def _linspace(start, stop, num):
# Theano linspace. Behaves similar to np.linspace
start = T.cast(start, theano.config.floatX)
stop = T.cast(stop, theano.config.floatX)
num = T.cast(num, theano.config.floatX)
step = (stop-start)/(num-1)
return T.arange(num, dtype=theano.config.floatX)*step+start
def get_output_for(self, input, **kwargs):
num_feature_maps = input.shape[self.axis]
num_pools = num_feature_maps // self.pool_size
pool_shape = ()
arange_shuffle_pattern = ()
for k in range(self.axis):
pool_shape += (input.shape[k],)
arange_shuffle_pattern += ('x',)
pool_shape += (num_pools, self.pool_size)
arange_shuffle_pattern += ('x', 0)
for k in range(self.axis + 1, input.ndim):
pool_shape += (input.shape[k],)
arange_shuffle_pattern += ('x',)
input_reshaped = input.reshape(pool_shape)
max_indices = T.argmax(input_reshaped, axis=self.axis + 1,
keepdims=True)
arange = T.arange(self.pool_size).dimshuffle(*arange_shuffle_pattern)
mask = T.eq(max_indices, arange).reshape(input.shape)
return input * mask
def _interpolate(im, x, y, out_height, out_width):
# *_f are floats
num_batch, height, width, channels = im.shape
height_f = T.cast(height, theano.config.floatX)
width_f = T.cast(width, theano.config.floatX)
# clip coordinates to [-1, 1]
x = T.clip(x, -1, 1)
y = T.clip(y, -1, 1)
# scale coordinates from [-1, 1] to [0, width/height - 1]
x = (x + 1) / 2 * (width_f - 1)
y = (y + 1) / 2 * (height_f - 1)
# obtain indices of the 2x2 pixel neighborhood surrounding the coordinates;
# we need those in floatX for interpolation and in int64 for indexing. for
# indexing, we need to take care they do not extend past the image.
x0_f = T.floor(x)
y0_f = T.floor(y)
x1_f = x0_f + 1
y1_f = y0_f + 1
x0 = T.cast(x0_f, 'int64')
y0 = T.cast(y0_f, 'int64')
x1 = T.cast(T.minimum(x1_f, width_f - 1), 'int64')
y1 = T.cast(T.minimum(y1_f, height_f - 1), 'int64')
# The input is [num_batch, height, width, channels]. We do the lookup in
# the flattened input, i.e [num_batch*height*width, channels]. We need
# to offset all indices to match the flat version
dim2 = width
dim1 = width*height
base = T.repeat(
T.arange(num_batch, dtype='int64')*dim1, out_height*out_width)
base_y0 = base + y0*dim2
base_y1 = base + y1*dim2
idx_a = base_y0 + x0
idx_b = base_y1 + x0
idx_c = base_y0 + x1
idx_d = base_y1 + x1
# use indices to lookup pixels for all samples
im_flat = im.reshape((-1, channels))
Ia = im_flat[idx_a]
Ib = im_flat[idx_b]
Ic = im_flat[idx_c]
Id = im_flat[idx_d]
# calculate interpolated values
wa = ((x1_f-x) * (y1_f-y)).dimshuffle(0, 'x')
wb = ((x1_f-x) * (y-y0_f)).dimshuffle(0, 'x')
wc = ((x-x0_f) * (y1_f-y)).dimshuffle(0, 'x')
wd = ((x-x0_f) * (y-y0_f)).dimshuffle(0, 'x')
output = T.sum([wa*Ia, wb*Ib, wc*Ic, wd*Id], axis=0)
return output
