def build_berry_alexnet(input_layer, num_classes, model_params={}):
from berry import BerryModel
from berry import layers
nn = BerryModel()
nn.add(input_layer)
nn.add(layers.Convolution2D(
nn.last, 96, 11, stride=4, pad='VALID', activation='relu',
W_stddev=1e-2, b_val=0.1))
nn.add(layers.MaxPooling2D(nn.last, 3, 2))
nn.add(layers.Convolution2D(nn.last, 256, 5, pad='SAME', activation='relu',
W_stddev=1e-2, b_val=0.1))
nn.add(layers.MaxPooling2D(nn.last, 3, 2))
nn.add(layers.Convolution2D(nn.last, 384, 3, pad='SAME', activation='relu',
W_stddev=1e-2, b_val=0.1))
nn.add(layers.Convolution2D(nn.last, 384, 3, pad='SAME', activation='relu',
W_stddev=1e-2, b_val=0.1))
nn.add(layers.Convolution2D(nn.last, 256, 3, pad='SAME', activation='relu',
W_stddev=1e-2, b_val=0.1))
nn.add(layers.MaxPooling2D(nn.last, 3, 2))
nn.add(layers.Flatten(nn.last))
nn.add(layers.Dense(
nn.last, 4096, activation='relu', W_stddev=5e-3, b_val=0.1))
nn.add(layers.Dropout(nn.last, .5))
nn.add(layers.Dense(
nn.last, 4096, activation='relu', W_stddev=5e-3, b_val=0.1))
nn.add(layers.Dropout(nn.last, .5))
nn.add(layers.Dense(
nn.last, num_classes, activation='softmax', W_stddev=5e-3, b_val=0.1))
return nn
def build_berry_lenet(input_layer, num_classes, model_params={}):
from berry import layers, BerryModel
nn = BerryModel()
nn.add(input_layer)
nn.add(
layers.Convolution2D(nn.last,
20,
5,
pad='VALID',
activation='sigmoid',
W_stddev=1e-1))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(
layers.Convolution2D(nn.last,
50,
5,
pad='SAME',
W_stddev=1e-1,
activation='sigmoid'))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(layers.Flatten(nn.last))
nn.add(layers.Dense(nn.last, 500, activation='relu', W_stddev=5e-3))
nn.add(
layers.Dense(nn.last, num_classes, activation='softmax',
W_stddev=5e-3))
return nn
def inference(i):
layer_dict = OrderedDict()
layer_dict['input'] = i
l = layers.Convolution2D(i,
96,
11,
stride=4,
pad='SAME',
activation='relu',
name='conv_1')
layer_dict[l.name] = l
l = layers.MaxPooling2D(l, 3, 2, name='pool_1')
layer_dict[l.name] = l
l = layers.Convolution2D(l,
256,
5,
pad='SAME',
activation='relu',
name='conv_2')
layer_dict[l.name] = l
l = layers.MaxPooling2D(l, 3, 2, name='pool_2')
layer_dict[l.name] = l
l = layers.Convolution2D(l,
384,
3,
pad='SAME',
activation='relu',
name='conv_3')
layer_dict[l.name] = l
l = layers.Convolution2D(l,
384,
3,
pad='SAME',
activation='relu',
name='conv_4')
layer_dict[l.name] = l
l = layers.Convolution2D(l,
256,
3,
pad='SAME',
activation='relu',
name='conv_5')
layer_dict[l.name] = l
l = layers.MaxPooling2D(l, 3, 2, name='pool_5')
layer_dict[l.name] = l
l = layers.Flatten(l, name='flat')
layer_dict[l.name] = l
l = layers.Dense(l, 4096, activation='relu', name='fc_1')
layer_dict[l.name] = l
l = layers.Dropout(l, 0.5, name='drop_1')
layer_dict[l.name] = l
l = layers.Dense(l, 4096, activation='relu', name='fc_2')
layer_dict[l.name] = l
l = layers.Dropout(l, 0.5, name='drop_2')
layer_dict[l.name] = l
l = layers.Dense(l, 50, activation='softmax', name='output')
layer_dict[l.name] = l
print layer_dict.keys()
return layer_dict, l.output
def inference(i, keep_prob):
print i.get_shape()
l = layers.Convolution2D(i, 64, 3, stride=1, pad='SAME', activation='relu')
print l.output_shape
l = layers.Convolution2D(l, 64, 3, stride=1, pad='SAME', activation='relu')
print l.output_shape
l = layers.MaxPooling2D(l, 2, 2)
print l.output_shape
l = layers.Convolution2D(l,
128,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
128,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.MaxPooling2D(l, 2, 2)
print l.output_shape
l = layers.Convolution2D(l,
256,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
256,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
256,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.MaxPooling2D(l, 2, 2)
print l.output_shape
l = layers.Convolution2D(l,
512,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
512,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
512,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.MaxPooling2D(l, 2, 2)
print l.output_shape
print l.output_shape
l = layers.Convolution2D(l,
512,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
512,
3,
stride=1,
pad='SAME',
activation='relu')
print l.output_shape
l = layers.Convolution2D(l,
512,
3,
stride=1,
pad='SAME',
activation='relu')
l = layers.MaxPooling2D(l, 2, 2)
print l.output_shape
l = layers.Flatten(l)
l = layers.Dense(l, 4096, activation='relu')
l = layers.Dropout(l, keep_prob)
l = layers.Dense(l, 4096, activation='relu')
l = layers.Dropout(l, keep_prob)
l = layers.Dense(l, 50, activation='softmax')
return l.output
def build_berry_vgg16(input_layer, num_classes, model_params={}):
from berry import layers, BerryModel
nn = BerryModel()
nn.add(input_layer)
# Expected input_shape = (224, 224, 3)
nn.add(
layers.Convolution2D(nn.last,
64,
5,
stride=2,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
64,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(
layers.Convolution2D(nn.last,
128,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
128,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(
layers.Convolution2D(nn.last,
256,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
256,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
256,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(
layers.Convolution2D(nn.last,
512,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
512,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
512,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(
layers.Convolution2D(nn.last,
512,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
512,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(
layers.Convolution2D(nn.last,
512,
3,
pad='SAME',
activation='relu',
init='deepnet',
b_val=0.0))
nn.add(layers.MaxPooling2D(nn.last, 2, 2))
nn.add(layers.Flatten(nn.last))
nn.add(
layers.Dense(nn.last,
4096,
activation='relu',
W_stddev=1e-2,
b_val=0.0))
nn.add(layers.Dropout(nn.last, 0.5))
nn.add(
layers.Dense(nn.last,
4096,
activation='relu',
W_stddev=1e-2,
b_val=0.0))
nn.add(layers.Dropout(nn.last, 0.5))
nn.add(
layers.Dense(nn.last,
num_classes,
activation='softmax',
W_stddev=1e-3,
b_val=0.0))
return nn