def create_cell_generator(num_nodes, reduction):
prefix = 'reduction' if reduction else 'normal'
cell_ops = [
SP1_ops('%s_op_%d_%d' % (prefix, i // 2, i % 2), reduction)
for i in range(2 * num_nodes)
]
connection_hparams = [
D(list(range(i // 2 + 2)),
name='%s_in_%d_%d' % (prefix, i // 2, i % 2))
for i in range(2 * num_nodes)
]
def generate(filters, cell_ratio):
return cell(
lambda: cell_input_fn(filters),
lambda in_id, node_id, op_num: intermediate_node_fn(
reduction, in_id, node_id, op_num, filters, cell_ratio,
cell_ops), lambda: add(2), combine_unused, num_nodes,
connection_hparams)
return generate
def generate_search_space(nodes_per_stage, filters_per_stage,
filter_size_per_stage):
search_space = []
for i in range(len(nodes_per_stage)):
search_space.append(
generate_stage(i, nodes_per_stage[i], filters_per_stage[i],
filter_size_per_stage[i]))
search_space.append(max_pool2d(D([3]), D([2]), D(['SAME'])))
search_space += [
flatten(),
fc_layer(D([1024])),
dropout(D([.5])),
fc_layer(D([10]))
]
return mo.siso_sequential(search_space)
def create_cell_generator(num_nodes):
h_connections = [
Bool(name='in_%d_%d' % (in_id, out_id))
for (in_id, out_id) in itertools.combinations(range(num_nodes + 2), 2)
]
cell_ops = [
D(['conv1', 'conv3', 'max3'], name='node_%d' % i)
for i in range(num_nodes)
]
def generate(filters):
return cell(
lambda channels: mo.siso_sequential(
[conv2d(D([channels]), D([1])),
batch_normalization(),
relu()]),
lambda num_inputs, node_id, channels: intermediate_node_fn(
num_inputs, node_id, channels, cell_ops), concat,
h_connections, 5, filters)
return generate
def stem(filters):
return mo.siso_sequential(
[conv2d(D([filters]), D([3])),
batch_normalization()])
def stem():
return mo.siso_sequential([
conv2d(D([128]), D([3])),
batch_normalization(),
relu(),
])
def relu():
return activation(D(['relu']))