def get_Params(self):
# warning : sometimes with the gpu mode the list
# of parameters can be from the output to the input layer
# we need in this case to inverse the list
# to check that we look at the list of parameters
# of the whole system
# if the first weight is a matrix then inverse
#return a dictionary of params
x = T.tensor4()
cost = self.apply(x).sum()
cg = ComputationGraph(cost)
W = VariableFilter(roles=[WEIGHT])(cg.variables)
B = VariableFilter(roles=[BIAS])(cg.variables)
if W[0].name[:5] == "layer":
return W, B
# find other parameters and retrieve them for the lists
gamma = []
beta = []
index_gamma = []
index_beta = []
for w, b, i in zip(W, B, range(len(W))):
if w.name == "log_gamma":
index_gamma.append(i)
gamma.append(w)
if b.name == "beta":
index_beta.append(i)
beta.append(b)
for i in index_gamma[::-1]:
W.pop(i)
for i in index_beta[::-1]:
B.pop(i)
if len(W) == 0:
import pdb
pdb.set_trace()
if W[0].ndim == 2:
W_ = []
for i in xrange(len(W)):
W_.append(W[len(W) - 1 - i])
W = W_
if B[0].ndim == 1:
B_ = []
for i in xrange(len(B)):
B_.append(B[len(B) - 1 - i])
B = B_
# if batch normalization has been introduced you need to reinject artificially
# the gamma and beta parameters so to fit with the actual protocol of dropout
if len(gamma) != len(beta):
raise Exception(
" gamma and beta parameters should be balanced : (%d, %d)",
len(gamma), len(beta))
if len(gamma) != 0:
if beta[0].shape.eval() != gamma[0].shape.eval():
beta.reverse()
W_new = []
B_new = []
for w, g in zip(W[:len(gamma)], gamma):
W_new.append(w)
W_new.append(g)
W_new += W[len(gamma):]
for b, b_ in zip(B[:len(gamma)], beta):
B_new.append(b)
B_new.append(b_)
B_new += B[len(gamma):]
W = W_new
B = B_new
for w, b, index in zip(W, B, range(len(W))):
w.name = "layer_" + str(index) + "_W"
b.name = "layer_" + str(index) + "_B"
return W, B
