def affine_backward(inputs, base_axis=1):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
w0 = inputs[2]
base_axis += inputs[0].ndim * (base_axis < 0)
ctx = nn.get_current_context()
dfx = AffineDataGrad(ctx, base_axis)
dfw = AffineFilterGrad(ctx, base_axis)
dfx.xshape = x0.shape
dfw.wshape = w0.shape
dx0 = dfx(dy, w0)
dw0 = dfw(dy, x0)
if len(inputs) == 4:
axes = [i for i in range(0, base_axis)]
db0 = F.sum(dy, axes, keepdims=False)
return dx0, dw0, db0
else:
return dx0, dw0
def fused_batch_normalization_backward(inputs,
axes=(1, ),
decay_rate=0.9,
eps=1e-05,
batch_stat=True,
nonlinearity='relu'):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
if nonlinearity not in ["", "relu"]:
raise ValueError("nonlinearity must be either '' or 'relu'.")
ctx = nn.get_current_context()
df = FusedBatchNormalizationBackward(ctx, axes, decay_rate, eps,
batch_stat, nonlinearity)
dy = inputs[0]
x0 = inputs[1]
b0 = inputs[2]
g0 = inputs[3]
rm = inputs[4]
rv = inputs[5]
z0 = inputs[6] if len(inputs) == 7 else None
df.is_add = True if z0 else False
y0 = get_output(x0, "FusedBatchNormalization")
if df.is_add:
dx0, db0, dg0, dz0 = df(dy, x0, b0, g0, rm, rv, y0, z0)
return dx0, db0, dg0, None, None, dz0
else:
dx0, db0, dg0 = df(dy, x0, b0, g0, rm, rv, y0)
return dx0, db0, dg0, None, None
def backward_impl(self, inputs, outputs, prop_down, accum):
# inputs: [inputs_fwd_graph] + [inputs_bwd_graph] or
# [inputs_fwd_graph] + [outputs_fwd_graph] + [inputs_bwd_graph]
# Args
kernel = self.forward_func.info.args["kernel"]
stride = self.forward_func.info.args["stride"]
ignore_border = self.forward_func.info.args["ignore_border"]
pad = self.forward_func.info.args["pad"]
channel_last = self.forward_func.info.args["channel_last"]
# Inputs
x0 = inputs[0].data
dy = inputs[1].data
# Outputs
dx0 = outputs[0].data
# Grads of inputs
g_x0 = inputs[0].grad
g_dy = inputs[1].grad
# Grads of outputs
g_dx0 = outputs[0].grad
# Compute
ctx = nn.get_current_context()
backward_func = nn.function.MaxPoolingBackward(
ctx, kernel, stride, ignore_border, pad, channel_last)
if prop_down[1]:
x0_ = nn.Variable(x0.shape).apply(
data=x0, grad=g_x0, need_grad=True)
dy_ = nn.Variable(dy.shape).apply(
data=dy, grad=g_dy, need_grad=True)
dx0_ = nn.Variable(dx0.shape).apply(data=dx0, grad=g_dx0)
backward_func.setup([x0_, dy_], [dx0_])
backward_func.backward([x0_, dy_], [dx0_], accum=accum)
def convolution_data_grad_backward(inputs,
base_axis=1,
pad=None,
stride=None,
dilation=None,
group=1,
channel_last=False):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
gdx = inputs[0]
dy = inputs[1]
w0 = inputs[2]
ctx = nn.get_current_context()
dfw = ConvolutionFilterGrad(ctx, base_axis, pad, stride, dilation, group,
channel_last)
dfw.wshape = w0.shape
gdy = F.convolution(gdx, w0, None, base_axis, pad, stride, dilation, group,
channel_last)
gw0 = dfw(dy, gdx)
return gdy, gw0
def _call_function(self, type_name, inputs, args):
import nnabla.function_bases as FB
function_expr = "FB.F.{type_name}(nn.{ctx}, **{args})".format(
type_name=type_name, ctx=nn.get_current_context(), args=args)
function = eval(function_expr)
o = function(*inputs)
return o
def deconvolution_filter_grad_backward(inputs,
base_axis=1,
pad=None,
stride=None,
dilation=None,
group=1,
channel_last=False,
output_padding=None):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
gdw = inputs[0]
dy = inputs[1]
x0 = inputs[2]
ctx = nn.get_current_context()
dfx = DeconvolutionDataGrad(ctx, base_axis, pad, stride, dilation, group,
channel_last, output_padding)
dfx.xshape = x0.shape
gdy = F.deconvolution(x0, gdw, None, base_axis, pad, stride, dilation,
group, channel_last, output_padding)
gx0 = dfx(dy, gdw)
return gdy, gx0
def _create_function(inputs, f, batch_size):
ctx = nn.get_current_context()
function_proto = f
# todo: arrange weight name for NNC
if function_proto.type == "Reshape": # if batch_size = -1, something wrong?
reshape_shape = resolve_reshape_params(
inputs, function_proto, batch_size)
function_instance = F.Reshape(
ctx, shape=reshape_shape, inplace=function_proto.reshape_param.inplace)
elif function_proto.type == 'Broadcast':
shape = resolve_broadcast_params(inputs, function_proto, batch_size)
function_instance = F.Broadcast(ctx, shape=shape)
elif function_proto.type == "RepeatStart":
raise NotImplementedError("Repeat not supported.")
function_instance = F.Identity(ctx)
elif function_proto.type == "RepeatEnd":
raise NotImplementedError("Repeat not supported.")
function_instance = F.Identity(ctx)
elif function_proto.type == "RecurrentOutput":
raise NotImplementedError("Recurrent not supported.")
function_instance = F.Stack(
ctx, axis=function_proto.recurrent_param.axis)
elif function_proto.type == "RecurrentInput":
raise NotImplementedError("Recurrent not supported.")
function_instance = F.Split(
ctx, axis=function_proto.recurrent_param.axis)
elif function_proto.type == "Delay":
raise NotImplementedError("Recurrent not supported.")
function_instance = F.Identity(ctx)
else:
function_instance = _create_function_instance(ctx, function_proto)
return function_instance
def scope_function():
# turn off auto forward mode
nn.set_auto_forward(False)
# clear all parameters
nn.clear_parameters()
# keep context
ctx = nn.get_current_context()
yield
# restore context
nn.set_default_context(ctx)
def batch_normalization_backward(inputs, axes=(1,), decay_rate=0.9, eps=1e-05,
batch_stat=True, no_scale=False, no_bias=False):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
ctx = nn.get_current_context()
df = BatchNormalizationBackward(
ctx, axes, decay_rate, eps, batch_stat, no_scale, no_bias)
d_inputs = df(*inputs)
return force_tuple(d_inputs) + (None, None)
def concatenate_backward(inputs, axis=None):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
axis = axis if axis is not None else len(dy.shape) - 1
ctx = nn.get_current_context()
df = ConcatenateDataGrad(ctx, axis=axis)
df.xshapes = [x.shape for x in inputs[1:]]
dx0 = df(dy)
return dx0
def unpooling_backward(inputs, kernel, channel_last=False):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
ctx = nn.get_current_context()
df = UnpoolingDataGrad(ctx, kernel, channel_last)
df.xshape = x0.shape
dx0 = df(dy)
return dx0
def global_average_pooling_backward(inputs):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
ctx = nn.get_current_context()
pool = GlobalAveragePoolingDataGrad(ctx)
pool.xshape = x0.shape
dx0 = pool(dy)
return dx0
def slice_backward(inputs, start=None, stop=None, step=None):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
ctx = nn.get_current_context()
df = SliceDataGrad(ctx, start, stop, step)
df.xshape = x0.shape
dx0 = df(dy)
return dx0
def deconvolution_backward(inputs,
base_axis=1,
pad=None,
stride=None,
dilation=None,
group=1,
channel_last=False,
output_padding=None):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
w0 = inputs[2]
base_axis += x0.ndim * (base_axis < 0)
# base_axis += inputs[0].ndim*(base_axis < 0)
ctx = nn.get_current_context()
dfx = DeconvolutionDataGrad(ctx, base_axis, pad, stride, dilation, group,
channel_last, output_padding)
dfw = DeconvolutionFilterGrad(ctx, base_axis, pad, stride, dilation, group,
channel_last, output_padding)
dfx.xshape = x0.shape
dfw.wshape = w0.shape
dx0 = dfx(dy, w0)
dw0 = dfw(dy, x0)
axes = [i for i in range(dy.ndim, base_axis)]
db0 = F.sum(dy, axes, keepdims=False) if len(inputs) == 4 else None
if len(inputs) == 4:
if channel_last:
axes = [i for i in range(dy.ndim - 1)]
else:
axes = [i for i in range(0, base_axis)] + \
[i for i in range(base_axis + 1, dy.ndim)]
db0 = F.sum(dy, axes, keepdims=False) if len(inputs) == 4 else None
return dx0, dw0, db0
else:
return dx0, dw0
def _create_function(self, function_proto):
inputs = self._create_inputs(function_proto.input)
function_instance = _create_function(
nn.get_current_context(), inputs, function_proto, self.batch_size)
outputs = function_instance(*inputs)
if not isinstance(outputs, tuple):
outputs = (outputs,)
for i, name in enumerate(function_proto.output):
try:
var, _ = self.vseen[name]
except:
self.vseen[name] = (outputs[i], [0])
continue
var.rewire_on(outputs[i])
def measure_cpu_gpu_instant_load():
# Get current cpu gpu load, as
# load = [rank, cpu_load, nvidia_device_id, gpu_load]
# result_arr: [load, load, ...]
gpu_load = []
if gpu_load_backend_ok:
global gpu_a_load
global gpu_m_count
gpu_m_count += 1
try:
comm = current_communicator()
if comm:
index = comm.local_rank
elif 'cuda' in str(nn.get_current_context().backend):
index = 0
else:
raise Exception
handler = pynvml.nvmlDeviceGetHandleByIndex(index)
gpu_load = [[
index,
pynvml.nvmlDeviceGetUtilizationRates(handler).gpu
]]
if index in gpu_a_load.keys():
gpu_a_load[index]['name'] = pynvml.nvmlDeviceGetName(
handler).decode("utf-8")
o_load = gpu_a_load[index]['load']
n_load = gpu_load[0][1]
gpu_a_load[index]['load'] = (
(gpu_m_count - 1) * o_load + n_load) / gpu_m_count
else:
gpu_a_load[index] = {
'name': pynvml.nvmlDeviceGetName(handler).decode("utf-8"),
'load': gpu_load[0][1]
}
except Exception:
gpu_load = []
if cpu_load_backend_ok:
global p_handler
cpu_load = p_handler.cpu_percent()
callback.update_status(
('cpu_gpu_load', collect_and_shape_result(cpu_load, gpu_load)))
def interpolate_backward(inputs, output_size, mode, align_corners=True,
half_pixel=False, half_pixel_for_nn=False, channel_last=False):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
ctx = nn.get_current_context()
df = InterpolateDataGrad(ctx, output_size, mode, align_corners,
half_pixel, half_pixel_for_nn, channel_last)
df.xshape = x0.shape
dx0 = df(dy)
return dx0
def embed_backward(inputs):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
w0 = inputs[2]
ctx = nn.get_current_context()
dfw = EmbedFilterGrad(ctx)
dfw.wshape = w0.shape
dw0 = dfw(dy, x0)
return None, dw0
def affine_filter_grad_backward(inputs, base_axis=1):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
gdw = inputs[0]
dy = inputs[1]
x0 = inputs[2]
ctx = nn.get_current_context()
dfx = AffineDataGrad(ctx, base_axis)
dfx.xshape = x0.shape
gdy = F.affine(x0, gdw, None, base_axis)
gx0 = dfx(dy, gdw)
return gdy, gx0
def scope_function():
# turn off auto forward mode
nn.set_auto_forward(False)
# clear all parameters
nn.clear_parameters()
# keep context
ctx = nn.get_current_context()
# use cached array
nn.prefer_cached_array(True)
# turn off re-computation
nn.set_global_recompute(False)
yield
# restore context
nn.set_default_context(ctx)
def affine_data_grad_backward(inputs, base_axis=1):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
gdx = inputs[0]
dy = inputs[1]
w0 = inputs[2]
ctx = nn.get_current_context()
dfw = AffineFilterGrad(ctx, base_axis)
dfw.wshape = w0.shape
gdy = F.affine(gdx, w0, None, base_axis)
gw0 = dfw(dy, gdx)
return gdy, gw0
def pad_backward(inputs, pad_width, mode='constant', constant_value=0):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
if mode != "constant":
raise NotImplementedError(
"{}_backward (mode!=constant) is not implemented.".format(func['snake_name']))
dy = inputs[0]
x0 = inputs[1]
ctx = nn.get_current_context()
# constant value is always zero after 1st-order derivative
df = PadDataGrad(ctx, pad_width, mode, constant_value=0)
df.xshape = x0.shape
dx0 = df(dy)
return dx0
def _connect_on_gradient_graph(self, grad_vars, f):
# 1. accumulate variables used more than one or do nothing
vf_vb_map = grad_vars.pop(f) # {VO_fwd: [VI_bwd]}
grad_inputs = []
for o in f.outputs:
# address `floating` variables; no function takes it as input.
# e.g., when dx, db, dg = dBN(...), (db, dg) are not used afterwards.
#v = vf_vb_map[o] if o in vf_vb_map else [None]
v = vf_vb_map[o] if o in vf_vb_map else [0]
if len(v) > 1:
grad_inputs += [sum(v)]
#grad_inputs += [F.add_n(v)]
else:
grad_inputs += v
# 2. lookup the backward function
f_fwd_name = f.info.type_name
if f_fwd_name not in registry:
raise ValueError(
"{} is not in the backward function registry".format(
f_fwd_name))
backward_func = registry[f_fwd_name]
# 3. connect
grad_inputs = grad_inputs + f.inputs
ctx = nn.get_current_context()
with nn.context_scope(ctx):
grad_outputs = backward_func(grad_inputs, **f.info.args)
grad_outputs = self._force_list(grad_outputs)
# 4. put grad_output as grad_input to a corresponding function
for inp, grad_out in zip(f.inputs, grad_outputs):
if grad_out is None:
continue
if inp.parent not in grad_vars:
grad_vars[inp.parent] = OrderedDict()
if inp not in grad_vars[inp.parent]:
grad_vars[inp.parent][inp] = [grad_out]
else:
grad_vars[inp.parent][inp] += [grad_out]
return grad_outputs
def max_pooling_backward_backward(inputs,
kernel,
stride=None,
ignore_border=True,
pad=None,
channel_last=False):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
gdx = inputs[0]
dy = inputs[1]
x0 = inputs[2]
ctx = nn.get_current_context()
df = MaxPoolingBackwardDataGrad(ctx, kernel, stride, ignore_border, pad,
channel_last)
df.yshape = dy.shape
gdy = df(gdx, x0)
return gdy, None
def average_pooling_backward(inputs,
kernel,
stride=None,
ignore_border=True,
pad=None,
channel_last=False,
including_pad=True):
"""
Args:
inputs (list of nn.Variable): Incomming grads/inputs to/of the forward function.
kwargs (dict of arguments): Dictionary of the corresponding function arguments.
Return:
list of Variable: Return the gradients wrt inputs of the corresponding function.
"""
dy = inputs[0]
x0 = inputs[1]
ctx = nn.get_current_context()
df = AveragePoolingDataGrad(ctx, kernel, stride, ignore_border, pad,
channel_last, including_pad)
df.xshape = x0.shape
dx0 = df(dy)
return dx0
# これで、 mnist-collection/dcgan.pyの内部にアクセスできるようになった。 今回の例ではハイ
# パーパラメータが設定されているのでそれに倣う。
source = inspect.getsource(I)
print(source[source.index("if __name__"):])
max_iter = 20000
learning_rate = 0.0002
batch_size = 64
weight_decay = 0.0001
# コンテキストを設定する。
context = get_extension_context("cudnn", device_id=0, type_config="float")
nn.set_default_context(context)
nn.get_current_context()
# Fakeパスの設定
z = nn.Variable([batch_size, 100, 1, 1])
fake = I.generator(z)
fake.persistent = True # Not to clear at backward
pred_fake = I.discriminator(fake)
loss_gen = F.mean(
F.sigmoid_cross_entropy(pred_fake, F.constant(1, pred_fake.shape)))
fake_dis = fake.get_unlinked_variable(need_grad=True)
fake_dis.need_grad = True # TODO: Workaround until v1.0.2
pred_fake_dis = I.discriminator(fake_dis)
loss_dis = F.mean(
F.sigmoid_cross_entropy(pred_fake_dis, F.constant(0, pred_fake_dis.shape)))
# Realパスの設定
def __init__(self, ):
ctx = nn.get_current_context()
if "half" in [x.split(":")[-1] for x in ctx.backend]:
raise ValueError(
"Half is not supported up to now, context = {}".format(ctx))
def abs_max_recorder(x, M, training=True):
ctx = nn.get_current_context()
func = AbsMaxRecorder(ctx, training)
return func(x, M)
def max_mva_recorder(x, M, decay=0.99, training=True):
ctx = nn.get_current_context()
func = MaxMvaRecorder(ctx, decay, training)
return func(x, M)
def minmax_minmax_recorder(x, m, M, training=True):
ctx = nn.get_current_context()
func = MinMaxMinMaxRecorder(ctx, training)
return func(x, m, M)
