def __init__(self, g, k, s, c, h_g, h_l, std, hidden_size, num_classes):
"""
Initialize the recurrent attention model and its
different components.
Args
----
- g: size of the square patches in the glimpses extracted
by the retina.
- k: number of patches to extract per glimpse.
- s: scaling factor that controls the size of successive patches.
- c: number of channels in each image.
- h_g: hidden layer size of the fc layer for `phi`.
- h_l: hidden layer size of the fc layer for `l`.
- std: standard deviation of the Gaussian policy.
- hidden_size: hidden size of the rnn.
- num_classes: number of classes in the dataset.
- num_glimpses: number of glimpses to take per image,
i.e. number of BPTT steps.
"""
super(RecurrentAttention, self).__init__()
self.std = std
self.sensor = glimpse_network(h_g, h_l, g, k, s, c)
self.rnn = core_network(hidden_size, hidden_size)
self.locator = location_network(hidden_size, 2, std)
self.classifier = action_network(hidden_size, num_classes)
self.baseliner = baseline_network(hidden_size, 1)
def test_glimpse():
config, unparsed = get_config()
train_loader, _ = get_train_valid_loader(config.data_dir,
config.batch_size,
config.random_seed,
config.valid_size, config.shuffle,
config.show_sample)
img, label = next(iter(train_loader))
r = retina(g=8, k=1, s=1)
g = glimpse_network(block=BasicBlock,
h_g=128,
h_l=32,
h_s=128,
g=8,
k=1,
s=1,
c=3)
l_t_prev = torch.rand(config.batch_size, 2).uniform_(-1, 1)
size_t_prev = torch.rand(config.batch_size, 5).uniform_(0, 1)
#print(g.feature_extractors["size_32"])
g_t = g(img, l_t_prev, size_t_prev)
return g_t
def __init__(self):
"""
Initialize the recurrent attention model and its
different components.
Args
----
- g: size of the square patches in the glimpses extracted
by the retina.
- k: number of patches to extract per glimpse. (Zoom Values)
- s: scaling factor that controls the size of successive patches.
- c: number of channels in each image.
- h_g: hidden layer size of the fc layer for `phi`.
- h_l: hidden layer size of the fc layer for `l`.
- std: standard deviation of the Gaussian policy.
- hidden_size: hidden size of the rnn.
- num_classes: number of classes in the dataset.
- num_glimpses: number of glimpses to take per image,
i.e. number of BPTT steps.
"""
super(RecurrentSpatialTransformer, self).__init__()
# Crop the image using differential STN at location and zoom
self.stn = stn_zoom(out_height=26, out_width=26)
# In glimpse Network we convert the cropped image in to feature sapce
# using CNN and FC Layer then concatenate it with the previous location
self.sensor = glimpse_network(glimpse_size = 26, h_g = 1024, h_l = 1024, c = 1)
self.rnn1 = nn.GRUCell(2048, 512)
self.fc_rnn = nn.Linear(512, 2048)
self.rnn2 = nn.GRUCell(2048, 512)
self.classifier = classification_network(input_size = 512, hidden_size = 1024, num_classes = 10)
self.locator = location_network(input_size = 512, hidden_size = 1024)
self.context = context_network(100,512)
def main():
# load images
imgs = []
paths = [data_dir + './lenna.jpg', data_dir + './cat.jpg']
for i in range(len(paths)):
img = img2array(paths[i], desired_size=[512, 512], expand=True)
imgs.append(torch.from_numpy(img))
imgs = torch.cat(imgs)
B, H, W, C = imgs.shape
loc = torch.Tensor([[-1., 1.], [-1., 1.]])
imgs, loc = Variable(imgs), Variable(loc)
sensor = glimpse_network(h_g=128, h_l=128, g=64, k=3, s=2, c=3)
g_t = sensor(imgs, loc)
rnn = core_network(input_size=256, hidden_size=256)
h_t = Variable(torch.zeros(g_t.shape[0], 256))
h_t = rnn(g_t, h_t)
classifier = action_network(256, 10)
a_t = classifier(h_t)
loc_net = location_network(256, 2, 0.11)
mu, l_t = loc_net(h_t)
base = baseline_network(256, 1)
b_t = base(h_t)
print("g_t: {}".format(g_t.shape))
print("h_t: {}".format(h_t.shape))
print("l_t: {}".format(l_t.shape))
print("a_t: {}".format(a_t.shape))
print("b_t: {}".format(b_t.shape))
def __init__(self):
super(RecurrentSpatialTransformer, self).__init__()
self.context = context_network(100, 512)
self.rnn = nn.GRU(512, 512, 1)
self.locator = location_network(input_size=512, hidden_size=1024)
# Crop the image using differential STN at location and zoom
self.stn = stn_zoom(out_height=26, out_width=26)
self.sensor = glimpse_network(glimpse_size=26, h_g=512, h_l=1024, c=1)
self.classifier = classification_network(input_size=512,
hidden_size=1024,
num_classes=10)
self.context_2 = context_network_2(12, 1024)
def __init__(self,
g,
k,
s,
c,
h_g,
h_l,
std,
hidden_size,
num_classes):
"""
Initialize the recurrent attention model and its
different components.
Args
----
- g: size of the square patches in the glimpses extracted
by the retina.
- k: number of patches to extract per glimpse.
- s: scaling factor that controls the size of successive patches.
- c: number of channels in each image.
- h_g: hidden layer size of the fc layer for `phi`.
- h_l: hidden layer size of the fc layer for `l`.
- std: standard deviation of the Gaussian policy.
- hidden_size: hidden size of the rnn.
- num_classes: number of classes in the dataset.
- num_glimpses: number of glimpses to take per image,
i.e. number of BPTT steps.
"""
super(RecurrentAttention, self).__init__()
self.std = std
# feature extraction on x at location l_t_prev, and combine the information
# of the image patches and their locations
self.sensor = glimpse_network(h_g, h_l, g, k, s, c)
# combine the information of current patch_info g_t and the hidden info from the last step h_t
self.rnn = core_network(hidden_size, hidden_size)
# Uses the internal state `h_t` of the core network to produce
# the location coordinates `l_t` for the next time step.
# only take the new h_t as input, without the old l_t_prev
self.locator = location_network(hidden_size, 2, std)
self.classifier = action_network(hidden_size, num_classes)
self.baseliner = baseline_network(hidden_size, 1)
def __init__(self,
g,
k,
s,
c,
h_g,
h_l,
std,
hidden_size,
num_classes,
kernel_size,
num_stacks,
stack_attn_mode):
"""
Initialize the recurrent attention model and its
different components.
Args
----
- g: size of the square patches in the glimpses extracted
by the retina.
- k: number of patches to extract per glimpse.
- s: scaling factor that controls the size of successive patches.
- c: number of channels in each image.
- h_g: hidden layer size of the fc layer for `phi`.
- h_l: hidden layer size of the fc layer for `l`.
- std: standard deviation of the Gaussian policy.
- hidden_size: hidden size of the rnn.
- num_classes: number of classes in the dataset.
- num_glimpses: number of glimpses to take per image,
i.e. number of BPTT steps.
- kernel_size: list of int, convolutional kernel size in stacked RAM
- num_stacks: int, number of layers in stacked RAM
- stack_attn_mode: str, values chosen from 'separate', 'concat', 'combine'
"""
super(RecurrentAttention, self).__init__()
self.std = std
self.num_stacks = num_stacks
self.stack_attn_mode = stack_attn_mode
self.sensor = nn.ModuleList([
glimpse_network(h_g, h_l, g, k, s, c, kernel_size)
for _ in range(num_stacks)
])
self.rnn = nn.ModuleList([
core_network(h_g + h_l, hidden_size)
for _ in range(num_stacks)
])
if stack_attn_mode == 'separate':
self.locator = nn.ModuleList([
location_network(hidden_size, 2, std)
for _ in range(num_stacks)
])
elif stack_attn_mode == 'concat':
self.locator = location_network(hidden_size * num_stacks, 2, std)
elif stack_attn_mode == 'combine':
self.locator = location_network(hidden_size * num_stacks, 2 * num_stacks, std)
else:
raise 'Unknown stack_attn_mode [%s]' % stack_attn_mode
self.baseliner = nn.ModuleList([
baseline_network(hidden_size, 1) for _ in range(num_stacks)
])
self.classifier = action_network(hidden_size * num_stacks, num_classes)
def __init__(self, g, c, image_size, std, hidden_size, num_classes,
config):
"""
Initialize the recurrent attention model and its
different components.
Args
----
- g: size of the square patches in the glimpses extracted
by the retina.
- c: number of channels in each image.
- image_size: a tuple: (H x W)
- std: standard deviation of the Gaussian policy.
- hidden_size: hidden size of the rnn.
- num_classes: number of classes in the dataset.
- num_glimpses: number of glimpses to take per image,
i.e. number of BPTT steps.
"""
super(RecurrentAttention, self).__init__()
# when the locations l is defined by a Gaussian distribution
self.std = std
# when the locations l is defined by a symmetry stable distribution
self.alpha = config.alpha
self.gamma = config.gamma
self.config = config
self.context = context_network(c, config.kernel_size, hidden_size)
self.sensor = glimpse_network(hidden_size, g, c, config)
self.rnn = core_network(hidden_size, hidden_size, config)
self.top_down_locator = location_network(hidden_size, 2, config)
self.bot_up_locator = Levy_bottom_up_generator(config.batch_size,
image_size, config)
self.combine_location = combine_location_network(hidden_size, config)
self.classifier = action_network(hidden_size, num_classes)
self.baseliner = baseline_network(hidden_size, 1)
# something for initialzing subroutine
dtype = (torch.cuda.FloatTensor
if self.config.use_gpu else torch.FloatTensor)
# derivative of Saliecy map
self.derivative_y = torch.tensor([-1, 0, 1]).reshape(1, 1, 3,
1).type(dtype)
self.derivative_x = torch.t(torch.tensor([-1, 0,
1])).reshape(1, 1, 1,
3).type(dtype)
# a weighted saliency s gauged at a fixation center
self.gaussian_kernel_sigma = math.floor(
image_size[0] /
12) # in the paper, /6 but pytorch does not accept such big kernel
gaussian_kernel_size = self.gaussian_kernel_sigma * 2 + 1
tmp_x, tmp_y = torch.meshgrid(
torch.arange(-self.gaussian_kernel_sigma,
self.gaussian_kernel_sigma + 1).type(dtype),
torch.arange(-self.gaussian_kernel_sigma,
self.gaussian_kernel_sigma + 1).type(dtype))
self.gaussian_kernel = torch.exp(
-(tmp_x.type(dtype)**2 + tmp_y.type(dtype)**2) /
self.gaussian_kernel_sigma**2).reshape(1, 1, gaussian_kernel_size,
gaussian_kernel_size)
