def create_model(input_dim):
row = sequence.input_variable(shape=input_dim)
col = sequence.input_variable(shape=input_dim)
rowh = Sequential([Embedding(opt.embed), Stabilizer(), Dropout(opt.dropout)])(row)
colh = Sequential([Embedding(opt.embed), Stabilizer(), Dropout(opt.dropout)])(col)
x = C.splice(rowh, colh, axis=-1)
x = lightlstm(opt.embed, opt.nhid)(x)
x = For(range(opt.layer-1), lambda: lightlstm(opt.nhid, opt.nhid))(x)
rowh = C.slice(x, -1, opt.nhid * 0, opt.nhid * 1)
colh = C.slice(x, -1, opt.nhid * 1, opt.nhid * 2)
row_predict = Sequential([Dropout(opt.dropout), Dense(input_dim)])(rowh)
col_predict = Sequential([Dropout(opt.dropout), Dense(input_dim)])(colh)
# variable : row label and col label
row_label = sequence.input_variable(shape=input_dim)
col_label = sequence.input_variable(shape=input_dim)
model = C.combine([row_predict, col_predict])
return {'row': row,
'col': col,
'row_label': row_label,
'col_label': col_label,
'model': model}
def create_model(input_sequence, label_sequence, vocab_dim, hidden_dim):
# Create the rnn that computes the latent representation for the next token.
rnn_with_latent_output = Sequential([
C.layers.Embedding(hidden_dim),
For(
range(num_layers), lambda: Sequential([
Stabilizer(),
Recurrence(LSTM(hidden_dim), go_backwards=False)
])),
])
# Apply it to the input sequence.
latent_vector = rnn_with_latent_output(input_sequence)
# Connect the latent output to (sampled/full) softmax.
if use_sampled_softmax:
weights = load_sampling_weights(token_frequencies_file_path)
smoothed_weights = np.float32(np.power(weights, alpha))
sampling_weights = C.reshape(C.Constant(smoothed_weights),
shape=(1, vocab_dim))
z, ce, errs = cross_entropy_with_sampled_softmax(
latent_vector, label_sequence, vocab_dim, hidden_dim,
softmax_sample_size, sampling_weights)
else:
z, ce, errs = cross_entropy_with_full_softmax(latent_vector,
label_sequence,
vocab_dim, hidden_dim)
return z, ce, errs
def create_model(output_dim):
return Sequential([
For(
range(num_layers), lambda: Sequential([
Stabilizer(),
Recurrence(LSTM(hidden_dim), go_backwards=False)
])),
Dense(output_dim)
])
def test_layers_stabilizer():
y = input(4)
p = Stabilizer()(y)
dat = np.array([[1.0,2.0,3.0,4.0]], dtype=np.float32)
res = p(y).eval({y: dat})
# a stabilizer starts with having no effect, hence input=output
np.testing.assert_array_almost_equal(res, dat, decimal=6, \
err_msg="Error in layer normalization computation")
X = np.eye(vocab_size, dtype=np.float32)[xi]
Y = np.eye(vocab_size, dtype=np.float32)[yi]
return [X], [Y]
get_sample(0)
input_sequence = sequence.input_variable(shape=vocab_size)
label_sequence = sequence.input_variable(shape=vocab_size)
model = Sequential([
For(
range(2), lambda: Sequential(
[Stabilizer(),
Recurrence(LSTM(256), go_backwards=False)])),
Dense(vocab_size)
])
z = model(input_sequence)
z_sm = cntk.softmax(z)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
lr_per_sample = learning_rate_schedule(0.001, UnitType.sample)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 5.0
gradient_clipping_with_truncation = True
learner = momentum_sgd(
X = np.eye(vocab_size, dtype=np.float32)[xi]
Y = np.eye(vocab_size, dtype=np.float32)[yi]
return [X], [Y]
sample(0)
input_seq_axis = Axis('inputAxis')
input_sequence = sequence.input_variable(shape=vocab_size, sequence_axis=input_seq_axis)
label_sequence = sequence.input_variable(shape=vocab_size, sequence_axis=input_seq_axis)
# model = Sequential([Dense(300),Dense(vocab_size)])
model = Sequential([
For(range(2), lambda:
Sequential([Stabilizer(), Recurrence(LSTM(256), go_backwards=False)])),
Dense(vocab_size)])
z = model(input_sequence)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
lr_per_sample = learning_rate_schedule(0.001, UnitType.sample)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 5.0
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters, lr_per_sample, momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
progress_printer = ProgressPrinter(freq=100, tag='Training')
def _RecurrentBlock(type, shape, cell_shape, activation, use_peepholes,
init, init_bias,
enable_self_stabilization, dropout_rate, seed,
name=''):
'''
Helper to create a recurrent block of type 'WeightDroppedLSTM', 'GRU', or RNNStep.
'''
has_projection = cell_shape is not None
shape = _as_tuple(shape)
cell_shape = _as_tuple(cell_shape) if cell_shape is not None else shape
if len(shape) != 1 or len(cell_shape) != 1:
raise ValueError("%s: shape and cell_shape must be vectors (rank-1 tensors)" % type)
# otherwise we'd need to fix slicing and Param initializers
stack_axis = -1 # for efficient computation, we stack multiple variables (along the fastest-changing one, to match BS)
# determine stacking dimensions
cell_shape_list = list(cell_shape)
stacked_dim = cell_shape_list[stack_axis]
cell_shape_list[stack_axis] = stacked_dim * {
'IndRNN': 1,
'IndyLSTM': 4,
'WeightDroppedLSTM': 4
}[type]
cell_shape_stacked = tuple(cell_shape_list) # patched dims with stack_axis duplicated 4 times
cell_shape_list[stack_axis] = stacked_dim * {
'IndRNN': 1,
'IndyLSTM': 4,
'WeightDroppedLSTM': 4
}[type]
cell_shape_stacked_H = tuple(cell_shape_list) # patched dims with stack_axis duplicated 4 times
# parameters
b = Parameter( cell_shape_stacked, init=init_bias, name='b') # bias
W = Parameter(_INFERRED + cell_shape_stacked, init=init, name='W') # input
H = Parameter(shape + cell_shape_stacked_H, init=init, name='H') # hidden-to-hidden
H1 = Parameter( cell_shape_stacked_H, init=init, name='H1') if type == 'IndyLSTM' else None # hidden-to-hidden
H2 = Parameter(shape , init=init, name='H2') if type == 'IndRNN' else None # hidden-to-hidden
Ci = Parameter( cell_shape, init=init, name='Ci') if use_peepholes else None # cell-to-hiddden {note: applied elementwise}
Cf = Parameter( cell_shape, init=init, name='Cf') if use_peepholes else None # cell-to-hiddden {note: applied elementwise}
Co = Parameter( cell_shape, init=init, name='Co') if use_peepholes else None # cell-to-hiddden {note: applied elementwise}
Wmr = Parameter(cell_shape + shape, init=init, name='P') if has_projection else None # final projection
# each use of a stabilizer layer must get its own instance
Sdh = Stabilizer(enable_self_stabilization=enable_self_stabilization, name='dh_stabilizer')
Sdc = Stabilizer(enable_self_stabilization=enable_self_stabilization, name='dc_stabilizer')
Sct = Stabilizer(enable_self_stabilization=enable_self_stabilization, name='c_stabilizer')
Sht = Stabilizer(enable_self_stabilization=enable_self_stabilization, name='P_stabilizer')
# DropConnect
dropout = C.layers.Dropout(dropout_rate=dropout_rate, seed=seed, name='h_dropout')
# define the model function itself
# general interface for Recurrence():
# (all previous outputs delayed, input) --> (outputs and state)
# where
# - the first output is the main output, e.g. 'h' for LSTM
# - the remaining outputs, if any, are additional state
# - if for some reason output != state, then output is still fed back and should just be ignored by the recurrent block
# LSTM model function
# in this case:
# (dh, dc, x) --> (h, c)
def weight_dropped_lstm(dh, dc, x):
dhs = Sdh(dh) # previous values, stabilized
dcs = Sdc(dc)
# note: input does not get a stabilizer here, user is meant to do that outside
# projected contribution from input(s), hidden, and bias
proj4 = b + times(x, W) + times(dhs, dropout(H))
it_proj = slice (proj4, stack_axis, 0*stacked_dim, 1*stacked_dim) # split along stack_axis
bit_proj = slice (proj4, stack_axis, 1*stacked_dim, 2*stacked_dim)
ft_proj = slice (proj4, stack_axis, 2*stacked_dim, 3*stacked_dim)
ot_proj = slice (proj4, stack_axis, 3*stacked_dim, 4*stacked_dim)
# helper to inject peephole connection if requested
def peep(x, c, C):
return x + C * c if use_peepholes else x
it = sigmoid(peep(it_proj, dcs, Ci)) # input gate(t)
# TODO: should both activations be replaced?
bit = it * activation(bit_proj) # applied to tanh of input network
ft = sigmoid(peep(ft_proj, dcs, Cf)) # forget-me-not gate(t)
bft = ft * dc # applied to cell(t-1)
ct = bft + bit # c(t) is sum of both
ot = sigmoid(peep(ot_proj, Sct(ct), Co)) # output gate(t)
ht = ot * activation(ct) # applied to tanh(cell(t))
c = ct # cell value
h = times(Sht(ht), Wmr) if has_projection else ht
return h, c
# LSTM model function
# in this case:
# (dh, dc, x) --> (h, c)
def indy_lstm(dh, dc, x):
dhs = Sdh(dh) # previous values, stabilized
dcs = Sdc(dc)
# note: input does not get a stabilizer here, user is meant to do that outside
# projected contribution from input(s), hidden, and bias
proj4 = b + times(x, W) + C.splice(dhs, dhs, dhs, dhs) * H1 # 4 is the number of stacked dim
it_proj = slice (proj4, stack_axis, 0*stacked_dim, 1*stacked_dim) # split along stack_axis
bit_proj = slice (proj4, stack_axis, 1*stacked_dim, 2*stacked_dim)
ft_proj = slice (proj4, stack_axis, 2*stacked_dim, 3*stacked_dim)
ot_proj = slice (proj4, stack_axis, 3*stacked_dim, 4*stacked_dim)
# helper to inject peephole connection if requested
def peep(x, c, C):
return x + C * c if use_peepholes else x
it = sigmoid(peep(it_proj, dcs, Ci)) # input gate(t)
# TODO: should both activations be replaced?
bit = it * activation(bit_proj) # applied to tanh of input network
ft = sigmoid(peep(ft_proj, dcs, Cf)) # forget-me-not gate(t)
bft = ft * dc # applied to cell(t-1)
ct = bft + bit # c(t) is sum of both
ot = sigmoid(peep(ot_proj, Sct(ct), Co)) # output gate(t)
ht = ot * activation(ct) # applied to tanh(cell(t))
c = ct # cell value
h = times(Sht(ht), Wmr) if has_projection else ht
return h, c
def ind_rnn(dh, x):
dhs = Sdh(dh) # previous value, stabilized
ht = activation(times(x, W) + dhs * H2 + b)
h = times(Sht(ht), Wmr) if has_projection else ht
return h
function = {
'IndRNN': ind_rnn,
'IndyLSTM': indy_lstm,
'WeightDroppedLSTM': weight_dropped_lstm
}[type]
# return the corresponding lambda as a CNTK Function
return BlockFunction(type, name)(function)
