output_cols = 0
if para.model_type == "segnet":
print('Initialising Segnet 4 Layers Encoder - Basic ...')
model, output_rows, output_cols = segnet.SegNet()
if para.model_type == "squeezed_segnet":
print('Initialising Squeezed Segnet. Very few parameters ...')
model, output_rows, output_cols = squeezed_segnet.squeezed_SegNet()
if para.model_type == "segnet_depool":
print('Initialising Segnet with Max Pooling Indices...')
model, output_rows, output_cols = segnet_depool.SegNet()
if para.model_type == "tiny_segnet_depool":
print('Initialising Tiny Segnet with Max Pooling Indices...')
model, output_rows, output_cols = tiny_segnet_depool.SegNet()
if para.model_type == "fcn":
print('Initialising FCN...')
model, output_rows, output_cols = fcn.fcn_8()
if para.model_type == "squeezed_fcn":
print('Initialising Squeezed FCN...')
model, output_rows, output_cols = squeezed_fcn.fcn_8()
# Compile the model using sgd optimizer
# =====================================
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=[dice_coef, precision, recall, 'accuracy'])
print(model.summary())
print(para.misc_dir_eval)
# Load the model weights
if para.compressed_SVD:
model.load_weights(para.misc_dir_eval + "/svd_" + str(para.SVD_ratio) +
"_weights.h5")
def keras2Tensorflow(model_path=para.misc_dir_eval + "/weights.h5",
model_path_pb=para.misc_dir_eval,
model_outputs=1,
K2T_prefix="output_node",
model_file_pb="model.pb"):
'''
model_path = The HDF5 Keras model you wish to convert to .pb
model_outputs = The number of outputs in the HDF5 model
model_path_pb = The directory to place the output .pb files for tensorflow
K2T_prefix = "k2tfout"
model_file_pb = "model.pb" #
'''
# Call model
# ==========
output_rows = 0
output_cols = 0
if para.model_type == "segnet_depool":
print('Initialising Segnet with Max Pooling Indices...')
model, output_rows, output_cols = segnet_depool.SegNet()
if para.model_type == "tiny_segnet_depool":
print('Initialising Tiny Segnet with Max Pooling Indices...')
model, output_rows, output_cols = tiny_segnet_depool.SegNet()
if para.model_type == "fcn":
print('Initialising FCN...')
model, output_rows, output_cols = fcn.fcn_8()
if para.model_type == "squeezed_fcn":
print('Initialising Squeezed FCN...')
model, output_rows, output_cols = squeezed_fcn.fcn_8()
if para.model_type == "sub_pixel":
print('Initialising sub_pixel ...')
model, output_rows, output_cols = sub_pixel.subPixelModel()
if para.model_type == "tiny_sub_pixel":
print('Initialising tiny_sub_pixel ...')
model, output_rows, output_cols = tiny_sub_pixel.tinySubPixelModel()
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=['accuracy'])
if not os.path.isdir(model_path_pb):
os.mkdir(model_path_pb)
K.set_learning_phase(0)
model.load_weights(para.misc_dir_eval + "/weights.h5")
#net_model = load_model(model_path)
# Alias the outputs in the model - this sometimes makes them easier to access in TF
pred = [None] * model_outputs
pred_node_names = [None] * model_outputs
for i in range(model_outputs):
pred_node_names[i] = K2T_prefix + str(i)
pred[i] = tf.identity(model.output[i], name=pred_node_names[i])
print('Output nodes names are: ', pred_node_names)
sess = K.get_session()
# Write the graph in human readable
f = 'graph_def_for_reference.pb.ascii'
tf.train.write_graph(sess.graph.as_graph_def(),
model_path_pb,
f,
as_text=True)
print('Saved the graph definition in ascii format at: ',
osp.join(model_path_pb, f))
# Write the graph in binary .pb file
from tensorflow.python.framework import graph_util
from tensorflow.python.framework import graph_io
constant_graph = graph_util.convert_variables_to_constants(
sess, sess.graph.as_graph_def(), pred_node_names)
graph_io.write_graph(constant_graph,
model_path_pb,
model_file_pb,
as_text=False)
print('Saved the constant graph (ready for inference) at: ',
osp.join(model_path_pb, model_file_pb))