def sparse_feature_net_61x61(batch_input_shape=(1, 2, 1080, 1280),
n_features=3,
reg=0.001,
init='he_normal',
weights_path=None):
model = Sequential()
# Keras 1
if (parse_version(keras_version) < parse_version('2.0.0')):
d = 1
model.add(
sparse_Convolution2D(64,
3,
3,
d=d,
init=init,
batch_input_shape=batch_input_shape,
border_mode='valid',
W_regularizer=l2(reg)))
model.add(Activation('relu'))
model.add(
sparse_Convolution2D(64,
4,
4,
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(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(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(200,
4,
4,
d=d,
init=init,
border_mode='valid',
W_regularizer=l2(reg)))
model.add(Activation('relu'))
model.add(TensorProd2D(200, 200, init=init, W_regularizer=l2(reg)))
model.add(Activation('relu'))
model.add(
TensorProd2D(200, n_features, init=init, W_regularizer=l2(reg)))
model.add(Activation(tensorprod_softmax))
# Keras 2
else:
d = 1
model.add(
sparse_Convolution2D(64,
3,
3,
d=d,
init=init,
batch_input_shape=batch_input_shape,
border_mode='valid',
W_regularizer=l2(reg)))
model.add(Activation('relu'))
model.add(
sparse_Convolution2D(64,
4,
4,
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(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(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(200,
4,
4,
d=d,
init=init,
border_mode='valid',
W_regularizer=l2(reg)))
model.add(Activation('relu'))
model.add(
TensorProd2D(200,
200,
kernel_initializer=init,
W_regularizer=l2(reg)))
model.add(Activation('relu'))
model.add(
TensorProd2D(200,
n_features,
kernel_initializer=init,
W_regularizer=l2(reg)))
model.add(Activation(tensorprod_softmax))
# read pre-computed weights if provided, and exit
if weights_path is not None:
model = set_weights(model, weights_path)
return model
im5 = np.ones(shape=(1, 2, 200, 200)) + 1
im5[:, :, 0:50, 0:50] = 3
# load one of the DeepCell models
basedatadir = os.path.normpath(
os.path.join(cytometer.__path__[0], '../data/deepcell'))
netdir = os.path.join(basedatadir, 'trained_networks')
model_dir = os.path.join(netdir, 'HeLa')
model_weights_file = '2017-06-21_HeLa_all_61x61_bn_feature_net_61x61_'
# instantiate model (same for all validation data)
j = 0 # model index
model = deepcell_models.sparse_bn_feature_net_61x61(batch_input_shape=im.shape)
model = deepcell.set_weights(
model, os.path.join(netdir, model_dir,
model_weights_file + str(j) + '.h5'))
im_out = model.predict(im)
im2_out = model.predict(im2)
im3_out = model.predict(im3)
im4_out = model.predict(im4)
im5_out = model.predict(im5)
# add padding to the outputs
im_out = np.pad(im_out,
pad_width=((0, 0), (0, 0), (30, 30), (30, 30)),
mode='constant',
constant_values=0)
im2_out = np.pad(im2_out,
pad_width=((0, 0), (0, 0), (30, 30), (30, 30)),
mode='constant',
def sparse_bn_feature_net_31x31(batch_input_shape=(1, 1, 1080, 1280),
n_features=3,
reg=1e-5,
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(BatchNormalization(axis=1, mode=2))
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(BatchNormalization(axis=1, mode=2))
model.add(Activation('relu'))
model.add(
sparse_Convolution2D(64,
3,
3,
d=d,
init=init,
border_mode='valid',
W_regularizer=l2(reg)))
model.add(BatchNormalization(axis=1, mode=2))
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(BatchNormalization(axis=1, mode=2))
model.add(Activation('relu'))
model.add(
sparse_Convolution2D(200,
3,
3,
d=d,
init=init,
border_mode='valid',
W_regularizer=l2(reg)))
model.add(BatchNormalization(axis=1, mode=2))
model.add(Activation('relu'))
model.add(TensorProd2D(200, 200, init=init, W_regularizer=l2(reg)))
model.add(BatchNormalization(axis=1, mode=2))
model.add(Activation('relu'))
model.add(TensorProd2D(200, n_features, init=init, W_regularizer=l2(reg)))
model.add(Activation(tensorprod_softmax))
if (weights_path != None):
model = set_weights(model, weights_path)
return model