class PtbMiniBatchesGenerator(object):
def __init__(self, ptb_train, ptb_valid, batch_size, sentence_max_len, device_id):
self.blocking_contexts = None
self.context = Context(device_id)
device_id = self.context.device_id
self.train_offsets = HomogeneousDataGenerator(ptb_train, batch_size, sentence_max_len, randomize=True, infinite=True)
self.valid_offsets = HomogeneousDataGenerator(ptb_valid, batch_size, sentence_max_len)
train_sentences = np.array([self.train_offsets.flatten_sentences])
valid_sentences = np.array([self.valid_offsets.flatten_sentences])
self.train_sents = Matrix.from_npa(train_sentences, 'int', device_id)
self.valid_sents = Matrix.from_npa(valid_sentences, 'int', device_id)
self._sent_lengths = np.empty((batch_size, 1), dtype=np.int32, order='F')[...]
self.sent_lengths = Matrix.from_npa(self._sent_lengths, device_id=device_id)
sentence_batch = Matrix.empty(batch_size, sentence_max_len, 'int', device_id)
self.sentence_batch = Connector(sentence_batch, self.context)
self.sentence_batch.sync_fill(0)
self._mask = Matrix.empty(sentence_batch.nrows, self.sentence_batch.ncols, 'float', device_id)
self.mask = List([Connector(self._mask[:, i]) for i in xrange(sentence_max_len)], self.sentence_batch.ncols)
self.train_offsets_iterator = iter(self.train_offsets)
self.valid_offsets_iterator = iter(self.valid_offsets)
self.training_mode = True
def set_training_mode(self):
self.training_mode = True
def set_testing_mode(self):
self.training_mode = False
def fprop(self):
if self.training_mode:
offsets = next(self.train_offsets_iterator)
sents = self.train_sents
else:
try:
offsets = next(self.valid_offsets_iterator)
sents = self.valid_sents
except StopIteration as e:
self.valid_offsets_iterator = iter(self.valid_offsets)
raise e
self.context.wait(*self.blocking_contexts)
self._sent_lengths = self._sent_lengths.base[:len(offsets)]
self.sentence_batch.nrows = len(offsets)
for k, offset in enumerate(offsets):
self.sentence_batch[k].assign(self.context, sents[:, offset[0]:offset[1]])
self._sent_lengths[k] = offset[1] - offset[0]
max_sent_len = int(np.max(self._sent_lengths))
self.sentence_batch.last_modification_context = self.context
self.sentence_batch.ncols = max_sent_len
self.sent_lengths.assign_npa(self.context, self._sent_lengths)
self._mask.mask_column_numbers_row_wise(self.context, self.sent_lengths)
for e in self.mask:
e.last_modification_context = self.context
self.sentence_batch.fprop()
self.mask.fprop()
class DataBlock(object):
def __init__(self, data, char_to_idx, batch_size, x_device_id,
y_device_id):
self.data = HomogeneousDataIterator(data, char_to_idx, batch_size,
True, True)
self.data_iterator = iter(self.data)
self.x_context = Context(x_device_id)
self.y_context = Context(y_device_id)
max_len = 0
for sub_line in data:
cur_len = len(sub_line)
if cur_len > max_len:
max_len = cur_len
print max_len
self.x = Connector(
Matrix.empty(batch_size, max_len - 1, 'int', x_device_id))
self._y = Matrix.empty(batch_size, max_len - 1, 'int', y_device_id)
self.y = List([Connector(self._y[:, i]) for i in xrange(max_len - 1)],
self.x.ncols)
self.lengths = Matrix.empty(self.x.nrows, 1, 'int', x_device_id)
self._mask = Matrix.empty(self.x.nrows, self.x.ncols, 'float',
x_device_id)
self.mask = List(
[Connector(self._mask[:, i]) for i in xrange(max_len)],
self.x.ncols)
self.blocking_contexts = None
def fprop(self):
self.x_context.wait(*self.blocking_contexts)
self.y_context.wait(*self.blocking_contexts)
data = next(self.data_iterator)
lengths_npa = np.array([[len(e) - 1] for e in data],
np.int32,
order='F')
x_npa = np.zeros((len(data), int(np.max(lengths_npa))), np.int32, 'F')
for k, e in enumerate(data):
x_npa[k, :len(e) - 1] = e[:-1]
self.x.assign_npa(self.x_context, x_npa)
y_npa = np.zeros((len(data), int(np.max(lengths_npa))), np.int32, 'F')
for k, e in enumerate(data):
y_npa[k, :len(e) - 1] = e[1:]
self._y.assign_npa(self.y_context, y_npa)
for e in self.y:
e.last_modification_context = self.y_context
self.lengths.assign_npa(self.x_context, lengths_npa)
self._mask.mask_column_numbers_row_wise(self.x_context, self.lengths)
for e in self.mask:
e.last_modification_context = self.x_context
self.x.fprop()
self.y.fprop()
self.mask.fprop()
class DataBlock(object):
def __init__(self, data, char_to_idx, batch_size, x_device_id, y_device_id):
self.data = HomogeneousDataIterator(data, char_to_idx, batch_size, True, True)
self.data_iterator = iter(self.data)
self.x_context = Context(x_device_id)
self.y_context = Context(y_device_id)
max_len = 0
for sub_line in data:
cur_len = len(sub_line)
if cur_len > max_len:
max_len = cur_len
print max_len
self.x = Connector(Matrix.empty(batch_size, max_len - 1, 'int', x_device_id))
self._y = Matrix.empty(batch_size, max_len - 1, 'int', y_device_id)
self.y = List([Connector(self._y[:, i]) for i in xrange(max_len - 1)], self.x.ncols)
self.lengths = Matrix.empty(self.x.nrows, 1, 'int', x_device_id)
self._mask = Matrix.empty(self.x.nrows, self.x.ncols, 'float', x_device_id)
self.mask = List([Connector(self._mask[:, i]) for i in xrange(max_len)], self.x.ncols)
self.blocking_contexts = None
def fprop(self):
self.x_context.wait(*self.blocking_contexts)
self.y_context.wait(*self.blocking_contexts)
data = next(self.data_iterator)
lengths_npa = np.array([[len(e) - 1] for e in data], np.int32, order='F')
x_npa = np.zeros((len(data), int(np.max(lengths_npa))), np.int32, 'F')
for k, e in enumerate(data):
x_npa[k, :len(e) - 1] = e[:-1]
self.x.assign_npa(self.x_context, x_npa)
y_npa = np.zeros((len(data), int(np.max(lengths_npa))), np.int32, 'F')
for k, e in enumerate(data):
y_npa[k, :len(e) - 1] = e[1:]
self._y.assign_npa(self.y_context, y_npa)
for e in self.y:
e.last_modification_context = self.y_context
self.lengths.assign_npa(self.x_context, lengths_npa)
self._mask.mask_column_numbers_row_wise(self.x_context, self.lengths)
for e in self.mask:
e.last_modification_context = self.x_context
self.x.fprop()
self.y.fprop()
self.mask.fprop()
class PtbMiniBatchesGenerator(object):
def __init__(self, ptb_train, ptb_valid, batch_size, sentence_max_len,
device_id):
self.blocking_contexts = None
self.context = Context(device_id)
device_id = self.context.device_id
self.train_offsets = HomogeneousDataGenerator(ptb_train,
batch_size,
sentence_max_len,
randomize=True,
infinite=True)
self.valid_offsets = HomogeneousDataGenerator(ptb_valid, batch_size,
sentence_max_len)
train_sentences = np.array([self.train_offsets.flatten_sentences])
valid_sentences = np.array([self.valid_offsets.flatten_sentences])
self.train_sents = Matrix.from_npa(train_sentences, 'int', device_id)
self.valid_sents = Matrix.from_npa(valid_sentences, 'int', device_id)
self._sent_lengths = np.empty((batch_size, 1),
dtype=np.int32,
order='F')[...]
self.sent_lengths = Matrix.from_npa(self._sent_lengths,
device_id=device_id)
sentence_batch = Matrix.empty(batch_size, sentence_max_len, 'int',
device_id)
self.sentence_batch = Connector(sentence_batch, self.context)
self.sentence_batch.sync_fill(0)
self._mask = Matrix.empty(sentence_batch.nrows,
self.sentence_batch.ncols, 'float',
device_id)
self.mask = List(
[Connector(self._mask[:, i]) for i in xrange(sentence_max_len)],
self.sentence_batch.ncols)
self.train_offsets_iterator = iter(self.train_offsets)
self.valid_offsets_iterator = iter(self.valid_offsets)
self.training_mode = True
def set_training_mode(self):
self.training_mode = True
def set_testing_mode(self):
self.training_mode = False
def fprop(self):
if self.training_mode:
offsets = next(self.train_offsets_iterator)
sents = self.train_sents
else:
try:
offsets = next(self.valid_offsets_iterator)
sents = self.valid_sents
except StopIteration as e:
self.valid_offsets_iterator = iter(self.valid_offsets)
raise e
self.context.wait(*self.blocking_contexts)
self._sent_lengths = self._sent_lengths.base[:len(offsets)]
self.sentence_batch.nrows = len(offsets)
for k, offset in enumerate(offsets):
self.sentence_batch[k].assign(self.context,
sents[:, offset[0]:offset[1]])
self._sent_lengths[k] = offset[1] - offset[0]
max_sent_len = int(np.max(self._sent_lengths))
self.sentence_batch.last_modification_context = self.context
self.sentence_batch.ncols = max_sent_len
self.sent_lengths.assign_npa(self.context, self._sent_lengths)
self._mask.mask_column_numbers_row_wise(self.context,
self.sent_lengths)
for e in self.mask:
e.last_modification_context = self.context
self.sentence_batch.fprop()
self.mask.fprop()
class MnistMiniBatchesGenerator(object):
def __init__(self, train_x, train_y, valid_x, valid_y, batch_size, device_id):
self.context = Context(device_id)
device_id = self.context.device_id
self.train_x = Matrix.from_npa(train_x.T.astype(np.float32), device_id=device_id)
self.valid_x = Matrix.from_npa(valid_x.T.astype(np.float32), device_id=device_id)
self.train_y = Matrix.from_npa(train_y[:, np.newaxis], 'int', device_id=device_id)
self.valid_y = Matrix.from_npa(valid_y[:, np.newaxis], 'int', device_id=device_id)
self.batch_size = batch_size
x = Matrix.empty(self.batch_size, self.train_x.nrows, device_id=device_id)
y = Matrix.empty(self.batch_size, 1, 'int', device_id)
self.x = Connector(x)
self.y = Connector(y)
self.train_indices = np.arange(int(self.train_x.ncols), dtype=np.int32)
self.valid_indices = np.arange(int(self.valid_x.ncols), dtype=np.int32)
self.indices = Matrix.empty(self.batch_size, 1, 'int', device_id)
self.rng = np.random.RandomState(42)
self.rng.shuffle(self.train_indices)
self.train_i = 0
self.valid_i = 0
self.training_mode = True
self.blocking_contexts = None
def set_training_mode(self):
self.training_mode = True
def set_testing_mode(self):
self.training_mode = False
def fprop(self):
indices = self.train_indices if self.training_mode else self.valid_indices
i = self.train_i if self.training_mode else self.valid_i
x = self.train_x if self.training_mode else self.valid_x
y = self.train_y if self.training_mode else self.valid_y
indices = indices[self.batch_size * i:self.batch_size * (i + 1)]
indices = np.asfortranarray(indices[:, np.newaxis])
if self.training_mode:
self.train_i += 1
else:
self.valid_i += 1
if indices.size:
self.indices.assign_npa(self.context, indices)
self.x.nrows = indices.size
self.y.nrows = indices.size
self.context.wait(*self.blocking_contexts)
x.slice_columns_and_transpose(self.context, self.indices, self.x)
y.slice_rows(self.context, self.indices, self.y)
self.x.fprop()
self.y.fprop()
else:
if self.training_mode:
self.train_i = 0
self.rng.shuffle(self.train_indices)
self.fprop()
else:
self.valid_i = 0
raise StopIteration()
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')
