def _maybe_calibrate_size(self, layers, out_filters, is_training):
"""Makes sure layers[0] and layers[1] have the same shapes."""
hw = [self._get_HW(layer) for layer in layers]
c = [self._get_C(layer) for layer in layers]
with tf.variable_scope("calibrate"):
x = layers[0]
if hw[0] != hw[1]:
assert hw[0] == 2 * hw[1]
with tf.variable_scope("pool_x"):
#x = tf.nn.relu(x)
x = self._factorized_reduction(x, out_filters, 2, is_training)
elif c[0] != out_filters:
with tf.variable_scope("pool_x"):
w = create_weight("w", [1, 1, c[0], out_filters])
#x = tf.nn.relu(x)
x = tf.nn.conv2d(x, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
y = layers[1]
if c[1] != out_filters:
with tf.variable_scope("pool_y"):
w = create_weight("w", [1, 1, c[1], out_filters])
#y = tf.nn.relu(y)
y = tf.nn.conv2d(y, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
y = batch_norm(y, is_training, data_format=self.data_format)
return [x, y]
def _enas_cell(self, x, curr_cell, prev_cell, op_id, out_filters):
"""Performs an enas operation specified by op_id."""
num_possible_inputs = curr_cell + 1
with tf.variable_scope("avg_pool"):
avg_pool = tf.layers.average_pooling2d(
x, [3, 3], [1, 1], "SAME", data_format=self.actual_data_format)
avg_pool_c = self._get_C(avg_pool)
if avg_pool_c != out_filters:
with tf.variable_scope("conv"):
w = create_weight(
"w", [num_possible_inputs, avg_pool_c * out_filters])
w = w[prev_cell]
w = tf.reshape(w, [1, 1, avg_pool_c, out_filters])
#avg_pool = tf.nn.relu(avg_pool)
avg_pool = tf.nn.conv2d(avg_pool, w, strides=[1, 1, 1, 1],
padding="SAME", data_format=self.data_format)
avg_pool = batch_norm(avg_pool, is_training=True,
data_format=self.data_format)
with tf.variable_scope("max_pool"):
max_pool = tf.layers.max_pooling2d(
x, [3, 3], [1, 1], "SAME", data_format=self.actual_data_format)
max_pool_c = self._get_C(max_pool)
if max_pool_c != out_filters:
with tf.variable_scope("conv"):
w = create_weight(
"w", [num_possible_inputs, max_pool_c * out_filters])
w = w[prev_cell]
w = tf.reshape(w, [1, 1, max_pool_c, out_filters])
#max_pool = tf.nn.relu(max_pool)
max_pool = tf.nn.conv2d(max_pool, w, strides=[1, 1, 1, 1],
padding="SAME", data_format=self.data_format)
max_pool = batch_norm(max_pool, is_training=True,
data_format=self.data_format)
x_c = self._get_C(x)
if x_c != out_filters:
with tf.variable_scope("x_conv"):
w = create_weight("w", [num_possible_inputs, x_c * out_filters])
w = w[prev_cell]
w = tf.reshape(w, [1, 1, x_c, out_filters])
#x = tf.nn.relu(x)
x = tf.nn.conv2d(x, w, strides=[1, 1, 1, 1], padding="SAME",
data_format=self.data_format)
x = batch_norm(x, is_training=True, data_format=self.data_format)
out = [
self._enas_conv(x, curr_cell, prev_cell, 3, out_filters),
self._enas_conv(x, curr_cell, prev_cell, 5, out_filters),
avg_pool,
max_pool,
x,
]
out = tf.stack(out, axis=0)
out = out[op_id, :, :, :, :]
return out
def _factorized_reduction(self, x, out_filters, stride, is_training):
"""Reduces the shape of x without information loss due to striding."""
assert out_filters % 2 == 0, (
"Need even number of filters when using this factorized reduction.")
if stride == 1:
with tf.variable_scope("path_conv"):
inp_c = self._get_C(x)
w = create_weight("w", [1, 1, inp_c, out_filters])
x = tf.nn.conv2d(x, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
return x
stride_spec = self._get_strides(stride)
# Skip path 1
path1 = tf.nn.avg_pool(
x, [1, 1, 1, 1], stride_spec, "VALID", data_format=self.data_format)
with tf.variable_scope("path1_conv"):
inp_c = self._get_C(path1)
w = create_weight("w", [1, 1, inp_c, out_filters // 2])
path1 = tf.nn.conv2d(path1, w, [1, 1, 1, 1], "VALID",
data_format=self.data_format)
# Skip path 2
# First pad with 0"s on the right and bottom, then shift the filter to
# include those 0"s that were added.
if self.data_format == "NHWC":
pad_arr = [[0, 0], [0, 1], [0, 1], [0, 0]]
path2 = tf.pad(x, pad_arr)[:, 1:, 1:, :]
concat_axis = 3
else:
pad_arr = [[0, 0], [0, 0], [0, 1], [0, 1]]
path2 = tf.pad(x, pad_arr)[:, :, 1:, 1:]
concat_axis = 1
path2 = tf.nn.avg_pool(
path2, [1, 1, 1, 1], stride_spec, "VALID", data_format=self.data_format)
with tf.variable_scope("path2_conv"):
inp_c = self._get_C(path2)
w = create_weight("w", [1, 1, inp_c, out_filters // 2])
path2 = tf.nn.conv2d(path2, w, [1, 1, 1, 1], "VALID",
data_format=self.data_format)
# Concat and apply BN
final_path = tf.concat(values=[path1, path2], axis=concat_axis)
final_path = batch_norm(final_path, is_training,
data_format=self.data_format)
return final_path
def _fixed_conv(self, x, f_size, out_filters, stride, is_training,
stack_convs=2):
"""Apply fixed convolution.
Args:
stacked_convs: number of separable convs to apply.
"""
for conv_id in range(stack_convs):
inp_c = self._get_C(x)
if conv_id == 0:
strides = self._get_strides(stride)
else:
strides = [1, 1, 1, 1]
with tf.variable_scope("sep_conv_{}".format(conv_id)):
w_depthwise = create_weight("w_depth", [f_size, f_size, inp_c, 1])
w_pointwise = create_weight("w_point", [1, 1, inp_c, out_filters])
#x = tf.nn.relu(x)
x = tf.nn.separable_conv2d(
x,
depthwise_filter=w_depthwise,
pointwise_filter=w_pointwise,
strides=strides, padding="SAME", data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
return x
def _maybe_calibrate_size(self, layers, out_filters, is_training):
"""Makes sure layers[0] and layers[1] have the same shapes."""
#hw = [self._get_HW(layer) for layer in layers]
c = [self._get_C(layer) for layer in layers]
print(c)
with tf.variable_scope("calibrate"):
x = layers[0]
if c[0] != out_filters:
with tf.variable_scope("shuffle_x"):
x = self._upsampler(x)
y = layers[1]
if c[1] != out_filters:
with tf.variable_scope("pool_y"):
w = create_weight("w", [1, 1, c[1], out_filters])
#y = tf.nn.relu(y)
y = tf.nn.conv2d(y,
w, [1, 1, 1, 1],
"SAME",
data_format=self.data_format)
y = batch_norm(y,
is_training,
data_format=self.data_format)
return [x, y]
def _enas_layer(self, layer_id, prev_layers, arc, out_filters):
"""
Args:
layer_id: current layer
prev_layers: cache of previous layers. for skip connections
start_idx: where to start looking at. technically, we can infer this
from layer_id, but why bother...
"""
assert len(prev_layers) == 2, "need exactly 2 inputs"
layers = [prev_layers[0], prev_layers[1]]
layers = self._maybe_calibrate_size(layers, out_filters, is_training=True)
used = []
for cell_id in range(self.num_cells):
prev_layers = tf.stack(layers, axis=0)
with tf.variable_scope("cell_{0}".format(cell_id)):
with tf.variable_scope("x"):
x_id = arc[4 * cell_id]
x_op = arc[4 * cell_id + 1]
x = prev_layers[x_id, :, :, :, :]
x = self._enas_cell(x, cell_id, x_id, x_op, out_filters)
x_used = tf.one_hot(x_id, depth=self.num_cells + 2, dtype=tf.int32)
with tf.variable_scope("y"):
y_id = arc[4 * cell_id + 2]
y_op = arc[4 * cell_id + 3]
y = prev_layers[y_id, :, :, :, :]
y = self._enas_cell(y, cell_id, y_id, y_op, out_filters)
y_used = tf.one_hot(y_id, depth=self.num_cells + 2, dtype=tf.int32)
out = x + y
used.extend([x_used, y_used])
layers.append(out)
used = tf.add_n(used)
indices = tf.where(tf.equal(used, 0))
indices = tf.to_int32(indices)
indices = tf.reshape(indices, [-1])
num_outs = tf.size(indices)
out = tf.stack(layers, axis=0)
out = tf.gather(out, indices, axis=0)
inp = prev_layers[0]
if self.data_format == "NHWC":
N = tf.shape(inp)[0]
H = tf.shape(inp)[1]
W = tf.shape(inp)[2]
C = tf.shape(inp)[3]
out = tf.transpose(out, [1, 2, 3, 0, 4])
out = tf.reshape(out, [N, H, W, num_outs * out_filters])
elif self.data_format == "NCHW":
N = tf.shape(inp)[0]
C = tf.shape(inp)[1]
H = tf.shape(inp)[2]
W = tf.shape(inp)[3]
out = tf.transpose(out, [1, 0, 2, 3, 4])
out = tf.reshape(out, [N, num_outs * out_filters, H, W])
else:
raise ValueError("Unknown data_format '{0}'".format(self.data_format))
with tf.variable_scope("final_conv"):
w = create_weight("w", [self.num_cells + 2, out_filters * out_filters])
w = tf.gather(w, indices, axis=0)
w = tf.reshape(w, [1, 1, num_outs * out_filters, out_filters])
#out = tf.nn.relu(out)
out = tf.nn.conv2d(out, w, strides=[1, 1, 1, 1], padding="SAME",
data_format=self.data_format)
out = batch_norm(out, is_training=True, data_format=self.data_format)
out = tf.reshape(out, tf.shape(prev_layers[0]))
return out
def _fixed_layer(self, layer_id, prev_layers, arc, out_filters, stride,
is_training, normal_or_reduction_cell="normal"):
"""
Args:
prev_layers: cache of previous layers. for skip connections
is_training: for batch_norm
"""
assert len(prev_layers) == 2
layers = [prev_layers[0], prev_layers[1]]
layers = self._maybe_calibrate_size(layers, out_filters,
is_training=is_training)
with tf.variable_scope("layer_base"):
x = layers[1]
inp_c = self._get_C(x)
w = create_weight("w", [1, 1, inp_c, out_filters])
#x = tf.nn.relu(x)
x = tf.nn.conv2d(x, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
layers[1] = x
used = np.zeros([self.num_cells + 2], dtype=np.int32)
f_sizes = [3, 5]
for cell_id in range(self.num_cells):
with tf.variable_scope("cell_{}".format(cell_id)):
x_id = arc[4 * cell_id]
used[x_id] += 1
x_op = arc[4 * cell_id + 1]
x = layers[x_id]
x_stride = stride if x_id in [0, 1] else 1
with tf.variable_scope("x_conv"):
if x_op in [0, 1]:
f_size = f_sizes[x_op]
x = self._fixed_conv(x, f_size, out_filters, x_stride, is_training)
elif x_op in [2, 3]:
inp_c = self._get_C(x)
if x_op == 2:
x = tf.layers.average_pooling2d(
x, [3, 3], [x_stride, x_stride], "SAME",
data_format=self.actual_data_format)
else:
x = tf.layers.max_pooling2d(
x, [3, 3], [x_stride, x_stride], "SAME",
data_format=self.actual_data_format)
if inp_c != out_filters:
w = create_weight("w", [1, 1, inp_c, out_filters])
#x = tf.nn.relu(x)
x = tf.nn.conv2d(x, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
else:
inp_c = self._get_C(x)
if x_stride > 1:
assert x_stride == 2
x = self._factorized_reduction(x, out_filters, 2, is_training)
if inp_c != out_filters:
w = create_weight("w", [1, 1, inp_c, out_filters])
#x = tf.nn.relu(x)
x = tf.nn.conv2d(x, w, [1, 1, 1, 1], "SAME", data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
if (x_op in [0, 1, 2, 3] and
self.drop_path_keep_prob is not None and
is_training):
x = self._apply_drop_path(x, layer_id)
y_id = arc[4 * cell_id + 2]
used[y_id] += 1
y_op = arc[4 * cell_id + 3]
y = layers[y_id]
y_stride = stride if y_id in [0, 1] else 1
with tf.variable_scope("y_conv"):
if y_op in [0, 1]:
f_size = f_sizes[y_op]
y = self._fixed_conv(y, f_size, out_filters, y_stride, is_training)
elif y_op in [2, 3]:
inp_c = self._get_C(y)
if y_op == 2:
y = tf.layers.average_pooling2d(
y, [3, 3], [y_stride, y_stride], "SAME",
data_format=self.actual_data_format)
else:
y = tf.layers.max_pooling2d(
y, [3, 3], [y_stride, y_stride], "SAME",
data_format=self.actual_data_format)
if inp_c != out_filters:
w = create_weight("w", [1, 1, inp_c, out_filters])
#y = tf.nn.relu(y)
y = tf.nn.conv2d(y, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
y = batch_norm(y, is_training, data_format=self.data_format)
else:
inp_c = self._get_C(y)
if y_stride > 1:
assert y_stride == 2
y = self._factorized_reduction(y, out_filters, 2, is_training)
if inp_c != out_filters:
w = create_weight("w", [1, 1, inp_c, out_filters])
#y = tf.nn.relu(y)
y = tf.nn.conv2d(y, w, [1, 1, 1, 1], "SAME",
data_format=self.data_format)
y = batch_norm(y, is_training, data_format=self.data_format)
if (y_op in [0, 1, 2, 3] and
self.drop_path_keep_prob is not None and
is_training):
y = self._apply_drop_path(y, layer_id)
out = x + y
layers.append(out)
out = self._fixed_combine(layers, used, out_filters, is_training,
normal_or_reduction_cell)
return out
def _model(self, x, is_training, reuse=False):
"""Compute the logits given the inputs."""
if self.fixed_arc is None:
is_training = True
with tf.variable_scope(self.name, reuse=reuse):
# the first two inputs
with tf.variable_scope("stem_conv"):
w = create_weight("w", [3, 3, 3, self.out_filters])
x = tf.nn.conv2d(
x, w, [1, 1, 1, 1], "SAME", data_format=self.data_format)
#x = batch_norm(x, is_training, data_format=self.data_format)
if self.data_format == "NHWC":
split_axis = 3
elif self.data_format == "NCHW":
split_axis = 1
else:
raise ValueError("Unknown data_format '{0}'".format(self.data_format))
layers = [x, x]
stem_out = x
# building layers in the micro space
out_filters = self.out_filters
for layer_id in range(self.num_layers + 2):
with tf.variable_scope("layer_{0}".format(layer_id)):
if layer_id not in self.pool_layers:
if self.fixed_arc is None:
x = self._enas_layer(
layer_id, layers, self.normal_arc, out_filters)
else:
x = self._fixed_layer(
layer_id, layers, self.normal_arc, out_filters, 1, is_training,
normal_or_reduction_cell="normal")
else:
#out_filters *= 2
if self.fixed_arc is None:
#x = self._factorized_reduction(x, out_filters, 2, is_training)
layers = [layers[-1], x]
x = self._enas_layer(
layer_id, layers, self.reduce_arc, out_filters)
else:
x = self._fixed_layer(
layer_id, layers, self.reduce_arc, out_filters, 2, is_training,
normal_or_reduction_cell="normal")
print("Layer {0:>2d}: {1}".format(layer_id, x))
layers = [layers[-1], x]
with tf.variable_scope("global_conv"):
w = create_weight("w", [3, 3, out_filters, out_filters])
x = tf.nn.conv2d(
x, w, [1, 1, 1, 1], "SAME", data_format=self.data_format)
x = batch_norm(x, is_training, data_format=self.data_format)
x = x + stem_out
with tf.variable_scope("upsample_1"):
w = create_weight("w", [3, 3, out_filters, out_filters * 4])
x = tf.nn.conv2d(
x, w, [1, 1, 1, 1], "SAME", data_format=self.data_format)
x = pixel_shuffler(x)
with tf.variable_scope("upsample_2"):
w = create_weight("w", [3, 3, out_filters, out_filters * 4])
x = tf.nn.conv2d(
x, w, [1, 1, 1, 1], "SAME", data_format=self.data_format)
x = pixel_shuffler(x)
with tf.variable_scope("output"):
w = create_weight("w", [3, 3, out_filters, 3])
x = tf.nn.conv2d(
x, w, [1, 1, 1, 1], "SAME", data_format=self.data_format)
# scale to [0, 255]
#x = tf.clip_by_value(x * self.std + self.mean, 0, 255)
#x = x * self.std + self.mean
return x