def sparse_bn_feature_net_multires_61x61(batch_input_shape = (1,2,1080,1280), n_features = 3, n_channels = 2, reg = 1e-5, weights_path = None):
print "Using multiresolution feature net 61x61 with batch normalization"
d = 1
inputs = Input(shape = batch_input_shape[1:])
layer1 = sparse_Convolution2D(64, 3, 3, d = d, init = 'he_normal', border_mode='valid', batch_input_shape = batch_input_shape, W_regularizer = l2(reg))(inputs)
norm1 = BatchNormalization(axis = 1)(layer1)
act1 = Activation('relu')(norm1)
layer2 = sparse_Convolution2D(64, 4, 4, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(act1)
norm2 = BatchNormalization(axis = 1)(layer2)
act2 = Activation('relu')(norm2)
pool1 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act2)
d *= 2
layer3 = sparse_Convolution2D(64, 3, 3, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(pool1)
norm3 = BatchNormalization(axis = 1)(layer3)
act3 = Activation('relu')(norm3)
layer4 = sparse_Convolution2D(64, 3, 3, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(act3)
norm4 = BatchNormalization(axis = 1)(layer4)
act4 = Activation('relu')(norm4)
pool2 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act4)
d *= 2
layer5 = sparse_Convolution2D(128, 3, 3, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(pool2)
norm5 = BatchNormalization(axis = 1)(layer5)
act5 = Activation('relu')(norm5)
layer6 = sparse_Convolution2D(128, 3, 3, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(act5)
norm6 = BatchNormalization(axis = 1)(layer6)
act6 = Activation('relu')(norm6)
pool3 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act6)
d *= 2
side_layer1 = sparse_Convolution2D(256,28,28, d = 2, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(pool1)
side_norm1 = BatchNormalization(axis = 1)(side_layer1)
side_act1 = Activation('relu')(side_norm1)
side_layer2 = sparse_Convolution2D(256,12,12, d = 4, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(pool2)
side_norm2 = BatchNormalization(axis = 1)(side_layer2)
side_act2 = Activation('relu')(side_norm2)
side_layer3 = sparse_Convolution2D(256,4,4, d = 8, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(pool3)
side_norm3 = BatchNormalization(axis = 1)(side_layer3)
side_act3 = Activation('relu')(side_norm3)
merge_layer1 = merge([side_act1, side_act2, side_act3], mode = 'concat', concat_axis = 1)
dense1 = TensorProd2D(768, 512, init = 'he_normal', W_regularizer = l2(reg))(merge_layer1)
dense_norm1 = BatchNormalization(axis = 1)(dense1)
dense_act1 = Activation('relu')(dense_norm1)
dense2 = TensorProd2D(512, n_features, init = 'he_normal', W_regularizer = l2(reg))(dense_act1)
dense_act2 = Activation(tensorprod_softmax)(dense2)
model = Model(input = inputs, output = dense_act2)
model = set_weights(model, weights_path)
return model
def sparse_feature_net_31x31(batch_input_shape = (1,2,1080,1280), n_features = 3, drop = 0.5, reg = 0.001, init = 'he_normal', weights_path = None):
model = Sequential()
d = 1
model.add(sparse_Convolution2D(32, 4, 4, d = d, init = init, batch_input_shape = batch_input_shape, border_mode='valid', W_regularizer = l2(reg)))
model.add(Activation('relu'))
model.add(sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d)))
d *= 2
model.add(sparse_Convolution2D(64, 3, 3, d = d, init = init, border_mode = 'valid', W_regularizer = l2(reg)))
model.add(Activation('relu'))
model.add(sparse_Convolution2D(64, 3, 3, d = d, init = init, border_mode='valid', W_regularizer = l2(reg)))
model.add(Activation('relu'))
model.add(sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d)))
d *= 2
model.add(sparse_Convolution2D(128, 3, 3, d = d, init = init, border_mode = 'valid', W_regularizer = l2(reg)))
model.add(Activation('relu'))
model.add(sparse_Convolution2D(200, 3, 3, d = d, init = init, border_mode='valid', W_regularizer = l2(reg)))
model.add(Dropout(drop))
model.add(Activation('relu'))
model.add(TensorProd2D(200, 200, init = init, W_regularizer = l2(reg)))
model.add(Dropout(drop))
model.add(Activation('relu'))
model.add(TensorProd2D(200, n_features, init = init, W_regularizer = l2(reg)))
model.add(Activation(tensorprod_softmax))
model = set_weights(model, weights_path)
return model
def sparse_feature_net_multires_101x101(batch_input_shape = (1,2,1080,1280), n_features = 3, reg = 0.001, weights_path = None):
d = 1
inputs = Input(shape = (2,101,101))
layer1 = sparse_Convolution2D(32, 4, 4, name = 'layer1', d = d, init = 'he_normal', border_mode='valid', input_shape=(2, 101, 101), W_regularizer = l2(reg))(inputs)
act1 = Activation('relu')(layer1)
pool1 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act1)
d *= 2
layer2 = sparse_Convolution2D(64, 4, 4, name = 'layer2', d = d, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(pool1)
act2 = Activation('relu')(layer2)
layer3 = sparse_Convolution2D(64, 3, 3, name = 'layer3', d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(act2)
act3 = Activation('relu')(layer3)
pool2 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act3)
d*= 2
layer4 = sparse_Convolution2D(128, 3, 3, name = 'layer4', d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(pool2)
act4 = Activation('relu')(layer4)
layer5 = sparse_Convolution2D(128, 3, 3, name = 'layer5', d = d, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(act4)
act5 = Activation('relu')(layer5)
layer6 = sparse_Convolution2D(128, 3, 3, name = 'layer6', d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(act5)
act6 = Activation('relu')(layer6)
pool3 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act6)
side_layer1 = sparse_Convolution2D(512,16,16, name = 'layer9', d = d, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(layer6)
dropside1 = Dropout(drop)(side_layer1)
d *= 2
layer7 = sparse_Convolution2D(256, 3, 3, name = 'layer7', d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(pool3)
act7 = Activation('relu')(layer7)
layer8 = sparse_Convolution2D(256, 3, 3, name = 'layer8', d = d, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(act7)
act8 = Activation('relu')(layer8)
layer9 = sparse_Convolution2D(512,4,4, name = 'layer9', d = d, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(act8)
drop1 = Dropout(drop)(layer9)
act9 = Activation('relu')(drop1)
merge_layer1 = merge([dropside1, act9], mode = 'concat')
dense1 = TensorProd2D(1024, 1024, name = 'layer10', init = 'he_normal', W_regularizer = l2(reg))(merge_layer1)
drop2 = Dropout(drop)(dense1)
act10 = Activation('relu')(drop2)
dense2 = TensorProd2D(1024, n_features, name = 'layer11', init = 'he_normal', W_regularizer = l2(reg))(act10)
act11 = Activation(tensorprod_softmax)(dense2)
model = Model(input = inputs, output = act11)
model = set_weights(model, weights_path)
return model
def sparse_bn_feature_net_multires_31x31(batch_input_shape = (1,2,1080,1280), n_features = 3, n_channels = 1, reg = 1e-5, weights_path = None):
print "Using multiresolution feature net 31x31 with batch normalization"
d = 1
inputs = Input(shape = batch_input_shape[1:])
layer1 = sparse_Convolution2D(32, 4, 4, d = d, init = 'he_normal', border_mode='valid', input_shape=(n_channels, 31, 31), W_regularizer = l2(reg))(inputs)
norm1 = BatchNormalization(axis = 1)(layer1)
act1 = Activation('relu')(norm1)
pool1 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act1)
d *= 2
layer2 = sparse_Convolution2D(64, 3, 3, d =d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(pool1)
norm2 = BatchNormalization(axis = 1)(layer2)
act2 = Activation('relu')(norm2)
layer3 = sparse_Convolution2D(64, 3, 3, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(act2)
norm3 = BatchNormalization(axis = 1)(layer3)
act3 = Activation('relu')(norm3)
pool2 = sparse_MaxPooling2D(pool_size=(2, 2), strides = (d,d))(act3)
d *= 2
layer4 = sparse_Convolution2D(128, 3, 3, d = d, init = 'he_normal', border_mode='valid', W_regularizer = l2(reg))(pool2)
norm4 = BatchNormalization(axis = 1)(layer4)
act4 = Activation('relu')(norm4)
side_layer1 = sparse_Convolution2D(200, 14, 14, d = 2, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(pool1)
side_norm1 = BatchNormalization(axis = 1)(side_layer1)
side_act1 = Activation('relu')(side_norm1)
side_layer2 = sparse_Convolution2D(200, 3, 3, d = 4, init = 'he_normal', border_mode = 'valid', W_regularizer = l2(reg))(norm4)
side_norm2 = BatchNormalization(axis = 1)(side_layer2)
side_act2 = Activation('relu')(side_norm2)
merge_layer1 = merge([side_act1, side_act2], mode = 'concat', concat_axis = 1)
flat = Flatten()(merge_layer1)
dense1 = TensorProd2D(400, 512, init = 'he_normal', W_regularizer = l2(reg))(flat)
dense_norm1 = BatchNormalization(axis = 1)(dense1)
dense_act1 = Activation('relu')(dense_norm1)
dense2 = TensorProd2D(512, n_features, init = 'he_normal', W_regularizer = l2(reg))(dense_act1)
dense_act2 = Activation(tensorprod_softmax)(dense2)
model = Model(input = inputs, output = dense_act2)
model = set_weights(model, weights_path)
return model