def multiFunc(self, arg1):
# load or create the inputs we need
multiIn = C.input(shape=arg1.shape, dynamic_axes = arg1.dynamic_axes)
bit_map = C.constant(self.bit_map)
max_bits = self.bit_map.max()
shape = multiIn.shape
reformed = C.reshape(multiIn, (-1,))
# lets compute the means we need
# carry over represents the remaining value that needs to binarized. For a single bit, this is just the input. For more bits,
# it is the difference between the previous bits approximation and the true value.
carry_over = multiIn
approx = C.element_times(multiIn, 0)
# iterate through the maximum number of bits specified by the bit maps, basically compute each level of binarization
for i in range(max_bits):
# determine which values of the input should be binarized to i bits or more
hot_vals = C.greater(bit_map, i)
# select only the values which we need to binarize
valid_vals = C.element_select(hot_vals, carry_over, 0)
# compute mean on a per kernel basis, reshaping is done to allow for sum reduction along only axis 0 (the kernels)
mean = C.element_divide(C.reduce_sum(C.reshape(C.abs(valid_vals), (valid_vals.shape[0], -1)), axis=1), C.reduce_sum(C.reshape(hot_vals, (hot_vals.shape[0], -1)), axis=1))
# reshape the mean to match the dimensionality of the input
mean = C.reshape(mean, (mean.shape[0], mean.shape[1], 1, 1))
# binarize the carry over
bits = C.greater(carry_over, 0)
bits = C.element_select(bits, bits, -1)
bits = C.element_select(hot_vals, bits, 0)
# add in the equivalent binary representation to the approximation
approx = C.plus(approx, C.element_times(mean, bits))
# compute the new carry over
carry_over = C.plus(C.element_times(C.element_times(-1, bits), mean), carry_over)
return approx, multiIn
def output_layer(self, query, match_context):
q_processed = C.placeholder(shape=(2*self.hidden_dim,))
mat_context = C.placeholder(shape=(2*self.hidden_dim,))
#output layer
r_q = question_pooling(q_processed, 2*self.hidden_dim) #shape n*(2*self.hidden_dim)
p1_logits = attention_weight(mat_context, r_q, 2*self.hidden_dim)
attention_pool = C.sequence.reduce_sum(p1_logits * mat_context)
state = C.layers.GRU(2*self.hidden_dim)(attention_pool, r_q)
p2_logits = attention_weight(mat_context, state, 2*self.hidden_dim)
@C.Function
def start_ave_point(p1_logits, p2_logits, point):
@C.Function
def start_ave(last, now):
now = now + last - last
new_start = now * C.sequence.gather(p2_logits, point)
point = C.sequence.future_value(point)
return new_start
start_logits_ave = C.layers.Recurrence(start_ave)(p1_logits)
return start_logits_ave
point = C.sequence.is_first(p1_logits)
point = C.layers.Sequential([For(range(2), lambda: C.layers.Recurrence(C.plus))])(point)
point = C.greater(C.constant(16), point)
start_logits_ave = start_ave_point(p1_logits, p2_logits, point)
@C.Function
def end_ave_point(p1_logits, p2_logits, point):
@C.Function
def end_ave(last, now):
now = now + last - last
new_end = now * C.sequence.gather(p2_logits, point)
point = C.sequence.past_value(point)
return new_end
end_logits_ave = C.layers.Recurrence(end_ave, go_backwards=True)(p2_logits)
return end_logits_ave
point = C.sequence.is_last(p1_logits)
point = C.layers.Sequential([For(range(2), lambda: C.layers.Recurrence(C.plus, go_backwards=True))])(point)
point = C.greater(C.constant(16),point)
end_logits_ave = end_ave_point(p1_logits, p2_logits, point)
start_logits = seq_hardmax(start_logits_ave)
end_logits = seq_hardmax(end_logits_ave)
'''
start_logits = seq_hardmax(p1_logits)
end_logits = seq_hardmax(p2_logits)
'''
return C.as_block(
C.combine([start_logits, end_logits]),
[(q_processed, query), (mat_context, match_context)],
'output_layer',
'output_layer')
def triangular_matrix_seq(mode: int = 1):
X = C.placeholder(1)
ones = C.ones_like(X[0])
perm_1 = C.layers.Recurrence(C.plus, return_full_state=True)(ones)
perm_2 = C.layers.Recurrence(C.plus,
go_backwards=True,
return_full_state=True)(ones)
arr_1 = C.sequence.unpack(perm_1, 0, True)
arr_2 = C.sequence.unpack(perm_2, 0, True)
mat = C.times_transpose(arr_1, arr_2)
mat_c = arr_1 * arr_2
diagonal_mat = mat - mat_c
final_mat = diagonal_mat
if mode == 0:
final_mat = C.equal(final_mat, 0)
elif mode == 1:
final_mat = C.less_equal(final_mat, 0)
elif mode == 2:
final_mat = C.less(final_mat, 0)
elif mode == -1:
final_mat = C.greater_equal(final_mat, 0)
elif mode == -2:
final_mat = C.greater(final_mat, 0)
result = C.as_block(final_mat, [(X, X)], 'triangular_matrix')
return C.stop_gradient(result)
def test_unfold():
from cntk.layers import UnfoldFrom
@Function
def double_up(s):
return s * 2
x = [[[0],[0],[0]],
[[0],[0],[0],[0],[0]]]
####################################################
# Test 1: simple unfold
####################################################
UF = UnfoldFrom(double_up)
@Function
@Signature(Sequence[Tensor[1]])
def FU(x):
return UF(Constant(1), x)
r = FU(x)
exp = [[[ 2 ], [ 4 ], [ 8 ]],
[[ 2 ], [ 4 ], [ 8 ], [ 16 ], [ 32 ]]]
assert_list_of_arrays_equal(r, exp, err_msg='Error in UnfoldFrom() forward')
####################################################
# Test 2: unfold with length increase and terminating condition
####################################################
UF = UnfoldFrom(double_up, until_predicate=lambda x: greater(x, 63), length_increase=1.6)
@Function
@Signature(Sequence[Tensor[1]])
def FU(x):
return UF(Constant(1), x)
r = FU(x)
exp = [[[ 2 ], [ 4 ], [ 8 ], [ 16 ], [ 32 ]], # tests length_increase
[[ 2 ], [ 4 ], [ 8 ], [ 16 ], [ 32 ], [ 64 ]]] # tests early cut-off due to until_predicate
assert_list_of_arrays_equal(r, exp, err_msg='Error in UnfoldFrom(..., until_predicate, length_increase, ...) forward')
def model(self):
c1_axis = C.Axis.new_unique_dynamic_axis('c1_axis')
c2_axis = C.Axis.new_unique_dynamic_axis('c2_axis')
b = C.Axis.default_batch_axis()
c1 = C.input_variable(self.word_dim,
dynamic_axes=[b, c1_axis],
name='c1')
c2 = C.input_variable(self.word_dim,
dynamic_axes=[b, c2_axis],
name='c2')
y = C.input_variable(1, dynamic_axes=[b], name='y')
c1_processed, c2_processed = self.input_layer(c1, c2).outputs
att_context = self.attention_layer(c2_processed, c1_processed,
'attention')
c2_len = C.layers.Fold(plus1)(c2_processed)
att_len = C.layers.Fold(plus1)(att_context)
cos = C.cosine_distance(
C.sequence.reduce_sum(c2_processed) / c2_len,
C.sequence.reduce_sum(att_context) / att_len)
prob = C.sigmoid(cos)
is_context = C.greater(prob, 0.5)
loss = C.losses.binary_cross_entropy(prob, y)
acc = C.equal(is_context, y)
return cos, loss, acc
def signFunc(self, arg):
# create an input variable that matches the dimension of the input argument
signIn = C.input(shape=arg.shape, dynamic_axes=arg.dynamic_axes)
# create the first stage of the sign function, check if input is greater than zero
actionfunc = C.greater(signIn, 0)
# return the second stage of the sign function, replace any 0s with -1s
return C.element_select(actionfunc, actionfunc, -1), signIn
def signFunc(self, arg):
# create an input variable that matches the dimension of the input argument
signIn = C.input(shape=arg.shape, dynamic_axes=arg.dynamic_axes)
# create the first stage of the sign function, check if input is greater than zero
actionfunc = C.greater(signIn, 0)
# return the second stage of the sign function, replace any 0s with -1s
return C.element_select(actionfunc, actionfunc, -1), signIn
def test_unfold():
from cntk.layers import UnfoldFrom
@Function
def double_up(s):
return s * 2
x = [[[0],[0],[0]],
[[0],[0],[0],[0],[0]]]
####################################################
# Test 1: simple unfold
####################################################
UF = UnfoldFrom(double_up)
@Function
@Signature(Sequence[Tensor[1]])
def FU(x):
return UF(Constant(1), x)
r = FU(x)
exp = [[ 2. , 4. , 8.],
[ 2. , 4. , 8. , 16. , 32. ]]
assert_list_of_arrays_equal(r, exp, err_msg='Error in UnfoldFrom() forward')
####################################################
# Test 2: unfold with length increase and terminating condition
####################################################
UF = UnfoldFrom(double_up, until_predicate=lambda x: greater(x, 63), length_increase=1.6)
@Function
@Signature(Sequence[Tensor[1]])
def FU(x):
return UF(Constant(1), x)
r = FU(x)
exp = [[ 2 , 4 , 8 , 16 , 32 ], # tests length_increase
[ 2 , 4 , 8 , 16 , 32 , 64 ]] # tests early cut-off due to until_predicate
assert_list_of_arrays_equal(r, exp, err_msg='Error in UnfoldFrom(..., until_predicate, length_increase, ...) forward')
def multiFunc(self, arg1):
multiIn = C.input(shape=arg1.shape, dynamic_axes=arg1.dynamic_axes)
bit_map = C.constant(self.bit_map)
max_bits = self.bit_map.max()
carry_over = multiIn
approx = C.element_times(multiIn, 0)
for i in range(max_bits):
hot_vals = C.greater(bit_map, i)
valid_vals = C.element_select(hot_vals, carry_over, 0)
mean = C.element_divide(C.reduce_sum(C.abs(valid_vals)),
C.reduce_sum(hot_vals))
bits = C.greater(carry_over, 0)
bits = C.element_select(bits, bits, -1)
bits = C.element_select(hot_vals, bits, 0)
approx = C.plus(approx, C.element_times(mean, bits))
carry_over = C.plus(
C.element_times(C.element_times(-1, bits), mean), carry_over)
return approx, multiIn
def multiFunc(self, arg1):
multiIn = C.input(shape=arg1.shape, dynamic_axes = arg1.dynamic_axes)
bit_map = C.constant(self.bit_map)
max_bits = self.bit_map.max()
shape = multiIn.shape
reformed = C.reshape(multiIn, (-1,))
carry_over = multiIn
approx = C.element_times(multiIn, 0)
for i in range(max_bits):
hot_vals = C.greater(bit_map, i)
valid_vals = C.element_select(hot_vals, carry_over, 0)
mean = C.element_divide(C.reduce_sum(C.abs(valid_vals)), C.reduce_sum(hot_vals))
bits = C.greater(carry_over, 0)
bits = C.element_select(bits, bits, -1)
bits = C.element_select(hot_vals, bits, 0)
approx = C.plus(approx, C.element_times(mean, bits))
carry_over = C.plus(C.element_times(C.element_times(-1, bits), mean), carry_over)
return approx, multiIn
def multiFunc(self, arg1):
# load or create the inputs we need
multiIn = C.input(shape=arg1.shape, dynamic_axes=arg1.dynamic_axes)
bit_map = C.constant(self.bit_map)
max_bits = self.bit_map.max()
shape = multiIn.shape
reformed = C.reshape(multiIn, (-1, ))
# lets compute the means we need
# carry over represents the remaining value that needs to binarized. For a single bit, this is just the input. For more bits,
# it is the difference between the previous bits approximation and the true value.
carry_over = multiIn
approx = C.element_times(multiIn, 0)
# iterate through the maximum number of bits specified by the bit maps, basically compute each level of binarization
for i in range(max_bits):
# determine which values of the input should be binarized to i bits or more
hot_vals = C.greater(bit_map, i)
# select only the values which we need to binarize
valid_vals = C.element_select(hot_vals, carry_over, 0)
# compute mean on a per kernel basis, reshaping is done to allow for sum reduction along only axis 0 (the kernels)
mean = C.element_divide(
C.reduce_sum(C.reshape(C.abs(valid_vals),
(valid_vals.shape[0], -1)),
axis=1),
C.reduce_sum(C.reshape(hot_vals, (hot_vals.shape[0], -1)),
axis=1))
# reshape the mean to match the dimensionality of the input
mean = C.reshape(mean, (mean.shape[0], mean.shape[1], 1, 1))
# binarize the carry over
bits = C.greater(carry_over, 0)
bits = C.element_select(bits, bits, -1)
bits = C.element_select(hot_vals, bits, 0)
# add in the equivalent binary representation to the approximation
approx = C.plus(approx, C.element_times(mean, bits))
# compute the new carry over
carry_over = C.plus(
C.element_times(C.element_times(-1, bits), mean), carry_over)
return approx, multiIn
def greater(left, right, name=''):
'''
Elementwise 'greater' comparison of two tensors. Result is 1 if left > right else 0.
Example:
>>> C.eval(C.greater([41., 42., 43.], [42., 42., 42.]))
[array([[0., 0., 1.]])]
>>> C.eval(C.greater([-1,0,1], [0]))
[array([[1., 0., 1.]])]
Args:
left: left side tensor
right: right side tensor
name (str): the name of the node in the network
Returns:
:class:`cntk.Function`
'''
from cntk import greater
left = sanitize_input(left, get_data_type(right))
right = sanitize_input(right, get_data_type(left))
return greater(left, right, name).output()
def greater(left, right, name=''):
'''
Elementwise 'greater' comparison of two tensors. Result is 1 if left > right else 0.
Example:
>>> C.eval(C.greater([41., 42., 43.], [42., 42., 42.]))
[array([[0., 0., 1.]])]
>>> C.eval(C.greater([-1,0,1], [0]))
[array([[1., 0., 1.]])]
Args:
left: left side tensor
right: right side tensor
name (str): the name of the node in the network
Returns:
:class:`cntk.Function`
'''
from cntk import greater
left = sanitize_input(left, get_data_type(right))
right = sanitize_input(right, get_data_type(left))
return greater(left, right, name).output()
def sample(self, n=1):
samples = C.random.uniform((n, 1))
indcies = C.argmax(C.greater(self.accum_prob - samples, 0), axis=1)
return C.squeeze(indcies)
def test_Greater(tmpdir):
model = C.greater([41., 42., 43.], [42., 42., 42.])
verify_no_input(model, tmpdir, 'Greater_0')
def test_Greater(tmpdir, dtype):
with C.default_options(dtype=dtype):
model = C.greater([41., 42., 43.], [42., 42., 42.])
verify_no_input(model, tmpdir, 'Greater_0')
def test_Greater(tmpdir, dtype):
with C.default_options(dtype = dtype):
model = C.greater([41., 42., 43.], [42., 42., 42.])
verify_no_input(model, tmpdir, 'Greater_0')
def validate_model(test_data, model, polymath):
begin_logits = model.outputs[0]
end_logits = model.outputs[1]
loss = model.outputs[2]
root = C.as_composite(loss.owner)
mb_source, input_map = create_mb_and_map(root,
test_data,
polymath,
randomize=False,
repeat=False)
begin_label = argument_by_name(root, 'ab')
end_label = argument_by_name(root, 'ae')
begin_prediction = C.sequence.input_variable(
1, sequence_axis=begin_label.dynamic_axes[1], needs_gradient=True)
end_prediction = C.sequence.input_variable(
1, sequence_axis=end_label.dynamic_axes[1], needs_gradient=True)
best_span_score = symbolic_best_span(begin_prediction, end_prediction)
predicted_span = C.layers.Recurrence(
C.plus)(begin_prediction - C.sequence.past_value(end_prediction))
true_span = C.layers.Recurrence(C.plus)(begin_label -
C.sequence.past_value(end_label))
common_span = C.element_min(predicted_span, true_span)
begin_match = C.sequence.reduce_sum(
C.element_min(begin_prediction, begin_label))
end_match = C.sequence.reduce_sum(C.element_min(end_prediction, end_label))
predicted_len = C.sequence.reduce_sum(predicted_span)
true_len = C.sequence.reduce_sum(true_span)
common_len = C.sequence.reduce_sum(common_span)
f1 = 2 * common_len / (predicted_len + true_len)
exact_match = C.element_min(begin_match, end_match)
precision = common_len / predicted_len
recall = common_len / true_len
overlap = C.greater(common_len, 0)
s = lambda x: C.reduce_sum(x, axis=C.Axis.all_axes())
stats = C.splice(s(f1), s(exact_match), s(precision), s(recall),
s(overlap), s(begin_match), s(end_match))
# Evaluation parameters
minibatch_size = 2048
num_sequences = 0
stat_sum = 0
loss_sum = 0
with tqdm(ncols=32) as progress_bar:
while True:
data = mb_source.next_minibatch(minibatch_size,
input_map=input_map)
if not data or not (begin_label in data
) or data[begin_label].num_sequences == 0:
break
out = model.eval(data,
outputs=[begin_logits, end_logits, loss],
as_numpy=False)
testloss = out[loss]
g = best_span_score.grad(
{
begin_prediction: out[begin_logits],
end_prediction: out[end_logits]
},
wrt=[begin_prediction, end_prediction],
as_numpy=False)
other_input_map = {
begin_prediction: g[begin_prediction],
end_prediction: g[end_prediction],
begin_label: data[begin_label],
end_label: data[end_label]
}
stat_sum += stats.eval((other_input_map))
loss_sum += np.sum(testloss.asarray())
num_sequences += data[begin_label].num_sequences
progress_bar.update(data[begin_label].num_sequences)
stat_avg = stat_sum / num_sequences
loss_avg = loss_sum / num_sequences
print(
"\nValidated {} sequences, loss {:.4f}, F1 {:.4f}, EM {:.4f}, precision {:4f}, recall {:4f} hasOverlap {:4f}, start_match {:4f}, end_match {:4f}"
.format(num_sequences, loss_avg, stat_avg[0], stat_avg[1], stat_avg[2],
stat_avg[3], stat_avg[4], stat_avg[5], stat_avg[6]))
return loss_avg
def validate_model(test_data, model, polymath):
begin_logits = model.outputs[0]
end_logits = model.outputs[1]
loss = model.outputs[2]
root = C.as_composite(loss.owner)
mb_source, input_map = create_mb_and_map(root, test_data, polymath, randomize=False, repeat=False)
begin_label = argument_by_name(root, 'ab')
end_label = argument_by_name(root, 'ae')
begin_prediction = C.sequence.input_variable(1, sequence_axis=begin_label.dynamic_axes[1], needs_gradient=True)
end_prediction = C.sequence.input_variable(1, sequence_axis=end_label.dynamic_axes[1], needs_gradient=True)
best_span_score = symbolic_best_span(begin_prediction, end_prediction)
predicted_span = C.layers.Recurrence(C.plus)(begin_prediction - C.sequence.past_value(end_prediction))
true_span = C.layers.Recurrence(C.plus)(begin_label - C.sequence.past_value(end_label))
common_span = C.element_min(predicted_span, true_span)
begin_match = C.sequence.reduce_sum(C.element_min(begin_prediction, begin_label))
end_match = C.sequence.reduce_sum(C.element_min(end_prediction, end_label))
predicted_len = C.sequence.reduce_sum(predicted_span)
true_len = C.sequence.reduce_sum(true_span)
common_len = C.sequence.reduce_sum(common_span)
f1 = 2*common_len/(predicted_len+true_len)
exact_match = C.element_min(begin_match, end_match)
precision = common_len/predicted_len
recall = common_len/true_len
overlap = C.greater(common_len, 0)
s = lambda x: C.reduce_sum(x, axis=C.Axis.all_axes())
stats = C.splice(s(f1), s(exact_match), s(precision), s(recall), s(overlap), s(begin_match), s(end_match))
# Evaluation parameters
minibatch_size = 20000
num_sequences = 0
stat_sum = 0
loss_sum = 0
while True:
data = mb_source.next_minibatch(minibatch_size, input_map=input_map)
if not data or not (begin_label in data) or data[begin_label].num_sequences == 0:
break
out = model.eval(data, outputs=[begin_logits,end_logits,loss], as_numpy=False)
testloss = out[loss]
g = best_span_score.grad({begin_prediction:out[begin_logits], end_prediction:out[end_logits]}, wrt=[begin_prediction,end_prediction], as_numpy=False)
other_input_map = {begin_prediction: g[begin_prediction], end_prediction: g[end_prediction], begin_label: data[begin_label], end_label: data[end_label]}
stat_sum += stats.eval((other_input_map))
loss_sum += np.sum(testloss.asarray())
num_sequences += data[begin_label].num_sequences
stat_avg = stat_sum / num_sequences
loss_avg = loss_sum / num_sequences
print("Validated {} sequences, loss {:.4f}, F1 {:.4f}, EM {:.4f}, precision {:4f}, recall {:4f} hasOverlap {:4f}, start_match {:4f}, end_match {:4f}".format(
num_sequences,
loss_avg,
stat_avg[0],
stat_avg[1],
stat_avg[2],
stat_avg[3],
stat_avg[4],
stat_avg[5],
stat_avg[6]))
return loss_avg
def word_level_drop(self, doc):
# doc [#, c][d]
seq_shape = C.sequence.is_first(doc)
u = C.random.uniform_like(seq_shape, seed=98052)
mask = C.element_select(C.greater(u, 0.08), 1.0, 0)
return doc * mask
def test_Greater(tmpdir):
model = C.greater([41., 42., 43.], [42., 42., 42.])
verify_no_input(model, tmpdir, 'Greater_0')
import cntk
A = [1, 3, 4]
B = [4, 3, 2]
print("less(A,B):")
less = cntk.less(A, B).eval()
print("{}\n".format(less))
print("equal(A,B):")
equal = cntk.equal(A, B).eval()
print("{}\n".format(equal))
print("greater(A,B)")
greater = cntk.greater(A, B).eval()
print("{}\n".format(greater))
print("greater_equal(A,B):")
greater_equal = cntk.greater_equal(A, B).eval()
print("{}\n".format(greater_equal))
print("not_equal(A,B):")
not_equal = cntk.not_equal(A, B).eval()
print("{}\n".format(not_equal))
print("less_equal(A,B):")
less_equal = cntk.less_equal(A, B).eval()
print("{}\n".format(less_equal))