def test_parametric_function_api():
"""
Testing :function:`nnabla.parametric_functions.parametric_function_api`.
"""
import nnabla as nn
import inspect
nn.clear_parameters()
shape = (2, 3, 4)
# Signature check
spec = inspect.getargspec(dummy_parametric_function)
assert spec.args == ['shape', 'f', 'i', 's', 'name']
assert spec.defaults == (10, 1, 'dummy', None)
assert dummy_parametric_function.__doc__.splitlines()[0] == 'Doc'
# Verify two different ways does the same thing.
# Using name argument
v = dummy_parametric_function(shape, name='group1')
# Using parameter_scope
with nn.parameter_scope('group1'):
v = dummy_parametric_function(shape)
params = nn.get_parameters()
assert len(params) == 2
assert list(iterkeys(params)) == ['group1/dummy/p1', 'group1/dummy/p2']
# No scope
v = dummy_parametric_function(shape)
params = nn.get_parameters()
len(params) == 4
assert list(iterkeys(params)) == ['group1/dummy/p1', 'group1/dummy/p2',
'dummy/p1', 'dummy/p2']
nn.clear_parameters()
def main():
batch_size, m, h, w = 4, 3, 32, 32
extension_module = "cpu"
device_id = 0
ctx = extension_context(extension_module, device_id=device_id)
x_l_data = np.random.randn(batch_size, m, h, w)
y_l_data = (np.random.rand(batch_size, 1) * 10).astype(np.int32)
x_l = nn.Variable(x_l_data.shape)
y_l = nn.Variable(y_l_data.shape)
x_l.d = x_l_data
y_l.d = y_l_data
# CNN
print("# CNN")
pred = cnn_model_003(ctx, x_l)
s = 0
for n, v in nn.get_parameters().iteritems():
n_params = np.prod(v.shape)
print(n, n_params)
s += n_params
print("n_params={}".format(s))
nn.clear_parameters()
# Resnet
print("# Resnet")
inmaps = 256
pred = resnet_model(ctx, x_l, inmaps=inmaps)
s = 0
for n, v in nn.get_parameters().iteritems():
n_params = np.prod(v.shape)
print(n, n_params)
s += n_params
print("n_params={}".format(s))
nn.clear_parameters()
def test_save_load_parameters():
v = nn.Variable([64, 1, 28, 28], need_grad=False)
with nn.parameter_scope("param1"):
with nn.parameter_scope("conv1"):
h = PF.convolution(v, 32, (3, 3))
b = PF.batch_normalization(h, batch_stat=True)
with nn.parameter_scope("conv2"):
h1 = PF.convolution(v, 32, (3, 3))
b2 = PF.batch_normalization(h1, batch_stat=True)
for k, v in iteritems(nn.get_parameters(grad_only=False)):
v.data.cast(np.float32)[...] = np.random.randn(*v.shape)
with nn.parameter_scope("param1"):
param1 = nn.get_parameters(grad_only=False)
nn.save_parameters("tmp.h5")
nn.save_parameters("tmp.protobuf")
with nn.parameter_scope("param2"):
nn.load_parameters('tmp.h5')
param2 = nn.get_parameters(grad_only=False)
with nn.parameter_scope("param3"):
nn.load_parameters('tmp.protobuf')
param3 = nn.get_parameters(grad_only=False)
for par2 in [param2, param3]:
assert param1.keys() == par2.keys() # Check order
for (n1, p1), (n2, p2) in zip(sorted(param1.items()), sorted(par2.items())):
assert n1 == n2
assert np.all(p1.d == p2.d)
assert p1.data.dtype == p2.data.dtype
assert p1.need_grad == p2.need_grad
def test_graph_model(model, seed):
np.random.seed(313)
rng = np.random.RandomState(seed)
x = nn.Variable([2, 3, 4, 4], need_grad=True)
t = nn.Variable([2, 1])
x.d = rng.randn(*x.shape)
t.d = rng.randint(0, 5, size=t.shape)
nn.set_default_context(nn.Context())
# Forwardprop by definintion
nn.clear_parameters()
if model == "mlp":
with nn.parameter_scope('fc1'):
z = PF.affine(x, 3)
z2 = F.relu(z, inplace=True)
with nn.parameter_scope('fc2'):
z3 = PF.affine(z2, 5)
elif model == "recurrent":
with nn.parameter_scope('fc1'):
z = PF.affine(x, 3)
z2 = F.relu(z, inplace=True)
h = z2
for _ in range(2):
with nn.parameter_scope('fc2'):
h = PF.affine(h, 3)
h = F.relu(h, inplace=True)
with nn.parameter_scope('fc3'):
z3 = PF.affine(h, 5)
elif model == "convolution":
with nn.parameter_scope('conv1'):
z = PF.convolution(x, 3, (2, 2))
z2 = F.relu(z, inplace=True)
with nn.parameter_scope('fc2'):
z3 = PF.affine(z2, 5)
else:
raise ValueError()
l = F.softmax_cross_entropy(z3, t, 1)
L = F.mean(l)
# Forwardprop
L.forward(clear_no_need_grad=True)
# Backprop
# Diff should be initialized since they are always accumulated
x.grad.zero()
L.backward(clear_buffer=True)
x.g = rng.randn(*x.shape)
parameters = nn.get_parameters()
for param in parameters.values():
param.grad.zero()
inputs = [x] + list(parameters.values())
from nbla_test_utils import \
compute_analytical_and_numerical_grad_graph as grads
agrad, ngrad = grads(L, inputs, 1e-3)
assert np.allclose(ngrad, agrad, atol=1.05e-2)
def test_parametric_function_1d(inshape, kernel, multiplier, outshape):
base_axis = len(inshape) - 2
sample_channels = inshape[base_axis]
outmap_channels = sample_channels * multiplier
x = nn.Variable(inshape)
y = PF.depthwise_convolution(x, kernel, multiplier=multiplier)
p = nn.get_parameters()
assert y.shape == outshape
assert p['depthwise_conv/W'].shape == (outmap_channels,) + kernel
assert p['depthwise_conv/b'].shape == (outmap_channels,)
nn.clear_parameters()
def decompose_network_and_set_params(model_load_path,
reference, slim, rrate=0.75):
# Parameters loaded globally, but call here for consistency
nn.load_parameters(model_load_path)
# Decompose
with nn.parameter_scope(reference):
trained_params = nn.get_parameters()
# original parameter
W = trained_params["fc3/affine/W"].d
# original maps
inmaps = W.shape[0]
outmaps0 = W.shape[1]
# new maps, R < N*M / (N+M) * rrate
outmaps1 = reduce_maps(inmaps, outmaps0, rrate)
# singular value decomposition
U, s, V = np.linalg.svd(W, full_matrices=False)
S = np.diag(s)
SV = S.dot(V)
U_approx = U[:, :outmaps1]
SV_approx = SV[:outmaps1, :outmaps0]
# Set trained parameters and decomposed parameters
# set trained parameters
with nn.parameter_scope(slim):
slim_params = nn.get_parameters()
for n, v in trained_params.items():
if not n in slim_params.keys():
continue
v_slim = slim_params[n]
v_slim.d = v.d
# set decomposed parameters and original bias
# a new bias is introduced due to decomposition
slim_params["fc-d0/affine/W"].d = U_approx
slim_params["fc-d1/affine/W"].d = SV_approx
b = trained_params["fc3/affine/b"]
slim_params["fc-d1/affine/b"].d = b.d
# Clear the parameters of the reference net
with nn.parameter_scope(reference):
nn.clear_parameters()
def test_graph_rewire(seed, clear_buffer):
nn.clear_parameters()
# A. defining graph definition utility
def mlp2(x, scope):
with nn.parameter_scope(scope):
h = F.tanh(PF.affine(x, 10, name='a1'))
h = F.tanh(PF.affine(h, 10, name='a1'))
return h
# A. Create a graph A.
xa = nn.Variable((2, 10), need_grad=True)
ya = mlp2(xa, 'a')
# B. Create a graph B.
xb = nn.Variable((2, 10), need_grad=True)
yb = mlp2(xb, 'b')
# C. Create directly connected graph.
xc = nn.Variable((2, 10))
yc = mlp2(mlp2(xc, 'a'), 'b')
# D. Rewire the graphs A and B.
xb.rewire_on(ya)
# E. Check whether the results are the same.
rng = np.random.RandomState(seed)
data = rng.randn(*xa.shape)
xa.d = data
xc.d = data
params = nn.get_parameters()
def zero_grad():
for p in params.values():
p.grad.zero()
def backup_params():
return [p.g.copy() for p in params.values()]
# Checking forward
yb.forward(clear_no_need_grad=clear_buffer)
yc.forward(clear_no_need_grad=clear_buffer)
assert_allclose(yb.d, yc.d)
# Checking backward
zero_grad()
yb.backward(clear_buffer=clear_buffer)
gb = backup_params()
zero_grad()
yc.backward(clear_buffer=clear_buffer)
gc = backup_params()
assert_allclose(xa.d, xc.d)
for b, c in zip(gb, gc):
assert_allclose(b, c)
def load_parameters(self, path, extension=".h5"):
"""Load parameters from a file into this module.
Args:
path: str or file-like object
"""
scope = OrderedDict()
with nn.parameter_scope('', scope):
nn.load_parameters(path, extension=extension)
params = nn.get_parameters()
self.set_parameters(params)
def modify(self, f, inputs):
params = [v.data for v in nn.get_parameters(grad_only=False).values()]
inputs_ = []
for inp in inputs:
if inp.data not in params:
inputs_.append(inp)
else:
inp = inp.get_unlinked_variable(need_grad=False)
inputs_.append(inp)
o = self._call_function(f.info.type_name, inputs_, f.info.args)
return o
def test_load_save_parameters():
module = MyModule(shape=(5, 5))
params = module.get_parameters()
if not os.path.exists('__nnabla_nas__'):
os.makedirs('__nnabla_nas__')
nn.save_parameters('__nnabla_nas__/params.h5', params)
nn.load_parameters('__nnabla_nas__/params.h5')
params0 = nn.get_parameters()
for k, v in params.items():
assert_allclose(v.d, params0[k].d)
def load_parameters(self, path, extension=".h5", raise_if_missing=True):
"""Load parameters from a file into this module.
Args:
path: str or file-like object
"""
scope = OrderedDict()
with nn.parameter_scope('', scope):
nn.load_parameters(path, extension=extension)
params = nn.get_parameters(grad_only=False)
self.set_parameters(params, raise_if_missing=raise_if_missing)
def load_parameters(self, path, raise_if_missing=False):
r"""Loads parameters from a file with the specified format.
Args:
path (str): The path to file.
raise_if_missing (bool, optional): Raise exception if some
parameters are missing. Defaults to `False`.
"""
with nn.parameter_scope('', OrderedDict()):
nn.load_parameters(path)
params = nn.get_parameters(grad_only=False)
self.set_parameters(params, raise_if_missing=raise_if_missing)
def train(self):
# variables for training
tx_in = nn.Variable(
[self._batch_size, self._x_input_length, self._cols_size])
tx_out = nn.Variable(
[self._batch_size, self._x_output_length, self._cols_size])
tpred = self.network(tx_in, self._lstm_unit_name, self._lstm_units)
tpred.persistent = True
loss = F.mean(F.squared_error(tpred, tx_out))
solver = S.Adam(self._learning_rate)
solver.set_parameters(nn.get_parameters())
# variables for validation
vx_in = nn.Variable(
[self._batch_size, self._x_input_length, self._cols_size])
vx_out = nn.Variable(
[self._batch_size, self._x_output_length, self._cols_size])
vpred = self.network(vx_in, self._lstm_unit_name, self._lstm_units)
# data iterators
tdata = self._load_dataset(self._training_dataset_path,
self._batch_size,
shuffle=True)
vdata = self._load_dataset(self._validation_dataset_path,
self._batch_size,
shuffle=True)
# monitors
from nnabla.monitor import Monitor, MonitorSeries, MonitorTimeElapsed
monitor = Monitor(self._monitor_path)
monitor_loss = MonitorSeries("Training loss", monitor, interval=10)
monitor_err = MonitorSeries("Training error", monitor, interval=10)
monitor_time = MonitorTimeElapsed("Training time",
monitor,
interval=100)
monitor_verr = MonitorSeries("Validation error", monitor, interval=10)
# Training loop
for i in range(self._max_iter):
if i % self._val_interval == 0:
ve = self._validate(vpred, vx_in, vx_out, vdata,
self._val_iter)
monitor_verr.add(i, ve / self._val_iter)
te = self._train(tpred, solver, loss, tx_in, tx_out, tdata.next(),
self._weight_decay)
monitor_loss.add(i, loss.d.copy())
monitor_err.add(i, te)
monitor_time.add(i)
ve = self._validate(vpred, vx_in, vx_out, vdata, self._val_iter)
monitor_verr.add(i, ve / self._val_iter)
# Save a best model parameters
nn.save_parameters(self._model_params_path)
def _get_variable_or_create(self, v, callback, current_scope):
if v.variable is not None:
return v.variable
v = callback._apply_generate_variable(v)
if v.variable is not None:
return v.variable
pvar = v.proto
name = pvar.name
shape = list(pvar.shape.dim)
if shape[0] < 0:
shape[0] = self.batch_size
shape = tuple(shape)
assert np.all(np.array(shape) > 0
), "Shape must be positive. Given {}.".format(shape)
if pvar.type != 'Parameter':
# Create a new variable and returns.
var = nn.Variable(shape)
v.variable = var
var.name = name
return var
# Trying to load the parameter from .nnp file.
callback.verbose('Loading parameter `{}` from .nnp.'.format(name))
try:
param = get_parameter(name)
if param is None:
logger.info(
'Parameter `{}` is not found. Initializing.'.format(name))
tmp = _create_variable(pvar, name, shape, self.rng)
param = tmp.variable_instance
set_parameter(name, param)
# Always copy param to current scope even if it already exists.
with nn.parameter_scope('', current_scope):
set_parameter(name, param)
except:
import sys
import traceback
raise ValueError(
'An error occurs during creation of a variable `{}` as a'
' parameter variable. The error was:\n----\n{}\n----\n'
'The parameters registered was {}'.format(
name, traceback.format_exc(), '\n'.join(
list(nn.get_parameters(grad_only=False).keys()))))
assert shape == param.shape
param = param.get_unlinked_variable(need_grad=v.need_grad)
v.variable = param
param.name = name
return param
def test_graph_clear_buffer(seed):
np.random.seed(313)
rng = np.random.RandomState(seed)
x = nn.Variable([2, 3, 4, 4])
t = nn.Variable([2, 1])
x.d = rng.randn(*x.shape)
t.d = rng.randint(0, 5, size=t.shape)
# Network definition
nn.set_default_context(nn.Context())
nn.clear_parameters()
x1 = x + 1
x2 = x1 - 1
with nn.parameter_scope('conv1'):
z = PF.convolution(x2, 3, (2, 2))
z2 = F.relu(z, inplace=True)
with nn.parameter_scope('fc2'):
z3 = PF.affine(z2, 5)
l = F.softmax_cross_entropy(z3, t, 1)
L = F.mean(l)
# Forwardprop
import tempfile
import os
tmpd = tempfile.mkdtemp()
nn.save_parameters(os.path.join(tmpd, 'parameter.h5'))
first = False
for cnng in [False, True]:
for cb in [False, True]:
_ = nn.load_parameters(os.path.join(tmpd, 'parameter.h5'))
for v in nn.get_parameters().values():
v.grad.zero()
L.forward(clear_no_need_grad=cnng)
L.backward(clear_buffer=cb)
if not first:
first = True
g = list(nn.get_parameters().values())[0].g.copy()
else:
g2 = list(nn.get_parameters().values())[0].g.copy()
assert np.all(g == g2)
def test_graph_clear_buffer(seed):
np.random.seed(313)
rng = np.random.RandomState(seed)
x = nn.Variable([2, 3, 4, 4])
t = nn.Variable([2, 1])
x.d = rng.randn(*x.shape)
t.d = rng.randint(0, 5, size=t.shape)
# Network definition
nn.set_default_context(nn.Context())
nn.clear_parameters()
x1 = x + 1
x2 = x1 - 1
with nn.parameter_scope('conv1'):
z = PF.convolution(x2, 3, (2, 2))
z2 = F.relu(z, inplace=True)
with nn.parameter_scope('fc2'):
z3 = PF.affine(z2, 5)
l = F.softmax_cross_entropy(z3, t, 1)
L = F.mean(l)
# Forwardprop
import tempfile
import os
tmpd = tempfile.mkdtemp()
nn.save_parameters(os.path.join(tmpd, 'parameter.h5'))
first = False
for cnng in [False, True]:
for cb in [False, True]:
_ = nn.load_parameters(os.path.join(tmpd, 'parameter.h5'))
for v in nn.get_parameters().values():
v.grad.zero()
L.forward(clear_no_need_grad=cnng)
L.backward(clear_buffer=cb)
if not first:
first = True
g = list(nn.get_parameters().values())[0].g.copy()
else:
g2 = list(nn.get_parameters().values())[0].g.copy()
assert np.all(g == g2)
def encode_text(text):
param_dict = nn.get_parameters()
embed_dim = param_dict['text_projection'].shape[1]
context_length = param_dict['positional_embedding'].shape[0]
vocab_size = param_dict['token_embedding/W'].shape[0]
transformer_width = param_dict['ln_final/W'].shape[0]
transformer_heads = transformer_width // 64
transformer_layers = len(
set(
k.split('/')[2] for k in param_dict.keys()
if k.startswith(f'transformer/resblocks')))
token_embedding = nn.parameter.get_parameter_or_create(
name='token_embedding/W', shape=(vocab_size, transformer_width))
x = F.embed(text, token_embedding) # [batch_size, n_ctx, d_model]
positional_embedding = nn.parameter.get_parameter_or_create(
name='positional_embedding',
shape=(context_length, transformer_width)).reshape(
(1, context_length, transformer_width))
x = x + positional_embedding
x = F.transpose(x, (1, 0, 2)) # NLD -> LND
x = transformer(x,
transformer_width,
transformer_layers,
transformer_heads,
attn_mask=build_attn_mask(context_length))
x = F.transpose(x, (1, 0, 2)) # LND -> NLD
ln_final_W = nn.parameter.get_parameter_or_create(
name='ln_final/W', shape=(transformer_width, )).reshape(
(1, 1, transformer_width))
ln_final_b = nn.parameter.get_parameter_or_create(
name='ln_final/b', shape=(transformer_width, )).reshape(
(1, 1, transformer_width))
x = F.layer_normalization(x, ln_final_b, ln_final_W, batch_axis=(0, 1))
idx = F.max(text, axis=-1, only_index=True)
idx.forward()
x = x[list(range(x.shape[0])), idx.d].reshape((1, x.shape[0], -1))
text_projection = nn.parameter.get_parameter_or_create(
name='text_projection', shape=(transformer_width, embed_dim)).reshape(
(1, transformer_width, embed_dim))
x = F.batch_matmul(x, text_projection)
x = x.reshape((-1, embed_dim))
return x
def get_parameters(self, recursive=True, grad_only=False, memo=None):
"""Obtain an OrderedDict object of all parameters in current Module.
For example,
.. code-block:: python
x = nn.Variable.from_numpy_array((np.random.random((8, 32, 256, 256))))
conv_bn = ConvBn(2)
y = conv_bn(x)
params = conv_bn.get_parameters()
for parameter_name, parameter_value in params.items():
print("{}:{}".format(parameter_name, parameter_value.shape))
The output looks like:
.. code-block:: none
conv/W:(2, 32, 1, 1)
bn/beta:(1, 2, 1, 1)
bn/gamma:(1, 2, 1, 1)
bn/mean:(1, 2, 1, 1)
bn/var:(1, 2, 1, 1)
Notice that the parameter name looks like a filepath, with splash separated
nested scope name. In addition, module name default is used with a prefix ``@``.
Args:
recursive (bool, optional, default=True):
Whether obtain the parameters of current module's submodules. Default is True.
grad_only (bool, optional, default=False):
Whether only obtain the grad. Default is False.
Returns:
OrderedDict:
Flattened parameter's name-value pairs of current Module.
"""
params = OrderedDict()
if memo is None:
memo = ParamMemo()
if recursive:
for name, module in self.submodules.items():
params.update(
insert_parent_name(
name,
module.get_parameters(recursive=recursive, grad_only=grad_only, memo=memo)))
with nn.parameter_scope('', self.parameter_scope):
found_params = nn.get_parameters(grad_only=grad_only)
filtered_params = memo.filter_and_update(found_params)
params.update(filtered_params)
return params
def __init__(self, black_list=[], params=None, name="identity"):
self.graph_info = None
self.entry_variables = None
self.black_list = black_list
self.params = params if params is not None else nn.get_parameters(
grad_only=False)
self.name = name
self.end_variable = None
self.outputs = []
# output of ref graph to output of new graph (TODO: change name)
self.input_map = {}
def training(steps, learning_rate):
solver = S.Sgd(learning_rate)
solver.set_parameters(
nn.get_parameters()) # Set parameter variables to be updated.
for i in range(steps):
x.d, t.d = data.next()
loss.forward()
solver.zero_grad() # Initialize gradients of all parameters to zero.
loss.backward()
solver.weight_decay(1e-5) # Applying weight decay as an regularization
solver.update()
if i % 100 == 0: # Print for each 10 iterations
print(i, loss.d)
def save_all_params(params_dict, c, k, j, bundle_size, step_size, save_dir,
epoch):
params_dict[c] = nn.get_parameters(grad_only=False).copy()
c += 1
if c == bundle_size or j == step_size - 1:
dn = os.path.join(save_dir, 'epoch%02d' % (epoch), 'weights')
ensure_dir(dn)
for cc, params in params_dict.items():
fn = '%s/model_step%04d.h5' % (dn, k + cc)
nn.save_parameters(fn, params=params, extension=".h5")
k += c
c = 0
params_dict = {}
return params_dict, c, k
def __init__(self,
batch_size=32,
learning_rate=1e-4,
max_iter=5086,
total_epochs=20,
monitor_path=None,
val_weight=None,
model_load_path=None):
"""
Construct all the necessary attributes for the attribute classifier.
Args:
batch_size (int): number of samples contained in each generated batch
learning_rate (float) : learning rate
max_iter (int) : maximum iterations for an epoch
total_epochs (int) : total epochs to train the model
val_weight : sample weights
monitor_path (str) : model parameter to be saved
model_load_path (str) : load the model
"""
self.batch_size = batch_size
# Resnet 50
# training graph
model = ResNet50()
self.input_image = nn.Variable((self.batch_size, ) + model.input_shape)
self.label = nn.Variable([self.batch_size, 1])
# fine tuning
pool = model(self.input_image, training=True, use_up_to='pool')
self.clf = clf_resnet50(pool)
self.clf.persistent = True
# loss
self.loss = F.mean(F.sigmoid_cross_entropy(self.clf, self.label))
# hyper parameters
self.solver = S.Adam(learning_rate)
self.solver.set_parameters(nn.get_parameters())
# validation graph
self.x_v = nn.Variable((self.batch_size, ) + model.input_shape)
pool_v = model(self.x_v, training=False, use_up_to='pool')
self.v_clf = clf_resnet50(pool_v, train=False)
self.v_clf_out = F.sigmoid(self.v_clf)
self.print_freq = 100
self.validation_weight = val_weight
# val params
self.acc = 0.0
self.total_epochs = total_epochs
self.max_iter = max_iter
self.monitor_path = monitor_path
if model_load_path is not None:
_ = nn.load_parameters(model_load_path)
def sample_arch_and_train(args, data_dict, controller_weights_dict):
"""
Execute these process.
1. For a certain number of times, let the controller construct sample architectures
and test their performances. (By calling get_sample_and_feedback)
2. By using the performances acquired by the previous process, train the controller.
3. Select one architecture with the best validation accuracy and train its parameters.
"""
solver = S.Momentum(args.control_lr) # create solver for the controller
solver.set_parameters(controller_weights_dict,
reset=False,
retain_state=True)
solver.zero_grad()
val_list = list()
arch_list = list()
with nn.auto_forward():
for c in range(args.num_candidate):
output_line = " Architecture {} / {} ".format((c + 1),
args.num_candidate)
print("{0:-^80s}".format(output_line))
# sample one architecture and get its feedback for RL as loss
loss, val_acc, sample_arch = get_sample_and_feedback(
args, data_dict)
val_list.append(val_acc)
arch_list.append(sample_arch)
loss.backward() # accumulate gradient each time
print("{0:-^80s}\n".format(" Reinforcement Learning Phase "))
print("current accumulated loss:", loss.d)
solver.weight_decay(0.025)
solver.update() # train the controller
print("\n{0:-^80s}\n".format(" CNN Learning Phase "))
best_idx = np.argmax(val_list)
sample_arch = arch_list[best_idx]
print("Train the model whose architecture is:")
show_arch(sample_arch)
print("and its accuracy is: {:.2f} %\n".format(100 * np.max(val_list)))
print("Learnable Parameters:", params_count(nn.get_parameters()))
# train a child network which achieves the best validation accuracy.
val_acc = CNN_run(args, sample_arch, data_dict, with_train=True)
return sample_arch, val_acc
def test_pf_prelu_execution(g_rng, inshape, base_axis, shared, slope_init,
fix_parameters):
slope_shape = tuple() if shared else (inshape[base_axis], )
slope_init = process_param_init(slope_init, slope_shape, g_rng)
kw = {}
insert_if_not_none(kw, 'slope_init', slope_init)
insert_if_not_default(kw, 'base_axis', base_axis, 1)
insert_if_not_default(kw, 'shared', shared, True)
insert_if_not_default(kw, 'fix_parameters', fix_parameters, False)
x = nn.Variable.from_numpy_array(g_rng.randn(*inshape))
# Check execution
y = PF.prelu(x, **kw)
y.forward()
y.backward()
# Check values
# TODO
# Check args
assert y.parent.info.type_name == 'PReLU'
args = y.parent.info.args
assert args['base_axis'] == base_axis
# Check created parameters
assert y.parent.inputs[0] == x
assert len(y.parent.inputs) == 2
assert len(nn.get_parameters()) == 1
slope = nn.get_parameters()['prelu/slope']
assert slope.shape == slope_shape
assert slope.need_grad
assert y.parent.inputs[1].need_grad == (not fix_parameters)
if isinstance(slope_init, np.ndarray):
assert np.allclose(slope_init, slope.d)
def main():
args = get_args()
nn.load_parameters(args.input)
params = nn.get_parameters(grad_only=False)
processed = False
# Convert memory layout
layout = get_memory_layout(params)
if args.memory_layout is None:
pass
if args.affine_to_conv:
rm_list = []
ret = affine_to_conv(params, args.memory_layout, rm_list)
for r in rm_list:
print(r)
nn.parameter.pop_parameter(r)
if ret:
logger.info('Converted affine to conv.')
processed |= ret
if args.memory_layout != layout:
logger.info(f'Converting memory layout to {args.memory_layout}.')
convert_memory_layout(params, args.memory_layout)
processed |= True
else:
logger.info('No need to convert memory layout.')
if args.force_4_channels:
ret = force_4_channels(params, args.memory_layout)
if ret:
logger.info('Converted first conv to 4-channel input.')
processed |= ret
if args.force_3_channels:
ret = force_3_channels(params, args.memory_layout)
if ret:
logger.info('Converted first conv to 3-channel input.')
processed |= ret
nn.clear_parameters()
for key, param in params.items():
print(key)
print(param.shape)
nn.parameter.set_parameter(key, param)
if not processed:
logger.info(
'No change has been made for the input. Not saving a new parameter file.')
return
logger.info(f'Save a new parameter file at {args.output}')
nn.save_parameters(args.output)
def test_save_load_parameters():
v = nn.Variable([64, 1, 28, 28], need_grad=False)
with nn.parameter_scope("param1"):
with nn.parameter_scope("conv1"):
h = PF.convolution(v, 32, (3, 3))
b = PF.batch_normalization(h, batch_stat=True)
with nn.parameter_scope("conv2"):
h1 = PF.convolution(v, 32, (3, 3))
b2 = PF.batch_normalization(h1, batch_stat=True)
for k, v in iteritems(nn.get_parameters(grad_only=False)):
v.data.cast(np.float32)[...] = np.random.randn(*v.shape)
with nn.parameter_scope("param1"):
param1 = nn.get_parameters(grad_only=False)
nn.save_parameters("tmp.h5")
nn.save_parameters("tmp.protobuf")
with nn.parameter_scope("param2"):
nn.load_parameters('tmp.h5')
param2 = nn.get_parameters(grad_only=False)
with nn.parameter_scope("param3"):
nn.load_parameters('tmp.protobuf')
param3 = nn.get_parameters(grad_only=False)
for par2 in [param2, param3]:
assert param1.keys() == par2.keys() # Check order
for (n1, p1), (n2, p2) in zip(sorted(param1.items()),
sorted(par2.items())):
assert n1 == n2
assert np.all(p1.d == p2.d)
if par2 is not param3:
# NOTE: data is automatically casted to fp32 in Protobuf
assert p1.data.dtype == p2.data.dtype
assert p1.need_grad == p2.need_grad
def create_nnabla_net(resnext=False):
# Create nnabla graph
from models import senet
x = nn.Variable((1, 3, 224, 224), need_grad=False)
if resnext:
y = senet.se_resnext50(x, 1000, test=True)
else:
y = senet.se_resnet50(x, 1000, test=True)
params = nn.get_parameters(grad_only=False)
param_dims = 0
for k, v in params.items():
param_dims += np.prod(v.shape)
print(k, v.shape, param_dims)
print('total parameters: ', param_dims)
return (x, y), params, param_dims
def test_compute_simple_hessian(ctx):
nn.clear_parameters()
# Network
state = nn.Variable((1, 2))
output = PF.affine(state,
1,
w_init=I.ConstantInitializer(value=1.),
b_init=I.ConstantInitializer(value=1.))
loss = F.sum(output**2)
# Input
state_array = np.array([[1.0, 0.5]])
state.d = state_array
# Grad of network
params = nn.get_parameters().values()
for param in params:
param.grad.zero()
grads = nn.grad([loss], params)
flat_grads = F.concatenate(*[F.reshape(grad, (-1,)) for grad in grads]) if len(grads) > 1 \
else F.reshape(grads[0], (-1,))
# Compute hessian
hessian = np.zeros((flat_grads.shape[0], flat_grads.shape[0]),
dtype=np.float32)
for i in range(flat_grads.shape[0]):
flat_grads_i = flat_grads[i]
flat_grads_i.forward()
for param in params:
param.grad.zero()
flat_grads_i.backward()
num_index = 0
for param in params:
grad = param.g.flatten() # grad of grad so this is hessian
hessian[i, num_index:num_index + len(grad)] = grad
num_index += len(grad)
actual = hessian
expected = np.array([[
2 * state_array[0, 0]**2, 2 * state_array[0, 0] * state_array[0, 1],
2 * state_array[0, 0]
],
[
2 * state_array[0, 0] * state_array[0, 1],
2 * state_array[0, 1]**2, 2 * state_array[0, 1]
], [2 * state_array[0, 0], 2 * state_array[0, 1],
2.]])
assert_allclose(actual, expected)
def data_distill(model, uniform_data_iterator, num_iter):
generated_img = []
for _ in range(uniform_data_iterator.size //
uniform_data_iterator.batch_size):
img, _ = uniform_data_iterator.next()
dst_img = nn.Variable(img.shape, need_grad=True)
dst_img.d = img
img_params = OrderedDict()
img_params['img'] = dst_img
init_lr = 0.5
solver = S.Adam(alpha=init_lr)
solver.set_parameters(img_params)
#scheduler = lr_scheduler.CosineScheduler(init_lr=0.5, max_iter=num_iter)
scheduler = ReduceLROnPlateauScheduler(init_lr=init_lr,
min_lr=1e-4,
verbose=False,
patience=100)
dummy_solver = S.Sgd(lr=0)
dummy_solver.set_parameters(nn.get_parameters())
for it in tqdm(range(num_iter)):
lr = scheduler.get_learning_rate()
solver.set_learning_rate(lr)
global outs
outs = []
global batch_stats
batch_stats = []
y = model(denormalize(dst_img),
force_global_pooling=True,
training=False) # denormalize to U(0, 255)
y.forward(function_post_hook=get_output)
assert len(outs) == len(batch_stats)
loss = zeroq_loss(batch_stats, outs, dst_img)
loss.forward()
solver.zero_grad()
dummy_solver.zero_grad()
loss.backward()
solver.weight_decay(1e-6)
solver.update()
scheduler.update_lr(loss.d)
generated_img.append(dst_img.d)
return generated_img
def mNextParam(self, idx):
if np.random.rand() > 0.8:
with nn.parameter_scope(net(7)):
param = nn.get_parameters()
for i, j in param.items():
self.mParam[net(idx)]["pre"][i] = self.mParam[net(
idx)]["next"][i].copy()
self.mParam[net(idx)]["next"][i] = param.get(i).d
else:
for i in self.mParam[net(idx)]["next"].keys():
self.mParam[net(idx)]["pre"][i] = self.mParam[net(
idx)]["next"][i].copy()
self.mParam[net(idx)]["next"][i] = np.random.randn(
*(self.mParam[net(idx)]["next"][i].shape))
return
return
def __init__(self, graph, device_id, ext_name, solver=None, n_run=100, max_measure_execution_time=1,
time_scale="m"):
self.graph = graph
# if solver is None, training time (forward + backward + update) is not calculated
self.solver = solver
self.n_run = n_run
self.device_id = str(device_id)
self.ext_name = ext_name
self.ext_module = import_extension_module(self.ext_name)
self.max_measure_execution_time = max_measure_execution_time
self.time_scale = time_scale
self.result = dict()
self.name2val = {v: k for k, v in nn.get_parameters().items()}
if self.n_run < 1:
raise AssertionError("n_run must be bigger than 1")
def create_graphviz_digraph(self, vleaf, format=None):
'''
Create a :obj:`graphviz.Digraph` object given the leaf variable of a
computation graph.
One of nice things of getting ``Digraph`` directly is that the drawn
graph can be displayed inline in a Jupyter notebook as described in
`Graphviz documentation <https://graphviz.readthedocs.io/en/stable/manual.html#jupyter-notebooks>`_.
Args:
vleaf (`nnabla.Variable`):
End variable. All variables and functions which can be
traversed from this variable are shown in the reuslt.
format (str):
Force overwrite ``format`` (``'pdf', 'png', ...)``) configuration.
Returns: graphviz.Digraph
'''
from nnabla import get_parameters
import copy
try:
from graphviz import Digraph
except:
raise ImportError("Install graphviz. `pip install graphviz.`")
if format is None:
format = self._format
graph = Digraph(format=format)
graph.attr("node", style="filled")
params = get_parameters(grad_only=False)
var2name = {v.data: k for k, v in params.items()}
fun2scope = {}
var2postname = copy.copy(var2name)
def fscope(f):
names = [var2name[v.data] for v in f.inputs if v.data in var2name]
if names:
c = os.path.commonprefix(names)
fun2scope[f] = c
for n in names:
var2postname[params[n].data] = n[len(c):]
vleaf.visit(fscope)
func = self.functor(graph, self._verbose,
fun2scope=fun2scope, var2name=var2postname)
vleaf.visit(func)
return graph
def get_parameters(self, recursive=True, grad_only=False, memo=None):
params = OrderedDict()
if memo is None:
memo = ParamMemo()
if recursive:
for name, module in self.submodules.items():
params.update(
insert_parent_name(
name,
module.get_parameters(recursive=recursive,
grad_only=grad_only,
memo=memo)))
with nn.parameter_scope('', self.parameter_scope):
found_params = nn.get_parameters(grad_only=grad_only)
filtered_params = memo.filter_and_update(found_params)
params.update(filtered_params)
return params
def train(max_iter=60000):
# Initialize data provider
di_l = I.data_iterator_mnist(batch_size, True)
di_t = I.data_iterator_mnist(batch_size, False)
# Network
shape_x = (1, 28, 28)
shape_z = (50, )
x = nn.Variable((batch_size, ) + shape_x)
loss_l = I.vae(x, shape_z, test=False)
loss_t = I.vae(x, shape_z, test=True)
# Create solver
solver = S.Adam(learning_rate)
solver.set_parameters(nn.get_parameters())
# Monitors for training and validation
path = cache_dir(os.path.join(I.name, "monitor"))
monitor = M.Monitor(path)
monitor_train_loss = M.MonitorSeries("train_loss", monitor, interval=600)
monitor_val_loss = M.MonitorSeries("val_loss", monitor, interval=600)
monitor_time = M.MonitorTimeElapsed("time", monitor, interval=600)
# Training Loop.
for i in range(max_iter):
# Initialize gradients
solver.zero_grad()
# Forward, backward and update
x.d, _ = di_l.next()
loss_l.forward(clear_no_need_grad=True)
loss_l.backward(clear_buffer=True)
solver.weight_decay(weight_decay)
solver.update()
# Forward for test
x.d, _ = di_t.next()
loss_t.forward(clear_no_need_grad=True)
# Monitor for logging
monitor_train_loss.add(i, loss_l.d.copy())
monitor_val_loss.add(i, loss_t.d.copy())
monitor_time.add(i)
return path
def test_propagate(self):
"""
The graph used bellow is from Fig.1 in the original paper (https://arxiv.org/pdf/1511.05493.pdf)
"""
edges = {"B": [(0, 1), (3, 2)], "C": [(2, 1), (1, 3)]}
with nn.parameter_scope("test_propagate"):
vertices = nn.Variable((4, 1))
outputs = L.propagate(vertices, edges)
params = nn.get_parameters()
self.assertEqual((4, 1), outputs.shape)
self.assertEqual(8, len(params))
self.assertEqual((1, 3, 1), params["W_zr/affine/W"].shape)
self.assertEqual((1, 2, 1), params["U_zr/affine/W"].shape)
self.assertEqual((1, 1), params["U/affine/W"].shape)
def main(args):
# Settings
device_id = args.device_id
batch_size = args.batch_size
batch_size_eval = args.batch_size_eval
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = int(n_train_data / batch_size)
n_iter = n_epoch * iter_epoch
extension_module = args.context
alpha = args.alpha
# Supervised Model
## ERM
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l_0 = nn.Variable((batch_size, m, h, w))
y_l_0 = nn.Variable((batch_size, 1))
pred = cnn_model_003(ctx, x_l_0)
loss_ce = ce_loss(ctx, pred, y_l_0)
loss_er = er_loss(ctx, pred)
loss_supervised = loss_ce + loss_er
## VRM (mixup)
x_l_1 = nn.Variable((batch_size, m, h, w))
y_l_1 = nn.Variable((batch_size, 1))
coef = nn.Variable()
coef_b = F.broadcast(coef.reshape([1]*x_l_0.ndim, unlink=True), x_l_0.shape)
x_l_m = coef_b * x_l_0 + (1 - coef_b) * x_l_1
coef_b = F.broadcast(coef.reshape([1]*pred.ndim, unlink=True), pred.shape)
y_l_m = coef_b * F.one_hot(y_l_0, (n_cls, )) \
+ (1-coef_b) * F.one_hot(y_l_1, (n_cls, ))
x_l_m.need_grad, y_l_m.need_grad = False, False
pred_m = cnn_model_003(ctx, x_l_m)
loss_er_m = er_loss(ctx, pred_m) #todo: need?
loss_ce_m = ce_loss_soft(ctx, pred, y_l_m)
loss_supervised_m = loss_ce_m #+ loss_er_m
# Semi-Supervised Model
## ERM
x_u0 = nn.Variable((batch_size, m, h, w))
x_u1 = nn.Variable((batch_size, m, h, w))
pred_x_u0 = cnn_model_003(ctx, x_u0)
pred_x_u1 = cnn_model_003(ctx, x_u1)
pred_x_u0.persistent, pred_x_u1.persistent = True, True
loss_sr = sr_loss(ctx, pred_x_u0, pred_x_u1)
loss_er0 = er_loss(ctx, pred_x_u0)
loss_er1 = er_loss(ctx, pred_x_u1)
loss_unsupervised = loss_sr + loss_er0 + loss_er1
## VRM (mixup)
x_u2 = nn.Variable((batch_size, m, h, w)) # not to overwrite x_u1.d
coef_u = nn.Variable()
coef_u_b = F.broadcast(coef_u.reshape([1]*x_u0.ndim, unlink=True), x_u0.shape)
x_u_m = coef_u_b * x_u0 + (1-coef_u_b) * x_u2
pred_x_u0_ = nn.Variable(pred_x_u0.shape) # unlink forward pass but reuse result
pred_x_u1_ = nn.Variable(pred_x_u1.shape)
pred_x_u0_.data = pred_x_u0.data
pred_x_u1_.data = pred_x_u1.data
coef_u_b = F.broadcast(coef_u.reshape([1]*pred_x_u0.ndim, unlink=True), pred_x_u0.shape)
y_u_m = coef_u_b * pred_x_u0_ + (1-coef_u_b) * pred_x_u1_
x_u_m.need_grad, y_u_m.need_grad = False, False
pred_x_u_m = cnn_model_003(ctx, x_u_m)
loss_er_u_m = er_loss(ctx, pred_x_u_m) #todo: need?
loss_ce_u_m = ce_loss_soft(ctx, pred_x_u_m, y_u_m)
loss_unsupervised_m = loss_ce_u_m #+ loss_er_u_m
# Evaluatation Model
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval = cnn_model_003(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
ve_best = 1.
save_path_prev = ""
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l_0.d, _ , y_l_0.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
## forward (supervised and its mixup)
loss_supervised.forward(clear_no_need_grad=True)
coef_data = np.random.beta(alpha, alpha)
coef.d = coef_data
x_l_1.d = np.random.permutation(x_l0_data)
y_l_1.d = np.random.permutation(y_l_data)
loss_supervised_m.forward(clear_no_need_grad=True)
## forward (unsupervised and its mixup)
loss_unsupervised.forward(clear_no_need_grad=True)
coef_data = np.random.beta(alpha, alpha)
coef_u.d = coef_data
x_u2.d = np.random.permutation(x_u1_data)
loss_unsupervised_m.forward(clear_no_need_grad=True)
## backward
solver.zero_grad()
loss_supervised.backward(clear_buffer=False)
loss_supervised_m.backward(clear_buffer=False)
loss_unsupervised.backward(clear_buffer=False)
loss_unsupervised_m.backward(clear_buffer=True)
solver.update()
# Evaluate
if int((i+1) % iter_epoch) == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
ve /= iter_val
msg = "Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve) * 100)
print(msg)
if ve < ve_best:
if not os.path.exists(args.model_save_path):
os.makedirs(args.model_save_path)
if save_path_prev != "":
os.remove(save_path_prev)
save_path = os.path.join(
args.model_save_path, 'params_%06d.h5' % epoch)
nn.save_parameters(save_path)
save_path_prev = save_path
ve_best = ve
st = time.time()
epoch +=1
def train():
"""
Main script.
"""
args = get_args()
# Get context.
from nnabla.contrib.context import extension_context
extension_module = args.context
if args.context is None:
extension_module = 'cpu'
logger.info("Running in %s" % extension_module)
ctx = extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
# Dataset
# We use Tiny ImageNet from Stanford CS231N class.
# https://tiny-imagenet.herokuapp.com/
# Tiny ImageNet consists of 200 categories, each category has 500 images
# in training set. The image size is 64x64. To adapt ResNet into 64x64
# image inputs, the input image size of ResNet is set as 56x56, and
# the stride in the first conv and the first max pooling are removed.
data = data_iterator_tiny_imagenet(args.batch_size, 'train')
vdata = data_iterator_tiny_imagenet(args.batch_size, 'val')
num_classes = 200
tiny = True # TODO: Switch ILSVRC2012 dataset and TinyImageNet.
t_model = get_model(
args, num_classes, test=False, tiny=tiny)
t_model.pred.persistent = True # Not clearing buffer of pred in backward
v_model = get_model(
args, num_classes, test=True, tiny=tiny)
v_model.pred.persistent = True # Not clearing buffer of pred in forward
# Create Solver.
solver = S.Momentum(args.learning_rate, 0.9)
solver.set_parameters(nn.get_parameters())
# Create monitor.
import nnabla.monitor as M
monitor = M.Monitor(args.monitor_path)
monitor_loss = M.MonitorSeries("Training loss", monitor, interval=10)
monitor_err = M.MonitorSeries("Training error", monitor, interval=10)
monitor_vloss = M.MonitorSeries("Validation loss", monitor, interval=10)
monitor_verr = M.MonitorSeries("Validation error", monitor, interval=10)
monitor_time = M.MonitorTimeElapsed("Training time", monitor, interval=10)
# Training loop.
for i in range(args.max_iter):
# Save parameters
if i % args.model_save_interval == 0:
nn.save_parameters(os.path.join(
args.model_save_path, 'param_%06d.h5' % i))
# Validation
if i % args.val_interval == 0:
# Clear all intermediate memory to save memory.
# t_model.loss.clear_recursive()
l = 0.0
e = 0.0
for j in range(args.val_iter):
images, labels = vdata.next()
v_model.image.d = images
v_model.label.d = labels
v_model.image.data.cast(np.uint8, ctx)
v_model.label.data.cast(np.int32, ctx)
v_model.loss.forward(clear_buffer=True)
l += v_model.loss.d
e += categorical_error(v_model.pred.d, v_model.label.d)
monitor_vloss.add(i, l / args.val_iter)
monitor_verr.add(i, e / args.val_iter)
# Clear all intermediate memory to save memory.
# v_model.loss.clear_recursive()
# Training
l = 0.0
e = 0.0
solver.zero_grad()
# Gradient accumulation loop
for j in range(args.accum_grad):
images, labels = data.next()
t_model.image.d = images
t_model.label.d = labels
t_model.image.data.cast(np.uint8, ctx)
t_model.label.data.cast(np.int32, ctx)
t_model.loss.forward(clear_no_need_grad=True)
t_model.loss.backward(clear_buffer=True) # Accumulating gradients
l += t_model.loss.d
e += categorical_error(t_model.pred.d, t_model.label.d)
solver.weight_decay(args.weight_decay)
solver.update()
monitor_loss.add(i, l / args.accum_grad)
monitor_err.add(i, e / args.accum_grad)
monitor_time.add(i)
# Learning rate decay at scheduled iter
if i in args.learning_rate_decay_at:
solver.set_learning_rate(solver.learning_rate() * 0.1)
nn.save_parameters(os.path.join(args.model_save_path,
'param_%06d.h5' % args.max_iter))
def main(args):
# Settings
device_id = args.device_id
batch_size = args.batch_size
batch_size_eval = args.batch_size_eval
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = int(n_train_data / batch_size)
n_iter = n_epoch * iter_epoch
extension_module = args.context
rampups_x = np.linspace(0, 1, 50)
rampups = np.exp(-5 * (1 - rampups_x)**2)
# Model
## supervised
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l = nn.Variable((batch_size, m, h, w))
y_l = nn.Variable((batch_size, 1))
pred = cnn_model_003(ctx, x_l)
loss_ce = ce_loss(ctx, pred, y_l)
loss_er = er_loss(ctx, pred)
loss_supervised = loss_ce + loss_er
## stochastic regularization
x_u0 = nn.Variable((batch_size, m, h, w))
x_u1 = nn.Variable((batch_size, m, h, w))
pred_x_u0 = cnn_model_003(ctx, x_u0)
pred_x_u1 = cnn_model_003(ctx, x_u1)
loss_sr = sr_loss(ctx, pred_x_u0, pred_x_u1)
loss_er0 = er_loss(ctx, pred_x_u0)
loss_er1 = er_loss(ctx, pred_x_u1)
coef = nn.Variable()
coef.d = 0
loss_unsupervised = coef * loss_sr + loss_er0 + loss_er1
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval = cnn_model_003(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
ve_best = 1.
save_path_prev = ""
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
loss_supervised.forward(clear_no_need_grad=True)
loss_unsupervised.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_supervised.backward(clear_buffer=True)
loss_unsupervised.backward(clear_buffer=True)
solver.update()
# Evaluate
if int((i+1) % iter_epoch) == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
ve /= iter_val
msg = "Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve) * 100)
print(msg)
st = time.time()
epoch +=1
coef.d = rampups[epoch]
def train(args):
"""
Main script.
"""
# Get context.
from nnabla.contrib.context import extension_context
extension_module = args.context
if args.context is None:
extension_module = 'cpu'
logger.info("Running in %s" % extension_module)
ctx = extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
# Create CNN network for both training and testing.
# TRAIN
# Fake path
z = nn.Variable([args.batch_size, 100, 1, 1])
fake = generator(z)
fake.persistent = True # Not to clear at backward
pred_fake = discriminator(fake)
loss_gen = F.mean(F.sigmoid_cross_entropy(
pred_fake, F.constant(1, pred_fake.shape)))
fake_dis = fake.unlinked()
pred_fake_dis = discriminator(fake_dis)
loss_dis = F.mean(F.sigmoid_cross_entropy(
pred_fake_dis, F.constant(0, pred_fake_dis.shape)))
# Real path
x = nn.Variable([args.batch_size, 1, 28, 28])
pred_real = discriminator(x)
loss_dis += F.mean(F.sigmoid_cross_entropy(pred_real,
F.constant(1, pred_real.shape)))
# Create Solver.
solver_gen = S.Adam(args.learning_rate, beta1=0.5)
solver_dis = S.Adam(args.learning_rate, beta1=0.5)
with nn.parameter_scope("gen"):
solver_gen.set_parameters(nn.get_parameters())
with nn.parameter_scope("dis"):
solver_dis.set_parameters(nn.get_parameters())
# Create monitor.
import nnabla.monitor as M
monitor = M.Monitor(args.monitor_path)
monitor_loss_gen = M.MonitorSeries("Generator loss", monitor, interval=10)
monitor_loss_dis = M.MonitorSeries(
"Discriminator loss", monitor, interval=10)
monitor_time = M.MonitorTimeElapsed("Time", monitor, interval=100)
monitor_fake = M.MonitorImageTile(
"Fake images", monitor, normalize_method=lambda x: x + 1 / 2.)
data = data_iterator_mnist(args.batch_size, True)
# Training loop.
for i in range(args.max_iter):
if i % args.model_save_interval == 0:
with nn.parameter_scope("gen"):
nn.save_parameters(os.path.join(
args.model_save_path, "generator_param_%06d.h5" % i))
with nn.parameter_scope("dis"):
nn.save_parameters(os.path.join(
args.model_save_path, "discriminator_param_%06d.h5" % i))
# Training forward
image, _ = data.next()
x.d = image / 255. - 0.5 # [0, 255] to [-1, 1]
z.d = np.random.randn(*z.shape)
# Generator update.
solver_gen.zero_grad()
loss_gen.forward(clear_no_need_grad=True)
loss_gen.backward(clear_buffer=True)
solver_gen.weight_decay(args.weight_decay)
solver_gen.update()
monitor_fake.add(i, fake)
monitor_loss_gen.add(i, loss_gen.d.copy())
# Discriminator update.
solver_dis.zero_grad()
loss_dis.forward(clear_no_need_grad=True)
loss_dis.backward(clear_buffer=True)
solver_dis.weight_decay(args.weight_decay)
solver_dis.update()
monitor_loss_dis.add(i, loss_dis.d.copy())
monitor_time.add(i)
nnp = os.path.join(
args.model_save_path, 'dcgan_%06d.nnp' % args.max_iter)
runtime_contents = {
'networks': [
{'name': 'Generator',
'batch_size': args.batch_size,
'outputs': {'G': fake},
'names': {'z': z}},
{'name': 'Discriminator',
'batch_size': args.batch_size,
'outputs': {'D': pred_real},
'names': {'x': x}}],
'executors': [
{'name': 'Generator',
'network': 'Generator',
'data': ['z'],
'output': ['G']},
{'name': 'Discriminator',
'network': 'Discriminator',
'data': ['x'],
'output': ['D']}]}
save.save(nnp, runtime_contents)
from cpp_forward_check import check_cpp_forward
check_cpp_forward(args.model_save_path, [z.d], [z], fake, nnp, "Generator")
def train(args):
"""
Main script.
"""
# Get context.
from nnabla.contrib.context import extension_context
extension_module = args.context
if args.context is None:
extension_module = 'cpu'
logger.info("Running in %s" % extension_module)
ctx = extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
# Create CNN network for both training and testing.
margin = 1.0 # Margin for contrastive loss.
# TRAIN
# Create input variables.
image0 = nn.Variable([args.batch_size, 1, 28, 28])
image1 = nn.Variable([args.batch_size, 1, 28, 28])
label = nn.Variable([args.batch_size])
# Create predition graph.
pred = mnist_lenet_siamese(image0, image1, test=False)
# Create loss function.
loss = F.mean(contrastive_loss(pred, label, margin))
# TEST
# Create input variables.
vimage0 = nn.Variable([args.batch_size, 1, 28, 28])
vimage1 = nn.Variable([args.batch_size, 1, 28, 28])
vlabel = nn.Variable([args.batch_size])
# Create predition graph.
vpred = mnist_lenet_siamese(vimage0, vimage1, test=True)
vloss = F.mean(contrastive_loss(vpred, vlabel, margin))
# Create Solver.
solver = S.Adam(args.learning_rate)
solver.set_parameters(nn.get_parameters())
# Create monitor.
import nnabla.monitor as M
monitor = M.Monitor(args.monitor_path)
monitor_loss = M.MonitorSeries("Training loss", monitor, interval=10)
monitor_time = M.MonitorTimeElapsed("Training time", monitor, interval=100)
monitor_vloss = M.MonitorSeries("Test loss", monitor, interval=10)
# Initialize DataIterator for MNIST.
rng = np.random.RandomState(313)
data = siamese_data_iterator(args.batch_size, True, rng)
vdata = siamese_data_iterator(args.batch_size, False, rng)
# Training loop.
for i in range(args.max_iter):
if i % args.val_interval == 0:
# Validation
ve = 0.0
for j in range(args.val_iter):
vimage0.d, vimage1.d, vlabel.d = vdata.next()
vloss.forward(clear_buffer=True)
ve += vloss.d
monitor_vloss.add(i, ve / args.val_iter)
if i % args.model_save_interval == 0:
nn.save_parameters(os.path.join(
args.model_save_path, 'params_%06d.h5' % i))
image0.d, image1.d, label.d = data.next()
solver.zero_grad()
# Training forward, backward and update
loss.forward(clear_no_need_grad=True)
loss.backward(clear_buffer=True)
solver.weight_decay(args.weight_decay)
solver.update()
monitor_loss.add(i, loss.d.copy())
monitor_time.add(i)
parameter_file = os.path.join(
args.model_save_path, 'params_%06d.h5' % args.max_iter)
nn.save_parameters(parameter_file)
nnp_file = os.path.join(
args.model_save_path, 'siamese_%06d.nnp' % (args.max_iter))
runtime_contents = {
'networks': [
{'name': 'Validation',
'batch_size': args.batch_size,
'outputs': {'y': vpred},
'names': {'x0': vimage0, 'x1': vimage1}}],
'executors': [
{'name': 'Runtime',
'network': 'Validation',
'data': ['x0', 'x1'],
'output': ['y']}]}
save.save(nnp_file, runtime_contents)
from cpp_forward_check import check_cpp_forward
check_cpp_forward(args.model_save_path, [vimage0.d, vimage1.d], [
vimage0, vimage1], vpred, nnp_file)
def test_data_parallel_communicator():
try:
import nnabla_ext
import nnabla_ext.cuda
from nnabla.contrib.context import extension_context
except:
pytest.skip("DataParallelCommunicator are only supported in CUDA now.")
n_devices = nnabla_ext.cuda.init.get_device_count()
if n_devices < 2:
pytest.skip("Number of cuda devices is less than 2.")
# Contexts and Computation Graph
extension_module = "cuda"
ctxs = []
for d in range(n_devices):
ctx = extension_context(extension_module,
device_id="{}".format(d))
ctxs.append(ctx)
with nn.context_scope(ctx):
x_data = np.random.rand(4, 5)
x = nn.Variable(x_data.shape)
with nn.parameter_scope("gpu{}".format(d)):
with nn.parameter_scope("affine1"):
z = PF.affine(x, 6)
with nn.parameter_scope("affine2"):
y = PF.affine(z, 5)
# Init w.g
grads = []
for d in range(n_devices):
with nn.parameter_scope("gpu{}".format(d)):
params = nn.get_parameters()
grad = []
for i, elm in enumerate(params.items()):
k, v = elm
grad_ = np.random.randn(*v.shape)
v.g = grad_
v.grad.cast(np.float32, ctxs[d])
grad.append(grad_)
grads.append(grad)
# Reference
ref_grads = []
with nn.parameter_scope("gpu{}".format(d)):
params = nn.get_parameters()
for i in range(len(params)):
ave_grad = 0
for d in range(n_devices):
ave_grad += grads[d][i]
ave_grad /= n_devices
ref_grads.append(ave_grad)
# Communicator
try:
comm = C.DataParalellCommunicator(ctxs[0])
except:
pytest.skip(
"DataParalellCommunicator is not supported in cpu or not linux platform.")
for d in range(n_devices):
with nn.parameter_scope("gpu{}".format(d)):
comm.add_context_and_parameters(
(ctxs[d], nn.get_parameters()))
comm.init()
comm.allreduce(division=True)
# Check
atol = 1e-6
for d in range(n_devices):
with nn.parameter_scope("gpu{}".format(d)):
params = nn.get_parameters()
for i, elm in enumerate(params.items()):
k, v = elm
assert np.allclose(ref_grads[i], v.g, atol=atol)
def train():
"""
Naive Multi-Device Training
NOTE: the communicator exposes low-level interfaces
* Parse command line arguments.
* Instantiate a communicator and set parameter variables.
* Specify contexts for computation.
* Initialize DataIterator.
* Construct a computation graph for training and one for validation.
* Initialize solver and set parameter variables to that.
* Create monitor instances for saving and displaying training stats.
* Training loop
* Computate error rate for validation data (periodically)
* Get a next minibatch.
* Execute forwardprop
* Set parameter gradients zero
* Execute backprop.
* Inplace allreduce (THIS IS THE MAIN difference from a single device training)
* Solver updates parameters by using gradients computed by backprop.
* Compute training error
"""
# Parse args
args = get_args()
n_train_samples = 50000
bs_valid = args.batch_size
# Communicator and Context
extension_module = "cuda.cudnn"
ctx = extension_context(extension_module)
comm = C.MultiProcessDataParalellCommunicator(ctx)
comm.init()
n_devices = comm.size
mpi_rank = comm.rank
device_id = mpi_rank
ctx = extension_context(extension_module, device_id=device_id)
# Create training graphs
test = False
image_train = nn.Variable((args.batch_size, 3, 32, 32))
label_train = nn.Variable((args.batch_size, 1))
pred_train = cifar100_resnet23_prediction(
image_train, ctx, test)
loss_train = cifar100_resnet32_loss(pred_train, label_train)
input_image_train = {"image": image_train, "label": label_train}
# add parameters to communicator
comm.add_context_and_parameters((ctx, nn.get_parameters()))
# Create validation graph
test = True
image_valid = nn.Variable((bs_valid, 3, 32, 32))
pred_valid = cifar100_resnet23_prediction(
image_valid, ctx, test)
input_image_valid = {"image": image_valid}
# Solvers
solver = S.Adam()
solver.set_parameters(nn.get_parameters())
base_lr = args.learning_rate
warmup_iter = int(1. * n_train_samples /
args.batch_size / n_devices) * args.warmup_epoch
warmup_slope = 1. * n_devices / warmup_iter
# Create monitor
from nnabla.monitor import Monitor, MonitorSeries, MonitorTimeElapsed
monitor = Monitor(args.monitor_path)
monitor_loss = MonitorSeries("Training loss", monitor, interval=10)
monitor_err = MonitorSeries("Training error", monitor, interval=10)
monitor_time = MonitorTimeElapsed("Training time", monitor, interval=100)
monitor_verr = MonitorSeries("Test error", monitor, interval=10)
with data_iterator_cifar100(args.batch_size, True) as tdata, \
data_iterator_cifar100(bs_valid, False) as vdata:
# Training-loop
for i in range(int(args.max_iter / n_devices)):
# Validation
if mpi_rank == 0:
if i % int(n_train_samples / args.batch_size / n_devices) == 0:
ve = 0.
for j in range(args.val_iter):
image, label = vdata.next()
input_image_valid["image"].d = image
pred_valid.forward()
ve += categorical_error(pred_valid.d, label)
ve /= args.val_iter
monitor_verr.add(i * n_devices, ve)
if i % int(args.model_save_interval / n_devices) == 0:
nn.save_parameters(os.path.join(
args.model_save_path, 'params_%06d.h5' % i))
# Forward/Zerograd/Backward
image, label = tdata.next()
input_image_train["image"].d = image
input_image_train["label"].d = label
loss_train.forward()
solver.zero_grad()
loss_train.backward()
# In-place Allreduce
comm.allreduce(division=True)
# Solvers update
solver.update()
# Linear Warmup
if i < warmup_iter:
lr = base_lr * n_devices * warmup_slope * i
solver.set_learning_rate(lr)
else:
lr = base_lr * n_devices
solver.set_learning_rate(lr)
if mpi_rank == 0:
e = categorical_error(
pred_train.d, input_image_train["label"].d)
monitor_loss.add(i * n_devices, loss_train.d.copy())
monitor_err.add(i * n_devices, e)
monitor_time.add(i * n_devices)
if mpi_rank == 0:
nn.save_parameters(os.path.join(
args.model_save_path,
'params_%06d.h5' % (args.max_iter / n_devices)))
def train():
"""
Main script.
Steps:
* Parse command line arguments.
* Specify a context for computation.
* Initialize DataIterator for MNIST.
* Construct a computation graph for training and validation.
* Initialize a solver and set parameter variables to it.
* Create monitor instances for saving and displaying training stats.
* Training loop
* Computate error rate for validation data (periodically)
* Get a next minibatch.
* Execute forwardprop on the training graph.
* Compute training error
* Set parameter gradients zero
* Execute backprop.
* Solver updates parameters by using gradients computed by backprop.
"""
args = get_args()
# Get context.
from nnabla.contrib.context import extension_context
extension_module = args.context
if args.context is None:
extension_module = 'cpu'
logger.info("Running in %s" % extension_module)
ctx = extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
# Create CNN network for both training and testing.
mnist_cnn_prediction = mnist_lenet_prediction
if args.net == 'resnet':
mnist_cnn_prediction = mnist_resnet_prediction
# TRAIN
# Create input variables.
image = nn.Variable([args.batch_size, 1, 28, 28])
label = nn.Variable([args.batch_size, 1])
# Create prediction graph.
pred = mnist_cnn_prediction(image, test=False)
pred.persistent = True
# Create loss function.
loss = F.mean(F.softmax_cross_entropy(pred, label))
# TEST
# Create input variables.
vimage = nn.Variable([args.batch_size, 1, 28, 28])
vlabel = nn.Variable([args.batch_size, 1])
# Create predition graph.
vpred = mnist_cnn_prediction(vimage, test=True)
# Create Solver.
solver = S.Adam(args.learning_rate)
solver.set_parameters(nn.get_parameters())
# Create monitor.
from nnabla.monitor import Monitor, MonitorSeries, MonitorTimeElapsed
monitor = Monitor(args.monitor_path)
monitor_loss = MonitorSeries("Training loss", monitor, interval=10)
monitor_err = MonitorSeries("Training error", monitor, interval=10)
monitor_time = MonitorTimeElapsed("Training time", monitor, interval=100)
monitor_verr = MonitorSeries("Test error", monitor, interval=10)
# Initialize DataIterator for MNIST.
data = data_iterator_mnist(args.batch_size, True)
vdata = data_iterator_mnist(args.batch_size, False)
# Training loop.
for i in range(args.max_iter):
if i % args.val_interval == 0:
# Validation
ve = 0.0
for j in range(args.val_iter):
vimage.d, vlabel.d = vdata.next()
vpred.forward(clear_buffer=True)
ve += categorical_error(vpred.d, vlabel.d)
monitor_verr.add(i, ve / args.val_iter)
if i % args.model_save_interval == 0:
nn.save_parameters(os.path.join(
args.model_save_path, 'params_%06d.h5' % i))
# Training forward
image.d, label.d = data.next()
solver.zero_grad()
loss.forward(clear_no_need_grad=True)
loss.backward(clear_buffer=True)
solver.weight_decay(args.weight_decay)
solver.update()
e = categorical_error(pred.d, label.d)
monitor_loss.add(i, loss.d.copy())
monitor_err.add(i, e)
monitor_time.add(i)
ve = 0.0
for j in range(args.val_iter):
vimage.d, vlabel.d = vdata.next()
vpred.forward(clear_buffer=True)
ve += categorical_error(vpred.d, vlabel.d)
monitor_verr.add(i, ve / args.val_iter)
parameter_file = os.path.join(
args.model_save_path, '{}_params_{:06}.h5'.format(args.net, args.max_iter))
nn.save_parameters(parameter_file)
def main(args):
# Settings
device_id = args.device_id
batch_size = args.batch_size
batch_size_eval = args.batch_size_eval
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = n_epoch * iter_epoch
extension_module = args.context
# Model
## supervised cnn
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l = nn.Variable((batch_size, m, h, w))
x_l.persistent = True
y_l = nn.Variable((batch_size, 1))
y_l.persistent = True
pred = cnn_model_003(ctx, "cnn", x_l)
loss_ce = ce_loss(ctx, pred, y_l)
loss_er = er_loss(ctx, pred)
loss_supervised = loss_ce + loss_er
## supervised resnet
pred_res = cifar10_resnet23_prediction(ctx, "resnet", x_l)
loss_res_ce = ce_loss(ctx, pred_res, y_l)
loss_res_supervised = loss_res_ce
## stochastic regularization for cnn
x_u0 = nn.Variable((batch_size, m, h, w))
x_u0.persistent = True
x_u1 = nn.Variable((batch_size, m, h, w))
pred_x_u0 = cnn_model_003(ctx, "cnn", x_u0)
pred_x_u0.persistent = True
pred_x_u1 = cnn_model_003(ctx, "cnn", x_u1)
loss_sr = sr_loss(ctx, pred_x_u0, pred_x_u1)
loss_er0 = er_loss(ctx, pred_x_u0)
loss_er1 = er_loss(ctx, pred_x_u1)
loss_unsupervised = loss_sr + loss_er0 + loss_er1
## knowledge transfer for resnet
pred_res_x_u0 = cifar10_resnet23_prediction(ctx, "resnet", x_u0)
loss_res_unsupervised = kl_divergence(ctx, pred_res_x_u0, pred_x_u0)
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
x_eval.persistent = True # reused
pred_eval = cnn_model_003(ctx, "cnn", x_eval, test=True)
pred_res_eval = cifar10_resnet23_prediction(ctx, "resnet", x_eval, test=True)
# Solver
with nn.context_scope(ctx):
with nn.parameter_scope("cnn"):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
with nn.parameter_scope("resnet"):
solver_res = S.Adam(alpha=learning_rate)
solver_res.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train for cnn
loss_supervised.forward(clear_no_need_grad=True)
loss_unsupervised.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_supervised.backward(clear_buffer=True)
loss_unsupervised.backward(clear_buffer=True)
solver.update()
# Train for resnet
loss_res_supervised.forward(clear_no_need_grad=True)
loss_res_unsupervised.forward(clear_no_need_grad=True)
solver_res.zero_grad()
loss_res_supervised.backward(clear_buffer=True)
pred_x_u0.need_grad = False # no need grad for teacher
loss_res_unsupervised.backward(clear_buffer=True)
solver_res.update()
pred_x_u0.need_grad = True
# Evaluate
if (i+1) % iter_epoch == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop for cnn
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
msg = "Model:cnn,Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve / iter_val) * 100)
print(msg)
# Evaluation loop for resnet
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_res_eval.forward(clear_buffer=True)
ve += categorical_error(pred_res_eval.d, label)
iter_val += 1
msg = "Model:resnet,Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve / iter_val) * 100)
print(msg)
st = time.time()
epoch +=1
def main():
# Get arguments
args = get_args()
data_file = "https://raw.githubusercontent.com/tomsercu/lstm/master/data/ptb.train.txt"
model_file = args.work_dir + "model.h5"
# Load Dataset
itow, wtoi, dataset = load_ptbset(data_file)
# Computation environment settings
from nnabla.contrib.context import extension_context
extension_module = args.context
if args.context is None:
extension_module = 'cpu'
logger.info("Running in %s" % extension_module)
ctx = extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
# Create data provider
n_word = len(wtoi)
n_dim = args.embed_dim
batchsize = args.batchsize
half_window = args.half_window_length
n_negative = args.n_negative_sample
di = DataIteratorForEmbeddingLearning(
batchsize=batchsize,
half_window=half_window,
n_negative=n_negative,
dataset=dataset)
# Create model
# - Real batch size including context samples and negative samples
size = batchsize * (1 + n_negative) * (2 * (half_window - 1))
# Model for learning
# - input variables
xl = nn.Variable((size,)) # variable for word
yl = nn.Variable((size,)) # variable for context
# Embed layers for word embedding function
# - f_embed : word index x to get y, the n_dim vector
# -- for each sample in a minibatch
hx = PF.embed(xl, n_word, n_dim, name="e1") # feature vector for word
hy = PF.embed(yl, n_word, n_dim, name="e1") # feature vector for context
hl = F.sum(hx * hy, axis=1)
# -- Approximated likelihood of context prediction
# pos: word context, neg negative samples
tl = nn.Variable([size, ], need_grad=False)
loss = F.sigmoid_cross_entropy(hl, tl)
loss = F.mean(loss)
# Model for test of searching similar words
xr = nn.Variable((1,), need_grad=False)
hr = PF.embed(xr, n_word, n_dim, name="e1") # feature vector for test
# Create solver
solver = S.Adam(args.learning_rate)
solver.set_parameters(nn.get_parameters())
# Create monitor.
monitor = M.Monitor(args.work_dir)
monitor_loss = M.MonitorSeries(
"Training loss", monitor, interval=args.monitor_interval)
monitor_time = M.MonitorTimeElapsed(
"Training time", monitor, interval=args.monitor_interval)
# Do training
max_epoch = args.max_epoch
for epoch in range(max_epoch):
# iteration per epoch
for i in range(di.n_batch):
# get minibatch
xi, yi, ti = di.next()
# learn
solver.zero_grad()
xl.d, yl.d, tl.d = xi, yi, ti
loss.forward(clear_no_need_grad=True)
loss.backward(clear_buffer=True)
solver.update()
# monitor
itr = epoch * di.n_batch + i
monitor_loss.add(itr, loss.d)
monitor_time.add(itr)
# Save model
nn.save_parameters(model_file)
# Evaluate by similarity
max_check_words = args.max_check_words
for i in range(max_check_words):
# prediction
xr.d = i
hr.forward(clear_buffer=True)
h = hr.d
# similarity calculation
w = nn.get_parameters()['e1/embed/W'].d
s = np.sqrt((w * w).sum(1))
w /= s.reshape((s.shape[0], 1))
similarity = w.dot(h[0]) / s[i]
# for understanding
output_similar_words(itow, i, similarity)
def main(args):
# Settings
device_id = args.device_id
batch_size = 100
batch_size_eval = 100
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = n_epoch * iter_epoch
extension_module = args.context
# Model
## supervised
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l = nn.Variable((batch_size, m, h, w))
y_l = nn.Variable((batch_size, 1))
pred = cnn_model_003(ctx, x_l)
loss_ce = ce_loss(ctx, pred, y_l)
loss_er = er_loss(ctx, pred)
loss_supervised = loss_ce + loss_er
## stochastic regularization
x_u0 = nn.Variable((batch_size, m, h, w), need_grad=False)
x_u1 = nn.Variable((batch_size, m, h, w), need_grad=False)
pred_x_u0 = cnn_model_003(ctx, x_u0)
pred_x_u1 = cnn_model_003(ctx, x_u1)
loss_sr = sr_loss(ctx, pred_x_u0, pred_x_u1)
loss_er0 = er_loss(ctx, pred_x_u0)
loss_er1 = er_loss(ctx, pred_x_u1)
loss_unsupervised = loss_sr + loss_er0 + loss_er1
## autoencoder
path = args.model_path
nn.load_parameters(path)
x_u0_rc = cnn_ae_model_000(ctx, x_u0, act=F.relu, test=True)
x_u1_rc = cnn_ae_model_000(ctx, x_u1, act=F.relu, test=True)
x_u0_rc.need_grad = False
x_u1_rc.need_grad = False
pred_x_u0_rc = cnn_model_003(ctx, x_u0_rc, test=False)
pred_x_u1_rc = cnn_model_003(ctx, x_u1_rc, test=False)
loss_sr_rc = sr_loss(ctx, pred_x_u0_rc, pred_x_u1_rc)
loss_er0_rc = er_loss(ctx, pred_x_u0_rc)
loss_er1_rc = er_loss(ctx, pred_x_u1_rc)
loss_unsupervised_rc = loss_sr_rc + loss_er0_rc + loss_er1_rc
loss_unsupervised += loss_unsupervised_rc
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval = cnn_model_003(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
loss_supervised.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_supervised.backward(clear_buffer=True)
solver.update()
loss_unsupervised.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_unsupervised.backward(clear_buffer=True)
solver.update()
# Evaluate
if (i+1) % iter_epoch == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = x_data[k:k+batch_size_eval, :]
label = y_data[k:k+batch_size_eval, :]
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
msg = "Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve / iter_val) * 100)
print(msg)
st = time.time()
epoch +=1
def train():
"""
Naive Multi-Device Training
NOTE: the communicator exposes low-level interfaces
* Parse command line arguments.
* Specify contexts for computation.
* Initialize DataIterator.
* Construct computation graphs for training and one for validation.
* Initialize solvers and set parameter variables to those.
* Instantiate a communicator and set parameter variables.
* Create monitor instances for saving and displaying training stats.
* Training loop
* Computate error rate for validation data (periodically)
* Get a next minibatch.
* Execute forwardprops
* Set parameter gradients zero
* Execute backprop.
* Inplace allreduce (THIS IS THE MAIN difference from a single device training)
* Solver updates parameters by using gradients computed by backprop.
* Compute training error
"""
# Parse args
args = get_args()
n_train_samples = 50000
bs_valid = args.batch_size
# Create contexts
extension_module = args.context
if extension_module != "cuda" and \
extension_module != "cuda.cudnn":
raise Exception("Use `cuda` or `cuda.cudnn` extension_module.")
n_devices = args.n_devices
ctxs = []
for i in range(n_devices):
ctx = extension_context(extension_module, device_id=i)
ctxs.append(ctx)
ctx = ctxs[-1]
# Create training graphs
input_image_train = []
preds_train = []
losses_train = []
test = False
for i in range(n_devices):
image = nn.Variable((args.batch_size, 3, 32, 32))
label = nn.Variable((args.batch_size, 1))
device_scope_name = "device{}".format(i)
pred = cifar100_resnet23_prediction(
image, ctxs[i], device_scope_name, test)
loss = cifar100_resnet32_loss(pred, label)
input_image_train.append({"image": image, "label": label})
preds_train.append(pred)
losses_train.append(loss)
# Create validation graph
test = True
device_scope_name = "device{}".format(0)
image_valid = nn.Variable((bs_valid, 3, 32, 32))
pred_valid = cifar100_resnet23_prediction(
image_valid, ctxs[i], device_scope_name, test)
input_image_valid = {"image": image_valid}
# Solvers
solvers = []
for i in range(n_devices):
with nn.context_scope(ctxs[i]):
solver = S.Adam()
device_scope_name = "device{}".format(i)
with nn.parameter_scope(device_scope_name):
params = nn.get_parameters()
solver.set_parameters(params)
solvers.append(solver)
# Communicator
comm = C.DataParalellCommunicator(ctx)
for i in range(n_devices):
device_scope_name = "device{}".format(i)
with nn.parameter_scope(device_scope_name):
ctx = ctxs[i]
params = nn.get_parameters()
comm.add_context_and_parameters((ctx, params))
comm.init()
# Create threadpools with one thread
pools = []
for _ in range(n_devices):
pool = ThreadPool(processes=1)
pools.append(pool)
# Once forward/backward to safely secure memory
for device_id in range(n_devices):
data, label = \
(np.random.randn(*input_image_train[device_id]["image"].shape),
(np.random.rand(*input_image_train[device_id]["label"].shape) * 10).astype(np.int32))
ret = pools[device_id].apply_async(forward_backward,
(input_image_train[device_id]["image"], data,
input_image_train[device_id]["label"], label,
losses_train[device_id], solvers[device_id]))
ret.get()
losses_train[device_id].d # sync to host
# Create monitor.
from nnabla.monitor import Monitor, MonitorSeries, MonitorTimeElapsed
monitor = Monitor(args.monitor_path)
monitor_loss = MonitorSeries("Training loss", monitor, interval=10)
monitor_err = MonitorSeries("Training error", monitor, interval=10)
monitor_time = MonitorTimeElapsed("Training time", monitor, interval=100)
monitor_verr = MonitorSeries("Test error", monitor, interval=10)
with data_iterator_cifar100(args.batch_size, True) as tdata, \
data_iterator_cifar100(bs_valid, False) as vdata:
# Training-loop
for i in range(int(args.max_iter / n_devices)):
# Validation
if i % int(n_train_samples / args.batch_size / n_devices) == 0:
ve = 0.
for j in range(args.val_iter):
image, label = vdata.next()
input_image_valid["image"].d = image
pred_valid.forward()
ve += categorical_error(pred_valid.d, label)
ve /= args.val_iter
monitor_verr.add(i * n_devices, ve)
if i % int(args.model_save_interval / n_devices) == 0:
nn.save_parameters(os.path.join(
args.model_save_path, 'params_%06d.h5' % i))
# Forwards/Zerograd/Backwards
fb_results = []
for device_id in range(n_devices):
image, label = tdata.next()
res = pools[device_id].apply_async(forward_backward,
(input_image_train[device_id]["image"], image,
input_image_train[device_id]["label"], label,
losses_train[device_id], solvers[device_id]))
fb_results.append(res)
for device_id in range(n_devices):
fb_results[device_id].get()
# In-place Allreduce
comm.allreduce()
# Solvers update
for device_id in range(n_devices):
solvers[device_id].update()
e = categorical_error(
preds_train[-1].d, input_image_train[-1]["label"].d)
monitor_loss.add(i * n_devices, losses_train[-1].d.copy())
monitor_err.add(i * n_devices, e)
monitor_time.add(i * n_devices)
nn.save_parameters(os.path.join(
args.model_save_path,
'params_%06d.h5' % (args.max_iter / n_devices)))
def main(args):
# Settings
device_id = args.device_id
batch_size = args.batch_size
batch_size_eval = args.batch_size_eval
n_l_train_data = args.n_label
n_train_data = 73257
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = args.epoch
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = int(n_epoch * iter_epoch)
extension_module = args.context
# Model
## supervised
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l = nn.Variable((batch_size, m, h, w))
y_l = nn.Variable((batch_size, 1))
pred = cnn_model_003(ctx, x_l)
loss_ce = ce_loss(ctx, pred, y_l)
loss_er = er_loss(ctx, pred)
loss_supervised = loss_ce + loss_er
## stochastic regularization
x_u0 = nn.Variable((batch_size, m, h, w))
x_u1 = nn.Variable((batch_size, m, h, w))
pred_x_u0 = cnn_model_003(ctx, x_u0)
pred_x_u1 = cnn_model_003(ctx, x_u1)
loss_sr = sr_loss(ctx, pred_x_u0, pred_x_u1)
loss_er0 = er_loss(ctx, pred_x_u0)
loss_er1 = er_loss(ctx, pred_x_u1)
loss_unsupervised = loss_sr + loss_er0 + loss_er1
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval = cnn_model_003(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/svhn/train.mat")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/svhn/l_train.mat")
u_train_path = os.path.join(home, "datasets/svhn/u_train.mat")
test_path = os.path.join(home, "datasets/svhn/test.mat")
# data reader
data_reader = SVHNDataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=False,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
ve_best = 1.
save_path_prev = ""
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
loss_supervised.forward(clear_no_need_grad=True)
loss_unsupervised.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_supervised.backward(clear_buffer=True)
loss_unsupervised.backward(clear_buffer=True)
solver.update()
# Evaluate
if int((i+1) % iter_epoch) == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
ve /= iter_val
msg = "Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve) * 100)
print(msg)
if ve < ve_best:
if not os.path.exists(args.model_save_path):
os.makedirs(args.model_save_path)
if save_path_prev != "":
os.remove(save_path_prev)
save_path = os.path.join(
args.model_save_path, 'params_%06d.h5' % epoch)
nn.save_parameters(save_path)
save_path_prev = save_path
ve_best = ve
st = time.time()
epoch +=1
def main(args):
# Settings
device_id = args.device_id
batch_sizes = [16, 32, 64]
batch_size_eval = 64
c, h, w = 3, 32, 32
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = n_train_data / int(np.mean(batch_sizes)) # approximate epoch
n_iter = n_epoch * iter_epoch
extension_module = args.context
# Model (Batch-Stochastic)
ctx = extension_context(extension_module, device_id=device_id)
## supervised
x_list, y_list, preds, losses_ce = batch_stochastic_supervised_network(
ctx, batch_sizes, c, h, w)
## stochastic regularization
x0_list, x1_list, _, losses_sr = batch_stochastic_unsupervised_network(
ctx, batch_sizes, c, h, w)
## evaluate
batch_size_eval, m, h, w = batch_size_eval, c, h, w
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval = cnn_model_003(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_sizes[0],
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
iter_ = 0
for i in range(n_iter):
idx = np.random.choice(np.arange(0, len(batch_sizes)))
idx_u = np.random.choice(np.arange(0, len(batch_sizes)))
# Get data
bs = batch_sizes[idx]
bs_u = batch_sizes[idx_u]
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch(bs)
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch(bs_u)
# Set it to the varaibles
x_l = x_list[idx]
y_l = y_list[idx]
x_u0 = x0_list[idx_u]
x_u1 = x1_list[idx_u]
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
loss_ce = losses_ce[idx]
loss_sr = losses_sr[idx_u]
loss_ce.forward(clear_no_need_grad=True)
loss_sr.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_ce.backward(clear_buffer=True)
loss_sr.backward(clear_buffer=True)
solver.update()
# Evaluate
if (i+1) % iter_epoch == 0: # approximate epoch
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
msg = "Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve / iter_val) * 100)
print(msg)
st = time.time()
epoch +=1
def main():
"""
Main script.
Steps:
* Get and set context.
* Load Dataset
* Initialize DataIterator.
* Create Networks
* Net for Labeled Data
* Net for Unlabeled Data
* Net for Test Data
* Create Solver.
* Training Loop.
* Test
* Training
* by Labeled Data
* Calculate Cross Entropy Loss
* by Unlabeled Data
* Estimate Adversarial Direction
* Calculate LDS Loss
"""
args = get_args()
# Get context.
from nnabla.contrib.context import extension_context
extension_module = args.context
if args.context is None:
extension_module = 'cpu'
logger.info("Running in %s" % extension_module)
ctx = extension_context(extension_module, device_id=args.device_id)
nn.set_default_context(ctx)
shape_x = (1, 28, 28)
n_h = args.n_units
n_y = args.n_class
# Load MNist Dataset
from mnist_data import MnistDataSource
with MnistDataSource(train=True) as d:
x_t = d.images
t_t = d.labels
with MnistDataSource(train=False) as d:
x_v = d.images
t_v = d.labels
x_t = np.array(x_t / 256.0).astype(np.float32)
x_t, t_t = x_t[:args.n_train], t_t[:args.n_train]
x_v, t_v = x_v[:args.n_valid], t_v[:args.n_valid]
# Create Semi-supervised Datasets
x_l, t_l, x_u, _ = split_dataset(x_t, t_t, args.n_labeled, args.n_class)
x_u = np.r_[x_l, x_u]
x_v = np.array(x_v / 256.0).astype(np.float32)
# Create DataIterators for datasets of labeled, unlabeled and validation
di_l = DataIterator(args.batchsize_l, [x_l, t_l])
di_u = DataIterator(args.batchsize_u, [x_u])
di_v = DataIterator(args.batchsize_v, [x_v, t_v])
# Create networks
# feed-forward-net building function
def forward(x, test=False):
return mlp_net(x, n_h, n_y, test)
# Net for learning labeled data
xl = nn.Variable((args.batchsize_l,) + shape_x, need_grad=False)
hl = forward(xl, test=False)
tl = nn.Variable((args.batchsize_l, 1), need_grad=False)
loss_l = F.mean(F.softmax_cross_entropy(hl, tl))
# Net for learning unlabeled data
xu = nn.Variable((args.batchsize_u,) + shape_x, need_grad=False)
r = nn.Variable((args.batchsize_u,) + shape_x, need_grad=True)
eps = nn.Variable((args.batchsize_u,) + shape_x, need_grad=False)
loss_u, yu = vat(xu, r, eps, forward, distance)
# Net for evaluating valiation data
xv = nn.Variable((args.batchsize_v,) + shape_x, need_grad=False)
hv = forward(xv, test=True)
tv = nn.Variable((args.batchsize_v, 1), need_grad=False)
# Create solver
solver = S.Adam(args.learning_rate)
solver.set_parameters(nn.get_parameters())
# Monitor trainig and validation stats.
import nnabla.monitor as M
monitor = M.Monitor(args.model_save_path)
monitor_verr = M.MonitorSeries("Test error", monitor, interval=240)
monitor_time = M.MonitorTimeElapsed("Elapsed time", monitor, interval=240)
# Training Loop.
t0 = time.time()
for i in range(args.max_iter):
# Validation Test
if i % args.val_interval == 0:
n_error = calc_validation_error(
di_v, xv, tv, hv, args.val_iter)
monitor_verr.add(i, n_error)
#################################
## Training by Labeled Data #####
#################################
# input minibatch of labeled data into variables
xl.d, tl.d = di_l.next()
# initialize gradients
solver.zero_grad()
# forward, backward and update
loss_l.forward(clear_no_need_grad=True)
loss_l.backward(clear_buffer=True)
solver.weight_decay(args.weight_decay)
solver.update()
#################################
## Training by Unlabeled Data ###
#################################
# input minibatch of unlabeled data into variables
xu.d, = di_u.next()
##### Calculate Adversarial Noise #####
# Sample random noise
n = np.random.normal(size=xu.shape).astype(np.float32)
# Normalize noise vector and input to variable
r.d = get_direction(n)
# Set xi, the power-method scaling parameter.
eps.data.fill(args.xi_for_vat)
# Calculate y without noise, only once.
yu.forward(clear_buffer=True)
# Do power method iteration
for k in range(args.n_iter_for_power_method):
# Initialize gradient to receive value
r.grad.zero()
# forward, backward, without update
loss_u.forward(clear_no_need_grad=True)
loss_u.backward(clear_buffer=True)
# Normalize gradinet vector and input to variable
r.d = get_direction(r.g)
##### Calculate loss for unlabeled data #####
# Clear remained gradients
solver.zero_grad()
# Set epsilon, the adversarial noise scaling parameter.
eps.data.fill(args.eps_for_vat)
# forward, backward and update
loss_u.forward(clear_no_need_grad=True)
loss_u.backward(clear_buffer=True)
solver.weight_decay(args.weight_decay)
solver.update()
##### Learning rate update #####
if i % args.iter_per_epoch == 0:
solver.set_learning_rate(
solver.learning_rate() * args.learning_rate_decay)
monitor_time.add(i)
# Evaluate the final model by the error rate with validation dataset
valid_error = calc_validation_error(di_v, xv, tv, hv, args.val_iter)
monitor_verr.add(i, valid_error)
monitor_time.add(i)
# Save the model.
nnp_file = os.path.join(
args.model_save_path, 'vat_%06d.nnp' % args.max_iter)
runtime_contents = {
'networks': [
{'name': 'Validation',
'batch_size': args.batchsize_v,
'outputs': {'y': hv},
'names': {'x': xv}}],
'executors': [
{'name': 'Runtime',
'network': 'Validation',
'data': ['x'],
'output': ['y']}]}
save.save(nnp_file, runtime_contents)
from cpp_forward_check import check_cpp_forward
check_cpp_forward(args.model_save_path, [xv.d], [xv], hv, nnp_file)
def main(args):
# Settings
device_id = args.device_id
batch_size = args.batch_size
batch_size_eval = args.batch_size_eval
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = int(n_train_data / batch_size)
n_iter = n_epoch * iter_epoch
extension_module = args.context
# Model
views = [global_view, spatial_view, feature_view]
## supervised
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l = nn.Variable((batch_size, m, h, w))
y_l = nn.Variable((batch_size, 1))
feature = cnn_model_003(ctx, x_l)
loss_supervised = []
for view in views:
pred = view(ctx, feature)
loss_ce = ce_loss(ctx, pred, y_l)
loss_er = er_loss(ctx, pred)
loss_supervised += [loss_ce, loss_er]
loss_supervised = reduce(lambda x, y: x+y, loss_supervised)
## cross view loss
x_u0 = nn.Variable((batch_size, m, h, w))
x_u1 = nn.Variable((batch_size, m, h, w))
feature_x_u0 = cnn_model_003(ctx, x_u0)
feature_x_u1 = cnn_model_003(ctx, x_u1)
pred_x_u0 = []
pred_x_u1 = []
loss_er = []
loss_unsupervised = []
for view in views:
pred = view(ctx, feature_x_u0)
pred_x_u0 += [pred]
loss_er +=[er_loss(ctx, pred)]
pred = view(ctx, feature_x_u1)
pred_x_u1 += [pred]
loss_er += [er_loss(ctx, pred)]
for pred_a, pred_b in itertools.product(pred_x_u0, pred_x_u1): # multi-view
if pred_a == pred_b:
continue
loss_unsupervised += [sr_loss(ctx, pred_a, pred_b)]
loss_unsupervised = reduce(lambda x, y: x+y, loss_unsupervised) \
+ reduce(lambda x, y: x+y, loss_er)
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
feature_eval = cnn_model_003(ctx, x_eval, test=True)
pred_eval = []
for view in views:
pred_eval += [view(ctx, feature_eval)]
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
ve_best = 1.
save_path_prev = ""
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
loss_supervised.forward(clear_no_need_grad=True)
loss_unsupervised.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_supervised.backward(clear_buffer=True)
loss_unsupervised.backward(clear_buffer=True)
solver.update()
# Evaluate
if int((i + 1) % iter_epoch) == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = [0., 0., 0.]
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
feature_eval.forward(clear_buffer=True)
for i in range(len(pred_eval)):
pred_eval[i].forward()
ve[i] += categorical_error(pred_eval[i].d, label)
iter_val += 1
for i, e in enumerate(ve):
e /= iter_val
msg = "Epoch-{}:{},ElapsedTime:{},Acc:{:02f}".format(
i,
epoch,
time.time() - st,
(1. - e) * 100)
print(msg)
st = time.time()
epoch +=1
def main(args):
# Settings
device_id = args.device_id
batch_size = args.batch_size
batch_size_eval = args.batch_size_eval
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 300
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = n_epoch * iter_epoch
extension_module = args.context
# Model
## supervised
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_l = nn.Variable((batch_size, m, h, w))
y_l = nn.Variable((batch_size, 1))
pred, log_var = cnn_model_003(ctx, x_l)
one = F.constant(1., log_var.shape)
loss_ce = ce_loss_with_uncertainty(ctx, pred, y_l, log_var)
reg_sigma = sigma_regularization(ctx, log_var, one)
loss_supervised = loss_ce + reg_sigma
## stochastic regularization
x_u0 = nn.Variable((batch_size, m, h, w))
x_u1 = nn.Variable((batch_size, m, h, w))
pred_x_u0, log_var0 = cnn_model_003(ctx, x_u0)
pred_x_u1, log_var1 = cnn_model_003(ctx, x_u1)
loss_sr = sr_loss_with_uncertainty(ctx,
pred_x_u0, pred_x_u1, log_var0, log_var1)
loss_er0 = er_loss(ctx, pred_x_u0)
loss_er1 = er_loss(ctx, pred_x_u1)
reg_sigma0 = sigma_regularization(ctx, log_var0, one)
reg_sigma1 = sigma_regularization(ctx, log_var1, one)
loss_unsupervised = loss_sr + loss_er0 + loss_er1 \
+ reg_sigma0 + reg_sigma1
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval, _ = cnn_model_003(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver_l= S.Adam(alpha=learning_rate)
solver_l.set_parameters(nn.get_parameters())
solver_u= S.Adam(alpha=learning_rate)
solver_u.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
for i in range(n_iter):
# Get data and set it to the varaibles
x_l0_data, x_l1_data, y_l_data = data_reader.get_l_train_batch()
x_u0_data, x_u1_data, y_u_data = data_reader.get_u_train_batch()
x_l.d, _ , y_l.d= x_l0_data, x_l1_data, y_l_data
x_u0.d, x_u1.d= x_u0_data, x_u1_data
# Train
## for supervised loss
loss_supervised.forward(clear_no_need_grad=True)
solver_l.zero_grad()
loss_supervised.backward(clear_buffer=True)
solver_l.update()
## for unsupervised loss
loss_unsupervised.forward(clear_no_need_grad=True)
solver_u.zero_grad()
loss_unsupervised.backward(clear_buffer=True)
solver_u.update()
# Evaluate
if (i+1) % iter_epoch == 0:
# Get data and set it to the varaibles
x_data, y_data = data_reader.get_test_batch()
# Evaluation loop
ve = 0.
iter_val = 0
for k in range(0, len(x_data), batch_size_eval):
x_eval.d = get_test_data(x_data, k, batch_size_eval)
label = get_test_data(y_data, k, batch_size_eval)
pred_eval.forward(clear_buffer=True)
ve += categorical_error(pred_eval.d, label)
iter_val += 1
msg = "Epoch:{},ElapsedTime:{},Acc:{:02f}".format(
epoch,
time.time() - st,
(1. - ve / iter_val) * 100)
print(msg)
st = time.time()
epoch +=1
def main(args):
# Settings
device_id = args.device_id
batch_size = 100
batch_size_eval = 100
n_l_train_data = 4000
n_train_data = 50000
n_cls = 10
learning_rate = 1. * 1e-3
n_epoch = 50
act = F.relu
iter_epoch = n_train_data / batch_size
n_iter = n_epoch * iter_epoch
extension_module = args.context
n_images = args.n_images
fname, _ = os.path.splitext(__file__)
dpath = "./{}_images_{}".format(fname, int(time.time()))
# Model
batch_size, m, h, w = batch_size, 3, 32, 32
ctx = extension_context(extension_module, device_id=device_id)
x_u = nn.Variable((batch_size, m, h, w))
pred = cnn_ae_model_001(ctx, x_u)
loss_recon = recon_loss(ctx, pred, x_u)
## evaluate
batch_size_eval, m, h, w = batch_size, 3, 32, 32
x_eval = nn.Variable((batch_size_eval, m, h, w))
pred_eval = cnn_ae_model_001(ctx, x_eval, test=True)
# Solver
with nn.context_scope(ctx):
solver = S.Adam(alpha=learning_rate)
solver.set_parameters(nn.get_parameters())
# Dataset
## separate dataset
home = os.environ.get("HOME")
fpath = os.path.join(home, "datasets/cifar10/cifar-10.npz")
separator = Separator(n_l_train_data)
separator.separate_then_save(fpath)
l_train_path = os.path.join(home, "datasets/cifar10/l_cifar-10.npz")
u_train_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
test_path = os.path.join(home, "datasets/cifar10/cifar-10.npz")
# data reader
data_reader = Cifar10DataReader(l_train_path, u_train_path, test_path,
batch_size=batch_size,
n_cls=n_cls,
da=True,
shape=True)
# Training loop
print("# Training loop")
epoch = 1
st = time.time()
acc_prev = 0.
for i in range(n_iter):
# Get data and set it to the varaibles
x_u_data, _, _ = data_reader.get_u_train_batch()
x_u.d = x_u_data
# Train
loss_recon.forward(clear_no_need_grad=True)
solver.zero_grad()
loss_recon.backward(clear_buffer=True)
solver.update()
# Evaluate
if (i+1) % iter_epoch == 0:
# Get data and forward
x_data, y_data = data_reader.get_test_batch()
pred_eval.forward(clear_buffer=True)
images = pred_eval.d
# Save n images
if not os.path.exists(dpath):
os.makedirs(dpath)
save_images(dpath, epoch, images[:n_images])
fpath = os.path.join(dpath, "epoch_{:05d}.h5".format(epoch))
nn.save_parameters(fpath)
st = time.time()
epoch +=1