# -*- coding: utf-8 -*-
# ELEKTRONN2 Toolkit
# Copyright (c) 2015 Marius F. Killinger
# All rights reserved
from __future__ import absolute_import, division, print_function
from builtins import filter, hex, input, int, map, next, oct, pow, range, \
super, zip
import numpy as np
from .. import neuromancer
from .utils_basic import pickleload
try:
from elektronn.training.config import Config
except:
Config = None
print('Warning: elektronn not installed. create_cnn will fail!')
[docs]def create_cnn(config_file, n_ch, param_file=None, mfp=False,
axis_order='theano', constant_weights=False,
imposed_input_size=None):
raise RuntimeError("Don't use this, rebuild the graph and import the "
"weights using load_params_into_model")
config = Config(config_file, None, None, use_existing_dir=True,
override_MFP_to_active=mfp,
imposed_input_size=imposed_input_size)
if config.mode!='img-img':
raise NotImplementedError()
if axis_order=='theano':
ps = config.patch_size
ndim = len(ps)
input_size = [None, ] * (2 + ndim)
input_size[0] = config.batch_size
if ndim==3:
tags = 'b,z,f,y,x'
input_size[1] = config.patch_size[0]
input_size[2] = n_ch
input_size[3] = config.patch_size[1]
input_size[4] = config.patch_size[2]
elif ndim==2:
tags = 'b,f,x,y'
input_size[1] = n_ch
input_size[2] = config.patch_size[0]
input_size[3] = config.patch_size[1]
if param_file is None:
param_file = config.paramfile
params = pickleload(param_file)
pool = params[-1]
f_shapes = params[0]
params = params[1:-1] # come in order W0, b0, W1, b1,...
neuromancer.node_basic.model_manager.newmodel('legacy')
inp = neuromancer.Input(input_size, tags)
conv = list(
zip(config.nof_filters, # doesn't have to be a list, does it?
config.filters, config.pool, config.activation_func,
config.pooling_mode, params[::2], params[1::2]))
for i, (n, f, p, act, p_m, W, b) in enumerate(conv):
W = [W, 'const'] if constant_weights else W
b = [b, 'const'] if constant_weights else b
inp = neuromancer.Conv(inp, n, f, p, mfp=mfp, activation_func=act,
w=W, b=b)
# last Layer
W = [params[-2], 'const'] if constant_weights else params[-2]
b = [params[-1], 'const'] if constant_weights else params[-1]
out = neuromancer.Conv(inp, config.n_lab, (1,) * ndim, (1,) * ndim,
activation_func='lin', w=W, b=b)
if mfp:
out = neuromancer.FragmentsToDense(out)
if config.target in ['affinity', 'malis']:
probs = neuromancer.Softmax(out, n_class=2, n_indep=3,
name='class_probabilities')
else:
probs = neuromancer.Softmax(out, n_class=config.n_lab,
name='class_probabilities')
elif axis_order=='dnn':
raise NotImplementedError()
model = neuromancer.model_manager.getmodel('legacy')
model.designate_nodes(input_node=inp, prediction_node=probs)
return model
[docs]def load_params_into_model(param_file, model):
"""
Loads parameters directly from save file into a graph manager (this requires
that the graph is identical to the cnn from the param file
:param param_file:
:param gm:
:return:
"""
params = pickleload(param_file)[1:-1]
i = 0
for node in model.nodes.values():
if hasattr(node, 'w'):
try:
w = np.transpose(params[i], (0, 2, 1, 3, 4))
node.w.set_value(w)
except:
node.w.set_value(params[i])
node.b.set_value(params[i + 1])
i += 2