid_hf) # bsz * C
return holistic_feat
def test_case(self):
import paddle
import paddle.nn.functional as F
import paddle.fluid.core as core
import paddle.fluid as fluid
paddle.enable_static()
input_data = np.array([[[[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12],
[13, 14, 15, 16]]]]).astype("float32")
x = fluid.data(name="x", shape=[1, 1, 4, 4], dtype="float32")
output, indices = F.max_pool2d(
x, kernel_size=2, stride=2, return_mask=True)
unpool_out = F.max_unpool2d(
output, indices, kernel_size=2, stride=None, output_size=(5, 5))
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
results = exe.run(paddle.fluid.default_main_program(),\
feed={"x":input_data},
fetch_list=[unpool_out],
return_numpy=True)
pool_out_np = np.array([[[[6., 8.], [14., 16.]]]]).astype("float32")
indices_np = np.array([[[[5, 7], [13, 15]]]]).astype("int32")
expect_res =unpool2dmax_forward_naive(pool_out_np, indices_np, [2,2], \
[2,2], [0,0], [5,5]).astype("float64")
self.assertTrue(np.allclose(results[0], expect_res))
def forward(self, x):
bottom_up_features = self.bottom_up(x)
bottom_up_features = bottom_up_features[self.output_b - 2:]
bottom_up_features = bottom_up_features[::-1]
results = []
prev_features = self.lateral_convs[0](bottom_up_features[0])
results.append(self.output_convs[0](prev_features))
for l_id, (features, lateral_conv, output_conv) in enumerate(
zip(bottom_up_features[1:], self.lateral_convs[1:],
self.output_convs[1:])):
top_down_features = F.interpolate(prev_features,
scale_factor=2,
mode="bilinear",
align_corners=False)
lateral_features = lateral_conv(features)
prev_features = lateral_features + top_down_features
results.append(output_conv(prev_features))
if (self.output_e == 6):
p6 = F.max_pool2d(results[0], kernel_size=1, stride=2, padding=0)
results.insert(0, p6)
elif (self.output_e == 7):
p6 = self.p6(results[0])
results.insert(0, p6)
p7 = self.p7(F.relu(results[0]))
results.insert(0, p7)
return results
def test_case(self):
import paddle
import paddle.nn.functional as F
import paddle.fluid.core as core
import paddle.fluid as fluid
import numpy as np
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
else:
place = core.CPUPlace()
with fluid.dygraph.guard(place):
input_data = np.array([[[[1, 2, 3, 4], [5, 6, 7,
8], [9, 10, 11, 12],
[13, 14, 15, 16]]]]).astype("float32")
input_x = paddle.to_tensor(input_data)
output, indices = F.max_pool2d(input_x,
kernel_size=2,
stride=2,
return_mask=True)
out_pp = F.max_unpool2d(output,
indices,
kernel_size=2,
stride=None,
output_size=(5, 5))
output_np = output.numpy()
indices_np = indices.numpy()
expect_res =unpool2dmax_forward_naive(output_np, indices_np, [2,2], \
[2,2], [0,0], [5,5]).astype("float64")
self.assertTrue(np.allclose(out_pp.numpy(), expect_res))
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 _simple_nms(self, heat, kernel=3):
"""
Use maxpool to filter the max score, get local peaks.
"""
pad = (kernel - 1) // 2
hmax = F.max_pool2d(heat, kernel, stride=1, padding=pad)
keep = paddle.cast(hmax == heat, 'float32')
return heat * keep
def run1():
with fluid.dygraph.guard():
input_np = np.random.uniform(-1, 1,
[2, 3, 32, 32]).astype(np.float32)
input_pd = fluid.dygraph.to_variable(input_np)
padding = [[0, 1], [0, 0], [0, 0], [0, 0]]
res_pd = max_pool2d(
input_pd, kernel_size=2, stride=2, padding=padding)
def build_model(self):
x = paddle.static.data(
name=self.feed_list[0], shape=self.feed_shape[0], dtype='float32')
# using fp32
x = paddle.static.nn.conv2d(input=x, num_filters=3, filter_size=3)
x = paddle.static.nn.batch_norm(x, act='relu')
x = F.max_pool2d(x, kernel_size=2, stride=2)
# using fp16
with paddle.static.amp.fp16_guard():
x = paddle.static.nn.conv2d(input=x, num_filters=6, filter_size=3)
x = paddle.static.nn.batch_norm(x, act='relu')
x = F.max_pool2d(x, kernel_size=2, stride=2)
# using fp32
x = paddle.static.nn.fc(x, size=10)
loss = paddle.mean(x)
self.fetch_list = [loss.name]
def forward(self, inputs):
x = inputs['image']
conv1 = self.conv1(x)
x = F.max_pool2d(conv1, kernel_size=3, stride=2, padding=1)
outs = []
for idx, stage in enumerate(self.res_layers):
x = stage(x)
if idx in self.return_idx:
outs.append(x)
return outs
def _forward(self, x):
branch3x3 = self.branch3x3(x)
branch3x3dbl = self.branch3x3dbl_1(x)
branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl)
branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl)
branch_pool = F.max_pool2d(x, kernel_size=3, stride=2)
outputs = [branch3x3, branch3x3dbl, branch_pool]
return outputs
def run3():
with fluid.dygraph.guard():
input_np = np.random.uniform(-1, 1,
[2, 3, 32, 32]).astype(np.float32)
input_pd = fluid.dygraph.to_variable(input_np)
padding = "padding"
res_pd = max_pool2d(input_pd,
kernel_size=2,
stride=2,
padding=padding,
data_format='NHWC')
def forward(self, inputs):
x = inputs['image']
res1 = self.res1(x)
x = F.max_pool2d(res1, kernel_size=3, stride=2, padding=1)
outs = []
for idx, stage in enumerate(self.res_layers):
x = stage(x)
if idx == self.freeze_at:
x.stop_gradient = True
if idx in self.return_idx:
outs.append(x)
return outs
def _forward(self, x):
branch3x3 = self.branch3x3_1(x)
branch3x3 = self.branch3x3_2(branch3x3)
branch7x7x3 = self.branch7x7x3_1(x)
branch7x7x3 = self.branch7x7x3_2(branch7x7x3)
branch7x7x3 = self.branch7x7x3_3(branch7x7x3)
branch7x7x3 = self.branch7x7x3_4(branch7x7x3)
branch_pool = F.max_pool2d(x, kernel_size=3, stride=2)
outputs = [branch3x3, branch7x7x3, branch_pool]
return outputs
def run9():
with fluid.dygraph.guard():
input_np = np.random.uniform(-1, 1,
[2, 3, 32, 32]).astype(np.float32)
input_pd = fluid.dygraph.to_variable(input_np)
res_pd = max_pool2d(input_pd,
kernel_size=2,
stride=2,
padding=0,
ceil_mode=False,
data_format='NHWC',
return_indices=True)
def forward(self, X):
h = F.relu(self.conv1_1(X))
h = F.relu(self.conv1_2(h))
relu1_2 = h
h = F.max_pool2d(h, kernel_size=2, stride=2)
h = F.relu(self.conv2_1(h))
h = F.relu(self.conv2_2(h))
relu2_2 = h
h = F.max_pool2d(h, kernel_size=2, stride=2)
h = F.relu(self.conv3_1(h))
h = F.relu(self.conv3_2(h))
h = F.relu(self.conv3_3(h))
relu3_3 = h
h = F.max_pool2d(h, kernel_size=2, stride=2)
h = F.relu(self.conv4_1(h))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
relu4_3 = h
return [relu1_2, relu2_2, relu3_3, relu4_3]
def forward(self, inputs):
y = paddle.reshape(
inputs, [-1, inputs.shape[2], inputs.shape[3], inputs.shape[4]])
y = self.conv(y)
y = F.max_pool2d(y, kernel_size=3, stride=2, padding=1)
for bottleneck_block in self.bottleneck_block_list:
y = bottleneck_block(y)
y = F.adaptive_avg_pool2d(y, output_size=1)
y = F.dropout(y, p=0.5)
y = paddle.reshape(y, [-1, self.seg_num, y.shape[1]])
y = paddle.mean(y, axis=1)
y = paddle.reshape(y, shape=[-1, 2048])
y = self.out(y)
return y
def forward(self, inputs):
x = inputs
conv1 = self.conv1(x)
x = F.max_pool2d(conv1, kernel_size=3, stride=2, padding=1)
outs = []
for idx, stage in enumerate(self.res_layers):
x = stage(x)
if idx == self.freeze_at:
x.stop_gradient = True
if idx in self.return_idx:
outs.append(x)
x = self.output_flatten(outs[-1])
x = self.output_fc(x)
x = self.bn_output(x)
return x
def forward(self, feats):
# feats: [P3 - P7]
lateral_feats = []
# up
up_feature = feats[-1]
for i, feature in enumerate(feats[::-1]):
if i == 0:
lateral_feats.append(feature)
else:
shape = paddle.shape(feature)
up_feature = F.interpolate(up_feature,
size=[shape[2], shape[3]])
lateral_feature = self._feature_fusion_cell(
self.conv_up[i - 1],
feature,
up_feature,
weights=self.up_weights[i - 1]
if self.use_weighted_fusion else None)
lateral_feats.append(lateral_feature)
up_feature = lateral_feature
out_feats = []
# down
down_feature = lateral_feats[-1]
for i, (lateral_feature,
route_feature) in enumerate(zip(lateral_feats[::-1], feats)):
if i == 0:
out_feats.append(lateral_feature)
else:
down_feature = F.max_pool2d(down_feature, 3, 2, 1)
if i == len(feats) - 1:
route_feature = None
weights = self.down_weights[
i - 1][:2] if self.use_weighted_fusion else None
else:
weights = self.down_weights[
i - 1] if self.use_weighted_fusion else None
out_feature = self._feature_fusion_cell(self.conv_down[i - 1],
lateral_feature,
down_feature,
route_feature,
weights=weights)
out_feats.append(out_feature)
down_feature = out_feature
return out_feats
def forward(self, body_feats):
laterals = []
used_backbone_levels = len(self.spatial_scale)
for i in range(used_backbone_levels):
laterals.append(self.lateral_convs[i](body_feats[i]))
used_backbone_levels = len(self.spatial_scale)
for i in range(used_backbone_levels - 1):
idx = used_backbone_levels - 1 - i
upsample = F.interpolate(
laterals[idx],
scale_factor=2.,
mode='nearest',
)
laterals[idx - 1] += upsample
fpn_output = []
for lvl in range(self.min_level, self.highest_backbone_level + 1):
i = lvl - self.min_level
fpn_output.append(self.fpn_convs[i](laterals[i]))
spatial_scales = self.spatial_scale
if self.num_outs > len(fpn_output):
# use max pool to get more levels on top of outputs (Faster R-CNN, Mask R-CNN)
if not self.has_extra_convs:
fpn_output.append(F.max_pool2d(fpn_output[-1], 1, stride=2))
spatial_scales = spatial_scales + [spatial_scales[-1] * 0.5]
# add extra conv levels for RetinaNet(use_c5)/FCOS(use_p5)
else:
if self.use_c5:
extra_source = body_feats[-1]
else:
extra_source = fpn_output[-1]
fpn_output.append(
self.fpn_convs[used_backbone_levels](extra_source))
spatial_scales = spatial_scales + [spatial_scales[-1] * 0.5]
for i in range(used_backbone_levels + 1, self.num_outs):
if self.relu_before_extra_convs:
fpn_output.append(self.fpn_convs[i](F.relu(
fpn_output[-1])))
else:
fpn_output.append(self.fpn_convs[i](fpn_output[-1]))
spatial_scales = spatial_scales + [
spatial_scales[-1] * 0.5
]
return fpn_output, spatial_scales
def forward(self, x):
if not self.training or self.keep_prob == 1:
return x
else:
gamma = (1. - self.keep_prob) / (self.block_size**2)
for s in x.shape[2:]:
gamma *= s / (s - self.block_size + 1)
matrix = paddle.cast(
paddle.rand(x.shape, x.dtype) < gamma, x.dtype)
mask_inv = F.max_pool2d(matrix,
self.block_size,
stride=1,
padding=self.block_size // 2)
mask = 1. - mask_inv
y = x * mask * (mask.numel() / mask.sum())
return y
def forward(self, x):
outputs = []
# stem
x = self.conv1(x)
x = self.bn1(x)
x = F.relu_(x)
x = F.max_pool2d(x, kernel_size=3, stride=2, padding=1)
# blocks
x = self.layer1(x)
outputs.append(x)
x = self.layer2(x)
outputs.append(x)
x = self.layer3(x)
outputs.append(x)
x = self.layer4(x)
outputs.append(x)
return outputs
def nms_hm(heat_map, kernel=3):
"""Do max_pooling for nms
Args:
heat_map (paddle.Tensor): pred cls heatmap
kernel (int, optional): max_pool kernel size. Defaults to 3.
Returns:
heatmap after nms
"""
pad = (kernel - 1) // 2
hmax = F.max_pool2d(heat_map,
kernel_size=(kernel, kernel),
stride=1,
padding=pad)
eq_index = (hmax == heat_map).astype("float32")
return heat_map * eq_index
def check_max_dygraph_ceilmode_results(self, place):
with fluid.dygraph.guard(place):
input_np = np.random.random([2, 3, 32, 32]).astype("float32")
input = fluid.dygraph.to_variable(input_np)
result = max_pool2d(
input, kernel_size=2, stride=2, padding=0, ceil_mode=True)
result_np = max_pool2D_forward_naive(
input_np,
ksize=[2, 2],
strides=[2, 2],
paddings=[0, 0],
ceil_mode=True)
self.assertTrue(np.allclose(result.numpy(), result_np))
max_pool2d_dg = paddle.nn.layer.MaxPool2D(
kernel_size=2, stride=2, padding=0, ceil_mode=True)
result = max_pool2d_dg(input)
self.assertTrue(np.allclose(result.numpy(), result_np))
def check_max_static_results(self, place):
with fluid.program_guard(fluid.Program(), fluid.Program()):
input = fluid.data(
name="input", shape=[2, 3, 32, 32], dtype="float32")
result = max_pool2d(input, kernel_size=2, stride=2, padding=0)
input_np = np.random.random([2, 3, 32, 32]).astype("float32")
result_np = pool2D_forward_naive(
input_np,
ksize=[2, 2],
strides=[2, 2],
paddings=[0, 0],
pool_type='max')
exe = fluid.Executor(place)
fetches = exe.run(fluid.default_main_program(),
feed={"input": input_np},
fetch_list=[result])
self.assertTrue(np.allclose(fetches[0], result_np))
def check_max_dygraph_nhwc_results(self, place):
with fluid.dygraph.guard(place):
input_np = np.random.random([2, 3, 32, 32]).astype("float32")
input = fluid.dygraph.to_variable(
np.transpose(input_np, [0, 2, 3, 1]))
result = max_pool2d(input,
kernel_size=2,
stride=2,
padding=0,
return_indices=False,
data_format="NHWC")
result_np = pool2D_forward_naive(input_np,
ksize=[2, 2],
strides=[2, 2],
paddings=[0, 0],
pool_type='max')
self.assertTrue(
np.allclose(np.transpose(result.numpy(), [0, 3, 1, 2]),
result_np))