def tf_net2deeper(model, weight, target_layer):
n2n = Net2Net()
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
net = load_teacher_net(sess, model, weight)
w1, b1 = get_weight_bias_of_layer(net, target_layer)
new_w, new_b = n2n.deeper(w1, True)
return new_w, new_b
def tf_net2wider_rand(model, weight, target_layer, next_layer, new_width):
n2n = Net2Net()
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
net = load_teacher_net(sess, model, weight)
w1, b1 = get_weight_bias_of_layer(net, target_layer)
w2, b2 = get_weight_bias_of_layer(net, next_layer)
nw1, nb1, nw2 = n2n.wider_rand(w1, b1, w2, new_width)
return nw1, nb1, nw2, b2
def get_deeper_weights(ref_layer):
'''
To calculate new weights to make the net deeper using Net2Net class,
one needs to swap the axes for the right order.
Dim of Keras conv layer: (OutChannel, InChannel, kH, kW)
conv layer Net2Net class accepts: (kH, kW, InChannel, OutChannel)
'''
parms = ref_layer.get_weights()
n2n = Net2Net()
if is_convolutional(ref_layer):
weights = parms[0].swapaxes(0, 2).swapaxes(1, 3).swapaxes(2, 3)
new_w, new_b = n2n.deeper(weights, True)
new_w = new_w.swapaxes(0, 2).swapaxes(1, 3)
else:
weights = parms[0]
new_w, new_b = n2n.deeper(weights, True)
return new_w, new_b
def net_to_deeper(model_teacher,model,l):
#Get weights of the teacher model
ws = []
bs= []
for layer_t in model_teacher.layers:
if layer_t.get_weights() != []:
ws.append(layer_t.get_weights()[0])
bs.append(layer_t.get_weights()[1])
n2n = Net2Net()
new_w, new_b = n2n.deeper(ws[l-1], True)
ws.insert(l,new_w)
bs.insert(l,new_b)
i = 0
for layer_s in model.layers:
if layer_s.get_weights() != []:
layer_s.set_weights([ws[i],bs[i]])
i += 1
return model
def net_to_wider(model_teacher,model,l,new_width_conv):
#Get weights of the teacher model
ws = []
bs = []
for layer in model_teacher.layers:
if layer.get_weights() != []:
ws.append(layer.get_weights()[0])
bs.append(layer.get_weights()[1])
w1 = ws[l]
b1 = bs[l]
w2 = ws[l+1]
n2n = Net2Net()
nw1, nb1, nw2 = n2n.wider(w1, b1, w2, new_width_conv, True)
ws[l] = nw1
bs[l] = nb1
ws[l+1] = nw2
i = 0
for layer in model.layers:
if layer.get_weights() != []:
layer.set_weights([ws[i],bs[i]])
i += 1
return model