def __init__(self,
image_emb_size,
input_num_channels,
image_height,
image_width,
text_emb_size,
using_recurrence=False):
super(ImageTextKernelResnetModule, self).__init__()
self.input_dims = (input_num_channels, image_height, image_width)
self.using_recurrence = using_recurrence
# Convert text embedding into 16 kernel of size 7x7 specific shape and size
self.dense_text_to_kernel = nn.Linear(text_emb_size, 16 * 3 * 7 * 7)
self.global_kernel = nn.Parameter(torch.FloatTensor(16, 3, 7, 7))
self.norm1 = nn.InstanceNorm2d(input_num_channels)
self.conv1 = nn.Conv2d(input_num_channels, 32, 7, stride=2, padding=3)
h, w = (image_height / 2, image_width / 2)
block1 = nn.Sequential(blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32))
h, w = h / 2, w / 2
block2 = nn.Sequential(blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32))
h, w = h / 2, w / 2
block3 = nn.Sequential(blocks.ResBlockStrided(32, 32), )
h, w = h / 2, w / 2
self.resnet_blocks = nn.Sequential(block1, block2, block3)
self.norm2 = nn.InstanceNorm2d(32)
self.dense = nn.Linear(h * w * 32, image_emb_size)
self.global_id = 0
def __init__(self, image_emb_size, input_num_channels,
image_height, image_width, using_recurrence=False):
super(ImagePositionResnetModule, self).__init__()
self.input_dims = (input_num_channels, image_height, image_width)
self.using_recurrence = using_recurrence
self.norm1 = nn.InstanceNorm2d(input_num_channels)
self.conv1 = nn.Conv2d(input_num_channels, 32, 7, stride=2, padding=3)
h, w = (image_height / 2, image_width / 2)
block1 = nn.Sequential(
blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32)
)
h, w = h / 2, w / 2
block2 = nn.Sequential(
blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32)
)
h, w = h / 2, w / 2
block3 = nn.Sequential(
blocks.ResBlockStrided(32, 32),
)
h, w = h / 2, w / 2
self.resnet_blocks = nn.Sequential(block1, block2, block3)
self.norm2 = nn.InstanceNorm2d(32)
self.dense = nn.Linear(h * w * 32, image_emb_size)
self.global_id = 0
self.ryan_top_layer_conv = nn.Conv2d(32, 1, 1, stride=1)
def __init__(self,
image_emb_size,
input_num_channels,
image_height,
image_width,
text_emb_size,
using_recurrence=False):
super(ImageAttentionResnetModule, self).__init__()
self.input_dims = (input_num_channels, image_height, image_width)
self.using_recurrence = using_recurrence
self.num_attention_heads = 5
self.norm1 = nn.InstanceNorm2d(input_num_channels)
self.conv1 = nn.Conv2d(input_num_channels, 32, 7, stride=2, padding=3)
h, w = (image_height / 2, image_width / 2)
block1 = nn.Sequential(blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32))
h, w = h / 2, w / 2
block2 = nn.Sequential(blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32))
h, w = h / 2, w / 2
block3 = nn.Sequential(blocks.ResBlockStrided(32, 32), )
h, w = h / 2, w / 2
self.num_feature_maps = 32
self.final_h = h
self.final_w = w
self.attention_dense_layers = nn.ModuleList()
for _ in xrange(self.num_attention_heads):
dense_layer = nn.Linear(text_emb_size, self.final_h * self.final_w)
self.attention_dense_layers.append(dense_layer)
self.resnet_blocks = nn.Sequential(block1, block2, block3)
self.final_dense = nn.Linear(h * w * self.num_attention_heads,
image_emb_size)
def __init__(self, image_emb_size, input_num_channels,
image_height, image_width, text_emb_size,
using_recurrence=False):
super(ImageGAResnetModule, self).__init__()
self.input_dims = (input_num_channels, image_height, image_width)
self.using_recurrence = using_recurrence
self.norm1 = nn.InstanceNorm2d(input_num_channels)
self.conv1 = nn.Conv2d(input_num_channels, 32, 7, stride=2, padding=3)
h, w = (image_height / 2, image_width / 2)
block1 = nn.Sequential(
blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32)
)
h, w = h / 2, w / 2
block2 = nn.Sequential(
blocks.ResBlockStrided(32, 32),
blocks.ResBlock(32)
)
h, w = h / 2, w / 2
block3 = nn.Sequential(
blocks.ResBlockStrided(32, 32),
)
h, w = h / 2, w / 2
self.num_feature_maps = 32
self.final_h = h
self.final_w = w
self.resnet_blocks = nn.Sequential(block1, block2, block3)
self.norm2 = nn.InstanceNorm2d(self.num_feature_maps)
self.text_dense = nn.Linear(text_emb_size, self.num_feature_maps)
self.final_dense = nn.Linear(h * w * self.num_feature_maps, image_emb_size)
self.global_id = 0
def __init__(self,
image_emb_size,
input_num_channels,
image_height,
image_width,
using_recurrence=False):
super(ImageResnetModule, self).__init__()
self.input_dims = (input_num_channels, image_height, image_width)
self.using_recurrence = using_recurrence
num_channels = 32
self.norm1 = nn.InstanceNorm2d(input_num_channels)
self.conv1 = nn.Conv2d(input_num_channels,
num_channels,
7,
stride=2,
padding=3)
h, w = (image_height // 2, image_width // 2)
block1 = nn.Sequential(
blocks.ResBlockStrided(num_channels, num_channels),
blocks.ResBlock(num_channels))
h, w = h // 2, w // 2
block2 = nn.Sequential(
blocks.ResBlockStrided(num_channels, num_channels),
blocks.ResBlock(num_channels))
h, w = h // 2, w // 2
block3 = nn.Sequential(
blocks.ResBlockStrided(num_channels, num_channels), )
h, w = h // 2, w // 2
self.resnet_blocks = nn.Sequential(block1, block2, block3)
self.norm2 = nn.InstanceNorm2d(num_channels)
self.dense = nn.Linear(h * w * num_channels, image_emb_size)
self.global_id = 0