class InputlessLstmBlock(object):
"""
Inputless A long short-term memory (LSTM) block.
Parameters
----------
R
b
grad_clipping
x
mask
prev_c
prev_h
device_id : int
Defines the device's id on which the computation will take place
Returns
-------
"""
def __init__(self,
R,
b,
grad_clipping,
mask,
prev_c,
prev_h,
device_id=None):
self.f_context = Context(device_id)
device_id = self.f_context.device_id
if R.bpropagable:
self.R, self.dL_dR = R.register_usage(device_id, device_id)
self.R_b_context = Context(device_id)
else:
self.R = R.register_usage(device_id)
if b.bpropagable:
self.b, self.dL_db = b.register_usage(device_id, device_id)
self.b_b_context = Context(device_id)
else:
self.b = b.register_usage(device_id)
self.grad_clipping = grad_clipping
if mask:
self.mask = mask.register_usage(device_id)
if prev_c.bpropagable:
self.prev_c, self.dL_dprev_c = prev_c.register_usage(
device_id, device_id)
else:
self.prev_c = prev_c.register_usage(device_id)
if prev_h.bpropagable:
self.prev_h, self.dL_dprev_h = prev_h.register_usage(
device_id, device_id)
else:
self.prev_h = prev_h.register_usage(device_id)
self.learning = R.bpropagable or prev_c.bpropagable or prev_h.bpropagable
if self.learning:
self.b_context = Context(device_id)
dim = self.R.nrows
batch_size = self.prev_c.nrows
self.zifo = Matrix.empty(batch_size, 4 * dim, device_id=device_id)
self.z = self.zifo[:, 0 * dim:1 * dim]
self.i = self.zifo[:, 1 * dim:2 * dim]
self.f = self.zifo[:, 2 * dim:3 * dim]
self.o = self.zifo[:, 3 * dim:4 * dim]
self.c = Matrix.empty_like(self.prev_c, device_id)
self.c = Connector(self.c, device_id if self.learning else None)
self.tanh_c = Matrix.empty_like(self.c, device_id)
self.h = Matrix.empty_like(self.c, device_id)
self.h = Connector(self.h, device_id if self.learning else None)
if self.learning:
self._dzifo_dpre_zifo = Matrix.empty_like(self.zifo)
self.dz_dpre_z = self._dzifo_dpre_zifo[:, 0 * dim:1 * dim]
self.di_dpre_i = self._dzifo_dpre_zifo[:, 1 * dim:2 * dim]
self.df_dpre_f = self._dzifo_dpre_zifo[:, 2 * dim:3 * dim]
self.do_dpre_o = self._dzifo_dpre_zifo[:, 3 * dim:4 * dim]
self.dL_dpre_zifo = self._dzifo_dpre_zifo
self.dL_dpre_z = self.dz_dpre_z
self.dL_dpre_i = self.di_dpre_i
self.dL_dpre_f = self.df_dpre_f
self.dL_dpre_o = self.do_dpre_o
self._dtanh_c_dc = Matrix.empty_like(self.c)
@property
def dzifo_dpre_zifo(self):
if self.learning:
return self._dzifo_dpre_zifo
@property
def dtanh_c_dc(self):
if self.learning:
return self._dtanh_c_dc
def fprop(self):
# zifo = tanh_sigm(h[t-1] * R + b)
self.zifo.assign_dot(self.f_context, self.prev_h, self.R)
self.zifo.add(self.f_context, self.b)
self.zifo.tanh_sigm(self.f_context,
self.zifo,
self.dzifo_dpre_zifo,
axis=1)
# c[t] = i[t] .* z[t] + f[t] .* c[t-1]
# h[t] = o[t] .* tanh(c[t])
self.c.assign_sum_hprod(self.f_context, self.i, self.z, self.f,
self.prev_c)
self.c.tanh(self.f_context, self.tanh_c, self.dtanh_c_dc)
self.h.assign_hprod(self.f_context, self.o, self.tanh_c)
if hasattr(self, 'mask'):
# s[t] = mask .* s[t] + (1 - mask) .* s[t-1]
self.c.assign_masked_addition(self.f_context, self.mask, self.c,
self.prev_c)
self.h.assign_masked_addition(self.f_context, self.mask, self.h,
self.prev_h)
self.c.fprop()
self.h.fprop()
def bprop(self):
if not self.learning:
return
dL_dc = self.c.backward_matrix
dL_dh = self.h.backward_matrix
if hasattr(self, 'mask'):
# dL/ds[t-1] = (1 - mask) .* dL/ds[t]
# dL/ds[t] = mask .* dL/ds[t]
if hasattr(self, 'dL_dprev_c'):
self.dL_dprev_c.add_hprod_one_minus_mask(
self.b_context, self.mask, dL_dc)
dL_dc.hprod(self.b_context, self.mask)
if hasattr(self, 'dL_dprev_h'):
self.dL_dprev_h.add_hprod_one_minus_mask(
self.b_context, self.mask, dL_dh)
dL_dh.hprod(self.b_context, self.mask)
# dL/dc[t] = dL[t+1]/dc[t] + dL/dh[t] .* o[t] .* dtanh(c[t])/dc[t]
dL_dc.add_hprod(self.b_context, dL_dh, self.o, self.dtanh_c_dc)
# dL/dpre_o[t] = dL/dh[t] .* tanh(c[t]) .* do[t]/dpre_o[t]
# dL/dpre_f[t] = dL/dc[t] .* c[t-1] .* df[t]/dpre_f[t]
# dL/dpre_i[t] = dL/dc[t] .* z[t] .* di[t]/dpre_i[t]
# dL/dpre_z[t] = dL/dc[t] .* i[t] .* dz[t]/dpre_z[t]
self.dL_dpre_o.assign_hprod(self.b_context, dL_dh, self.tanh_c,
self.do_dpre_o)
self.dL_dpre_f.assign_hprod(self.b_context, dL_dc, self.prev_c,
self.df_dpre_f)
self.dL_dpre_i.assign_hprod(self.b_context, dL_dc, self.z,
self.di_dpre_i)
self.dL_dpre_z.assign_hprod(self.b_context, dL_dc, self.i,
self.dz_dpre_z)
self.dL_dpre_zifo.last_modif_context = self.b_context
if self.grad_clipping:
self.dL_dpre_zifo.clip(self.b_context, -self.grad_clipping,
self.grad_clipping)
if hasattr(self, 'dL_dR'):
# dL_dR += h[t-1].T * dL/dpre_zifo[t]
self.dL_dR.add_dot(self.R_b_context, self.prev_h,
self.dL_dpre_zifo, 'T')
if hasattr(self, 'dL_db'):
# dL_db += sum(dL/dpre_zifo[t], axis=0)
self.dL_db.add_repeat_derivative(self.b_b_context,
self.dL_dpre_zifo,
self.dL_dpre_zifo.nrows,
axis=0)
if hasattr(self, 'dL_dprev_c'):
# dL/dc[t-1] = f[t] .* dL/dc[t]
self.dL_dprev_c.add_hprod(self.b_context, self.f, dL_dc)
if hasattr(self, 'dL_dprev_h'):
# dL/dh[t-1] = dL/dpre_zifo[t] * R.T
self.dL_dprev_h.add_dot(self.b_context, self.dL_dpre_zifo, self.R,
'N', 'T')
class LstmBlock(object):
"""
A long short-term memory (LSTM) block.
Parameters
----------
W
R
b
grad_clipping
x
mask
prev_c
prev_h
device_id : int
Defines the device's id on which the computation will take place
Returns
-------
"""
def __init__(self, W, R, b, grad_clipping, x, mask, prev_c, prev_h, device_id=None):
self.f_context = Context(device_id)
device_id = self.f_context.device_id
if W.bpropagable:
self.W, self.dL_dW = W.register_usage(device_id, device_id)
self.W_b_context = Context(device_id)
else:
self.W = W.register_usage(device_id)
if R.bpropagable:
self.R, self.dL_dR = R.register_usage(device_id, device_id)
self.R_b_context = Context(device_id)
else:
self.R = R.register_usage(device_id)
if b.bpropagable:
self.b, self.dL_db = b.register_usage(device_id, device_id)
self.b_b_context = Context(device_id)
else:
self.b = b.register_usage(device_id)
self.grad_clipping = grad_clipping
if x.bpropagable:
self.x, self.dL_dx = x.register_usage(device_id, device_id)
self.x_b_context = Context(device_id)
else:
self.x = x.register_usage(device_id)
if mask:
self.mask = mask.register_usage(device_id)
if prev_c.bpropagable:
self.prev_c, self.dL_dprev_c = prev_c.register_usage(device_id, device_id)
self.prev_c_b_context = Context(device_id)
else:
self.prev_c = prev_c.register_usage(device_id)
if prev_h.bpropagable:
self.prev_h, self.dL_dprev_h = prev_h.register_usage(device_id, device_id)
self.prev_h_b_context = Context(device_id)
else:
self.prev_h = prev_h.register_usage(device_id)
self.learning = W.bpropagable or R.bpropagable or x.bpropagable or \
prev_c.bpropagable or prev_h.bpropagable
if self.learning:
self.b_context = Context(device_id)
dim = self.R.nrows
batch_size = self.x.nrows
self.zifo = Matrix.empty(batch_size, 4 * dim, device_id=device_id)
self.z = self.zifo[:, 0*dim:1*dim]
self.i = self.zifo[:, 1*dim:2*dim]
self.f = self.zifo[:, 2*dim:3*dim]
self.o = self.zifo[:, 3*dim:4*dim]
self.c = Matrix.empty_like(self.prev_c, device_id)
self.c = Connector(self.c, device_id if self.learning else None)
self.tanh_c = Matrix.empty_like(self.c, device_id)
self.h = Matrix.empty_like(self.c, device_id)
self.h = Connector(self.h, device_id if self.learning else None)
if self.learning:
self._dzifo_dpre_zifo = Matrix.empty_like(self.zifo)
self.dz_dpre_z = self._dzifo_dpre_zifo[:, 0*dim:1*dim]
self.di_dpre_i = self._dzifo_dpre_zifo[:, 1*dim:2*dim]
self.df_dpre_f = self._dzifo_dpre_zifo[:, 2*dim:3*dim]
self.do_dpre_o = self._dzifo_dpre_zifo[:, 3*dim:4*dim]
self.dL_dpre_zifo = self._dzifo_dpre_zifo
self.dL_dpre_z = self.dz_dpre_z
self.dL_dpre_i = self.di_dpre_i
self.dL_dpre_f = self.df_dpre_f
self.dL_dpre_o = self.do_dpre_o
self._dtanh_c_dc = Matrix.empty_like(self.c)
@property
def dzifo_dpre_zifo(self):
if self.learning:
return self._dzifo_dpre_zifo
@property
def dtanh_c_dc(self):
if self.learning:
return self._dtanh_c_dc
def fprop(self):
# zifo = tanh_sigm(x[t] * W + h[t-1] * R + b)
self.zifo.assign_dot(self.f_context, self.x, self.W)
self.zifo.add_dot(self.f_context, self.prev_h, self.R)
self.zifo.add(self.f_context, self.b)
self.zifo.tanh_sigm(self.f_context, self.zifo, self.dzifo_dpre_zifo, axis=1)
# c[t] = i[t] .* z[t] + f[t] .* c[t-1]
# h[t] = o[t] .* tanh(c[t])
self.c.assign_sum_hprod(self.f_context, self.i, self.z, self.f, self.prev_c)
self.c.tanh(self.f_context, self.tanh_c, self.dtanh_c_dc)
self.h.assign_hprod(self.f_context, self.o, self.tanh_c)
if hasattr(self, 'mask'):
# s[t] = mask .* s[t] + (1 - mask) .* s[t-1]
self.c.assign_masked_addition(self.f_context, self.mask, self.c, self.prev_c)
self.h.assign_masked_addition(self.f_context, self.mask, self.h, self.prev_h)
self.c.fprop()
self.h.fprop()
def bprop(self):
dL_dc = self.c.backward_matrix
dL_dh = self.h.backward_matrix
if hasattr(self, 'mask'):
# dL/ds[t-1] = (1 - mask) .* dL/ds[t]
# dL/ds[t] = mask .* dL/ds[t]
if hasattr(self, 'dL_dprev_c'):
self.dL_dprev_c.add_hprod_one_minus_mask(self.prev_c_b_context, self.mask, dL_dc)
dL_dc.hprod(self.prev_c_b_context, self.mask)
if hasattr(self, 'dL_dprev_h'):
self.dL_dprev_h.add_hprod_one_minus_mask(self.prev_h_b_context, self.mask, dL_dh)
dL_dh.hprod(self.prev_h_b_context, self.mask)
# dL/dc[t] = dL[t+1]/dc[t] + dL/dh[t] .* o[t] .* dtanh(c[t])/dc[t]
dL_dc.add_hprod(self.b_context, dL_dh, self.o, self.dtanh_c_dc)
# self.dzifo_dpre_zifo was calculated in self.f_context,
# now we have to explicitly wait it in context self.b_context, because
# self.dx_dpre_x does not have proper last_modif_context
self.b_context.wait(self.f_context)
# dL/dpre_o[t] = dL/dh[t] .* tanh(c[t]) .* do[t]/dpre_o[t]
# dL/dpre_f[t] = dL/dc[t] .* c[t-1] .* df[t]/dpre_f[t]
# dL/dpre_i[t] = dL/dc[t] .* z[t] .* di[t]/dpre_i[t]
# dL/dpre_z[t] = dL/dc[t] .* i[t] .* dz[t]/dpre_z[t]
self.dL_dpre_o.assign_hprod(self.b_context, dL_dh, self.tanh_c, self.do_dpre_o)
self.dL_dpre_f.assign_hprod(self.b_context, dL_dc, self.prev_c, self.df_dpre_f)
self.dL_dpre_i.assign_hprod(self.b_context, dL_dc, self.z, self.di_dpre_i)
self.dL_dpre_z.assign_hprod(self.b_context, dL_dc, self.i, self.dz_dpre_z)
if self.grad_clipping:
self.dL_dpre_zifo.clip(self.b_context, -self.grad_clipping, self.grad_clipping)
else:
self.dL_dpre_zifo.last_modif_context = self.b_context
if hasattr(self, 'dL_dW'):
# dL_dW += x[t].T * dL/dpre_zifo[t]
self.dL_dW.add_dot(self.W_b_context, self.x, self.dL_dpre_zifo, 'T')
if hasattr(self, 'dL_dR'):
# dL_dR += h[t-1].T * dL/dpre_zifo[t]
self.dL_dR.add_dot(self.R_b_context, self.prev_h, self.dL_dpre_zifo, 'T')
if hasattr(self, 'dL_db'):
# dL_db += sum(dL/dpre_zifo[t], axis=0)
self.dL_db.add_repeat_derivative(self.b_b_context, self.dL_dpre_zifo, self.dL_dpre_zifo.nrows, axis=0)
if hasattr(self, 'dL_dx'):
# dL/dx[t] = dL/dpre_zifo[t] * W.T
self.dL_dx.add_dot(self.x_b_context, self.dL_dpre_zifo, self.W, 'N', 'T')
if hasattr(self, 'dL_dprev_c'):
# dL/dc[t-1] = f[t] .* dL/dc[t]
self.dL_dprev_c.add_hprod(self.prev_c_b_context, self.f, dL_dc)
if hasattr(self, 'dL_dprev_h'):
# dL/dh[t-1] = dL/dpre_zifo[t] * R.T
self.dL_dprev_h.add_dot(self.prev_h_b_context, self.dL_dpre_zifo, self.R, 'N', 'T')
