def reward_progressive_match_label(sample, batch, sample_score=None):
seq_len = util.find_first_min_zero(sample)
mask, __ = util.masked_full_like(sample,
1,
num_non_padding=seq_len + 1,
dtype=np.int32)
mask = mask * (seq_len > 0)
sample, _sample = sample * mask, sample
label = _label = batch.labels.label
pad_width = abs(len(sample) - len(label))
pad_width = ((0, pad_width), (0, 0))
if len(label) < len(sample):
label = np.pad(label, pad_width, 'constant', constant_values=0)
elif len(sample) < len(label):
sample = np.pad(sample, pad_width, 'constant', constant_values=0)
diff = np.abs(sample - label)
match = (diff == 0).astype(np.float32) # / batch.features.dec_seq_len
if len(_label) < len(_sample):
match[len(_label) - 1:, :] = 0
elif len(_sample) < len(_label):
match = match[:len(_sample), :]
avgmatch = np.sum(match * mask) / np.sum(mask)
summatch = np.sum(match, axis=0)
mismatch = np.argmin(match, axis=0)
mismatch_mask, __ = util.masked_full_like(match,
1,
num_non_padding=mismatch)
match = match * mismatch_mask
for ib in range(sample.shape[1]):
if summatch[ib] > 0 and mismatch[ib] == 0:
continue
match[mismatch[ib], ib] = -0.1
return match, avgmatch
def reward_match_label(sample, batch, partial_match=True, sample_score=None):
seq_len = util.find_first_min_zero(sample)
mask, __ = util.masked_full_like(sample,
1,
num_non_padding=seq_len + 1,
dtype=np.int32)
mask = mask * (seq_len > 0)
sample, _sample = sample * mask, sample
label = _label = batch.labels.label
pad_width = abs(len(sample) - len(label))
pad_width = ((0, pad_width), (0, 0))
if len(label) < len(sample):
label = np.pad(label, pad_width, 'constant', constant_values=0)
elif len(sample) < len(label):
sample = np.pad(sample, pad_width, 'constant', constant_values=0)
diff = np.abs(sample - label)
if partial_match:
match = (diff == 0).astype(np.float32) / (seq_len + 1)
if len(_label) < len(_sample):
match[len(_label) - 1:, :] = 0
elif len(_sample) < len(_label):
match = match[:len(_sample), :]
else:
sumdiff = np.sum(diff, axis=0)
match = np.zeros_like(sample, dtype=np.float32)
for ib in range(seq_len.shape[0]):
if sumdiff[ib] == 0:
match[seq_len[ib], ib] = 1
match = match * mask
avgmatch = np.sum(match) / np.sum(seq_len > 0)
return match, avgmatch
def reward_constant(sample, batch, constant=-0.1, sample_score=None):
# return batch.labels.label_weight, np.mean(batch.labels.label_weight)
seq_len = util.find_first_min_zero(sample)
mask, __ = util.masked_full_like(sample,
1,
num_non_padding=seq_len + 1,
dtype=np.int32)
mask = mask * (seq_len > 0) * constant
return mask * constant, np.mean(seq_len + 1) * constant
def reward_global_ngram_stat(sample,
batch,
global_count,
current_count,
update_fn,
ngram_fn,
sample_score=None):
seq_len = util.find_first_min_zero(sample)
scores = np.zeros_like(sample, dtype=np.float32)
batch_ngrams = []
for ib in range(sample.shape[1]):
seq = sample[:seq_len[ib], ib]
ngrams = tuple(ngram_fn(seq))
for it, ngram in enumerate(ngrams):
scores[it, ib] = -np.log(
(current_count[ngram] + 1) / (global_count[ngram] + 1))
batch_ngrams.append(ngrams)
update_fn(batch, batch_ngrams)
return scores, np.sum(scores) / np.sum(seq_len + 1)
def reward_bleu(sample,
batch,
ref_fn,
reward_incomplete=False,
sample_score=None):
seq_len = util.find_first_min_zero(sample) + 1
scores = np.zeros_like(sample, dtype=np.float32)
b_refs = ref_fn(batch)
c = 0
for ib in range(len(seq_len)):
step = seq_len[ib]
refs = b_refs[ib]
completed = step <= sample.shape[0]
if refs is not None:
c += 1
if refs is not None and (completed or reward_incomplete):
if not completed:
step = sample.shape[0]
scores[step - 1, ib] = util.sentence_bleu(refs, sample[:step, ib])
return scores, np.sum(scores) / c