def _pooling_function(self, inputs, pool_size, strides, border_mode, dim_ordering):
'''
This pooling operation is basically the sum of both MaxPooling and AveragePooling outputs.
'''
output1 = K.pool2d(inputs, pool_size, strides, border_mode, dim_ordering, pool_mode='max')
output2 = K.pool2d(inputs, pool_size, strides, border_mode, dim_ordering, pool_mode='avg')
return output1 + output2
def _pooling_function(self, back_end, inputs, pool_size, strides,
border_mode, dim_ordering):
output1 = K.pool2d(inputs, pool_size, strides,
border_mode, dim_ordering, pool_mode='max')
output2 = K.pool2d(inputs, pool_size, strides,
border_mode, dim_ordering, pool_mode='ave')
return output1 + output2
def weighted_bce_dice_loss(y_true, y_pred):
y_true = K.cast(y_true, 'float32')
y_pred = K.cast(y_pred, 'float32')
# if we want to get same size of output, kernel size must be odd number
averaged_mask = K.pool2d(
y_true, pool_size=(11, 11), strides=(1, 1), padding='same', pool_mode='avg')
border = K.cast(K.greater(averaged_mask, 0.005), 'float32') * K.cast(K.less(averaged_mask, 0.995), 'float32')
weight = K.ones_like(averaged_mask)
w0 = K.sum(weight)
weight += border * 2
w1 = K.sum(weight)
weight *= (w0 / w1)
loss = weighted_bce_loss(y_true, y_pred, weight) + \
weighted_dice_loss(y_true, y_pred, weight)
return loss
def call(self, x, mask=None):
if K.image_dim_ordering == "th":
_, f, r, c = self.shape
else:
_, r, c, f = self.shape
half_n = self.n // 2
squared = K.square(x)
pooled = K.pool2d(squared, (half_n, half_n), strides=(1, 1),
padding="same", pool_mode="avg")
if K.image_dim_ordering == "th":
summed = K.sum(pooled, axis=1, keepdims=True)
averaged = (self.alpha / self.n) * K.repeat_elements(summed, f, axis=1)
else:
summed = K.sum(pooled, axis=3, keepdims=True)
averaged = (self.alpha / self.n) * K.repeat_elements(summed, f, axis=3)
denom = K.pow(self.k + averaged, self.beta)
return x / denom
def get_output(self, train=False):
print "OccMaxPool", self.output_shape, self.input_shape
num_tracks = (
self.input_shape[1] if self.num_tracks == 'full'
else self.num_tracks
)
num_bases = (
self.input_shape[3] if self.num_bases == 'full'
else self.num_bases
)
X = self.get_input(train)
X = K.permute_dimensions(X, (0,2,1,3))
rv = K.pool2d(
X,
pool_size=(num_tracks, num_bases),
strides=(num_tracks, num_bases),
pool_mode='max'
)
rv = K.permute_dimensions(rv, (0,2,1,3))
return rv
def call(self, x, mask=None):
if K.image_dim_ordering == "th":
_, f, r, c = self.shape
else:
_, r, c, f = self.shape
half_n = self.n // 2
squared = K.square(x)
pooled = K.pool2d(squared, (half_n, half_n),
strides=(1, 1),
padding="same",
pool_mode="avg")
if K.image_dim_ordering == "th":
summed = K.sum(pooled, axis=1, keepdims=True)
averaged = (self.alpha / self.n) * K.repeat_elements(
summed, f, axis=1)
else:
summed = K.sum(pooled, axis=3, keepdims=True)
averaged = (self.alpha / self.n) * K.repeat_elements(
summed, f, axis=3)
denom = K.pow(self.k + averaged, self.beta)
return x / denom
def routing_step(feature_map, pool_size, strides, score_map):
# Compute average vector
if score_map is None:
# in first iteration we cannot compute weights average
# just take input feature map
weighted_fm = feature_map
else:
# normalize scores
normalized_scores = CapsulePooling2D.normalize_pool_map(
score_map, pool_size=pool_size, strides=strides, temperature=2)
weighted_fm = normalized_scores * feature_map
avg_feature_map = K.pool2d(weighted_fm,
pool_size=pool_size,
strides=strides,
pool_mode='avg',
padding='same')
avg_feature_map = CapsulePooling2D.squash_fm(avg_feature_map)
avg_feature_columns = K.resize_images(avg_feature_map,
pool_size[0],
pool_size[1],
data_format='channels_last')
# fit caps columns to feature map
input_shape = K.int_shape(feature_map)
avg_feature_columns = avg_feature_columns[:, :input_shape[
1], :input_shape[2], :input_shape[3]]
# compute V * Vs
new_score_map = K.sum(feature_map * avg_feature_columns,
-1,
keepdims=True)
if score_map is None:
score_map = new_score_map
else:
score_map = score_map + new_score_map
return score_map, weighted_fm
def _do_max_pooling(feature_matrix):
"""Performs max-pooling on 2-D feature map.
M = number of rows before pooling
N = number of columns before pooling
m = number of rows after pooling
n = number of columns after pooling
:param feature_matrix: Feature map as M-by-N numpy array.
:return: feature_matrix: New feature map (m-by-n).
"""
this_feature_matrix = numpy.expand_dims(feature_matrix, axis=0)
this_feature_matrix = numpy.expand_dims(this_feature_matrix, axis=-1)
feature_tensor = K.pool2d(x=K.variable(this_feature_matrix),
pool_mode='max',
pool_size=(2, 2),
strides=(2, 2),
padding='valid',
data_format='channels_last')
return feature_tensor.eval(session=K.get_session())[0, ..., 0]
def _batch_gen(self, traces, spikes, shape, batch_size, nb_steps, margin):
# Apply the error margin by max pooling the spikes once up-front.
lens = [len(x) for x in spikes]
for l in np.unique(lens):
idxs, = np.where(lens == l)
x = np.vstack([spikes[i] for i in idxs])
x = K.variable(x.astype(np.float32))
x = K.expand_dims(K.expand_dims(x, axis=0), axis=-1)
x = K.pool2d(x, (1, margin + 1), padding='same')
x = K.get_value(x[0, :, :, 0])
for i in range(x.shape[0]):
spikes[idxs[i]] = x[i]
while True:
idxs = np.arange(len(traces))
cidxs = cycle(rng.choice(idxs, len(idxs), replace=False))
for _ in range(nb_steps):
# Empty batches (traces and spikes).
tb = np.zeros((batch_size, ) + shape, dtype=np.float64)
sb = np.zeros((batch_size, ) + shape, dtype=np.uint8)
for bidx in range(batch_size):
# Dataset and sample indices.
idx = next(cidxs)
# Pick start and end point around positive spike index.
x0 = rng.randint(0, len(spikes[idx]) - shape[0])
x1 = x0 + shape[0]
# Populate batch.
tb[bidx] = traces[idx][x0:x1]
sb[bidx] = spikes[idx][x0:x1]
yield tb, sb
def _get_laplacian_pyramid(self, inputs: plaidml.tile.Value) -> List[plaidml.tile.Value]:
""" Obtain the Laplacian Pyramid.
Parameters
----------
inputs: :class:`plaidml.tile.Value`
The input batch of images to run through the Laplacian Pyramid
Returns
-------
list
The tensors produced from the Laplacian Pyramid
"""
pyramid = []
current = inputs
for _ in range(self._max_levels):
gauss = self._conv_gaussian(current)
diff = current - gauss
pyramid.append(diff)
current = K.pool2d(gauss, (2, 2), strides=(2, 2), padding="valid", pool_mode="avg")
pyramid.append(current)
return pyramid
def call(self, inputs):
# Input = [X^H, X^L]
assert len(inputs) == 2
high_input, low_input = inputs
# High -> High conv
high_to_high = K.conv2d(high_input,
self.high_to_high_kernel,
strides=self.strides,
padding=self.padding,
data_format="channels_last")
# High -> Low conv
high_to_low = K.pool2d(high_input, (2, 2),
strides=(2, 2),
pool_mode="avg")
high_to_low = K.conv2d(high_to_low,
self.high_to_low_kernel,
strides=self.strides,
padding=self.padding,
data_format="channels_last")
# Low -> High conv
low_to_high = K.conv2d(low_input,
self.low_to_high_kernel,
strides=self.strides,
padding=self.padding,
data_format="channels_last")
low_to_high = K.repeat_elements(low_to_high, 2,
axis=1) # Nearest Neighbor Upsampling
low_to_high = K.repeat_elements(low_to_high, 2, axis=2)
# Low -> Low conv
low_to_low = K.conv2d(low_input,
self.low_to_low_kernel,
strides=self.strides,
padding=self.padding,
data_format="channels_last")
# Cross Add
high_add = high_to_high + low_to_high
low_add = high_to_low + low_to_low
return [high_add, low_add]
def weighted_dice_loss(y_true, y_pred):
y_true = K.cast(y_true, 'float32')
y_pred = K.cast(y_pred, 'float32')
if K.int_shape(y_pred)[1] == 128:
kernel_size = 11
elif K.int_shape(y_pred)[1] == 256:
kernel_size = 21
elif K.int_shape(y_pred)[1] == 512:
kernel_size = 21
elif K.int_shape(y_pred)[1] == 1024:
kernel_size = 41
else:
raise ValueError('Unexpected image size')
averaged_mask = K.pool2d(
y_true, pool_size=(kernel_size, kernel_size), strides=(1, 1), padding='same', pool_mode='avg')
border = K.cast(K.greater(averaged_mask, 0.005), 'float32') * K.cast(K.less(averaged_mask, 0.995), 'float32')
weight = K.ones_like(averaged_mask)
w0 = K.sum(weight)
weight += border * 2
w1 = K.sum(weight)
weight *= (w0 / w1)
loss = 1 - weighted_dice_coeff(y_true, y_pred, weight)
return loss
def my_loss(x_true, merged_decode): #batch, n = merged_decode.get_shape() x_decode = merged_decode[:, :n] z = merged_decode[:, n:] #h_xi_zj = gaussian_cond_ent(x_true, z, invert_sigmoid = True, subtract_log_det =True) # NOW this is a conditional entropy not MI mi_ji = gaussian_mi(x_true, z) # prev #OLD calculation of h(Xi - g(Z)) term within screening loss #xi_entropy_est = error_entropy(x_true, 0, skew = skew, kurt = kurt) #mi_xi_z = error_entropy(x_true, x_decode, invert_sigmoid = True, subtract_log_det=True, skew = skew, kurt = kurt) max_mi = K.pool2d(x=K.reshape(mi_ji, (1, 1, K.cast(z.shape[1], 'int32'), n)), pool_size=(z.shape[1], 1), strides=(1, 1), padding='valid', data_format='channels_first', pool_mode='max') max_mi = K.reshape(max_mi, (n,)) # TAKING MIN using smooth min (-alpha) #max_mi = tf.divide(K.sum(tf.multiply(mi_ji, K.exp(alpha*mi_ji)), axis = 0), K.sum(K.exp(alpha*mi_ji), axis = 0)) return K.sum(mi_ji) - K.sum(max_mi)
def _pooling_function(self, inputs, pool_size, strides,
padding, data_format):
channels = inputs.get_shape()[-1]
holder = []
for c in range(channels):
init = tf.expand_dims(inputs[:,:,:,c], axis=-1)
if self.rank == 2:
outputs = K.conv2d(
init,
tf.expand_dims(self.kernel[:,:,c,:], 2),
strides=self.strides,
padding=self.padding,
data_format=self.data_format)
if self.with_softmax:
exp = tf.exp(outputs)
prob = exp / tf.reduce_sum(exp, -1, keep_dims=True)
outputs = outputs * prob
holder.append(tf.expand_dims(K.mean(outputs, axis=-1), axis=-1))
output = K.pool2d(inputs, self.kernel_size, strides,
padding, data_format, pool_mode='max')
concat = tf.concat(holder, axis=-1)
out = concat + output
return out
def call(self, x, **kwargs):
#make sure the inputs is a list?
out_shape = K.shape(x)
input_shape = K.int_shape(x)
#note the input_shape = (samples,cols,rows,channels)
num_rows = input_shape[1]
num_cols = input_shape[2]
#divided pooling_list
row_length = [num_rows / i for i in self.pooling_list]
col_length = [num_cols / i for i in self.pooling_list]
outputs = []
for i, pooling_size in enumerate(self.pooling_list):
#这里需要注意了，看看原文
h_wid = math.ceil(row_length[i])
w_wid = math.ceil(col_length[i])
h_str = math.floor(row_length[i])
w_str = math.floor(col_length[i])
out = K.pool2d(x=x,
pool_size=(h_wid, w_wid),
strides=(h_str, w_str))
out = K.reshape(x=out, shape=(out_shape[0], -1))
outputs.append(out)
return K.concatenate(outputs, 1)
def call(self, inputs, **kwargs):
context = inputs[0]
x, y = (inputs[1][:, 0], inputs[1][:, 1])
# preprocessing states to get memory
# Need this for later computations
batch_size = tf.shape(context)[0]
# global read operation
# r_t = read(M_t) ; output dimension of r_t = self.units
first_conv = K.conv2d(self.neural_map,
self.conv_kernel1,
strides=(1, 1),
padding='valid')
#second_conv = K.conv2d(first_conv, self.conv_kernel2, strides=(1, 1), padding='valid')
pool_conv = K.pool2d(first_conv,
pool_size=(2, 2),
strides=(1, 1),
padding='valid',
pool_mode='avg')
flatten_conv = K.reshape(pool_conv, (1, -1))
dense1_conv = K.dot(flatten_conv, self.conv_dense1)
#dense2_conv = K.dot(dense1_conv, self.conv_dense2)
r_t = dense1_conv
r_t = K.tile(r_t, K.stack([batch_size, 1]))
#r_t = K.print_tensor(r_t, "r_t: ")
# context read operation
# c_t = context(M_t, s_t, r_t)
q_t = K.concatenate([context, r_t]) # [1x(s+c)]
q_t = K.dot(q_t, self.context_kernel) # [1xc]
#q_t = K.print_tensor(q_t, "q_t: ")
# at = K.exp(K.dot(q_t, memory)) # [1xhxw]
# at_sum = K.sum(at)
# at_sum_repeated = K.repeat (at_sum, (self.memory_size[0], self.memory_size[1]))
# at /= at_sum # [1xhxw]
# Compute attention with softmax, use batch_dot function
# Keras is garbage, and so am I
at = tf.tensordot(q_t, self.neural_map, axes=[[1], [3]])
at /= tf.norm(at, keepdims=True)
at = K.exp(at)
#at = tf.Print(at, [at], "at1: ", summarize=1000)
# Annoying conversion
neural_map_tensor = tf.convert_to_tensor(self.neural_map)
#neural_map_tensor = tf.Print(neural_map_tensor, [neural_map_tensor], "neural_map_tensor: ", summarize=1000)
#at = K.sum(q_t_repeated * neural_map_tensor, axis=3)
at_sum = K.sum(at, axis=2, keepdims=True)
at_sum = K.sum(at_sum, axis=3, keepdims=True)
at_sum = K.repeat_elements(at_sum, self.memory_size[0], 2)
at_sum = K.repeat_elements(at_sum, self.memory_size[1], 3)
#at_sum = tf.Print(at_sum, [at_sum], "at_sum: ", summarize=1000)
at = at / at_sum
at = K.repeat_elements(at, self.units, 1)
at = K.permute_dimensions(at, (0, 2, 3, 1))
#at = K.print_tensor(at, "at: ")
#at = tf.Print(at, [at], "at2: ", summarize=1000)
c_t = at * neural_map_tensor
#c_t = tf.Print(c_t, [c_t], "c_t1: ", summarize=1000)
c_t = K.sum(c_t, axis=1)
#c_t = tf.Print(c_t, [c_t], "c_t2: ", summarize=1000)
c_t = K.sum(c_t, axis=1)
#c_t = tf.Print(c_t, [c_t], "c_t3: ", summarize=1000)
# Computing write value
# m_t+1_x,y = write(s_t, r_t, c_t, m_t_x,y)
s_t = K.dot(context, self.write_kernel) # [1xc]
#s_t = tf.Print(s_t, [s_t], "s_t: ", summarize=1000)
global_imp = K.batch_dot(s_t, r_t, axes=1) # [1x1]
#global_imp = tf.Print(global_imp, [global_imp], "global_imp: ", summarize=1000)
local_imp = K.batch_dot(s_t, c_t, axes=1) # [1x1]
#local_imp = tf.Print(local_imp, [local_imp], "local_imp: ", summarize=1000)
# Convert batched x,y coordinates into indices for memory
# We want to index into memory by doing something like
# memory[sample_number, y, x] so that we get a {units} dimensional
# vector out. When we do this for every sample in the batch, we'll
# end up with a {number_batches, units} dimensional matrix.
# First we copy and reshape the memory into {batch_size x H x W x units}
zeros = K.zeros_like(y)
idx = K.stack([zeros, y, x], axis=1) # {batch_size x 3}
# And index into the memory using tf.gather_nd
mem_t = tf.gather_nd(neural_map_tensor, idx) # {batch_size x units}
#mem_t = tf.Print(mem_t, [mem_t], "mem_t before update: ", summarize=1000)
# This line has a divide-by-zero case, in (local_imp + global_imp).
#mem_t += (local_imp / (local_imp + global_imp)) * K.dot((mem_t - s_t), self.write_update)
#Check where local_imp + global_imp is zero, then just increase it by 1.
#local_imp = tf.Print(local_imp, [local_imp], "local_imp: ", summarize=1000)
#global_imp = tf.Print(global_imp, [global_imp], "global_imp: ", summarize=1000)
combined_imp = local_imp + global_imp
more_zeros = K.zeros_like(combined_imp)
eq = tf.equal(combined_imp, more_zeros)
combined_imp = tf.where(eq, combined_imp + 1., combined_imp)
#combined_imp = tf.Print(combined_imp, [combined_imp], "combined_imp: ", summarize=1000)
mem_t += (local_imp / combined_imp) * K.dot(
(mem_t - s_t), self.write_update)
mem_t /= tf.norm(mem_t, keepdims=True)
#mem_t = tf.Print(mem_t, [mem_t], "mem_t after update: ", summarize=1000)
# Update memory.
updates = []
neural_map_updated = tf.scatter_nd_update(self.neural_map, idx, mem_t)
updates.append((self.neural_map, neural_map_updated))
self.add_update(updates, inputs)
#neural_map_updated = K.print_tensor(neural_map_updated, "neural_map_updated: ")
# memory[:, :] = mem_t
return c_t + mem_t
def conv2d_offset_average_pool_eval(_, padding, x):
"""Perform an average pooling operation on x"""
return K.pool2d(x, (1, 1),
strides=(3, 3),
padding=padding,
pool_mode='avg')
def call(self, input_):
"""
Forward pass through the layer.
Args:
input_: the input tensor to pass through the pyramid pooling module
Returns:
the output tensor from the pyramid pooling module
"""
# the shape to up-sample pooled tensors to
if self.data_format == 'channels_last':
channel_axis = -1
output_shape = K.int_shape(input_)[1:-1]
if self.data_format == 'channels_first':
channel_axis = 1
output_shape = K.int_shape(input_)[2:]
# create a list of output tensors to concatenate together
output_tensors = [input_]
# iterate over the bin sizes in the pooling module
for (level, bin_size) in enumerate(self.bin_sizes):
# pass the inputs through the pooling layer with the given bin
# size, i.e., a square kernel with side matching the bin size and
# a matching stride
x = K.pool2d(
input_,
tuple(dim // bin_size for dim in output_shape),
padding=self.padding,
pool_mode=self.pool_mode,
data_format=self.data_format,
)
# pass the pooled valued through a 1 x 1 convolution
x = K.conv2d(
x,
self.kernels[level],
strides=(1, 1),
padding=self.padding,
data_format=self.data_format,
)
# if use bias, apply the bias to the convolved outputs
if self.use_bias:
x = K.bias_add(
x,
self.biases[level],
data_format=self.data_format,
)
# apply the activation function if there is one
if self.activation is not None:
x = self.activation(x)
# if data format is channels first, have to permute before resizing
# because resize expects channels last
if self.data_format == 'channels_first':
x = K.permute_dimensions(x, [0, 2, 3, 1])
# up-sample the outputs back to the input shape
x = tf.compat.v1.image.resize_bilinear(x, output_shape)
# if data format is channels first, have to permute after resizing
# because resize output channels last
if self.data_format == 'channels_first':
x = K.permute_dimensions(x, [0, 3, 1, 2])
# concatenate the output tensor with the stack of output tensors
output_tensors += [x]
# concatenate the output tensors before returning
return K.concatenate(output_tensors, axis=channel_axis)
def pooling_function(inputs, pool_size, strides, padding, data_format):
return -K.pool2d(
-inputs, pool_size, strides, padding, data_format, pool_mode='max')
def l2_norm(x):
x = K.square(x)
x = K.pool2d(x, pool_size=(3, 3), strides=(1, 1), padding='same', data_format='channels_last', pool_mode='avg')
x = x * 9
x = K.sqrt(x)
return x
def call(self, x):
output = K.pool2d(x, self.pool_size, self.strides,
self.padding, self.data_format,
pool_mode='max')
return output
def _pooling_function(self, inputs, pool_size, strides,
border_mode, dim_ordering):
output = K.pool2d(inputs, pool_size, strides,
border_mode, dim_ordering, pool_mode='max')
return output
def conv2d_offset_average_pool_eval(_, padding, x):
"""Perform an average pooling operation on x"""
return K.pool2d(x, (1, 1), strides=(3, 3), padding=padding, pool_mode='avg')
def min_pool2d(x):
min_x = -K.pool2d(-x, pool_size=(2, 2), strides=(1, 1), padding='same')
return min_x
def mean2d(y):
y = K.pool2d(y, (self.kernel_size[0], 1), pool_mode = 'avg', padding = 'same')
y = K.pool2d(y, (1, self.kernel_size[1]), pool_mode = 'avg', padding = 'same')
return y
def call(self, inputs, states, training=None):
# inputs are going to come in as a tuple of size 2
# where inputs[0] is the context prior to the neural
# map, and inputs[1] will be another tuple - (x, y)
context = inputs
x, y = (self.pos_input[:, 0], self.pos_input[:, 1])
# x, y = (K.expand_dims(x, -1), K.expand_dims(y, -1))
# preprocessing states to get memory
memory = states[1:((self.memory_size[0] * self.memory_size[1]) + 1)]
memory = K.transpose(memory)
memory = K.reshape(
memory, (-1, self.units, self.memory_size[0], self.memory_size[1]))
# Need this for later computations
batch_size = K.shape(memory)[0]
# global read operation
# r_t = read(M_t) ; output dimension of r_t = self.units
first_conv = K.conv2d(memory,
self.conv_kernel1,
strides=(1, 1),
padding='same',
data_format='channels_first')
second_conv = K.conv2d(first_conv,
self.conv_kernel2,
strides=(1, 1),
padding='valid',
data_format='channels_first')
pool_conv = K.pool2d(second_conv,
pool_size=(2, 2),
strides=(1, 1),
padding='valid',
data_format='channels_first',
pool_mode='avg')
flatten_conv = K.batch_flatten(pool_conv)
dense1_conv = K.dot(flatten_conv, self.conv_dense1)
dense2_conv = K.dot(dense1_conv, self.conv_dense2)
r_t = dense2_conv
# context read operation
# c_t = context(M_t, s_t, r_t)
q_t = K.concatenate([context, r_t]) # [1x(s+c)]
q_t = K.dot(q_t, self.context_kernel) # [1xc]
# at = K.exp(K.dot(q_t, memory)) # [1xhxw]
# at_sum = K.sum(at)
# at_sum_repeated = K.repeat (at_sum, (self.memory_size[0], self.memory_size[1]))
# at /= at_sum # [1xhxw]
# Compute attention with softmax, use batch_dot function
# Keras is garbage, and so am I
q_t_repeated = K.expand_dims(q_t, axis=2)
q_t_repeated = K.expand_dims(q_t_repeated, axis=3)
q_t_repeated = K.repeat_elements(q_t_repeated, self.memory_size[0], 2)
q_t_repeated = K.repeat_elements(q_t_repeated, self.memory_size[1], 3)
at = K.sum(q_t_repeated * memory, axis=1)
at_sum = K.sum(at, axis=1, keepdims=True)
at_sum = K.sum(at_sum, axis=2, keepdims=True)
at_sum = K.repeat_elements(at_sum, self.memory_size[0], 1)
at_sum = K.repeat_elements(at_sum, self.memory_size[1], 2)
at = at / at_sum
at = K.expand_dims(at, axis=1)
at = K.repeat_elements(at, self.units, 1)
c_t = K.sum(at * memory, axis=2)
c_t = K.sum(c_t, axis=2)
# Computing write value
# m_t+1_x,y = write(s_t, r_t, c_t, m_t_x,y)
s_t = K.dot(context, self.write_kernel) # [1xc]
global_imp = K.batch_dot(s_t, r_t, axes=1) # [1x1]
local_imp = K.batch_dot(s_t, c_t, axes=1) # [1x1]
# Convert batched x,y coordinates into indices for memory
# We want to index into memory by doing something like
# memory[sample_number, y, x] so that we get a {units} dimensional
# vector out. When we do this for every sample in the batch, we'll
# end up with a {number_batches, units} dimensional matrix.
# First we copy and reshape the memory into {batch_size x H x W x units}
mem_t = K.reshape(
memory,
(batch_size, self.memory_size[0], self.memory_size[1], self.units))
# Now we generate the indices
B = tf.range(0, batch_size) # {batch_size}
idx = tf.stack([B, y, x], axis=1) # {batch_size x 3}
# And index into the memory using tf.gather_nd
mem_t = tf.gather_nd(mem_t, idx) # {batch_size x units}
mem_t += (local_imp / (local_imp + global_imp)) * K.dot(
(mem_t - s_t), self.write_update)
# We have to reshape the memory b/c of how we calculated indices:
memory = K.reshape(
memory,
(batch_size, self.memory_size[0], self.memory_size[1], self.units))
# While loop to update memory.
i = tf.constant(0)
c = lambda i: tf.less(i, batch_size)
def body(i):
memory[i, y[i], x[i]] = mem_t[i]
return tf.add(i, 1)
tf.while_loop(c, body, [i])
# Would love to just use this, but tf has a stick in its butt
# memory = tf.scatter_nd_update(memory, idx, mem_t)
# memory[:, :] = mem_t
# update states
mem_t = K.reshape(mem_t, (1, self.units))
new_states = states[:((self.memory_size[0] * self.memory_size[1]) + 1)]
new_states[0] = r_t
new_states[((x - 1) * self.units) + y] = mem_t
return c_t, new_states
def conv2d_offset_max_pool_eval(_, padding, x):
"""Perform a max pooling operation on x"""
return K.pool2d(x, (1, 1), strides=(3, 3), padding=padding, pool_mode='max')
def call(self, PMandHazeIMGandA):
patchMap = PMandHazeIMGandA[0]
hazeImg = PMandHazeIMGandA[1]
A = PMandHazeIMGandA[2]
x = tf.cast(patchMap, tf.int32)
s = K.int_shape(patchMap)
mapList = []
for i in range(1, 121):
if i == 1:
mapList.append(K.greater(1.3, patchMap))
else:
mapList.append(K.equal(x, i))
pmbox = K.concatenate(mapList, axis=3)
pmbox = tf.cast(pmbox, tf.float32)
patchConvList = []
patchMinList = []
"""
hazeImg=-hazeImg
for i in range(1,121):
recoveri=K.pool2d(hazeImg, pool_size=(i,i), strides=(1,1),
padding='same', data_format=None,
pool_mode='max')
recoveri=-recoveri
recoveri=K.min(recoveri,axis=3,keepdims=True)
patchMinList.append(recoveri)
"""
hazeImgM = K.min(hazeImg, axis=3, keepdims=True)
hazeImgM = -hazeImgM
for i in range(1, 121):
if i <= 5:
recoveri = K.pool2d(hazeImgM,
pool_size=(i, i),
strides=(1, 1),
padding='same',
data_format=None,
pool_mode='max')
#recoveri=-recoveri
#recoveri=K.min(recoveri,axis=3,keepdims=True)
patchMinList.append(recoveri)
else:
#print(len(patchMinList),'len patchMinList')
#recoveri=-patchMinList[i-2]
recoveri = K.pool2d(patchMinList[i - 3],
pool_size=(3, 3),
strides=(1, 1),
padding='same',
data_format=None,
pool_mode='max')
#recoveri=-recoveri
#recoveri=K.min(recoveri,axis=3,keepdims=True)
patchMinList.append(recoveri)
#for i in range(120):
# patchMinList[i]=K.min(-patchMinList[i],axis=3,keepdims=True)
#hazeImg=-hazeImg
rcbox = K.concatenate(patchMinList, axis=3)
#print(K.int_shape(rcbox),'int shape')
rcbox = -rcbox
A2 = K.expand_dims(A, 2)
A2 = K.expand_dims(A2, 3)
A2 = K.tile(A2, [1, 480, 640, 120])
t = rcbox / A2
w = 31 / 32
t = w * t
t = 1 - t
pm_rc0 = pmbox * t
pm_rc = K.sum(pm_rc0, axis=3, keepdims=True)
pm_rc = K.clip(pm_rc, 0.0, 1.0)
return [pm_rc]
def call(self, ins):
u = ins[0]
img = ins[1]
### g learned by net
g1 = K.conv2d(self.img, self.kernel_1, padding='same')
g1 = K.bias_add(g1, self.bias_1)
g1 = K.relu(g1)
g1 = K.conv2d(g1, self.kernel_2, padding='same')
g1 = K.bias_add(g1, self.bias_2)
g1 = K.relu(g1)
g2 = K.pool2d(g1, (2,2), padding='same')
g2 = K.conv2d(g2, self.kernel_3, padding='same')
g2 = K.bias_add(g2, self.bias_3)
g2 = K.relu(g2)
g2 = K.conv2d(g2, self.kernel_4, padding='same')
g2 = K.bias_add(g2, self.bias_4)
g2 = K.relu(g2)
g3 = K.pool2d(g2, (2,2), padding='same')
g3 = K.conv2d(g3, self.kernel_5, padding='same')
g3 = K.bias_add(g3, self.bias_5)
g3 = K.relu(g3)
g3 = K.conv2d(g3, self.kernel_6, padding='same')
g3 = K.bias_add(g3, self.bias_6)
g3 = K.relu(g3)
g3 = K.conv2d(g3, self.kernel_7, padding='same')
g3 = K.bias_add(g3, self.bias_7)
g3 = K.relu(g3)
g3 = K.conv2d(g3, self.kernel_8, padding='same')
g3 = K.bias_add(g3, self.bias_8)
g3 = K.relu(g3)
g4 = K.resize_images(g3, 2, 2, data_format='channels_last', interpolation='bilinear')
g4 = K.concatenate([g2, g4])
g4 = K.conv2d(g4, self.kernel_9, padding='same')
g4 = K.bias_add(g4, self.bias_9)
g4 = K.relu(g4)
g4 = K.conv2d(g4, self.kernel_10, padding='same')
g4 = K.bias_add(g4, self.bias_10)
g4 = K.relu(g4)
g5 = K.resize_images(g4, 2, 2, data_format='channels_last', interpolation='bilinear')
g5 = K.concatenate([g1, g5])
g5 = K.conv2d(g5, self.kernel_11, padding='same')
g5 = K.bias_add(g5, self.bias_11)
g5 = K.relu(g5)
g5 = K.conv2d(g5, self.kernel_12, padding='same')
g5 = K.bias_add(g5, self.bias_12)
g = K.sigmoid(g5)
### grad(g)
g_x = K.conv2d(g, kXC, padding='same')
g_x = scale(g_x, self.shp, self.rhp, self.dx)
g_y = K.conv2d(g, kYC, padding='same')
g_y = scale(g_y, self.shp, self.rhp, self.dy)
### transport - upwind
xp = K.conv2d(u, xKP, padding='same')
xn = K.conv2d(u, xKN, padding='same')
yp = K.conv2d(u, xYP, padding='same')
yn = K.conv2d(u, xYN, padding='same')
fxp = K.relu( g_x)
fxn = -1.0 * K.relu( -1.0 * g_x)
fyp = K.relu( g_y)
fyn = -1.0 * K.relu( -1.0 * g_y)
xpp = fxp*xp
xnn = fxn*xn
ypp = fyp*yp
ynn = fyn*yn
xterms = xpp + xnn
xterms = scale(xterms, self.shp, self.rhp, self.dx)
yterms = ypp + ynn
yterms = scale(yterms, self.shp, self.rhp, self.dy)
transport = xterms + yterms
### curvature - learned
grad_u = K.conv2d(u, self.kernel_k1, padding='same')
norm_grad_u = K.sqrt( K.epsilon() + K.sum( K.square(grad_u), axis=-1, keepdims=True) )
grad_u = grad_u / (norm_grad_u + K.epsilon())
kappa = K.conv2d(grad_u, self.kernel_k2, padding='same')
curvature = g*kappa*norm_grad_u
### balloon
balloon = g*norm_grad_u
return u + K.constant(self.dt)*( curvature * self.alpha \
+ transport * self.beta \
+ balloon * self.gamma )
def _pooling_function(self, inputs, pool_size, strides, padding,
data_format, dilation_rate):
"""We illustrate how this method works with an example. Let's have a
dilated kernel with pool_size=(3,2), dilation_rate=(2,4). The *
represent the non-zero elements
* 0 0 0 *
0 0 0 0 0
* 0 0 0 *
0 0 0 0 0
* 0 0 0 *
Let's have the following image (each dot represents a pixel)
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
· · · · · · · · ·
The pooling kernel will slide over the image, similarly to convolution,
computing a pooling result at each step. For example, for the vertical
slide
dr = 0 dr = 1 dr = 2
X · · · X · · · · · · · · · · · · · · · · · · · · · ·
· · · · · · · · · X · · · X · · · · · · · · · · · · ·
X · · · X · · · · · · · · · · · · · X · · · X · · · ·
· · · · · · · · · X · · · X · · · · · · · · · · · · ·
X · · · X · · · · · · · · · · · · · X · · · X · · · ·
· · · · · · · · · , X · · · X · · · · , · · · · · · · · · , etc
· · · · · · · · · · · · · · · · · · X · · · X · · · ·
· · · · · · · · · · · · · · · · · · · · · · · · · · ·
· · · · · · · · · · · · · · · · · · · · · · · · · · ·
· · · · · · · · · · · · · · · · · · · · · · · · · · ·
· · · · · · · · · · · · · · · · · · · · · · · · · · ·
· · · · · · · · · · · · · · · · · · · · · · · · · · ·
Ideally, this process should be implemented at low level in Theano and
TensorFlow. This here is a high-level implementation, but in order to
preserve some performance, we reuse K.pool2d() as much as possible.
For that, we are going to subsample each axis of the input D times if
the dilation factor is D. For the example above (subsampling marked
with X):
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · , · · · · · · · · · , · · · · · · · · · , · · · · · · · · · ,
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · , · · · · · · · · · , · · · · · · · · · , · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
· · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · ·
X · · · X · · · X · X · · · X · · · · · X · · · X · · · · · X · · · X ·
We then apply K.pool2d() to each of the subsamplings above (with
padding='same' so that each block produces a block of the same size).
The results are restacked to form an array of the same size as inputs.
Finally, if the user wanted only padding='valid', we crop out the
borders of the array.
"""
if (strides != 1) and (strides != (1, 1)):
raise Exception('Only implemented for strides=1 or strides=(1,1)')
sz = K.shape(inputs)
if data_format == 'channels_first': # (batch,chan,row,col)
nbatch = sz[0]
nchan = sz[1]
nrows = sz[2]
ncols = sz[3]
elif data_format == 'channels_last': # (batch,row,col,chan)
nbatch = sz[0]
nrows = sz[1]
ncols = sz[2]
nchan = sz[3]
else:
raise ValueError(
'Expected data_format to be channels_first or channels_last')
# allocate space for output
outputs = K.zeros(shape=sz.eval())
# compute slice objects. Each slice object will be used to split the
# input into a block. Each block will be pooled with dilation=1 (no
# dilation). The overall effect is like pooling the whole image with
# dilation>1
# first, we compute the slices to split the input in horizontal and
# vertical blocks, separately
block_slices_row = [
slice(i, nrows, dilation_rate[0]) for i in range(dilation_rate[0])
]
block_slices_col = [
slice(j, ncols, dilation_rate[1]) for j in range(dilation_rate[1])
]
# then, we make all combinations of row-blocks and col-blocks, to get
# all the blocks that split the inputs. Now we can iterate all the
# blocks and process them separately
block_slices_combinatorial = itertools.product(
*[block_slices_row, block_slices_col])
total_row_padding_before = 0
total_row_padding_after = 0
total_col_padding_before = 0
total_col_padding_after = 0
for sl in block_slices_combinatorial:
# DEBUG: sl = list(itertools.islice(block_slices_combinatorial, 1))[0]
# extend the slice with the batch and channel
if data_format == 'channels_first': # (batch,chan,row,col)
block_slice = list((slice(0, nbatch, 1), slice(0, nchan, 1)) +
sl)
elif data_format == 'channels_last': # (batch,row,col,chan)
block_slice = list((slice(0, nbatch, 1), ) + sl +
(slice(0, nchan, 1), ))
# extract block
block = inputs[block_slice]
#print(block.eval()) # DEBUG
# pool each block without dilation
block_pooled = K.pool2d(block,
pool_size,
strides=(1, 1),
padding='same',
data_format=data_format,
pool_mode='max')
#print(block_pooled.eval()) # DEBUG
# compute amount of padding that this blocks contributes
((row_padding_before, row_padding_after),
(col_padding_before, col_padding_after)) = self._block_padding(
block=block_pooled,
pool_size=pool_size,
data_format=data_format)
total_row_padding_before += row_padding_before
total_row_padding_after += row_padding_after
total_col_padding_before += col_padding_before
total_col_padding_after += col_padding_after
# put block into output
if (K.backend() == 'theano'):
outputs = T.set_subtensor(outputs[block_slice], block_pooled)
elif (K.backend() == 'tensorflow'):
outputs[block_slice].assign(block_pooled)
else:
raise Exception('not implemented')
# row blocks contribute redundantly to the amount of padding columns,
# and column blocks to padding rows. We divide by the number of blocks
# alon each axis to obtain the actual amount of padding in the input
total_row_padding_before /= dilation_rate[1]
total_row_padding_after /= dilation_rate[1]
total_col_padding_before /= dilation_rate[0]
total_col_padding_after /= dilation_rate[0]
total_row_padding_before = int(total_row_padding_before)
total_row_padding_after = int(total_row_padding_after.eval())
total_col_padding_before = int(total_col_padding_before)
total_col_padding_after = int(total_col_padding_after.eval())
# remove the border where the kernel didn't fully overlap the input
if padding == 'valid':
if data_format == 'channels_first': # (batch,chan,row,col)
outputs = outputs[:, :, total_row_padding_before:(
nrows - total_row_padding_after),
total_col_padding_before:(
ncols - total_col_padding_after)]
elif data_format == 'channels_last': # (batch,row,col,chan)
outputs = outputs[:, total_row_padding_before:(
nrows - total_row_padding_after),
total_col_padding_before:(
ncols - total_col_padding_after), :]
# return tensor
return outputs
def _pooling_function(self, inputs, pool_size, strides,
padding, data_format):
output = K.pool2d(K.square(inputs), pool_size, strides,
padding, data_format, pool_mode='avg')
return K.sqrt(output + K.epsilon())
def _nms(heat, kernel=3):
hmax = K.pool2d(heat, (kernel, kernel), padding='same', pool_mode='max')
keep = K.cast(K.equal(hmax, heat), K.floatx())
return heat * keep
def _pooling_function(self, x, name=None):
#b = K.shape(x)[0]
input_shape = K.int_shape(x)
b, r, c, channel = input_shape[0], input_shape[1], input_shape[
2], input_shape[3]
pr, pc = self.pool_size
sr, sc = self.strides
# compute number of windows
num_r = math.ceil(r / sr) if self.padding == 'SAME' else r // sr
num_c = math.ceil(c / sc) if self.padding == 'SAME' else c // sc
def _mid_pool(inputs, is_train):
input_shape = inputs.shape
batch = input_shape[0]
# reshape
w = np.transpose(inputs, (0, 3, 1, 2))
w = np.reshape(w, (batch * channel, r, c))
def pool(x):
# sort
size = pr * pc
x_flat = np.reshape(x, [size])
x_sorted = np.argsort(-x_flat)
p = x_sorted[math.ceil(size / 2)]
mid_matrix = np.zeros((pr, pc))
mid_matrix[p // pr, p % pc] = 1.0
return mid_matrix
re = np.zeros((batch * channel, num_r, num_c), dtype=np.float32)
# extract with pool_size
for i in range(num_r):
for j in range(num_c):
crop = np.zeros((batch * channel, pr, pc))
crop[:, :, :] = w[:, i * sr:i * sr + pr,
j * sc:j * sc + pc]
# pool
outs = np.array(list(map(pool, crop)))
if is_train:
re[:, i, j] = (crop * outs).max(axis=(1, 2))
else:
re[:, i, j] = (crop * outs).sum(axis=(1, 2))
# reshape
re = np.reshape(re, (batch, channel, num_r, num_c))
re = np.transpose(re, (0, 2, 3, 1))
return re
def custom_grad(op, grad):
if self.data_format == 'NHWC':
ksizes = [1, self.pool_size[0], self.pool_size[1], 1]
strides = [1, self.strides[0], self.strides[1], 1]
else:
ksizes = [1, 1, self.pool_size[0], self.pool_size[1]]
strides = [1, 1, self.strides[0], self.strides[1]]
#return gen_nn_ops.max_pool_grad(
return gen_nn_ops.avg_pool_grad(
array_ops.shape(op.inputs[0]),
grad,
ksizes,
strides,
self.padding,
data_format=self.data_format), K.tf.constant(0.0)
def py_func(func, inp, Tout, stateful=True, name=None, grad=None):
# Need to generate a unique name to avoid duplicates:
rnd_name = 'MidPooling2DGrad' + str(np.random.randint(0, 1E+8))
K.tf.RegisterGradient(rnd_name)(grad)
g = K.tf.get_default_graph()
with g.gradient_override_map({"PyFunc": rnd_name}):
return K.tf.py_func(func,
inp,
Tout,
stateful=stateful,
name=name)
def _mid_range_pool(x, name=None):
with ops.name_scope(name, "mod", [x]) as name:
z = py_func(_mid_pool, [x, K.learning_phase()], [K.tf.float32],
name=name,
grad=custom_grad)[0]
z.set_shape((b, num_r, num_c, channel))
return z
"""
Mixed Pooling
"""
max_pool = K.pool2d(x,
self.pool_size,
strides=self.strides,
padding=self.padding.lower(),
pool_mode="max")
mid_pool = _mid_range_pool(x, name)
def _train_pool(max_pool, mid_pool):
self.alpha = random.uniform(0, 1)
self.alpha_frequencies[0] += self.alpha
self.alpha_frequencies[1] += 1 - self.alpha
return self.alpha * max_pool + (1 - self.alpha) * mid_pool
def _test_pool(max_pool, mid_pool):
return K.switch(
K.less(self.alpha_frequencies[0], self.alpha_frequencies[1]),
mid_pool, max_pool)
outs = K.in_train_phase(_train_pool(max_pool, mid_pool),
_test_pool(max_pool, mid_pool))
return outs
def call(self, inputs):
boxes = inputs[0]
# Feature Maps. List of feature maps from different level of the
# feature pyramid. Each is [batch, height, width, channels]
score_maps = inputs[1:]
# Assign each ROI to a level in the pyramid based on the ROI area.
y1, x1, y2, x2 = tf.split(boxes, 4, axis=2)
h = y2 - y1
w = x2 - x1
# Equation 1 in the Feature Pyramid Networks paper. Account for
# the fact that our coordinates are normalized here.
# e.g. a 224x224 ROI (in pixels) maps to P4
image_area = tf.cast(self.image_shape[0] * self.image_shape[1],
tf.float32)
roi_level = log2_graph(tf.sqrt(h * w) / (224.0 / tf.sqrt(image_area)))
roi_level = tf.minimum(
5, tf.maximum(2, 4 + tf.cast(tf.round(roi_level), tf.int32)))
roi_level = tf.squeeze(roi_level, 2)
# Loop through levels and apply ROI pooling to each. P2 to P5.
pooled = []
box_to_level = []
for i, level in enumerate(range(2, 6)):
ix = tf.where(tf.equal(roi_level, level))
level_boxes = tf.gather_nd(boxes, ix)
# Box indicies for crop_and_resize.
box_indices = tf.cast(ix[:, 0], tf.int32)
# Keep track of which box is mapped to which level
box_to_level.append(ix)
# Stop gradient propogation to ROI proposals
level_boxes = tf.stop_gradient(level_boxes)
box_indices = tf.stop_gradient(box_indices)
# Here we use the simplified approach of a single value per bin,
# which is how it's done in tf.crop_and_resize()
# Result: [batch * num_boxes, pool_height, pool_width, channels]
pooled.append(
tf.image.crop_and_resize(
score_maps[i],
level_boxes,
box_indices,
[self.pool_shape * self.k, self.pool_shape * self.k],
method="bilinear"))
# Pack pooled features into one tensor
pooled = tf.concat(pooled, axis=0)
# position-sensitive ROI pooling + classify
score_map_bins = []
for channel_step in range(self.k * self.k):
bin_x = K.variable(int(channel_step % self.k) * self.pool_shape,
dtype='int32')
bin_y = K.variable(int(channel_step / self.k) * self.pool_shape,
dtype='int32')
channel_indices = K.variable(list(
range(channel_step * self.channel_num,
(channel_step + 1) * self.channel_num)),
dtype='int32')
croped = tf.image.crop_to_bounding_box(
tf.gather(pooled, indices=channel_indices, axis=-1), bin_y,
bin_x, self.pool_shape, self.pool_shape)
# [pool_shape, pool_shape, channel_num] ==> [1,1,channel_num] ==> [1, channel_num]
croped_mean = K.pool2d(croped, (self.pool_shape, self.pool_shape),
strides=(1, 1),
padding='valid',
data_format="channels_last",
pool_mode='avg')
# [batch * num_rois, 1,1,channel_num] ==> [batch * num_rois, 1, channel_num]
croped_mean = K.squeeze(croped_mean, axis=1)
score_map_bins.append(croped_mean)
# [batch * num_rois, k^2, channel_num]
score_map_bins = tf.concat(score_map_bins, axis=1)
# [batch * num_rois, k*k, channel_num] ==> [batch * num_rois,channel_num]
# because "keepdims=False", the axis 1 will not keep. else will be [batch * num_rois,1,channel_num]
pooled = K.sum(score_map_bins, axis=1)
# Pack box_to_level mapping into one array and add another
# column representing the order of pooled boxes
box_to_level = tf.concat(box_to_level, axis=0)
box_range = tf.expand_dims(tf.range(tf.shape(box_to_level)[0]), 1)
box_to_level = tf.concat([tf.cast(box_to_level, tf.int32), box_range],
axis=1)
# Rearrange pooled features to match the order of the original boxes
# Sort box_to_level by batch then box index
# TF doesn't have a way to sort by two columns, so merge them and sort.
sorting_tensor = box_to_level[:, 0] * 100000 + box_to_level[:, 1]
ix = tf.nn.top_k(sorting_tensor,
k=tf.shape(box_to_level)[0]).indices[::-1]
ix = tf.gather(box_to_level[:, 2], ix)
pooled = tf.gather(pooled, ix)
# Re-add the batch dimension
pooled = tf.expand_dims(pooled, 0)
return pooled
def min_max_pool2d(x):
max_x = K.pool2d(x, pool_size=(2, 2), strides=(2, 2))
min_x = min_pool2d(x)
return K.concatenate([max_x, min_x], axis=3) # concatenate on channel
def conv2d_offset_max_pool_eval(_, padding, x):
"""Perform a max pooling operation on x"""
return K.pool2d(x, (1, 1),
strides=(3, 3),
padding=padding,
pool_mode='max')
def Kpool(x, pool_size):
return K.pool2d(x, pool_size, pool_mode='avg')
def points_nms(heat, kernel=2):
# kernel must be 2
hmax = K.pool2d(heat, (kernel, kernel), padding='same', pool_mode='max')
keep = K.cast(K.equal(hmax, heat), K.floatx())
return heat * keep
def min_max_pool2d(x):
max_x = K.pool2d(x, pool_size=(7, 7), strides=(2, 2))
min_x = -K.pool2d(-x, pool_size=(7, 7), strides=(2, 2))
return K.concatenate([max_x, min_x], axis=1) # concatenate on channel
