def random_conv_layer_plus(size, num, prob=0.8):
def if_cond(switch, size, num):
if switch > prob:
return [[size, 1, num], [1, size, num]]
return [[size, size, num]]
return spec.merge([spec.uniform(), size, num], if_cond)
def uniform_snap(start, end):
node = spec.uniform(start, end)
def snap_to_ends(value):
delta = (end - start) / 10.0
if abs(value - start) < delta: return start
if abs(value - end) < delta: return end
return value
return spec.wrap(node, snap_to_ends)
return [[size, size, num]]
def random_conv_layer_plus(size, num, prob=0.8):
def if_cond(switch, size, num):
if switch > prob:
return [[size, 1, num], [1, size, num]]
return [[size, size, num]]
return spec.merge([spec.uniform(), size, num], if_cond)
activations = ['relu', 'relu6', 'elu', 'prelu', 'leaky_relu']
hyper_params_spec = {
'init_stdev': spec.uniform(0.04, 0.06),
'augment': {
'crop_size': spec.choice(range(2)),
},
'optimizer': {
'learning_rate': 10**spec.uniform(-2.5, -3.5),
'beta1': 0.9,
'beta2': 0.999,
'epsilon': 1e-8,
},
'conv': {
'layers_num': 3,
1: {
'filters':
random_conv_layer(size=spec.choice(range(3, 8)),
num=spec.choice(range(24, 41))),
return [[size, 1, num], [1, size, num]]
def random_conv_layer_plus(size, num, prob=0.8):
def if_cond(switch, size, num):
if switch > prob:
return [[size, 1, num], [1, size, num]]
return [[size, size, num]]
return spec.merge([spec.uniform(), size, num], if_cond)
activations = ['relu', 'relu6', 'elu', 'prelu', 'leaky_relu']
hyper_params_spec_2_0 = {
'init_stdev': 10**spec.uniform(-1.5, -1),
'optimizer': {
'learning_rate': 10**spec.uniform(-3.2, -3),
'beta1': 0.9,
'beta2': 0.999,
'epsilon': 1e-8,
},
'conv': {
'layers_num': 3,
1: {
'filters':
random_conv_layer_plus(size=spec.choice(range(3, 8)),
num=spec.choice(range(24, 41))),
'pools': [2, 2],
'activation':
spec.choice(activations),