def forward(self, x):
out = F.relu(F.max_pool2d(self.conv1(x), 2))
out = F.relu(F.max_pool2d(self.conv2(out), 2))
out = out.view(-1, 320)
out = F.relu(self.fc1(out))
out = self.fc2(out)
return F.log_softmax(out, dim=1)
def forward(self, x):
x = self.conv1(x)
x = F.max_pool2d(x, 2) + F.avg_pool2d(x, 2)
x = self.block1(x)
x = self.group1(x)
x = F.max_pool2d(x, 2) + F.avg_pool2d(x, 2)
x = self.block2(x)
x = self.group2(x)
x = F.max_pool2d(x, 2) + F.avg_pool2d(x, 2)
x = self.block3(x)
x = self.group3(x)
x = self.block4(x)
x = self.group4(x)
x = F.max_pool2d(x, 2) + F.avg_pool2d(x, 2)
x = x.view(x.size(0), -1)
fc = self.fc(x)
x = F.dropout(fc, training=self.training)
output = list()
for name, fun in self.fc_dict.iteritems():
out = fun(x)
output.append(out)
return output, fc
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), 2)
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = x.view(-1, 64 * 7 * 7) # reshape Variable
x = F.relu(self.fc1(x))
x = self.fc2(x)
return F.log_softmax(x, dim=-1)
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
x = x.view(-1, 320)
x = F.relu(self.fc1(x))
x = F.dropout(x, training=self.training)
return F.log_softmax(self.fc2(x))
def forward(self, X):
h = F.relu(self.conv1_1(X), inplace=True)
h = F.relu(self.conv1_2(h), inplace=True)
# relu1_2 = h
h = F.max_pool2d(h, kernel_size=2, stride=2)
h = F.relu(self.conv2_1(h), inplace=True)
h = F.relu(self.conv2_2(h), inplace=True)
# relu2_2 = h
h = F.max_pool2d(h, kernel_size=2, stride=2)
h = F.relu(self.conv3_1(h), inplace=True)
h = F.relu(self.conv3_2(h), inplace=True)
h = F.relu(self.conv3_3(h), inplace=True)
# relu3_3 = h
h = F.max_pool2d(h, kernel_size=2, stride=2)
h = F.relu(self.conv4_1(h), inplace=True)
h = F.relu(self.conv4_2(h), inplace=True)
h = F.relu(self.conv4_3(h), inplace=True)
# relu4_3 = h
h = F.relu(self.conv5_1(h), inplace=True)
h = F.relu(self.conv5_2(h), inplace=True)
h = F.relu(self.conv5_3(h), inplace=True)
relu5_3 = h
return relu5_3
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2(x), 2))
x = x.view(-1, 320)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return x
def forward(self, x):
if self.transform_input:
x = x.clone()
x[:, 0] = x[:, 0] * (0.229 / 0.5) + (0.485 - 0.5) / 0.5
x[:, 1] = x[:, 1] * (0.224 / 0.5) + (0.456 - 0.5) / 0.5
x[:, 2] = x[:, 2] * (0.225 / 0.5) + (0.406 - 0.5) / 0.5
else: warn("Input isn't transformed")
x = self.Conv2d_1a_3x3(x)
x = self.Conv2d_2a_3x3(x)
x = self.Conv2d_2b_3x3(x)
x = F.max_pool2d(x, kernel_size=3, stride=2)
x = self.Conv2d_3b_1x1(x)
x = self.Conv2d_4a_3x3(x)
x = F.max_pool2d(x, kernel_size=3, stride=2)
x = self.Mixed_5b(x)
x = self.Mixed_5c(x)
x = self.Mixed_5d(x)
x = self.Mixed_6a(x)
x = self.Mixed_6b(x)
x = self.Mixed_6c(x)
x = self.Mixed_6d(x)
x = self.Mixed_6e(x)
x = self.Mixed_7a(x)
x = self.Mixed_7b(x)
x_for_attn = x = self.Mixed_7c(x)
# 8 x 8 x 2048
x = F.avg_pool2d(x, kernel_size=8)
# 1 x 1 x 2048
x_for_capt = x = x.view(x.size(0), -1)
# 2048
x = self.fc(x)
# 1000 (num_classes)
return x_for_attn, x_for_capt, x
def forward(self, x):
x1 = self.conv1(x)
x1 = F.max_pool2d(x1, 3, stride=2)
x2 = self.fire2(x1)
x3 = self.fire3(x2)
if self.bypass:
x3 = x3 + x2
x4 = self.fire4(x3)
x4 = F.max_pool2d(x4, 3, stride=2)
x5 = self.fire5(x4)
if self.bypass:
x5 = x5 + x4
x6 = self.fire6(x5)
x7 = self.fire7(x6)
if self.bypass:
x7 = x7 + x6
x8 = self.fire8(x7)
x8 = F.max_pool2d(x8, 3, stride=2)
x9 = self.fire9(x8)
if self.bypass:
x9 = x9 + x8
x9 = F.dropout(x9, training=self.training)
x10 = F.relu(self.conv10(x9))
f = F.avg_pool2d(x10, x10.size()[2:]).view(x10.size(0), -1)
if not self.training:
return f
if self.loss == {'xent'}:
return f
elif self.loss == {'xent', 'htri'}:
return f, f
else:
raise KeyError("Unsupported loss: {}".format(self.loss))
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2(x), 2))
x = x.view(-1, 7*7*64)
x = F.relu(self.fc1(x))
x = F.dropout(x, 0.4)
x = self.fc2(x)
return F.log_softmax(x, dim=1)
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2(x), 2))
x = x.view(-1, 1600)
x = F.relu(self.fc1(x))
x = F.dropout(x, training=self.training)
x = self.fc2(x)
return th.abs(10 - x)
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2)) # max pooling over a 2x2 window
x = F.max_pool2d(F.relu(self.conv2(x)), 2) # square x can only specify single number
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def forward(self, x):
steps=x.shape[0]
batch=x.shape[1]
x=x.view(x.shape[0]*x.shape[1],1,-1,11)
out = F.max_pool2d(self.conv1(x), (2, 1))
out = F.max_pool2d(self.conv2(out), (2, 2))
out= out.view(steps,batch,-1)
return out
def forward(self, x, y, z):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2(x), 2))
x = x.view(-1, 1600)
x = F.relu(self.fc1(x))
x = F.dropout(x, training=self.training)
x = self.fc2(x)
return F.log_softmax(x), F.log_softmax(x), F.log_softmax(x)
def forward(self, x):
x = F.max_pool2d(F.relu(self.convolution_0(x)), (2, 2))
x = F.max_pool2d(F.relu(self.convolution_1(x)), (2, 2))
x = x.view(-1, 32 * 5 * 5)
x = F.relu(self.fully_connected_0(x))
x = F.relu(self.fully_connected_1(x))
x = self.fully_connected_2(x)
return x
def forward(self,x): x = F.max_pool2d(F.relu(self.conv1(x)),(2,2)) x = F.max_pool2d(F.relu(self.conv2(x)),2) x = x.view(x.size()[0],-1) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = self.fc3(x) return x
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
x = F.max_pool2d(F.relu(self.conv2(x)), 2) # when shape is square, a single number is fine.
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2)) # Max pooling over a (2, 2) window
x = F.max_pool2d(F.relu(self.conv2(x)), 2) # If the size is a square you can only specify a single number
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def siam(patch, params):
o = conv2d(patch, params, 'conv0', stride=3)
o = F.max_pool2d(F.relu(o), 2, 2)
o = conv2d(o, params, 'conv1')
o = F.max_pool2d(F.relu(o), 2, 2)
o = conv2d(o, params, 'conv2')
o = F.relu(o)
return o.view(o.size(0), -1)
def deepcompare_2ch(input, params):
o = conv2d(input, params, 'conv0', stride=3)
o = F.max_pool2d(F.relu(o), 2, 2)
o = conv2d(o, params, 'conv1')
o = F.max_pool2d(F.relu(o), 2, 2)
o = conv2d(o, params, 'conv2')
o = F.relu(o).view(o.size(0), -1)
return linear(o, params, 'fc')
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, kernel_size=2, stride=2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, kernel_size=2, stride=2)
x = x.view([-1, 21 * 21 * 64])
x = self.fc1(x)
x = F.dropout(x, p=0.4, training=self.training)
return F.softmax(self.fc2(x))
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, 2, 2)
x = x.view(-1, 4 * 4 * 50)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return F.log_softmax(x, dim=1)
def forward(self, x):
if self.transform_input:
x = x.clone()
x[:, 0] = x[:, 0] * (0.229 / 0.5) + (0.485 - 0.5) / 0.5
x[:, 1] = x[:, 1] * (0.224 / 0.5) + (0.456 - 0.5) / 0.5
x[:, 2] = x[:, 2] * (0.225 / 0.5) + (0.406 - 0.5) / 0.5
# 299 x 299 x 3
x = self.Conv2d_1a_3x3(x)
# 149 x 149 x 32
x = self.Conv2d_2a_3x3(x)
# 147 x 147 x 32
x = self.Conv2d_2b_3x3(x)
# 147 x 147 x 64
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 73 x 73 x 64
x = self.Conv2d_3b_1x1(x)
# 73 x 73 x 80
x = self.Conv2d_4a_3x3(x)
# 71 x 71 x 192
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 35 x 35 x 192
x = self.Mixed_5b(x)
# 35 x 35 x 256
x = self.Mixed_5c(x)
# 35 x 35 x 288
x = self.Mixed_5d(x)
# 35 x 35 x 288
x = self.Mixed_6a(x)
# 17 x 17 x 768
x = self.Mixed_6b(x)
# 17 x 17 x 768
x = self.Mixed_6c(x)
# 17 x 17 x 768
x = self.Mixed_6d(x)
# 17 x 17 x 768
x = self.Mixed_6e(x)
# 17 x 17 x 768
if self.training and self.aux_logits:
aux = self.AuxLogits(x)
# 17 x 17 x 768
x = self.Mixed_7a(x)
# 8 x 8 x 1280
x = self.Mixed_7b(x)
# 8 x 8 x 2048
x = self.Mixed_7c(x)
# 8 x 8 x 2048
x = F.avg_pool2d(x, kernel_size=8)
# 1 x 1 x 2048
x = F.dropout(x, training=self.training)
# 1 x 1 x 2048
x = x.view(x.size(0), -1)
# 2048
x = self.fc(x)
# 1000 (num_classes)
if self.training and self.aux_logits:
return x, aux
return x
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, 2)
out = x.view(x.size(0), -1)
out = F.relu(self.fc1(out))
out = self.fc2(out)
return out
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
#x = x.view(-1, 320)
x = x.view((x.size(0), -1))
x = F.relu(self.fc1(x))
#x = F.dropout(x, training=self.training)
x = self.fc2(x)
return x
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(self.conv2(x))
x = F.relu(F.max_pool2d(self.conv3(x), 2))
x = F.relu(self.conv4(x))
x = x.view(-1, 4200)
x = F.relu(self.fc1(x))
action_scores = self.fc2(x)
return F.softmax(action_scores)
def forward(self, src, lengths=None):
"""See :func:`onmt.encoders.encoder.EncoderBase.forward()`"""
batch_size = src.size(0)
# (batch_size, 64, imgH, imgW)
# layer 1
src = F.relu(self.layer1(src[:, :, :, :] - 0.5), True)
# (batch_size, 64, imgH/2, imgW/2)
src = F.max_pool2d(src, kernel_size=(2, 2), stride=(2, 2))
# (batch_size, 128, imgH/2, imgW/2)
# layer 2
src = F.relu(self.layer2(src), True)
# (batch_size, 128, imgH/2/2, imgW/2/2)
src = F.max_pool2d(src, kernel_size=(2, 2), stride=(2, 2))
# (batch_size, 256, imgH/2/2, imgW/2/2)
# layer 3
# batch norm 1
src = F.relu(self.batch_norm1(self.layer3(src)), True)
# (batch_size, 256, imgH/2/2, imgW/2/2)
# layer4
src = F.relu(self.layer4(src), True)
# (batch_size, 256, imgH/2/2/2, imgW/2/2)
src = F.max_pool2d(src, kernel_size=(1, 2), stride=(1, 2))
# (batch_size, 512, imgH/2/2/2, imgW/2/2)
# layer 5
# batch norm 2
src = F.relu(self.batch_norm2(self.layer5(src)), True)
# (batch_size, 512, imgH/2/2/2, imgW/2/2/2)
src = F.max_pool2d(src, kernel_size=(2, 1), stride=(2, 1))
# (batch_size, 512, imgH/2/2/2, imgW/2/2/2)
src = F.relu(self.batch_norm3(self.layer6(src)), True)
# # (batch_size, 512, H, W)
all_outputs = []
for row in range(src.size(2)):
inp = src[:, :, row, :].transpose(0, 2) \
.transpose(1, 2)
row_vec = torch.Tensor(batch_size).type_as(inp.data) \
.long().fill_(row)
pos_emb = self.pos_lut(row_vec)
with_pos = torch.cat(
(pos_emb.view(1, pos_emb.size(0), pos_emb.size(1)), inp), 0)
outputs, hidden_t = self.rnn(with_pos)
all_outputs.append(outputs)
out = torch.cat(all_outputs, 0)
return hidden_t, out, lengths
def forward(self, input, lengths=None):
"See :obj:`onmt.modules.EncoderBase.forward()`"
batch_size = input.size(0)
# (batch_size, 64, imgH, imgW)
# layer 1
input = F.relu(self.layer1(input[:, :, :, :]-0.5), True)
# (batch_size, 64, imgH/2, imgW/2)
input = F.max_pool2d(input, kernel_size=(2, 2), stride=(2, 2))
# (batch_size, 128, imgH/2, imgW/2)
# layer 2
input = F.relu(self.layer2(input), True)
# (batch_size, 128, imgH/2/2, imgW/2/2)
input = F.max_pool2d(input, kernel_size=(2, 2), stride=(2, 2))
# (batch_size, 256, imgH/2/2, imgW/2/2)
# layer 3
# batch norm 1
input = F.relu(self.batch_norm1(self.layer3(input)), True)
# (batch_size, 256, imgH/2/2, imgW/2/2)
# layer4
input = F.relu(self.layer4(input), True)
# (batch_size, 256, imgH/2/2/2, imgW/2/2)
input = F.max_pool2d(input, kernel_size=(1, 2), stride=(1, 2))
# (batch_size, 512, imgH/2/2/2, imgW/2/2)
# layer 5
# batch norm 2
input = F.relu(self.batch_norm2(self.layer5(input)), True)
# (batch_size, 512, imgH/2/2/2, imgW/2/2/2)
input = F.max_pool2d(input, kernel_size=(2, 1), stride=(2, 1))
# (batch_size, 512, imgH/2/2/2, imgW/2/2/2)
input = F.relu(self.batch_norm3(self.layer6(input)), True)
# # (batch_size, 512, H, W)
all_outputs = []
for row in range(input.size(2)):
inp = input[:, :, row, :].transpose(0, 2)\
.transpose(1, 2)
row_vec = torch.Tensor(batch_size).type_as(inp.data)\
.long().fill_(row)
pos_emb = self.pos_lut(Variable(row_vec))
with_pos = torch.cat(
(pos_emb.view(1, pos_emb.size(0), pos_emb.size(1)), inp), 0)
outputs, hidden_t = self.rnn(with_pos)
all_outputs.append(outputs)
out = torch.cat(all_outputs, 0)
return hidden_t, out
def stream(input, name):
o = conv2d(input, params, name + '.conv0')
o = F.max_pool2d(F.relu(o), 2, 2)
o = conv2d(o, params, name + '.conv1')
o = F.max_pool2d(F.relu(o), 2, 2)
o = conv2d(o, params, name + '.conv2')
o = F.relu(o)
o = conv2d(o, params, name + '.conv3')
o = F.relu(o)
return o.view(o.size(0), -1)
def forward(self, x):
out = F.relu(self.conv1(x))
out = F.max_pool2d(out, 2)
out = F.relu(self.conv2(out))
out = F.max_pool2d(out, 2)
out = out.view(out.size(0), -1)
out = F.relu(self.fc1(out))
out = F.relu(self.fc2(out))
out = self.fc3(out)
return out
def count_neurons(self, image_dim):
x = Variable(torch.rand(1, *image_dim))
x = F.relu(F.max_pool2d(self.convolution1(x), 3, 2))
x = F.relu(F.max_pool2d(self.convolution2(x), 3, 2))
x = F.relu(F.max_pool2d(self.convolution3(x), 3, 2))
return x.data.view(1, -1).size(1)
def ConvBlockFunction(x, w, b, w_bn, b_bn):
x = F.conv2d(x, w, b, padding=1)
x = F.batch_norm(x, running_mean=None, running_var=None, weight=w_bn, bias=b_bn, training=True)
x = F.relu(x)
x = F.max_pool2d(x, kernel_size=2, stride=2)
return x
def forward(self, x, loc, label_class=None, label_box=None):
'''
Param:
x: FloatTensor(batch_num, 3, H, W)
loc: FloatTensor(batch_num, 4)
label_class: LongTensor(batch_num, N_max) or None
label_box: FloatTensor(batch_num, N_max, 4) or None
Return 1:
loss: FloatTensor(batch_num)
Return 2:
cls_i_preds: LongTensor(batch_num, topk)
cls_p_preds: FloatTensor(batch_num, topk)
reg_preds: FloatTensor(batch_num, topk, 4)
'''
w_2_in = self.w_relu(self.w_2_in)
w_3_in = self.w_relu(self.w_3_in)
w_2_in /= torch.sum(w_2_in, dim=0) + self.eps
w_3_in /= torch.sum(w_3_in, dim=0) + self.eps
C3, C4, C5 = self.backbone(x)
D5 = self.bi1_prj1_5(C5)
D5 = self.bi1_prj1_5bn(D5)
D4 = self.bi1_prj1_4(C4)
D4 = self.bi1_prj1_4bn(D4)
D3 = self.bi1_prj1_3(C3)
D3 = self.bi1_prj1_3bn(D3)
E4 = self.bi1_prj2_4(
(w_2_in[0][0] * D4 + w_2_in[1][0] * self.upsample(D5)))
F3 = self.bi1_prj3_3(
(w_2_in[0][1] * D3 + w_2_in[1][1] * self.upsample(E4)))
F4 = self.bi1_prj3_4((w_3_in[0][0] * self.downsample(F3) +
w_3_in[1][0] * E4 + w_3_in[2][0] * D4))
F5 = self.bi1_prj3_5(
(w_2_in[0][2] * D5 + w_2_in[1][2] * self.downsample(F4)))
G4 = self.bi2_prj2_4(
(w_2_in[0][3] * F4 + w_2_in[1][3] * self.upsample(F5)))
H3 = self.bi2_prj3_3(
(w_2_in[0][4] * F3 + w_2_in[1][4] * self.upsample(G4)))
H4 = self.bi2_prj3_4((w_3_in[0][1] * self.downsample(H3) +
w_3_in[1][1] * G4 + w_3_in[2][1] * F4))
H5 = self.bi2_prj3_5(
(w_2_in[0][5] * F5 + w_2_in[1][5] * self.downsample(H4)))
P6 = self.conv_out6(C5)
P7 = self.conv_out7(self.relu(P6))
P3, P4, P5, P6, P7 = H3, H4, H5, P6, P7
# log = [w_2_in.data.cpu().numpy(), w_3_in.data.cpu().numpy()]
# np.save('./bifpn_weight/weight_log', log)
# print('w_2_in = {0}, w_3_in={1}'.format(w_2_in.data.cpu().numpy(), w_3_in.data.cpu().numpy()))
if self.balanced_fpn:
# kernel_size, stride, padding, dilation, False, False
P3 = F.max_pool2d(P3, 3, 2, 1, 1, False, False)
P5 = self.upsample(P5)
P4 = (P3 + P4 + P5) / 3.0
# kernel_size, stride, padding, False, False
P5 = F.avg_pool2d(P4, 3, 2, 1, False, False)
P3 = self.upsample(P4)
pred_list = [P3, P4, P5, P6, P7]
assert len(pred_list) == self.scales
# assert len(pred_list) == len(self.scales)
cls_out = []
reg_out = []
for item in pred_list:
cls_i = self.conv_cls(item)
reg_i = self.conv_reg(item)
# cls_i: [b, an*classes, H, W] -> [b, H*W*an, classes]
cls_i = cls_i.permute(0, 2, 3, 1).contiguous()
cls_i = cls_i.view(cls_i.shape[0], -1, self.classes)
# reg_i: [b, an*4, H, W] -> [b, H*W*an, 4]
reg_i = reg_i.permute(0, 2, 3, 1).contiguous()
reg_i = reg_i.view(reg_i.shape[0], -1, 4)
cls_out.append(cls_i)
reg_out.append(reg_i)
# cls_out[b, hwan, classes]
# reg_out[b, hwan, 4]
cls_out = torch.cat(cls_out, dim=1)
reg_out = torch.cat(reg_out, dim=1)
if (label_class is not None) and (label_box is not None):
targets_cls, targets_reg = self._encode(
label_class, label_box, loc) # (b, hwan), (b, hwan, 4)
mask_cls = targets_cls > -1 # (b, hwan)
mask_reg = targets_cls > 0 # (b, hwan)
num_pos = torch.sum(mask_reg, dim=1).clamp_(min=self.scales) # (b)
# num_pos = torch.sum(mask_reg, dim=1).clamp_(min=len(self.scales)) # (b)
loss = []
for b in range(targets_cls.shape[0]):
cls_out_b = cls_out[b][mask_cls[b]] # (S+-, classes)
reg_out_b = reg_out[b][mask_reg[b]] # (S+, 4)
targets_cls_b = targets_cls[b][mask_cls[b]] # (S+-)
targets_reg_b = targets_reg[b][mask_reg[b]] # # (S+, 4)
loss_cls_b = sigmoid_focal_loss(cls_out_b, targets_cls_b, 2.0,
0.25).sum().view(1)
loss_reg_b = smooth_l1_loss(reg_out_b, targets_reg_b,
0.11).sum().view(1)
loss.append((loss_cls_b + loss_reg_b) / float(num_pos[b]))
return torch.cat(loss, dim=0) # (b
else:
return self._decode(cls_out, reg_out, loc)
def forward(self, x):
x = F.max_pool2d(F.pad(x, (0, 1, 0, 1), mode='replicate'), 2, stride=1)
return x
def forward(self, x):
if self.transform_input:
x_ch0 = torch.unsqueeze(x[:, 0],
1) * (0.229 / 0.5) + (0.485 - 0.5) / 0.5
x_ch1 = torch.unsqueeze(x[:, 1],
1) * (0.224 / 0.5) + (0.456 - 0.5) / 0.5
x_ch2 = torch.unsqueeze(x[:, 2],
1) * (0.225 / 0.5) + (0.406 - 0.5) / 0.5
x = torch.cat((x_ch0, x_ch1, x_ch2), 1)
# 299 x 299 x 3
x = self.Conv2d_1a_3x3(x)
# 149 x 149 x 32
x = self.Conv2d_2a_3x3(x)
# 147 x 147 x 32
x = self.Conv2d_2b_3x3(x)
# 147 x 147 x 64
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 73 x 73 x 64
x = self.Conv2d_3b_1x1(x)
# 73 x 73 x 80
x = self.Conv2d_4a_3x3(x)
# 71 x 71 x 192
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 35 x 35 x 192
x = self.Mixed_5b(x)
# 35 x 35 x 256
x = self.Mixed_5c(x)
# 35 x 35 x 288
x = self.Mixed_5d(x)
# 35 x 35 x 288
y = self.Mixed_6a(x)
# 17 x 17 x 768
y = self.Mixed_6b(y)
# 17 x 17 x 768
y = self.Mixed_6c(y)
# 17 x 17 x 768
y = self.Mixed_6d(y)
# 17 x 17 x 768
y = self.Mixed_6e(y)
# 17 x 17 x 768
if self.training and self.aux_logits:
aux = self.AuxLogits(y)
# 17 x 17 x 768
z = self.Mixed_7a(y)
# 8 x 8 x 1280
z = self.Mixed_7b(z)
# 8 x 8 x 2048
z = self.Mixed_7c(z)
# 8 x 8 x 2048
if self.is_fusion:
return x, y, z
z = F.avg_pool2d(z, kernel_size=8)
# 1 x 1 x 2048
z = F.dropout(z, training=self.training)
# 1 x 1 x 2048
z = z.view(z.size(0), -1)
# 2048
z = self.fc_(z)
# 1000 (num_classes)
if self.training and self.aux_logits:
return z, aux
return z
def forward(self, x):
out = F.relu(
self.en1_bn(F.max_pool2d(self.en1(self.encoder1(x)), 2, 2)))
out1 = F.relu(
self.enf1_bn(
F.interpolate(self.enf1(self.encoderf1(x)),
scale_factor=(2, 2),
mode='bilinear')))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.inte1_1bn(self.intere1_1(out1))),
scale_factor=(0.25, 0.25),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.inte1_2bn(self.intere1_2(tmp))),
scale_factor=(4, 4),
mode='bilinear'))
u1 = out
o1 = out1
out = F.relu(
self.en2_bn(F.max_pool2d(self.en2(self.encoder2(out)), 2, 2)))
out1 = F.relu(
self.enf2_bn(
F.interpolate(self.enf2(self.encoderf2(out1)),
scale_factor=(2, 2),
mode='bilinear')))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.inte2_1bn(self.intere2_1(out1))),
scale_factor=(0.0625, 0.0625),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.inte2_2bn(self.intere2_2(tmp))),
scale_factor=(16, 16),
mode='bilinear'))
u2 = out
o2 = out1
out = F.relu(
self.en3_bn(F.max_pool2d(self.en3(self.encoder3(out)), 2, 2)))
out1 = F.relu(
self.enf3_bn(
F.interpolate(self.enf3(self.encoderf3(out1)),
scale_factor=(2, 2),
mode='bilinear')))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.inte3_1bn(self.intere3_1(out1))),
scale_factor=(0.015625, 0.015625),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.inte3_2bn(self.intere3_2(tmp))),
scale_factor=(64, 64),
mode='bilinear'))
### End of encoder block
# print(out.shape,out1.shape)
out = F.relu(
self.de1_bn(
F.interpolate(self.de1(self.decoder1(out)),
scale_factor=(2, 2),
mode='bilinear')))
out1 = F.relu(
self.def1_bn(F.max_pool2d(self.def1(self.decoderf1(out1)), 2, 2)))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.intd1_1bn(self.interd1_1(out1))),
scale_factor=(0.0625, 0.0625),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.intd1_2bn(self.interd1_2(tmp))),
scale_factor=(16, 16),
mode='bilinear'))
out = torch.add(out, u2)
out1 = torch.add(out1, o2)
out = F.relu(
self.de2_bn(
F.interpolate(self.de2(self.decoder2(out)),
scale_factor=(2, 2),
mode='bilinear')))
out1 = F.relu(
self.def2_bn(F.max_pool2d(self.def2(self.decoderf2(out1)), 2, 2)))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.intd2_1bn(self.interd2_1(out1))),
scale_factor=(0.25, 0.25),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.intd2_2bn(self.interd2_2(tmp))),
scale_factor=(4, 4),
mode='bilinear'))
out = torch.add(out, u1)
out1 = torch.add(out1, o1)
out = F.relu(
self.de3_bn(
F.interpolate(self.de3(self.decoder3(out)),
scale_factor=(2, 2),
mode='bilinear')))
out1 = F.relu(
self.def3_bn(F.max_pool2d(self.def3(self.decoderf3(out1)), 2, 2)))
out = torch.add(out, out1)
out = F.relu(self.final(out))
# out = self.soft(out)
# print(out.shape)
return out
def forward(ctx, input, return_aggregation, win_size, peak_filter):
ctx.num_flags = 4
# assert win_size % 2 == 1, 'Window size for peak finding must be odd.'
# offset = (win_size - 1) // 2
# padding = torch.nn.ConstantPad2d(offset, float('-inf'))
# padded_maps = padding(input)
# batch_size, num_channels, h, w = padded_maps.size()
# element_map = torch.arange(0, h * w).long().view(1, 1, h, w)[:, :, offset: -offset, offset: -offset]
# element_map = element_map.to(input.device)
# _, indices = F.max_pool2d(
# padded_maps,
# kernel_size=win_size,
# stride=1,
# return_indices=True)
# peak_map = (indices == element_map)
peak_map = None
if type(win_size) == type([1, 2, 3]):
for win in win_size:
offset = (win - 1) // 2
padding = torch.nn.ConstantPad2d(offset, float('-inf'))
padded_maps = padding(input)
batch_size, num_channels, h, w = padded_maps.size()
element_map = torch.arange(0, h * w).long().view(
1, 1, h, w)[:, :, offset:-offset, offset:-offset]
element_map = element_map.to(input.device)
_, indices = F.max_pool2d(padded_maps,
kernel_size=win,
stride=1,
return_indices=True)
if peak_map is None:
peak_map = (indices == element_map)
else:
peak_map += peak_map
else:
offset = (win_size - 1) // 2
padding = torch.nn.ConstantPad2d(offset, float('-inf'))
padded_maps = padding(input)
batch_size, num_channels, h, w = padded_maps.size()
element_map = torch.arange(0, h * w).long().view(1, 1, h,
w)[:, :,
offset:-offset,
offset:-offset]
element_map = element_map.to(input.device)
_, indices = F.max_pool2d(padded_maps,
kernel_size=win_size,
stride=1,
return_indices=True)
peak_map = (indices == element_map)
np_input_person = padded_maps.cpu().data.numpy()[0, 16, :, :]
person_indices = indices.cpu().data.numpy()[0, 16, :, :]
np_peak_map_person = peak_map.cpu().data.numpy()[0, 16, :, :]
# peak filtering
if peak_filter:
mask = input >= peak_filter(input)
peak_map = (peak_map * mask)
peak_list = torch.nonzero(peak_map)
ctx.mark_non_differentiable(peak_list)
# peak aggregation
if return_aggregation:
peak_map = peak_map.float()
ctx.save_for_backward(input, peak_map)
return peak_list, (input * peak_map).view(
batch_size, num_channels, -1).sum(2) / peak_map.view(
batch_size, num_channels, -1).sum(2)
else:
return peak_list
def forward(self, x):
out = self.conv(F.relu(self.bn(x)))
out = F.max_pool2d(out, 2)
return out
def forward(self, x):
p = int(np.floor((self.kernel_size - 1) / 2))
p2d = (p, p, p, p)
return F.max_pool2d(F.pad(x, p2d), self.kernel_size, stride=2)
def forward(self, x):
# print('x',x.size())
y = self.encode(x) # y,x尺寸一致
# print('y',y.size())
y_small = F.max_pool2d(y, kernel_size=2, stride=2)
return y, y_small
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
out = self.layers(x)
return out, F.max_pool2d(out, 2)
1000,
),
("matinv", MF.matinv, torch.inverse, [(10, 10)], [(30, 30)], True, 1000),
(
"matmul",
MF.matmul,
torch.matmul,
[(64, 32), (32, 64)],
[(2048, 1024), (1024, 2048)],
True,
1000,
),
(
"max_pool2d",
lambda x: MF.max_pool2d(x, 2),
lambda x: TF.max_pool2d(x, 2),
[(2, 32, 16, 16)],
[(64, 512, 16, 16)],
True,
1000,
),
(
"normal",
lambda x: mge.random.normal(0, 1, x.shape),
lambda x: torch.randn(x.shape, device="cuda"),
[(100, 100)],
[(64, 512, 16, 16)],
True,
1000,
),
(
def forward(self, x):
return [F.max_pool2d(x, 1, 2, 0)]
def forward(self, rgb_inputs, depth_inputs):
######## DEPTH ENCODER ########
# Stage 1
x = self.conv11d(depth_inputs)
x_1 = self.CBR1_D(x)
x, id1_d = F.max_pool2d(x_1,
kernel_size=2,
stride=2,
return_indices=True)
# Stage 2
x_2 = self.CBR2_D(x)
x, id2_d = F.max_pool2d(x_2,
kernel_size=2,
stride=2,
return_indices=True)
# Stage 3
x_3 = self.CBR3_D(x)
x, id3_d = F.max_pool2d(x_3,
kernel_size=2,
stride=2,
return_indices=True)
x = self.dropout3_d(x)
# Stage 4
x_4 = self.CBR4_D(x)
x, id4_d = F.max_pool2d(x_4,
kernel_size=2,
stride=2,
return_indices=True)
x = self.dropout4_d(x)
# Stage 5
x_5 = self.CBR5_D(x)
######## RGB ENCODER ########
# Stage 1
y = self.CBR1_RGB(rgb_inputs)
y = torch.add(y, x_1)
y, id1 = F.max_pool2d(y, kernel_size=2, stride=2, return_indices=True)
# Stage 2
y = self.CBR2_RGB(y)
y = torch.add(y, x_2)
y, id2 = F.max_pool2d(y, kernel_size=2, stride=2, return_indices=True)
# Stage 3
y = self.CBR3_RGB(y)
y = torch.add(y, x_3)
y, id3 = F.max_pool2d(y, kernel_size=2, stride=2, return_indices=True)
y = self.dropout3(y)
# Stage 4
y = self.CBR4_RGB(y)
y = torch.add(y, x_4)
y, id4 = F.max_pool2d(y, kernel_size=2, stride=2, return_indices=True)
y = self.dropout4(y)
# Stage 5
y = self.CBR5_RGB(y)
y = torch.add(y, x_5)
y_size = y.size()
y, id5 = F.max_pool2d(y, kernel_size=2, stride=2, return_indices=True)
y = self.dropout5(y)
######## DECODER ########
# Stage 5 dec
y = F.max_unpool2d(y, id5, kernel_size=2, stride=2, output_size=y_size)
y = self.CBR5_Dec(y)
# Stage 4 dec
y = F.max_unpool2d(y, id4, kernel_size=2, stride=2)
y = self.CBR4_Dec(y)
# Stage 3 dec
y = F.max_unpool2d(y, id3, kernel_size=2, stride=2)
y = self.CBR3_Dec(y)
# Stage 2 dec
y = F.max_unpool2d(y, id2, kernel_size=2, stride=2)
y = self.CBR2_Dec(y)
# Stage 1 dec
y = F.max_unpool2d(y, id1, kernel_size=2, stride=2)
y = self.CBR1_Dec(y)
return y
def forward(self, x):
# Stage 1
x11 = F.relu(self.conv1_1(x))
x12 = F.relu(self.conv1_2(x11))
x1p, id1 = F.max_pool2d(x12,
kernel_size=(2, 2),
stride=(2, 2),
return_indices=True)
# Stage 2
x21 = F.relu(self.conv2_1(x1p))
x22 = F.relu(self.conv2_2(x21))
x2p, id2 = F.max_pool2d(x22,
kernel_size=(2, 2),
stride=(2, 2),
return_indices=True)
# Stage 3
x31 = F.relu(self.conv3_1(x2p))
x32 = F.relu(self.conv3_2(x31))
x33 = F.relu(self.conv3_3(x32))
x3p, id3 = F.max_pool2d(x33,
kernel_size=(2, 2),
stride=(2, 2),
return_indices=True)
# Stage 4
x41 = F.relu(self.conv4_1(x3p))
x42 = F.relu(self.conv4_2(x41))
x43 = F.relu(self.conv4_3(x42))
x4p, id4 = F.max_pool2d(x43,
kernel_size=(2, 2),
stride=(2, 2),
return_indices=True)
# Stage 5
x51 = F.relu(self.conv5_1(x4p))
x52 = F.relu(self.conv5_2(x51))
x53 = F.relu(self.conv5_3(x52))
x5p, id5 = F.max_pool2d(x53,
kernel_size=(2, 2),
stride=(2, 2),
return_indices=True)
# Stage 6
x61 = F.relu(self.conv6_1(x5p))
# Stage 6d
x61d = F.relu(self.deconv6_1(x61))
# Stage 5d
x5d = F.max_unpool2d(x61d, id5, kernel_size=2, stride=2)
x51d = F.relu(self.deconv5_1(x5d))
# Stage 4d
x4d = F.max_unpool2d(x51d, id4, kernel_size=2, stride=2)
x41d = F.relu(self.deconv4_1(x4d))
# Stage 3d
x3d = F.max_unpool2d(x41d, id3, kernel_size=2, stride=2)
x31d = F.relu(self.deconv3_1(x3d))
# Stage 2d
x2d = F.max_unpool2d(x31d, id2, kernel_size=2, stride=2)
x21d = F.relu(self.deconv2_1(x2d))
# Stage 1d
x1d = F.max_unpool2d(x21d, id1, kernel_size=2, stride=2)
x12d = F.relu(self.deconv1_1(x1d))
# Should add sigmoid? github repo add so.
raw_alpha = self.deconv1(x12d)
pred_mattes = F.sigmoid(raw_alpha)
if self.stage <= 1:
return pred_mattes, 0
# Stage2 refine conv1
refine0 = torch.cat((x[:, :3, :, :], pred_mattes), 1)
refine1 = F.relu(self.refine_conv1(refine0))
refine2 = F.relu(self.refine_conv2(refine1))
refine3 = F.relu(self.refine_conv3(refine2))
# Should add sigmoid?
# sigmoid lead to refine result all converge to 0...
#pred_refine = F.sigmoid(self.refine_pred(refine3))
pred_refine = self.refine_pred(refine3)
pred_alpha = F.sigmoid(raw_alpha + pred_refine)
#print(pred_mattes.mean(), pred_alpha.mean(), pred_refine.sum())
return pred_mattes, pred_alpha
def forward(self, im_data, im_info, gt_boxes, num_boxes):
batch_size = im_data.size(0)
im_info = im_info.data
gt_boxes = gt_boxes.data
num_boxes = num_boxes.data
# feed image data to base model to obtain base feature map
base_feat = self.RCNN_base(im_data)
# feed base feature map tp RPN to obtain rois
rois, rpn_loss_cls, rpn_loss_bbox = self.RCNN_rpn(base_feat, im_info, gt_boxes, num_boxes)
# if it is training phrase, then use ground trubut bboxes for refining
if self.training:
roi_data = self.RCNN_proposal_target(rois, gt_boxes, num_boxes)
rois, rois_label, rois_target, rois_inside_ws, rois_outside_ws = roi_data
rois_label = Variable(rois_label.view(-1).long())
rois_target = Variable(rois_target.view(-1, rois_target.size(2)))
rois_inside_ws = Variable(rois_inside_ws.view(-1, rois_inside_ws.size(2)))
rois_outside_ws = Variable(rois_outside_ws.view(-1, rois_outside_ws.size(2)))
else:
rois_label = None
rois_target = None
rois_inside_ws = None
rois_outside_ws = None
rpn_loss_cls = torch.zeros(1)
rpn_loss_bbox = torch.zeros(1)
rois = Variable(rois)
# do roi pooling based on predicted rois
if cfg.POOLING_MODE == 'crop':
# pdb.set_trace()
# pooled_feat_anchor = _crop_pool_layer(base_feat, rois.view(-1, 5))
grid_xy = _affine_grid_gen(rois.view(-1, 5), base_feat.size()[2:], self.grid_size)
grid_yx = torch.stack([grid_xy.data[:,:,:,1], grid_xy.data[:,:,:,0]], 3).contiguous()
pooled_feat = self.RCNN_roi_crop(base_feat, Variable(grid_yx).detach())
if cfg.CROP_RESIZE_WITH_MAX_POOL:
pooled_feat = F.max_pool2d(pooled_feat, 2, 2)
elif cfg.POOLING_MODE == 'align':
pooled_feat = self.RCNN_roi_align(base_feat, rois.view(-1, 5))
elif cfg.POOLING_MODE == 'pool':
pooled_feat = self.RCNN_roi_pool(base_feat, rois.view(-1,5))
# feed pooled features to top model
pooled_feat = self._head_to_tail(pooled_feat)
# compute bbox offset
bbox_pred = self.RCNN_bbox_pred(pooled_feat)
if self.training and not self.class_agnostic:
# select the corresponding columns according to roi labels
bbox_pred_view = bbox_pred.view(bbox_pred.size(0), int(bbox_pred.size(1) / 4), 4)
bbox_pred_select = torch.gather(bbox_pred_view, 1, rois_label.view(rois_label.size(0), 1, 1).expand(rois_label.size(0), 1, 4))
bbox_pred = bbox_pred_select.squeeze(1)
# compute object classification probability
cls_score = self.RCNN_cls_score(pooled_feat)
cls_prob = F.softmax(cls_score, 1)
RCNN_loss_cls = torch.zeros(1)
RCNN_loss_bbox = torch.zeros(1)
if self.training:
# classification loss
RCNN_loss_cls = F.cross_entropy(cls_score, rois_label)
# bounding box regression L1 loss
RCNN_loss_bbox = _smooth_l1_loss(bbox_pred, rois_target, rois_inside_ws, rois_outside_ws)
cls_prob = cls_prob.view(batch_size, rois.size(1), -1)
bbox_pred = bbox_pred.view(batch_size, rois.size(1), -1)
# return rois, cls_prob, bbox_pred, rpn_loss_cls, rpn_loss_bbox, RCNN_loss_cls, RCNN_loss_bbox, rois_label
return rois, cls_prob, bbox_pred
def forward(self, x):
out = F.relu(self.en1_bn(F.max_pool2d(self.encoder1(x), 2,
2))) #U-Net branch
out1 = F.relu(
self.enf1_bn(
F.interpolate(self.encoderf1(x),
scale_factor=(2, 2),
mode='bilinear'))) #Ki-Net branch
tmp = out
out = torch.add(out,
F.interpolate(F.relu(
self.inte1_1bn(self.intere1_1(out1))),
scale_factor=(0.25, 0.25),
mode='bilinear')) #CRFB
out1 = torch.add(out1,
F.interpolate(F.relu(
self.inte1_2bn(self.intere1_2(tmp))),
scale_factor=(4, 4),
mode='bilinear')) #CRFB
u1 = out #skip conn
o1 = out1 #skip conn
out = F.relu(self.en2_bn(F.max_pool2d(self.encoder2(out), 2, 2)))
out1 = F.relu(
self.enf2_bn(
F.interpolate(self.encoderf2(out1),
scale_factor=(2, 2),
mode='bilinear')))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.inte2_1bn(self.intere2_1(out1))),
scale_factor=(0.0625, 0.0625),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.inte2_2bn(self.intere2_2(tmp))),
scale_factor=(16, 16),
mode='bilinear'))
u2 = out
o2 = out1
out = F.relu(self.en3_bn(F.max_pool2d(self.encoder3(out), 2, 2)))
out1 = F.relu(
self.enf3_bn(
F.interpolate(self.encoderf3(out1),
scale_factor=(2, 2),
mode='bilinear')))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.inte3_1bn(self.intere3_1(out1))),
scale_factor=(0.015625, 0.015625),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.inte3_2bn(self.intere3_2(tmp))),
scale_factor=(64, 64),
mode='bilinear'))
### End of encoder block
### Start Decoder
out = F.relu(
self.de1_bn(
F.interpolate(self.decoder1(out),
scale_factor=(2, 2),
mode='bilinear'))) #U-NET
out1 = F.relu(self.def1_bn(F.max_pool2d(self.decoderf1(out1), 2,
2))) #Ki-NET
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.intd1_1bn(self.interd1_1(out1))),
scale_factor=(0.0625, 0.0625),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.intd1_2bn(self.interd1_2(tmp))),
scale_factor=(16, 16),
mode='bilinear'))
out = torch.add(out, u2) #skip conn
out1 = torch.add(out1, o2) #skip conn
out = F.relu(
self.de2_bn(
F.interpolate(self.decoder2(out),
scale_factor=(2, 2),
mode='bilinear')))
out1 = F.relu(self.def2_bn(F.max_pool2d(self.decoderf2(out1), 2, 2)))
tmp = out
out = torch.add(
out,
F.interpolate(F.relu(self.intd2_1bn(self.interd2_1(out1))),
scale_factor=(0.25, 0.25),
mode='bilinear'))
out1 = torch.add(
out1,
F.interpolate(F.relu(self.intd2_2bn(self.interd2_2(tmp))),
scale_factor=(4, 4),
mode='bilinear'))
out = torch.add(out, u1)
out1 = torch.add(out1, o1)
out = F.relu(
self.de3_bn(
F.interpolate(self.decoder3(out),
scale_factor=(2, 2),
mode='bilinear')))
out1 = F.relu(self.def3_bn(F.max_pool2d(self.decoderf3(out1), 2, 2)))
out = torch.add(out, out1) # fusion of both branches
out = F.relu(self.final(out)) #1*1 conv
# out = self.soft(out)
return out
def forward(self, x):
return F.max_pool2d(x, (x.size(-2), x.size(-1)))
def net_forward(x, layer_by_layer=False, from_layer=0):
x = x - nrms_mean # cannot be inplace
x.div_(nrms_std)
if not layer_by_layer:
return net(x)
cldr = list(net.children())
if args.model.startswith('resnet'):
x = net.conv1(x)
x = net.bn1(x)
x = net.relu(x)
x = net.maxpool(x)
x = net.layer1(x)
x = net.layer2(x)
x = net.layer3(x)
x = net.layer4(x)
outputs = [net.avgpool(x)]
flat_features = outputs[-1].view(x.size(0), -1)
outputs.append(net.fc(flat_features))
elif args.model.startswith('inception'):
# 299 x 299 x 3
x = net.Conv2d_1a_3x3(x)
# 149 x 149 x 32
x = net.Conv2d_2a_3x3(x)
# 147 x 147 x 32
x = net.Conv2d_2b_3x3(x)
# 147 x 147 x 64
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 73 x 73 x 64
x = net.Conv2d_3b_1x1(x)
# 73 x 73 x 80
x = net.Conv2d_4a_3x3(x)
# 71 x 71 x 192
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 35 x 35 x 192
x = net.Mixed_5b(x)
# 35 x 35 x 256
x = net.Mixed_5c(x)
# 35 x 35 x 288
x = net.Mixed_5d(x)
# 35 x 35 x 288
x = net.Mixed_6a(x)
# 17 x 17 x 768
x = net.Mixed_6b(x)
# 17 x 17 x 768
x = net.Mixed_6c(x)
# 17 x 17 x 768
x = net.Mixed_6d(x)
# 17 x 17 x 768
x = net.Mixed_6e(x)
# 17 x 17 x 768
x = net.Mixed_7a(x)
# 8 x 8 x 1280
x = net.Mixed_7b(x)
# 8 x 8 x 2048
x = net.Mixed_7c(x)
# 8 x 8 x 2048
x = F.avg_pool2d(x, kernel_size=8)
# 1 x 1 x 2048
outputs = [F.dropout(x, training=net.training)]
# 1 x 1 x 2048
flat_features = outputs[-1].view(x.size(0), -1)
# 2048
outputs.append(net.fc(flat_features))
# 1000 (num_classes)
else:
outputs = [net.features(x)]
for cidx, c in enumerate(net.classifier.children()):
flat_features = outputs[-1].view(x.size(0), -1)
outputs.append(c(flat_features))
return outputs
def forward(self, x):
features = None
# --> fixed-size input: batch x 3 x 299 x 299
x = nn.functional.interpolate(x,
size=(299, 299),
mode='bilinear',
align_corners=False) #上采样或者下采样至给定size
# 299 x 299 x 3
x = self.Conv2d_1a_3x3(x)
# 149 x 149 x 32
x = self.Conv2d_2a_3x3(x)
# 147 x 147 x 32
x = self.Conv2d_2b_3x3(x)
# 147 x 147 x 64
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 73 x 73 x 64
x = self.Conv2d_3b_1x1(x)
# 73 x 73 x 80
x = self.Conv2d_4a_3x3(x)
# 71 x 71 x 192
x = F.max_pool2d(x, kernel_size=3, stride=2)
# 35 x 35 x 192
x = self.Mixed_5b(x)
# 35 x 35 x 256
x = self.Mixed_5c(x)
# 35 x 35 x 288
x = self.Mixed_5d(x)
# 35 x 35 x 288
x = self.Mixed_6a(x)
# 17 x 17 x 768
x = self.Mixed_6b(x)
# 17 x 17 x 768
x = self.Mixed_6c(x)
# 17 x 17 x 768
x = self.Mixed_6d(x)
# 17 x 17 x 768
x = self.Mixed_6e(x)
# 17 x 17 x 768
# image region features
# features = x
# 17 x 17 x 768
x = self.Mixed_7a(x)
# 8 x 8 x 1280
x = self.Mixed_7b(x)
# 8 x 8 x 2048
x = self.Mixed_7c(x)
# 8 x 8 x 2048
# x = F.avg_pool2d(x, kernel_size=8)
# 1 x 1 x 2048
# x = F.dropout(x, training=self.training)
# 1 x 1 x 2048
# x = x.view(x.size(0), -1) # for visual_feature_extraction.py use this as the output
x = x.mean(dim=(2, 3))
# 2048
# global image features
# cnn_code = self.emb_cnn_code(x)
# 512
# if features is not None:
# features = self.emb_features(features)
return x #cnn_code #1024
def forward(self, x):
#print (x.size())
x = self.init_conv1(x)
x = self.init_bn1(x)
x = F.relu(x, True)
#conv_x1 = x
x = self.init_res1(x)
x = self.init_conv2(x)
#print ("Skip", x.size())
#print ("Skip", conv_x.size())
x = self.init_bn2(x)
x = F.relu(x, True)
conv_x2 = x
x = F.max_pool2d(x, 2)
### can be on 1x1 cov
x = self.init_res2(x)
x = self.init_res3(x)
x = self.core_hourglass(x)
#x = self.core_hourglass1(x)
#x = self.core_hourglass2(x)
x = self.tail_res0(x)
x = self.tail_res00(x)
x = self.tail_deconv1(x)
x = self.tail_bn1(x)
x = F.relu(x, True)
#print ("Dce 1", x.size())
#split = x
#print x.shape
x = self.tail_res1(x)
#x = self.tail_res1_1(x)
###################
x = torch.cat((x, conv_x2),1)
x= self.init_conv_union(x)
x= self.init_bn_union(x)
x = F.relu(x, True)
###################
x = self.tail_resx(x)
x = self.tail_deconv2(x)
x = self.tail_bn2(x)
x = F.relu(x, True)
x = self.dropout(x)
#print ("Dce 2", x.size())
#split = x
x = self.tail_res2(x)
#out1 = x
###################
#x = torch.cat((x,split),1)
#x= self.init_conv_union1(x)
#x= self.init_bn_union1(x)
#x = F.relu(x, True)
###################
# Right tail
#x = self.rtail_deconv2(split)
#x = self.rtail_bn2(x)
#x = F.relu(x, True)
#x = self.rtail_res2(x)
x = self.rtail_res2_2(x)
#x = self.rtail_res2_3(x)
#out2 = x
#out = torch.cat((out1,out2),1)
#out1 = self.init_conv3(out1)
out2 = self.init_conv4(x)
return out2
def forward(self, x):
x = self.pre(x)
x = F.max_pool2d(x, 2)
return self.conv_bn2(self.conv_bn1(x)) + x
def non_maximum_suppression(a):
ap = F.max_pool2d(a, 3, stride=1, padding=1)
mask = (a == ap).float().clamp(min=0.0)
return a * mask
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 3))
x = x.view(-1, 192)
x = self.fc(x)
return F.log_softmax(x, dim=1)
def get_conv_output(self, x):
# Layer 1: Convolutional + 3x3 Max Pooling + Batch Norm
conv1_out = self.conv1(x)
pool1_out = F.max_pool2d(conv1_out, 3)
bn1_out = self.bn1(pool1_out)
# Layer 2: Convolutional + Batch Norm
conv2_out = self.conv2(bn1_out)
bn2_out = self.bn2(conv2_out)
# Layer 3: Convolutional + 2x2 Max Pooling + Batch Norm
conv3_out = self.conv3(bn2_out)
pool3_out = F.max_pool2d(conv3_out, 2)
bn3_out = self.bn3(pool3_out)
# Layer 4: Convolutional + Batch Norm
conv4_out = self.conv4(bn3_out)
bn4_out = self.bn4(conv4_out)
# Layer 5: Convolutional + 2x2 Max Pooling + Batch Norm
conv5_out = self.conv5(bn4_out)
pool5_out = F.max_pool2d(conv5_out, 2)
bn5_out = self.bn5(pool5_out)
# Layer 6: Convolutional + Batch Norm
conv6_out = self.conv6(bn5_out)
bn6_out = self.bn6(conv6_out)
# Layer 7: Convolutional + 2x2 Max Pooling + Batch Norm
conv7_out = self.conv7(bn6_out)
pool7_out = F.max_pool2d(conv7_out, 2)
bn7_out = HF.modified_bn(self.bn7, pool7_out)
# Layer 8: Convolutional + Batch Norm
conv8_out = self.conv8(bn7_out)
bn8_out = HF.modified_bn(self.bn8, conv8_out)
# Build dictionary containing outputs of each layer
conv_out = {
self.CONV1: conv1_out,
self.POOL1: pool1_out,
self.BN1: bn1_out,
self.CONV2: conv2_out,
self.BN2: bn2_out,
self.CONV3: conv3_out,
self.POOL3: pool3_out,
self.BN3: bn3_out,
self.CONV4: conv4_out,
self.BN4: bn4_out,
self.CONV5: conv5_out,
self.POOL5: pool5_out,
self.BN5: bn5_out,
self.CONV6: conv6_out,
self.BN6: bn6_out,
self.CONV7: conv7_out,
self.POOL7: pool7_out,
self.BN7: bn7_out,
self.CONV8: conv8_out,
self.BN8: bn8_out,
}
return conv_out
def forward(self, x):
x = self.conv(x)
x = F.max_pool2d(x, kernel_size=x.size()[2:]).squeeze()
return self.fc(x)
def forward(self, x):
nsample = x.size()[3]
x = self.activation(self.bn_avg(self.conv_avg(x)))
x = F.max_pool2d(x, kernel_size=(1, nsample)).squeeze(3)
return x
def forward(self, x):
x = self.static_padding(x)
x = F.max_pool2d(x, self.kernel_size, self.stride, self.padding,
self.dilation, self.ceil_mode, self.return_indices)
return x
def max_pool2d(self, in_features, kernel_size, stride, padding):
return F.max_pool2d(in_features, kernel_size=kernel_size, stride=stride, padding=padding)
def max_filter(img):
assert img.dim() == 3
img = img.unsqueeze(0)
img = tnf.max_pool2d(img, kernel_size=3, stride=1, padding=1)
return img.squeeze(0)