def _one_directional_loop(di):
# di=0, forward GRU
# di=1, backward GRU
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
gru_x = linear.linear(x, xws[layer_idx], xbs[layer_idx])
gru_h = linear.linear(h, hws[layer_idx], hbs[layer_idx])
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
h_bar = (1 - z) * h_bar + z * h
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def _one_directional_loop(di):
# di=0, forward RNN
# di=1, backward RNN
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
rnn_in = (
linear.linear(x, xws[layer_idx], xbs[layer_idx]) +
linear.linear(h, hws[layer_idx], hbs[layer_idx]))
if activation == 'tanh':
h_bar = tanh.tanh(rnn_in)
elif activation == 'relu':
h_bar = relu.relu(rnn_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def _one_directional_loop(di):
# di=0, forward GRU
# di=1, backward GRU
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
gru_x = linear.linear(x, xws[layer_idx], xbs[layer_idx])
gru_h = linear.linear(h, hws[layer_idx], hbs[layer_idx])
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
h_bar = (1 - z) * h_bar + z * h
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def _one_directional_loop(di):
# di=0, forward RNN
# di=1, backward RNN
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
rnn_in = (linear.linear(x, xws[layer_idx],
xbs[layer_idx]) +
linear.linear(h, hws[layer_idx], hbs[layer_idx]))
if activation == 'tanh':
h_bar = tanh.tanh(rnn_in)
elif activation == 'relu':
h_bar = relu.relu(rnn_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def _lstm(x, h, c, w, b):
xw = _stack_weight([w[2], w[0], w[1], w[3]])
hw = _stack_weight([w[6], w[4], w[5], w[7]])
xb = _stack_weight([b[2], b[0], b[1], b[3]])
hb = _stack_weight([b[6], b[4], b[5], b[7]])
lstm_in = linear.linear(x, xw, xb) + linear.linear(h, hw, hb)
c_bar, h_bar = lstm.lstm(c, lstm_in)
return h_bar, c_bar
def f(x, h, c, w, b):
xw, hw = w
xb, hb = b
rnn_in = linear.linear(x, xw, xb) + linear.linear(h, hw, hb)
if activation == 'tanh':
return tanh.tanh(rnn_in), None
elif activation == 'relu':
return relu.relu(rnn_in), None
def _lstm(x, h, c, w, b):
xw = _stack_weight([w[2], w[0], w[1], w[3]])
hw = _stack_weight([w[6], w[4], w[5], w[7]])
xb = _stack_weight([b[2], b[0], b[1], b[3]])
hb = _stack_weight([b[6], b[4], b[5], b[7]])
lstm_in = linear.linear(x, xw, xb) + linear.linear(h, hw, hb)
c_bar, h_bar = lstm.lstm(c, lstm_in)
return h_bar, c_bar
def f(x, h, c, w, b):
xw, hw = w
xb, hb = b
rnn_in = linear.linear(x, xw, xb) + linear.linear(h, hw, hb)
if activation == 'tanh':
return tanh.tanh(rnn_in), None
elif activation == 'relu':
return relu.relu(rnn_in), None
def __call__(self, x):
if self.W.data is None:
in_size = functools.reduce(operator.mul, x.shape[1:], 1)
self._initialize_params(in_size)
self.W.data[self.triu_indices] = 0
return linear.linear(x, self.W, self.b)
def _gru(x, h, c, w, b):
xw = concat.concat([w[0], w[1], w[2]], axis=0)
hw = concat.concat([w[3], w[4], w[5]], axis=0)
xb = concat.concat([b[0], b[1], b[2]], axis=0)
hb = concat.concat([b[3], b[4], b[5]], axis=0)
gru_x = linear.linear(x, xw, xb)
gru_h = linear.linear(h, hw, hb)
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
return (1 - z) * h_bar + z * h, None
def lz3_Linear(_in, _out):
print(_in.link.__dict__['_params'])
_out.data[0] = np.zeros(_out.data[0].shape)
my_linear_out = linear(_in.args[0], _in.link.__dict__['W'],
_in.link.__dict__['b'])
my_linear_out = my_linear_out.reshape(_out.data[0].shape)
_out.data[0] = my_linear_out.data
def ly2_Linear(_in, _out):
if var.n < 0: # No fault injection for Normal System case
assert _in.args[
0].dtype == np.float32, 'Unsupport input type {}'.format(
_out.dtype)
assert _out.dtype == np.float32, 'Unsupport out type {}'.format(
_out.dtype)
return
batch, in_size = _in.args[0].shape
batch, go_size = _out.data.shape
detect_flag, layer, node, bit, sa01 = var.faultpat[var.n]
if layer == 1:
# Update _in with sa01
g = _in.args[0].data.copy()
for i in range(batch):
normal = g[i][node % in_size]
v_float, v_uint = bitChange(normal, bit, sa01)
g[i][node % in_size] = np.float32(v_float)
if 0:
print("{:8d} faultpattern={}".format(var.n, var.faultpat[var.n]))
print(' ' * 8, np.max(_in.args[0]), '=>', np.max(g),
np.min(_in.args[0]), '=>', np.min(g))
_in.args[0].data = g
# Calculate Linear Layer after fault insertion
this_linear_out = linear(_in.args[0], _in.link.__dict__['W'],
_in.link.__dict__['b'])
this_linear_out = this_linear_out.reshape(_out.data.shape)
_out.data = this_linear_out.data
def _gru(x, h, c, w, b):
xw = concat.concat([w[0], w[1], w[2]], axis=0)
hw = concat.concat([w[3], w[4], w[5]], axis=0)
xb = concat.concat([b[0], b[1], b[2]], axis=0)
hb = concat.concat([b[3], b[4], b[5]], axis=0)
gru_x = linear.linear(x, xw, xb)
gru_h = linear.linear(h, hw, hb)
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
return (1 - z) * h_bar + z * h, None
def test_t_is_10_nonzero_c_sequence_output():
np.random.seed(2)
N = 1
T = 10
C1 = 128
C2 = 64
vx = np.random.normal(size=(N, T, C1)).astype(np.float32)
vw_input = np.random.normal(size=(C1, C2 * 4)).astype(np.float32)
vw_hidden = np.random.normal(size=(C2, C2 * 4)).astype(np.float32)
vb = np.random.normal(size=(C2 * 4,)).astype(np.float32)
vc_in = np.random.normal(size=(N, C2)).astype(np.float32)
vc_out = vc_in.copy()
vh_in = np.random.normal(size=(N, C2)).astype(np.float32)
vh = vh_in
vw_input_c = _convert_to_chainer_order(vw_input)
vw_hidden_c = _convert_to_chainer_order(vw_hidden)
vb_c = _convert_to_chainer_order(vb[None, :])
vh_sequence = []
for i in range(T):
vc_out, vh = lstm(vc_out, linear(vx[:, i, :], vw_input_c.T) + linear(vh, vw_hidden_c.T) + vb_c)
vh_sequence.append(vh.data)
vh = np.array(vh_sequence).transpose((1, 0, 2)) # TNC -> NTC
vc_out = vc_out.data
x = Variable(vx.shape, order=OrderNTC)
c_in = ConstantVariable(vc_in, order=OrderNC)
vh_in = ConstantVariable(vh_in, order=OrderNC)
w_input = ConstantVariable(vw_input, order=OrderCN)
w_hidden = ConstantVariable(vw_hidden, order=OrderCN)
b = ConstantVariable(vb, order=OrderC)
y, c_out = LSTM(None, return_sequences=True, use_bias=True, use_initial_c=True, use_initial_h=True,
activation="tanh", recurrent_activation="sigmoid")(x, w_input, w_hidden, b, initial_c=c_in,
initial_h=vh_in)
generate_kernel_test_case(
description=f"LSTM t=10 initial_c,initial_h=nonzero sequence_out",
backend=["webassembly", "webgpu"],
graph=Graph([x], [y, c_out]),
inputs={x: vx},
expected={y: vh, c_out: vc_out},
EPS=1e-3,
ABS_EPS=1e-7
)
def __call__(self, x):
"""Applies the linear layer. However, I checked this code for simple data, It does not work...
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.U.data is None or self.V.data is not None:
in_size = x.shape[1]
self._initialize_params(in_size)
# x: (batch_size, CxHxW)
# V: (CxHxW, k)
# W: (k, CxHxW)
# (V*(U*x))+b = Wx + b
W1 = linear.linear(x, self.U)
return linear.linear(W1, self.V, self.b)
def __call__(self, x, W=None, b=None):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.has_uninitialized_params:
with cuda.get_device(self._device_id):
self._initialize_params(x.size // x.shape[0])
if W is not None:
return linear.linear(x, W, b)
return linear.linear(x, self.W, self.b)
def __call__(self, x, W=None, b=None):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.has_uninitialized_params:
with cuda.get_device_from_id(self._device_id):
self._initialize_params(x.size // x.shape[0])
if W is not None:
return linear.linear(x, W, b)
return linear.linear(x, self.W, self.b)
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.data is None:
self._initialize_params(x.size // x.shape[0])
return linear.linear(x, self.W_bar, self.b)
def _one_directional_loop(di):
# di=0, forward LSTM
# di=1, backward LSTM
h_list = []
c_list = []
layer_idx = direction * layer + di
h = hx[layer_idx]
c = cx[layer_idx]
if di == 0:
xs_list = xs_next
else:
xs_list = reversed(xs_next)
counter = 0
for x in xs_list:
counter += 1
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
c_rest = None
if layer != 0:
x = dropout.dropout(x, ratio=dropout_ratio)
if counter == 4:
lstm_in = linear.linear(x, xws[layer_idx],
xbs[layer_idx])
else:
lstm_in = linear.linear(
x, xws[layer_idx], xbs[layer_idx]) + linear.linear(
h, hws[layer_idx], hbs[layer_idx])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_list.append(h_bar)
c_list.append(c_bar)
return h, c, h_list, c_list
def __call__(self, x, W_lateral, b_lateral):
lstm_in = x
if self.h is not None:
lstm_in += linear.linear(self.h, W_lateral, b_lateral)
if self.c is None:
xp = self.xp
self.c = variable.Variable(
xp.zeros((len(x.data), self.state_size), dtype=x.data.dtype))
self.c, self.h = lstm.lstm(self.c, lstm_in)
return self.h
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
return linear.linear(x, self.W, self.b)
def __call__(self, x, W_lateral, W_peephole, b_peephole):
lstm_in = x
if self.h is not None:
lstm_in += linear.linear(self.h, W_lateral)
if self.c is None:
xp = self.xp
self.c = variable.Variable(xp.zeros((len(x.data), self.state_size),
dtype=x.data.dtype),
volatile='auto')
self.c, self.h = lstm_peephole(self.c, lstm_in, W_peephole, b_peephole)
return self.h
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
return linear.linear(x, self.W, self.b)
def max_singular_value(W, u=None, Ip=1):
"""
Apply power iteration for the weight parameter
"""
xp = cuda.get_array_module(W.data)
if u is None:
u = xp.random.normal(size=(1, W.shape[0])).astype(xp.float32)
_u = u
for _ in range(Ip):
_v = _l2normalize(xp.dot(_u, W.data), eps=1e-12)
_u = _l2normalize(xp.dot(_v, W.data.transpose()), eps=1e-12)
sigma = sum.sum(linear.linear(_u, transpose.transpose(W)) * _v)
return sigma, _u, _v
def __call__(self, xnext, eow, h1, h2, h3, W1, W2, W3, b1, b2, b3, prob_bias):
"""
xnext : next state of a pen. ndim=(batchsize,3)
h : input vector
W1, W2, W3: (h.shape[1], 1 + mix_size * 6)
b1, b2, b3: (1, 1 + mix_size * 6)
prob_bias: probability bias
"""
mix_size = self.mix_size
y = linear.linear(h1, W1, b1)
y += linear.linear(h2, W2, b2)
y += linear.linear(h3, W3, b3)
eos_hat, pi_hat, mu1_hat, mu2_hat, sg1_hat, sg2_hat, rho_hat = chainer.functions.split_axis(
y,
numpy.asarray([1, 1+mix_size, 1+2*mix_size, 1+3*mix_size, 1+4*mix_size, 1+5*mix_size]), axis=1)
self.loss, self.xpred, self.eos, self.pi_, self.mux, self.muy, self.sgx, self.sgy, self.rho = mixture_density_outputs(
xnext, eow, eos_hat, pi_hat * (1. + prob_bias), mu1_hat, mu2_hat, sg1_hat - prob_bias, sg2_hat - prob_bias, rho_hat)
return self.loss, self.xpred, self.eos, self.pi_, self.mux, self.muy, self.sgx, self.sgy, self.rho
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.data is None:
self._initialize_params(x.size // x.shape[0])
return linear.linear(x, self.W, self.b)
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.has_uninitialized_params:
self._initialize_params(x.shape[1])
return linear.linear(x, self.W, self.b)
def _one_directional_loop(di):
# di=0, forward LSTM
# di=1, backward LSTM
h_list = []
c_list = []
layer_idx = direction * layer + di
h = hx[layer_idx]
c = cx[layer_idx]
if di == 0:
xs_list = xs_next
else:
xs_list = reversed(xs_next)
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
c_rest = None
if layer != 0:
x = dropout.dropout(x, ratio=dropout_ratio,
train=train)
lstm_in = linear.linear(x, xws[layer_idx],
xbs[layer_idx]) + \
linear.linear(h, hws[layer_idx], hbs[layer_idx])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_list.append(h_bar)
c_list.append(c_bar)
return h, c, h_list, c_list
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.data is None:
in_size = functools.reduce(operator.mul, x.shape[1:], 1)
self._initialize_params(in_size)
return linear.linear(x, self.W, self.b)
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.has_uninitialized_params:
with cuda.get_device(self._device_id):
self._initialize_params(x.size // len(x.data))
return linear.linear(x, self.W, self.b)
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.data is None:
in_size = functools.reduce(operator.mul, x.shape[1:], 1)
self._initialize_params(in_size)
return linear.linear(x, self.W, self.b)
def l2_Linear(_in, _out):
_name = 'l2_Linear'
owner_layer = 3
# _in.args[0].shape : batch, in_size (1024, 300)
# _out.shape : batch, go_size (1024, 10)
batch, in_size = _in.args[0].data.shape
batch, go_size = _out.data.shape
if var.n < 0: # No fault injection for Normal System case
print(_name, 'in/out =', _in.args[0].data.shape,
_out.data.shape)
assert _in.args[
0].dtype == np.float32, 'Unsupport input type {}'.format(
_out.dtype)
assert _out.dtype == np.float32, 'Unsupport out type {}'.format(
_out.dtype)
# return
detect_flag, layer, node, bit, sa01 = var.faultpat[
var.n] if var.n >= 0 else [False, -1, -1, -1, -1]
if layer == owner_layer or var.pi == _name:
g = _in.args[0].data.copy()
# Update _in as PI
if var.pi == _name:
for i in range(np.prod(g.shape)):
g.reshape(-1)[i] = var.PIpat.reshape(-1)[i]
# Update _in with sa01
if layer == owner_layer:
for i in range(batch):
normal = g.reshape(batch, -1)[i, node]
v_float, v_uint = bitChange(normal, bit, sa01)
g.reshape(batch, -1)[i, node] = np.float32(v_float)
#_in.args[0].data = g
# Calculate Linear Layer after fault insertion
this_linear_out = linear(g, _in.link.__dict__['W'],
_in.link.__dict__['b'])
this_linear_out = this_linear_out.reshape(_out.data.shape)
_out.data = this_linear_out.data
if layer == 4:
g = _out.data.copy()
for i in range(batch):
normal = g.reshape(batch, -1)[i, node]
v_float, v_uint = bitChange(normal, bit, sa01)
g.reshape(batch, -1)[i, node] = np.float32(v_float)
if 0:
print("{:8d} faultpattern={}".format(var.n, var.faultpat[var.n]))
print(' ' * 8, np.max(_out.data), '=>', np.max(g),
np.min(_out.data), '=>', np.min(g))
_out.data = g
def __call__(self, x):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.has_uninitialized_params:
with cuda.get_device(self._device_id):
self._initialize_params(x.size // len(x.data))
return linear.linear(x, self.W, self.b)
def lz3_Linear(_in, _out):
_out.data[0] = np.zeros(_out.data[0].shape)
#print("_in",_in.args[0].shape)
#_in.args = np.zeros(_in.args[0].shape,dtype=np.float32)
print(_out.data[0].shape)
my_linear_out = linear(_in.args[0], _in.link.__dict__['W'],
_in.link.__dict__['b'])
my_linear_out = my_linear_out.reshape(_out.data[0].shape)
print(my_linear_out.shape)
print(my_linear_out)
set_trace()
_out.data[0] = my_linear_out.data
print(_out)
pass
def lx1_Linear(_in,_out):
# n=0
n=var.n
if n<0: return # No fault injection for Normal System case
num = 1
d=mylist()
print("mylist:",n,d[n])
x=d[n][0]
y=d[n][1]
g = func(_in.args[0], x, y, num, 1)
_in.args[0].data = g
my_linear_out = linear(_in.args[0], _in.link.__dict__['W'], _in.link.__dict__['b'])
my_linear_out = my_linear_out.reshape(_out.data.shape)
_out.data = my_linear_out.data
def __call__(self, cs, ls, h, W, b):
"""
cs : one-hot-encoding of a length U character sequence
h : input vector (summation of affine transformation of outputs from hidden layers)
"""
mix_size = self.mix_size
y = linear.linear(h, W, b)
a_hat, b_hat, k_hat = chainer.functions.split_axis(
y, numpy.asarray([mix_size, 2 * mix_size]), axis=1)
if self.k_prev is None:
xp = self.xp
self.k_prev = variable.Variable(xp.zeros_like(k_hat.data))
self.ws, self.k_prev, self.eow = soft_window(cs, ls, a_hat, b_hat,
k_hat, self.k_prev)
return self.ws, self.eow
def forward(self, x, n_batch_axes=1):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
n_batch_axes (int): The number of batch axes. The default is 1. The
input variable is reshaped into
(:math:`{\\rm n\\_batch\\_axes} + 1`)-dimensional tensor.
This should be greater than 0.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.array is None:
in_size = utils.size_of_shape(x.shape[1:])
self._initialize_params(in_size)
return linear.linear(x, self.W, self.b, n_batch_axes=n_batch_axes)
def forward(self, x, n_batch_axes=1):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Batch of input vectors.
n_batch_axes (int): The number of batch axes. The default is 1. The
input variable is reshaped into
(:math:`{\\rm n\\_batch\\_axes} + 1`)-dimensional tensor.
This should be greater than 0.
Returns:
~chainer.Variable: Output of the linear layer.
"""
if self.W.array is None:
in_size = functools.reduce(operator.mul, x.shape[1:], 1)
self._initialize_params(in_size)
return linear.linear(x, self.W, self.b, n_batch_axes=n_batch_axes)
def lx1_Linear(_in,_out):
# n=0
n=var.n
if n<0: return # No fault injection for Normal System case
num = 1
d=mylist()
print("mylist:",n,d[n])
x=d[n][0]
y=d[n][1]
#g = func(_in.args[0].get(stream=None), x, y, num, 1)
#g = _in.args[0].get(stream=None)
#_in.args[0][:,x,y,0]=_in.args[0][:,x,y,0]
for i in range(10):_in.args[0][0,x,y,0]=1
#_in.args[0].set(G,stream=None)
my_linear_out = linear(_in.args[0], _in.link.__dict__['W'], _in.link.__dict__['b'])
my_linear_out = my_linear_out.reshape(_out.data.shape)
_out.data.set(my_linear_out.data.get(stream=None).copy())
def __call__(self, x):
if self.mu_w.data is None:
in_size = functools.reduce(operator.mul, x.shape[1:], 1)
self._initialize_params(in_size)
dtype = self.mu_w.dtype
out_size, in_size = self.mu_w.shape
if self.factorized:
eps_i = self._eps(in_size, dtype)
eps_j = self._eps(out_size, dtype)
eps_w = self.xp.outer(eps_j, eps_i)
else:
eps_w = self._eps((out_size, in_size), dtype)
eps_j = self._eps(out_size, dtype)
W = self.mu_w + self.sigma_w * eps_w
if self.mu_b is None:
b = None
else:
b = self.mu_b.reshape((out_size, )) + self.sigma_b * eps_j
return linear.linear(x, W, b)
def n_step_lstm(n_layers,
dropout_ratio,
hx,
cx,
ws,
bs,
xs,
train=True,
use_cudnn=True):
"""Stacked Long Short-Term Memory function for sequence inputs.
This function calculates stacked LSTM with sequences. This function gets
an initial hidden state :math:`h_0`, an initial cell state :math:`c_0`,
an input sequence :math:`x`, weight matrices :math:`W`, and bias vectors
:math:`b`.
This function calculates hidden states :math:`h_t` and :math:`c_t` for each
time :math:`t` from input :math:`x_t`.
.. math::
i_t = \sigma(W_0 x_t + W_4 h_{t-1} + b_0 + b_4)
f_t = \sigma(W_1 x_t + W_5 h_{t-1} + b_1 + b_5)
o_t = \sigma(W_2 x_t + W_6 h_{t-1} + b_2 + b_6)
a_t = \tanh(W_3 x_t + W_7 h_{t-1} + b_3 + b_7)
c_t = f_t \dot c_{t-1} + i_t \dot a_t
h_t = o_t \dot \tanh(c_t)
As the function accepts a sequence, it calculates :math:`h_t` for all
:math:`t` with one call. Eight weight matrices and eight bias vectors are
required for each layers. So, when :math:`S` layers exists, you need to
prepare :math:`8S` weigth matrices and :math:`8S` bias vectors.
If the number of layers ``n_layers`` is greather than :math:`1`, input
of ``k``-th layer is hidden state ``h_t`` of ``k-1``-th layer.
Note that all input variables except first layer may have different shape
from the first layer.
Args:
n_layers(int): Number of layers.
dropout_ratio(float): Dropout ratio.
hx (chainer.Variable): Variable holding stacked hidden states.
Its shape is ``(S, B, N)`` where ``S`` is number of layers and is
equal to ``n_layers``, ``B`` is mini-batch size, and ``N`` is
dimention of hidden units.
cx (chainer.Variable): Variable holding stacked cell states.
It has the same shape as ``hx``.
ws (list of list of chainer.Variable): Weight matrices. ``ws[i]``
represents weights for i-th layer.
Each ``ws[i]`` is a list containing eight matrices.
``ws[i][j]`` is corresponding with ``W_j`` in the equation.
Only ``ws[0][j]`` where ``0 <= j < 4`` is ``(I, N)`` shape as they
are multiplied with input variables. All other matrices has
``(N, N)`` shape.
bs (list of list of chainer.Variable): Bias vectors. ``bs[i]``
represnents biases for i-th layer.
Each ``bs[i]`` is a list containing eight vectors.
``bs[i][j]`` is corresponding with ``b_j`` in the equation.
Shape of each matrix is ``(N,)`` where ``N`` is dimention of
hidden units.
xs (list of chainer.Variable): A list of :class:`chainer.Variable`
holding input values. Each element ``xs[t]`` holds input value
for time ``t``. Its shape is ``(B_t, I)``, where ``B_t`` is
mini-batch size for time ``t``, and ``I`` is size of input units.
Note that this functions supports variable length sequences.
When sequneces has different lengths, sort sequences in descending
order by length, and transpose the sorted sequence.
:func:`~chainer.functions.transpose_sequence` transpose a list
of :func:`~chainer.Variable` holding sequence.
So ``xs`` needs to satisfy
``xs[t].shape[0] >= xs[t + 1].shape[0]``.
train (bool): If ``True``, this function executes dropout.
use_cudnn (bool): If ``True``, this function uses cuDNN if available.
Returns:
tuple: This functions returns a tuple concaining three elements,
``hy``, ``cy`` and ``ys``.
- ``hy`` is an updated hidden states whose shape is same as ``hx``.
- ``cy`` is an updated cell states whose shape is same as ``cx``.
- ``ys`` is a list of :class:~chainer.Variable. Each element
``ys[t]`` holds hidden states of the last layer corresponding
to an input ``xs[t]``. Its shape is ``(B_t, N)`` where ``B_t`` is
mini-batch size for time ``t``, and ``N`` is size of hidden
units. Note that ``B_t`` is the same value as ``xs[t]``.
.. seealso::
:func:`chainer.functions.lstm`
"""
xp = cuda.get_array_module(hx, hx.data)
if use_cudnn and xp is not numpy and cuda.cudnn_enabled and \
_cudnn_version >= 5000:
states = get_random_state().create_dropout_states(dropout_ratio)
# flatten all input variables
inputs = tuple(
itertools.chain((hx, cx), itertools.chain.from_iterable(ws),
itertools.chain.from_iterable(bs), xs))
rnn = NStepLSTM(n_layers, states, train=train)
ret = rnn(*inputs)
hy, cy = ret[:2]
ys = ret[2:]
return hy, cy, ys
else:
hx = split_axis.split_axis(hx, n_layers, axis=0, force_tuple=True)
hx = [reshape.reshape(h, h.shape[1:]) for h in hx]
cx = split_axis.split_axis(cx, n_layers, axis=0, force_tuple=True)
cx = [reshape.reshape(c, c.shape[1:]) for c in cx]
xws = [_stack_weight([w[2], w[0], w[1], w[3]]) for w in ws]
hws = [_stack_weight([w[6], w[4], w[5], w[7]]) for w in ws]
xbs = [_stack_weight([b[2], b[0], b[1], b[3]]) for b in bs]
hbs = [_stack_weight([b[6], b[4], b[5], b[7]]) for b in bs]
ys = []
for x in xs:
batch = x.shape[0]
h_next = []
c_next = []
for layer in six.moves.range(n_layers):
h = hx[layer]
c = cx[layer]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
x = dropout.dropout(x, ratio=dropout_ratio, train=train)
h = dropout.dropout(h, ratio=dropout_ratio, train=train)
lstm_in = linear.linear(x, xws[layer], xbs[layer]) + \
linear.linear(h, hws[layer], hbs[layer])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_next.append(h)
c_next.append(c)
x = h_bar
hx = h_next
cx = c_next
ys.append(x)
hy = stack.stack(hx)
cy = stack.stack(cx)
return hy, cy, tuple(ys)
def fixed_length_n_step_lstm(
n_layers,
dropout_ratio,
hx,
cx,
ws,
bs,
xs,
train=True,
):
xp = cuda.get_array_module(hx, hx.data)
if xp is not numpy and cuda.cudnn_enabled and _cudnn_version >= 5000:
states = get_random_state().create_dropout_states(dropout_ratio)
# flatten all input variables
inputs = tuple(
itertools.chain((hx, cx), itertools.chain.from_iterable(ws),
itertools.chain.from_iterable(bs), (xs, )))
rnn = FixedLengthNStepLSTMFunction(n_layers, states, train=train)
ret = rnn(*inputs)
hy, cy, ys = ret
_, batch_size, dim = hy.shape
ys_reshape = F.reshape(ys,
(-1, batch_size, dim)) # (length, batch, dim)
return hy, cy, ys_reshape
else:
hx = split_axis.split_axis(hx, n_layers, axis=0, force_tuple=True)
hx = [reshape.reshape(h, h.shape[1:]) for h in hx]
cx = split_axis.split_axis(cx, n_layers, axis=0, force_tuple=True)
cx = [reshape.reshape(c, c.shape[1:]) for c in cx]
xws = [_stack_weight([w[2], w[0], w[1], w[3]]) for w in ws]
hws = [_stack_weight([w[6], w[4], w[5], w[7]]) for w in ws]
xbs = [_stack_weight([b[2], b[0], b[1], b[3]]) for b in bs]
hbs = [_stack_weight([b[6], b[4], b[5], b[7]]) for b in bs]
ys = []
for x in xs:
batch = x.shape[0]
h_next = []
c_next = []
for layer in six.moves.range(n_layers):
h = hx[layer]
c = cx[layer]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
x = dropout.dropout(x, ratio=dropout_ratio)
h = dropout.dropout(h, ratio=dropout_ratio)
lstm_in = linear.linear(x, xws[layer], xbs[layer]) + \
linear.linear(h, hws[layer], hbs[layer])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_next.append(h)
c_next.append(c)
x = h_bar
hx = h_next
cx = c_next
ys.append(x)
hy = stack.stack(hx)
cy = stack.stack(cx)
#return hy, cy, tuple(ys)
ys_concat = F.concat(ys, axis=0)
ys_reshape = F.reshape(
ys_concat,
(-1, ys[0].shape[0], ys[0].shape[1])) # (length, batch, dim)
return hy, cy, ys_reshape
def nonlinear(x, W, b=None):
y = linear.linear(x, W, b)
return y * y
def n_step_lstm(
n_layers, dropout_ratio, hx, cx, ws, bs, xs, train=True,
use_cudnn=True):
"""Stacked Long Short-Term Memory function for sequence inputs.
This function calculates stacked LSTM with sequences. This function gets
an initial hidden state :math:`h_0`, an initial cell state :math:`c_0`,
an input sequence :math:`x`, weight matrices :math:`W`, and bias vectors
:math:`b`.
This function calculates hidden states :math:`h_t` and :math:`c_t` for each
time :math:`t` from input :math:`x_t`.
.. math::
i_t &= \\sigma(W_0 x_t + W_4 h_{t-1} + b_0 + b_4) \\\\
f_t &= \\sigma(W_1 x_t + W_5 h_{t-1} + b_1 + b_5) \\\\
o_t &= \\sigma(W_2 x_t + W_6 h_{t-1} + b_2 + b_6) \\\\
a_t &= \\tanh(W_3 x_t + W_7 h_{t-1} + b_3 + b_7) \\\\
c_t &= f_t \\dot c_{t-1} + i_t \\dot a_t \\\\
h_t &= o_t \\dot \\tanh(c_t)
As the function accepts a sequence, it calculates :math:`h_t` for all
:math:`t` with one call. Eight weight matrices and eight bias vectors are
required for each layers. So, when :math:`S` layers exists, you need to
prepare :math:`8S` weigth matrices and :math:`8S` bias vectors.
If the number of layers ``n_layers`` is greather than :math:`1`, input
of ``k``-th layer is hidden state ``h_t`` of ``k-1``-th layer.
Note that all input variables except first layer may have different shape
from the first layer.
Args:
n_layers(int): Number of layers.
dropout_ratio(float): Dropout ratio.
hx (chainer.Variable): Variable holding stacked hidden states.
Its shape is ``(S, B, N)`` where ``S`` is number of layers and is
equal to ``n_layers``, ``B`` is mini-batch size, and ``N`` is
dimention of hidden units.
cx (chainer.Variable): Variable holding stacked cell states.
It has the same shape as ``hx``.
ws (list of list of chainer.Variable): Weight matrices. ``ws[i]``
represents weights for i-th layer.
Each ``ws[i]`` is a list containing eight matrices.
``ws[i][j]`` is corresponding with ``W_j`` in the equation.
Only ``ws[0][j]`` where ``0 <= j < 4`` is ``(I, N)`` shape as they
are multiplied with input variables. All other matrices has
``(N, N)`` shape.
bs (list of list of chainer.Variable): Bias vectors. ``bs[i]``
represnents biases for i-th layer.
Each ``bs[i]`` is a list containing eight vectors.
``bs[i][j]`` is corresponding with ``b_j`` in the equation.
Shape of each matrix is ``(N,)`` where ``N`` is dimention of
hidden units.
xs (list of chainer.Variable): A list of :class:`~chainer.Variable`
holding input values. Each element ``xs[t]`` holds input value
for time ``t``. Its shape is ``(B_t, I)``, where ``B_t`` is
mini-batch size for time ``t``, and ``I`` is size of input units.
Note that this functions supports variable length sequences.
When sequneces has different lengths, sort sequences in descending
order by length, and transpose the sorted sequence.
:func:`~chainer.functions.transpose_sequence` transpose a list
of :func:`~chainer.Variable` holding sequence.
So ``xs`` needs to satisfy
``xs[t].shape[0] >= xs[t + 1].shape[0]``.
train (bool): If ``True``, this function executes dropout.
use_cudnn (bool): If ``True``, this function uses cuDNN if available.
Returns:
tuple: This functions returns a tuple concaining three elements,
``hy``, ``cy`` and ``ys``.
- ``hy`` is an updated hidden states whose shape is same as ``hx``.
- ``cy`` is an updated cell states whose shape is same as ``cx``.
- ``ys`` is a list of :class:`~chainer.Variable` . Each element
``ys[t]`` holds hidden states of the last layer corresponding
to an input ``xs[t]``. Its shape is ``(B_t, N)`` where ``B_t`` is
mini-batch size for time ``t``, and ``N`` is size of hidden
units. Note that ``B_t`` is the same value as ``xs[t]``.
.. seealso::
:func:`chainer.functions.lstm`
"""
xp = cuda.get_array_module(hx, hx.data)
if use_cudnn and xp is not numpy and cuda.cudnn_enabled and \
_cudnn_version >= 5000:
states = get_random_state().create_dropout_states(dropout_ratio)
# flatten all input variables
inputs = tuple(itertools.chain(
(hx, cx),
itertools.chain.from_iterable(ws),
itertools.chain.from_iterable(bs),
xs))
rnn = NStepLSTM(n_layers, states, train=train)
ret = rnn(*inputs)
hy, cy = ret[:2]
ys = ret[2:]
return hy, cy, ys
else:
hx = split_axis.split_axis(hx, n_layers, axis=0, force_tuple=True)
hx = [reshape.reshape(h, h.shape[1:]) for h in hx]
cx = split_axis.split_axis(cx, n_layers, axis=0, force_tuple=True)
cx = [reshape.reshape(c, c.shape[1:]) for c in cx]
xws = [_stack_weight([w[2], w[0], w[1], w[3]]) for w in ws]
hws = [_stack_weight([w[6], w[4], w[5], w[7]]) for w in ws]
xbs = [_stack_weight([b[2], b[0], b[1], b[3]]) for b in bs]
hbs = [_stack_weight([b[6], b[4], b[5], b[7]]) for b in bs]
ys = []
for x in xs:
batch = x.shape[0]
h_next = []
c_next = []
for layer in six.moves.range(n_layers):
h = hx[layer]
c = cx[layer]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
x = dropout.dropout(x, ratio=dropout_ratio, train=train)
h = dropout.dropout(h, ratio=dropout_ratio, train=train)
lstm_in = linear.linear(x, xws[layer], xbs[layer]) + \
linear.linear(h, hws[layer], hbs[layer])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_next.append(h)
c_next.append(c)
x = h_bar
hx = h_next
cx = c_next
ys.append(x)
hy = stack.stack(hx)
cy = stack.stack(cx)
return hy, cy, tuple(ys)
