def forward(self, x, y, step):
x = self.x_op(x)
y = self.y_op(y)
X_DROP = False
Y_DROP = False
if self.search_space == 'small':
if self.drop_path_keep_prob is not None and self.training:
X_DROP = True
if self.drop_path_keep_prob is not None and self.training:
Y_DROP = True
elif self.search_space == 'middle':
if self.drop_path_keep_prob is not None and self.training:
X_DROP = True
if self.drop_path_keep_prob is not None and self.training:
Y_DROP = True
elif self.search_space == 'large':
if self.drop_path_keep_prob is not None and self.training:
X_DROP = True
if self.drop_path_keep_prob is not None and self.training:
Y_DROP = True
if X_DROP:
x = apply_drop_path(x, self.drop_path_keep_prob, self.layer_id,
self.layers, step, self.steps)
if Y_DROP:
y = apply_drop_path(y, self.drop_path_keep_prob, self.layer_id,
self.layers, step, self.steps)
out = x + y
return out
def forward(self, x, x_id, x_op, y, y_id, y_op, step, bn_train=False):
stride = self.stride if x_id in [0, 1] else 1
x = self.x_op[x_op](x, x_id, stride, bn_train)
stride = self.stride if y_id in [0, 1] else 1
y = self.y_op[y_op](y, y_id, stride, bn_train)
X_DROP = False
Y_DROP = False
if self.search_space == 'small':
if x_op not in [
4
] and self.drop_path_keep_prob is not None and self.training:
X_DROP = True
if y_op not in [
4
] and self.drop_path_keep_prob is not None and self.training:
Y_DROP = True
elif self.search_space == 'middle':
if x_op not in [
0, 1
] and self.drop_path_keep_prob is not None and self.training:
X_DROP = True
if y_op not in [
0, 1
] and self.drop_path_keep_prob is not None and self.training:
Y_DROP = True
if X_DROP:
x = apply_drop_path(x, self.drop_path_keep_prob, self.layer_id,
self.layers, step, self.steps)
if Y_DROP:
y = apply_drop_path(y, self.drop_path_keep_prob, self.layer_id,
self.layers, step, self.steps)
return x + y
def forward(self, x, y, step):
if self.x_op_id in [10, 13]:
x = F.pad(x, [0, 1, 0, 1])
x = self.x_op(x)
if self.x_id_fact_reduce is not None:
x = self.x_id_fact_reduce(x)
if self.drop_path_keep_prob is not None and self.training:
x = apply_drop_path(x, self.drop_path_keep_prob, self.layer_id, self.layers, step, self.steps)
if self.y_op_id in [10, 13]:
y = F.pad(y, [0, 1, 0, 1])
y = self.y_op(y)
if self.y_id_fact_reduce is not None:
y = self.y_id_fact_reduce(y)
if self.drop_path_keep_prob is not None and self.training:
y = apply_drop_path(y, self.drop_path_keep_prob, self.layer_id, self.layers, step, self.steps)
out = x + y
return out
def call(self, x, y=None, step=None, training=False, **kwargs):
x = self.x_op(x, training=training)
if self.x_id_fact_reduce is not None:
x = self.x_id_fact_reduce(x, training=training)
if self.drop_path_keep_prob is not None and training:
x = apply_drop_path(x, self.drop_path_keep_prob, self.cell_id,
self.num_cells, step, self.steps)
y = self.y_op(y)
if self.y_id_fact_reduce is not None:
y = self.y_id_fact_reduce(y, training=training)
if self.drop_path_keep_prob is not None and training:
y = apply_drop_path(y, self.drop_path_keep_prob, self.cell_id,
self.num_cells, step, self.steps)
out = x + y
return out
def forward(self, x, x_id, x_op, y, y_id, y_op, step, bn_train=False):
stride = self.stride if x_id in [0, 1] else 1
if x_op == 0:
x = self.x_sep_conv_3(x, x_id, stride, bn_train=bn_train)
elif x_op == 1:
x = self.x_sep_conv_5(x, x_id, stride, bn_train=bn_train)
elif x_op == 2:
x = self.x_avg_pool(x, stride)
elif x_op == 3:
x = self.x_max_pool(x, stride)
else:
assert x_op == 4
if stride > 1:
assert stride == 2
if x_id == 0:
x = self.x_id_reduce_1(x, bn_train=bn_train)
else:
x = self.x_id_reduce_2(x, bn_train=bn_train)
stride = self.stride if y_id in [0, 1] else 1
if y_op == 0:
y = self.y_sep_conv_3(y, y_id, stride, bn_train=bn_train)
elif y_op == 1:
y = self.y_sep_conv_5(y, y_id, stride, bn_train=bn_train)
elif y_op == 2:
y = self.y_avg_pool(y, stride)
elif y_op == 3:
y = self.y_max_pool(y, stride)
else:
assert y_op == 4
if stride > 1:
assert stride == 2
if y_id == 0:
y = self.x_id_reduce_1(y, bn_train=bn_train)
else:
y = self.x_id_reduce_2(y, bn_train=bn_train)
if x_op != 4 and self.drop_path_keep_prob is not None and self.training:
x = apply_drop_path(x, self.drop_path_keep_prob, self.layer_id,
self.layers, step, self.steps)
if y_op != 4 and self.drop_path_keep_prob is not None and self.training:
y = apply_drop_path(y, self.drop_path_keep_prob, self.layer_id,
self.layers, step, self.steps)
return x + y
