def from_json_loads(json_data):
net_info = CellNetworkInfo()
if isinstance(json_data, list):
net_info['master'] = LayerInfoList.from_json_loads(json_data)
net_info.cell_names.append('master')
else:
for key in ['master', 'normal', 'reduction']:
jd = json_data.get(key, None)
if jd:
net_info[key] = LayerInfoList.from_json_loads(jd)
net_info.cell_names.append(key)
return net_info
def darts_rnn_base_cell_info(next_id=0,
input_ids=[0, 1],
end_merge=LayerTypes.MERGE_WITH_CAT):
"""
See implementation of PetridishRNNCell to see an example of
how this list of info is used.
The first two info are x_and_h and init_layer, which is a
projected x_and_h multiplied with gate.
The rest of layers use specified operation to morph the layers.
This is DARTS from the paper writing
"""
LT = LayerTypes
l_info = LayerInfoList([
LayerInfo(input_ids[0]), # next_id + 0
LayerInfo(input_ids[1]), # next_id + 1
LayerInfo(next_id + 2,
inputs=[input_ids[1]],
operations=[LT.FC_RELU_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 3,
inputs=[next_id + 2],
operations=[LT.FC_RELU_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 4,
inputs=[next_id + 3],
operations=[LT.FC_TANH_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 5,
inputs=[next_id + 4],
operations=[LT.FC_RELU_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 6,
inputs=[next_id + 5],
operations=[LT.FC_RELU_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 7,
inputs=[next_id + 2],
operations=[LT.FC_IDEN_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 8,
inputs=[next_id + 6],
operations=[LT.FC_RELU_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 9,
inputs=[next_id + 2],
operations=[LT.FC_RELU_MUL_GATE, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 10,
inputs=[
next_id + 1,
next_id + 2,
next_id + 3,
next_id + 4,
next_id + 5,
next_id + 6,
next_id + 7,
next_id + 8,
next_id + 9,
],
operations=[
LT.IDENTITY, LT.IDENTITY, LT.IDENTITY, LT.IDENTITY,
LT.IDENTITY, LT.IDENTITY, LT.IDENTITY, LT.IDENTITY,
LT.IDENTITY, LT.MERGE_WITH_AVG
])
])
return l_info
def fully_connected_resnet_cell_info(next_id=0,
input_ids=[0, 1],
end_merge=LayerTypes.MERGE_WITH_CAT):
LT = LayerTypes
l_info = LayerInfoList([
LayerInfo(input_ids[0]),
LayerInfo(input_ids[1]),
LayerInfo(next_id + 2,
inputs=[input_ids[1]],
operations=[LT.FC_SGMD_MUL_GATE, LT.MERGE_WITH_SUM])
])
return ensure_end_merge(l_info, end_merge)
def nasnata_reduction_cell_info(next_id=0,
input_ids=[0, 1],
end_merge=LayerTypes.MERGE_WITH_CAT):
LT = LayerTypes
l_info = LayerInfoList()
l_info.extend([
LayerInfo(input_ids[0]),
LayerInfo(input_ids[1]), # most recent layer
LayerInfo(next_id + 2,
inputs=[input_ids[1], input_ids[0]],
operations=[
LT.SEPARABLE_CONV_5_2, LT.SEPARABLE_CONV_7_2,
LT.MERGE_WITH_SUM
]),
LayerInfo(next_id + 3,
inputs=[input_ids[1], input_ids[0]],
operations=[
LT.MAXPOOL_3x3, LT.SEPARABLE_CONV_7_2, LT.MERGE_WITH_SUM
]),
LayerInfo(next_id + 4,
inputs=[input_ids[1], input_ids[0]],
operations=[
LT.AVGPOOL_3x3, LT.SEPARABLE_CONV_5_2, LT.MERGE_WITH_SUM
]),
LayerInfo(next_id + 5,
inputs=[next_id + 3, next_id + 2],
operations=[LT.IDENTITY, LT.AVGPOOL_3x3, LT.MERGE_WITH_SUM]),
LayerInfo(next_id + 6,
inputs=[next_id + 2, input_ids[1]],
operations=[
LT.SEPARABLE_CONV_3_2, LT.MAXPOOL_3x3, LT.MERGE_WITH_SUM
]),
])
l_info.append(cat_unused(l_info, next_id + 7, end_merge))
return l_info
def separable_resnet_cell_info(next_id=0,
input_ids=[0, 1],
end_merge=LayerTypes.MERGE_WITH_CAT):
LT = LayerTypes
l_info = LayerInfoList([
LayerInfo(input_ids[0]),
LayerInfo(input_ids[1]),
LayerInfo(
next_id + 2,
inputs=[input_ids[1], input_ids[1]],
operations=[LT.SEPARABLE_CONV_3_2, LT.IDENTITY, LT.MERGE_WITH_SUM])
])
return ensure_end_merge(l_info, end_merge)
def basic_resnet_cell_info(next_id=0,
input_ids=[0, 1],
end_merge=LayerTypes.MERGE_WITH_CAT):
LT = LayerTypes
l_info = LayerInfoList([
LayerInfo(input_ids[0]),
LayerInfo(input_ids[1]),
LayerInfo(next_id + 2,
inputs=[input_ids[1]],
operations=[LT.CONV_3, LT.MERGE_WITH_NOTHING]),
LayerInfo(next_id + 3,
inputs=[next_id + 2, input_ids[1]],
operations=[LT.CONV_3, LT.IDENTITY, LT.MERGE_WITH_SUM])
])
return ensure_end_merge(l_info, end_merge)
def default_master(n_normal_inputs=2,
n_reduction_inputs=2,
num_cells=18,
num_reduction_layers=2,
num_init_reductions=0,
skip_reduction_layer_input=0,
use_aux_head=1):
reduction_layers = CellNetworkInfo.calc_reduction_layers(
num_cells, num_reduction_layers, num_init_reductions)
master = LayerInfoList()
layer_id = 0
n_inputs = n_normal_inputs if num_init_reductions == 0 else n_reduction_inputs
for _ in range(n_inputs):
master.append(LayerInfo(layer_id=layer_id))
layer_id += 1
# true_num_cells counts cells from the first non-input with 0-based index
true_num_cells = num_cells + num_init_reductions + num_reduction_layers
for ci in range(true_num_cells):
info = LayerInfo(layer_id)
if ci in reduction_layers:
info.inputs = list(
range(layer_id - n_reduction_inputs, layer_id))
n_in = len(info.inputs)
info.operations = [LayerTypes.IDENTITY] * n_in + ['reduction']
info.down_sampling = 1
else:
if (skip_reduction_layer_input and ci - 1 in reduction_layers
and ci > num_init_reductions):
# imagenet : do not take the input of regular reduction as skip connection.
info.inputs = (list(
range(layer_id - n_normal_inputs - 1, layer_id - 2)) +
[layer_id - 1])
else:
info.inputs = list(
range(layer_id - n_normal_inputs, layer_id))
n_in = len(info.inputs)
info.operations = [LayerTypes.IDENTITY] * n_in + ['normal']
master.append(info)
layer_id += 1
# aux_weight at the last cell before the last reduction
if use_aux_head and len(reduction_layers) > 0:
master[reduction_layers[-1] - 1 + n_inputs].aux_weight = 0.4
master[-1].aux_weight = 1.0
return master