def __init__(self,):
super().__init__()
self.layers = nn.Sequential(
Conv2D(3, 32, 3, 2, 1),
Conv2D(32, 64, 3, 2, 1),
Conv2D(64, 128, 3, 2, 1),
Conv2D(128, 128, 3, 2, 1)
)
self.reset_parameters()
def __init__(self, answer_num, in_dim, proj_dim, fc_dim, feat_h, norm_type, add_coord):
super().__init__()
if add_coord:
proj_l = Conv2D(in_dim, proj_dim, 1, norm_type=norm_type, orders=('coord', 'conv', 'act'))
else:
proj_l = Conv2D(in_dim, proj_dim, 1, norm_type=norm_type, orders=('conv', 'act'))
self.layers = nn.Sequential(
proj_l,
nn.MaxPool2d(feat_h, feat_h, 0),
Flatten(),
Linear(proj_dim if proj_dim > 0 else in_dim, fc_dim, norm_type=norm_type, orders=('linear', 'norm', 'act')),
Linear(fc_dim, answer_num, orders=('linear',))
)
self.reset_parameters()
def __init__(self, answer_num, img_dim, query_dim, proj_dim, fc_dim, feat_h, norm_type, act_type, add_coord):
super().__init__()
from .layers.film_layers import HardAttnSumLayer
if add_coord:
proj_l = Conv2D(img_dim + query_dim, proj_dim, 1, orders=('coord', 'conv', 'norm', 'act'), act_type=act_type,
norm_type=norm_type)
else:
proj_l = Conv2D(img_dim + query_dim, proj_dim, 1, orders=('conv', 'norm', 'act'), act_type=act_type,
norm_type=norm_type)
self.layers = nn.Sequential(
proj_l,
HardAttnSumLayer(),
Linear(proj_dim, fc_dim, norm_type=norm_type,
orders=('linear', 'norm', 'act'), act_type=act_type
),
Linear(fc_dim, answer_num, orders=('linear',))
)
self.reset_parameters()
def __init__(self):
super().__init__()
import torchvision
resnet = torchvision.models.resnet101(False)
layers = [
nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool),
resnet.layer1,
resnet.layer2,
resnet.layer3,
Conv2D(1024, 128, 3, padding=1, orders=('coord', 'conv', 'norm', 'act'))
]
self.layers = nn.Sequential(*layers)
self.reset_parameters()
def __init__(self, in_dim, layer_dims, layer_norm_types, layer_act_types, layer_coord_types,
layer_se_types, layer_orders=None):
super().__init__()
self.layers = nn.ModuleList()
layer_orders = layer_orders if layer_orders is not None else (self.default_orders,) * len(layer_dims)
for idx, out_dim in enumerate(layer_dims):
se_type = layer_se_types[idx]
orders = layer_orders[idx]
if se_type is not None:
orders = list(orders)
orders.insert(orders.index('act'), 'se')
layer = Conv2D(in_dim, out_dim, 3, padding=1, orders=orders, norm_type=layer_norm_types[idx],
act_type=layer_act_types[idx], coord_type=layer_coord_types[idx], se_type=se_type)
self.layers.append(layer)
in_dim = out_dim
self.reset_parameters()