def test_get_batch_data(self):
data = np.array([[12, 10, 10, 0], [12, 8, 7, 0], [13, 4, 11, 0],
[8, 13, 9, 0], [7, 5, 11, 0], [0, 12, 12, 0],
[0, 10, 0, 0], [0, 0, 0, 0]])
seq_len = np.array([6, 8, 7, 0])
features = dstruct.Seq2SeqFeatureTuple(*(data, seq_len, None, None))
labels = dstruct.SeqLabelTuple(*(None, None, None))
batch = dstruct.BatchTuple(features, labels, None, False)
new_batch = generator.get_batch_data(
batch,
data,
start_id=1,
seq_len_idx=1,
input_key='dec_inputs',
seq_len_key='dec_seq_len',
unmasked_token_weight=np.ones_like(data) * 2)
_f, _l, _n, _k = new_batch
self.assertIs(data, _f.enc_inputs, 'enc data is the same object.')
self.assertIs(seq_len, _f.enc_seq_len,
'enc seq len is the same object.')
np.testing.assert_array_equal(data[:-1, :],
_f.dec_inputs[1:, :],
err_msg='dec input is shifted data.')
np.testing.assert_array_equal(data,
_l.label,
err_msg='dec output is data.')
np.testing.assert_array_equal(seq_len,
_f.dec_seq_len,
err_msg='dec seq len is correct.')
self.assertEqual(sum(seq_len), _n, 'num tokens is correct.')
w = np.array([[1, 1, 1, 0], [1, 1, 1, 0], [1, 1, 1, 0], [1, 1, 1, 0],
[1, 1, 1, 0], [1, 1, 1, 0], [0, 1, 1, 0], [0, 1, 0, 0]])
np.testing.assert_array_equal(w * 2,
new_batch.labels.label_weight,
err_msg='token weight')
def get_batch_data(batch,
y_arr,
unmasked_token_weight=None,
unmasked_seq_weight=None,
start_id=1,
seq_len_idx=1,
input_key='inputs',
seq_len_key='seq_len'):
y_len = np.argmin(y_arr, axis=0) + 1
y_len[batch.features[seq_len_idx] <= 0] = 0
seq_weight = np.where(y_len > 0, 1, 0).astype(np.float32)
if unmasked_seq_weight is not None:
seq_weight *= unmasked_seq_weight
token_weight, num_tokens = util.masked_full_like(y_arr, 1, y_len)
if unmasked_token_weight is not None:
token_weight *= unmasked_token_weight
start = np.full((1, len(y_len)), start_id, dtype=np.int32) * seq_weight
x_arr = np.vstack((start.astype(np.int32), y_arr))[:-1, :]
features = batch.features._replace(**{
input_key: x_arr,
seq_len_key: y_len
})
labels = ds.SeqLabelTuple(y_arr, token_weight, seq_weight)
batch = ds.BatchTuple(features, labels, num_tokens, batch.keep_state)
return batch
def test_reward_match_label(self):
data = np.array([[12, 10, 10, 0], [12, 8, 7, 0], [13, 4, 11, 0],
[8, 13, 9, 0], [7, 5, 11, 0], [0, 12, 12, 0],
[0, 10, 0, 0], [0, 0, 0, 0]])
features = dstruct.Seq2SeqFeatureTuple(*(None, None, None, None))
labels = dstruct.SeqLabelTuple(*(data, None, None))
batch = dstruct.BatchTuple(features, labels, None, False)
sample = np.array([[12, 10, 10, 0], [12, 8, 7, 0], [11, 4, 11, 0],
[8, 13, 9, 0], [7, 5, 11, 0], [0, 12, 12, 0],
[0, 1, 0, 0], [0, 1, 0, 0], [0, 0, 0, 0]])
exact_match = np.array([[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 1, 0], [0, 0, 0, 0], [0, 0, 0, 0]])
parti_match = np.array([[1, 1, 1, 0], [1, 1, 1, 0], [0, 1, 1, 0],
[1, 1, 1, 0], [1, 1, 1, 0], [1, 1, 1, 0],
[0, 0, 1, 0], [0, 0, 0, 0], [0, 0, 0, 0]])
sample_len = np.array([6., 9., 7., 0.])
_sample_len = np.array([6., 9., 7., 1.]) # for division
m, avg = generator.reward_match_label(sample,
batch,
partial_match=False)
np.testing.assert_array_equal(m, exact_match,
'label exact match reward')
self.assertEqual(avg, 1 / 3, 'average correct')
m, avg = generator.reward_match_label(sample,
batch,
partial_match=True)
np.testing.assert_array_equal(m, parti_match / _sample_len,
'label match reward')
self.assertEqual(
avg,
np.sum(parti_match / _sample_len) / np.sum(sample_len > 0),
'average correct')
def lseq2seq_batch_iter(enc_data,
dec_data,
label_data,
mask_data,
batch_size=1,
shuffle=True):
"""same as seq2seq_batch_iter, just add label"""
data_tuple = (enc_data, dec_data, label_data, mask_data)
for x, y, L, M in batch_iter(batch_size,
shuffle,
*data_tuple,
pad=[[], [], 0, 2]):
enc, enc_len = util.hstack_list(x)
dec, dec_len = util.hstack_list(y)
label = np.array(L, dtype=np.int32)
mask = np.array(M, dtype=np.int32)
in_dec = dec[:-1, :]
out_dec = dec[1:, :]
seq_weight = np.where(dec_len > 0, 1, 0)
dec_len -= seq_weight
token_weight, num_tokens = util.masked_full_like(
out_dec, 1, num_non_padding=dec_len)
seq_weight = seq_weight.astype(np.float32)
features = ds.LSeq2SeqFeatureTuple(enc, enc_len, in_dec, dec_len,
label, mask)
labels = ds.SeqLabelTuple(out_dec, token_weight, seq_weight)
yield ds.BatchTuple(features, labels, num_tokens, False)
def seq_batch_iter(in_data,
out_data,
weights,
batch_size=1,
shuffle=True,
keep_sentence=True):
"""wrapper of batch_iter to format seq data"""
keep_state = not keep_sentence
# add one more argumennt and pass it to "batch_iter" below
# also add 0 for the padding
if weights:
# import pdb; pdb.set_trace()
for x, y, w in batch_iter(batch_size,
shuffle,
in_data,
out_data,
weights,
pad=[[], [], 0]):
x_arr, x_len = util.hstack_list(x)
y_arr, y_len = util.hstack_list(y)
# w_arr, w_len = util.hstack_list(w)
# change seq_weight to be the input weight
seq_weight = np.where(y_len > 0, w, 0).astype(np.float32)
# import pdb; pdb.set_trace()
token_weight, num_tokens = util.masked_full_like(
y_arr, w, num_non_padding=y_len)
features = ds.SeqFeatureTuple(x_arr, x_len)
labels = ds.SeqLabelTuple(y_arr, token_weight, seq_weight)
yield ds.BatchTuple(features, labels, num_tokens, keep_state)
else:
for x, y in batch_iter(batch_size,
shuffle,
in_data,
out_data,
pad=[[], []]):
x_arr, x_len = util.hstack_list(x)
y_arr, y_len = util.hstack_list(y)
seq_weight = np.where(y_len > 0, 1, 0).astype(np.float32)
token_weight, num_tokens = util.masked_full_like(
y_arr, 1, num_non_padding=y_len)
features = ds.SeqFeatureTuple(x_arr, x_len)
labels = ds.SeqLabelTuple(y_arr, token_weight, seq_weight)
yield ds.BatchTuple(features, labels, num_tokens, keep_state)
def concat_seq_batch(batch1, batch2):
_f1, _l1, _n1, _k1 = batch1
_f2, _l2, _n2, _k2 = batch2
inputs = util.hstack_with_padding(_f1.inputs, _f2.inputs)
seq_len = np.concatenate((_f1.seq_len, _f2.seq_len))
f = ds.SeqFeatureTuple(inputs, seq_len)
label = util.hstack_with_padding(_l1.label, _l2.label)
label_weight = util.hstack_with_padding(_l1.label_weight, _l2.label_weight)
_l2.seq_weight[:] = 0
seq_weight = np.concatenate((_l1.seq_weight, _l2.seq_weight))
l = ds.SeqLabelTuple(label, label_weight, seq_weight)
return ds.BatchTuple(f, l, _n1 + _n2, _k1)
def _format_word2def(x, w, c, y, sw):
enc, enc_len = util.hstack_list(x)
dec, dec_len = util.hstack_list(y)
word = np.array(w, dtype=np.int32)
char, char_len = util.vstack_list(c)
in_dec = dec[:-1, :]
out_dec = dec[1:, :]
seq_weight = np.array(sw, dtype=np.float32)
dec_len -= np.where(dec_len > 0, 1, 0)
token_weight, num_tokens = util.masked_full_like(out_dec,
1,
num_non_padding=dec_len)
seq_weight = seq_weight.astype(np.float32)
features = ds.Word2DefFeatureTuple(enc, enc_len, word, char, char_len,
in_dec, dec_len)
labels = ds.SeqLabelTuple(out_dec, token_weight, seq_weight)
return ds.BatchTuple(features, labels, num_tokens, False)
def concat_word2def_batch(batch1, batch2):
_f1, _l1, _n1, _k1 = batch1
_f2, _l2, _n2, _k2 = batch2
enc_inputs = util.hstack_with_padding(_f1.enc_inputs, _f2.enc_inputs)
enc_seq_len = np.concatenate((_f1.enc_seq_len, _f2.enc_seq_len))
words = np.concatenate((_f1.words, _f2.words))
chars = util.vstack_with_padding(_f1.chars, _f2.chars)
char_len = np.concatenate((_f1.char_len, _f2.char_len))
dec_inputs = util.hstack_with_padding(_f1.dec_inputs, _f2.dec_inputs)
dec_seq_len = np.concatenate((_f1.dec_seq_len, _f2.dec_seq_len))
f = ds.Word2DefFeatureTuple(enc_inputs, enc_seq_len, words, chars,
char_len, dec_inputs, dec_seq_len)
label = util.hstack_with_padding(_l1.label, _l2.label)
label_weight = util.hstack_with_padding(_l1.label_weight, _l2.label_weight)
seq_weight = np.concatenate((_l1.seq_weight, _l2.seq_weight))
l = ds.SeqLabelTuple(label, label_weight, seq_weight)
return ds.BatchTuple(f, l, _n1 + _n2, False)
def seq2seq_batch_iter(enc_data, dec_data, batch_size=1, shuffle=True):
"""wrapper of batch_iter to format seq2seq data"""
for x, y in batch_iter(batch_size,
shuffle,
enc_data,
dec_data,
pad=[[], []]):
enc, enc_len = util.hstack_list(x)
dec, dec_len = util.hstack_list(y)
in_dec = dec[:-1, :]
out_dec = dec[1:, :]
seq_weight = np.where(dec_len > 0, 1, 0)
dec_len -= seq_weight
token_weight, num_tokens = util.masked_full_like(
out_dec, 1, num_non_padding=dec_len)
seq_weight = seq_weight.astype(np.float32)
features = ds.Seq2SeqFeatureTuple(enc, enc_len, in_dec, dec_len)
labels = ds.SeqLabelTuple(out_dec, token_weight, seq_weight)
yield ds.BatchTuple(features, labels, num_tokens, False)
def seq_batch_iter(in_data,
out_data,
batch_size=1,
shuffle=True,
keep_sentence=True):
"""wrapper of batch_iter to format seq data"""
keep_state = not keep_sentence
for x, y in batch_iter(batch_size,
shuffle,
in_data,
out_data,
pad=[[], []]):
x_arr, x_len = util.hstack_list(x)
y_arr, y_len = util.hstack_list(y)
seq_weight = np.where(y_len > 0, 1, 0).astype(np.float32)
token_weight, num_tokens = util.masked_full_like(y_arr,
1,
num_non_padding=y_len)
features = ds.SeqFeatureTuple(x_arr, x_len)
labels = ds.SeqLabelTuple(y_arr, token_weight, seq_weight)
yield ds.BatchTuple(features, labels, num_tokens, keep_state)
def fill_data(pos):
return ds.BatchTuple(pos, labels, batch_size, keep_state)
def fill_data(pos):
return ds.BatchTuple(pos, labels, np.sum(num_tokens[list(pos)]),
keep_state)
