def pretrain_main(db, params):
dim = params["dim"]
n_layer = len(dim) - 1
w_down = []
b_down = []
for i in xrange(n_layer):
rate = params["rate"][0]
layers = [("data", ("data", "label", db + "_train", db + "_test", 1.0))]
str_x = "data"
for j in xrange(i):
str_h = "inner%d" % (j + 1)
layers.extend([("inner_lr", (str_h, str_x, dim[j + 1], 0.05, 0.0, 0.0)), ("relu", (str_h,))])
str_x = str_h
if i == n_layer - 1:
str_h = "output"
else:
str_h = "inner%d" % (i + 1)
if i != 0:
layers.extend(
[
("drop_copy", (str_x + "_drop", str_x, params["drop"][0])),
("inner_init", (str_h, str_x + "_drop", dim[i + 1], 0.01)),
]
)
else:
layers.extend(
[
("drop_copy", (str_x + "_drop", str_x, 0.0)),
("inner_init", (str_h, str_x + "_drop", dim[i + 1], 0.01)),
]
)
if i != n_layer - 1:
layers.append(("relu", (str_h,)))
layers.append(("drop", (str_h, params["drop"][0])))
layers.append(("inner_init", ("d_" + str_x, str_h, dim[i], 0.01)))
if i != 0:
layers.append(("relu", ("d_" + str_x,)))
layers.append(("euclid", ("pt_loss%d" % (i + 1), "d_" + str_x, str_x)))
else:
layers.append(("euclid", ("pt_loss%d" % (i + 1), "d_" + str_x, str_x)))
with open("stack_net.prototxt", "w") as fnet:
dec.make_net(fnet, layers)
with open("pt_solver.prototxt", "w") as fsolver:
fsolver.write(
"""net: "stack_net.prototxt"
base_lr: {0}
lr_policy: "step"
gamma: 0.1
stepsize: {1}
display: 1000
test_iter: 100
test_interval: 10000
max_iter: {2}
momentum: 0.9
momentum_burnin: 1000
weight_decay: {3}
snapshot: 10000
snapshot_prefix: "exp/{4}/save"
snapshot_after_train:true
solver_mode: GPU
debug_info: false
device_id: 0""".format(
rate, params["step"][0], params["pt_iter"][0], params["decay"][0], db
)
)
if i > 0:
model = "exp/" + db + "/save_iter_%d.caffemodel" % params["pt_iter"][0]
else:
model = None
mean, net = dec.extract_feature("stack_net.prototxt", model, [str_x], 1, train=True, device=0)
net.save("stack_init.caffemodel")
os.system("caffe train --solver=pt_solver.prototxt --weights=stack_init.caffemodel")
net = caffe.Net("stack_net.prototxt", "exp/" + db + "/save_iter_%d.caffemodel" % params["pt_iter"][0])
w_down.append(net.params["d_" + str_x][0].data.copy())
b_down.append(net.params["d_" + str_x][1].data.copy())
del net
net = caffe.Net("pt_net.prototxt", "exp/" + db + "/save_iter_%d.caffemodel" % params["pt_iter"][0])
for i in xrange(n_layer):
if i == 0:
k = "d_data"
else:
k = "d_inner%d" % i
net.params[k][0].data[...] = w_down[i]
net.params[k][1].data[...] = b_down[i]
net.save("stack_init_final.caffemodel")
def main(db, params):
n_layer = params["n_layer"][0]
drop = params["drop"][0]
encoder_layers = [("data", ("data", "label", db + "_train", db + "_test", 1.0))]
decoder_layers = [("euclid", ("pt_loss", "d_data", "data"))]
last_dim = params["dim"][0]
niter = params["iter"][0]
rate = params["rate"][0]
for i in xrange(n_layer):
str_h1 = "inner%d" % (i + 1)
str_h2 = "d_inner%d" % (i + 1)
str_x = "inner%d" % (i)
str_y = "d_inner%d" % (i)
dim = params["dim"][i + 1]
if i == 0:
str_x = "data"
str_y = "d_data"
if i == n_layer - 1:
str_h1 = "output"
str_h2 = "output"
if i != n_layer - 1:
encoder_layers.extend(
[
("inner_init", (str_h1, str_x, dim, np.sqrt(1.0 / last_dim))),
("relu", (str_h1,)),
("drop", (str_h1, drop)),
]
)
else:
encoder_layers.extend(
[("inner_init", (str_h1, str_x, dim, np.sqrt(1.0 / last_dim))), ("drop", (str_h1, drop))]
)
if i != 0:
decoder_layers.append(("drop", (str_y, drop)))
decoder_layers.extend([("relu", (str_y,)), ("inner_init", (str_y, str_h2, last_dim, np.sqrt(1.0 / dim)))])
else:
decoder_layers.extend([("inner_init", (str_y, str_h2, last_dim, np.sqrt(1.0 / dim)))])
last_dim = dim
with open("pt_net.prototxt", "w") as fnet:
dec.make_net(fnet, encoder_layers + decoder_layers[::-1])
with open("ft_solver.prototxt", "w") as fsolver:
fsolver.write(
"""net: "pt_net.prototxt"
base_lr: {0}
lr_policy: "step"
gamma: 0.1
stepsize: {1}
display: 1000
test_iter: 100
test_interval: 10000
max_iter: {2}
momentum: 0.9
momentum_burnin: 1000
weight_decay: {3}
snapshot: 10000
snapshot_prefix: "exp/{4}/save"
snapshot_after_train:true
solver_mode: GPU
debug_info: false
device_id: 0""".format(
rate, params["step"][0], niter, params["decay"][0], db
)
)
