def _vectorize(self, strings, method='first'):
method = method.lower()
if method not in ['first', 'last', 'mean', 'word']:
raise ValueError(
"method not supported, only support 'first', 'last', 'mean' and 'word'"
)
input_ids, input_masks, _, s_tokens = bert_tokenization(
self._tokenizer, strings)
v = self._sess.run(
self._vectorizer,
feed_dict={
self._X: input_ids,
self._input_masks: input_masks
},
)
if method == 'first':
v = v[:, 0]
elif method == 'last':
v = v[:, -1]
elif method == 'mean':
v = np.mean(v, axis=1)
else:
v = [
merge_sentencepiece_tokens(
list(zip(s_tokens[i], v[i][:len(s_tokens[i])])),
weighted=False,
vectorize=True,
) for i in range(len(v))
]
return v
def _translate(self, strings, beam_search=True):
input_ids, input_masks, input_segments, _ = bert_tokenization(
self._tokenizer, strings, cleaning=translation_textcleaning)
if beam_search:
output = self._beam
else:
output = self._greedy
p = sess.run(
output,
feed_dict={
model.X: batch_x,
model.input_masks: batch_mask,
model.segment_ids: batch_segment,
},
)
result = []
for output in p:
output = [i for i in output if i > 1]
output = self._tokenizer.convert_ids_to_tokens(output)
output = [(t, 1) for t in output]
output = merge_wordpiece_tokens(output)
output = [t[0] for t in output]
results.append(' '.join(output))
return result
def _vectorize(self, strings, method='first'):
method = method.lower()
if method not in ['first', 'last', 'mean', 'word']:
raise ValueError(
"method not supported, only support 'first', 'last', 'mean' and 'word'"
)
input_ids, input_masks, _, s_tokens = bert_tokenization(
self._tokenizer, strings)
r = self._execute(
inputs=[input_ids, input_masks],
input_labels=['Placeholder', 'Placeholder_1'],
output_labels=['vectorizer'],
)
v = r['vectorizer']
if method == 'first':
v = v[:, 0]
elif method == 'last':
v = v[:, -1]
elif method == 'mean':
v = np.mean(v, axis=1)
else:
v = [
merge_sentencepiece_tokens(
list(zip(s_tokens[i], v[i][:len(s_tokens[i])])),
weighted=False,
vectorize=True,
) for i in range(len(v))
]
return v
def vectorize(self, strings):
"""
Vectorize string inputs using bert attention.
Parameters
----------
strings : str / list of str
Returns
-------
array: vectorized strings
"""
if isinstance(strings, list):
if not isinstance(strings[0], str):
raise ValueError('input must be a list of strings or a string')
else:
if not isinstance(strings, str):
raise ValueError('input must be a list of strings or a string')
if isinstance(strings, str):
strings = [strings]
batch_x, _, _, _ = bert_tokenization(
self._tokenizer, strings, cls = self._cls, sep = self._sep
)
return self._sess.run(self.logits, feed_dict = {self.X: batch_x})
def vectorize(self, strings: List[str]):
"""
Vectorize string inputs using bert attention.
Parameters
----------
strings : List[str]
Returns
-------
result: np.array
"""
batch_x, batch_masks, batch_segments, _ = bert_tokenization(
self._tokenizer, strings
)
return self._sess.run(
self.logits,
feed_dict = {
self.X: batch_x,
self.MASK: batch_masks,
self.segment_ids: batch_segments,
},
)
def _attention(self, strings):
batch_x, _, _, s_tokens = bert_tokenization(
self._tokenizer, strings, cls = self._cls, sep = self._sep
)
maxlen = max([len(s) for s in s_tokens])
s_tokens = padding_sequence(s_tokens, maxlen, pad_int = self._sep)
attentions = self._sess.run(self.attns, feed_dict = {self.X: batch_x})
return attentions, s_tokens
def _classify(self, strings):
input_ids, input_masks, _, _ = bert_tokenization(
self._tokenizer, strings
)
return self._sess.run(
self._softmax,
feed_dict = {self._X: input_ids, self._input_masks: input_masks},
)
def _classify(self, strings):
input_ids, input_masks, _, _ = bert_tokenization(
self._tokenizer, strings)
r = self._execute(
inputs=[input_ids, input_masks],
input_labels=['Placeholder', 'Placeholder_1'],
output_labels=['logits'],
)
return softmax(r['logits'], axis=-1)
def _attention(self, strings):
batch_x, batch_masks, _, s_tokens = bert_tokenization(
self._tokenizer, strings
)
maxlen = max([len(s) for s in s_tokens])
s_tokens = padding_sequence(s_tokens, maxlen, pad_int = '[SEP]')
attentions = self._sess.run(
self.attns, feed_dict = {self.X: batch_x, self.MASK: batch_masks}
)
return attentions, s_tokens, batch_masks
def _classify(self, strings):
input_ids, _, _, _ = bert_tokenization(self._tokenizer, strings)
input_ids = tf.keras.preprocessing.sequence.pad_sequences(
input_ids, padding='post', maxlen=self._maxlen)
r = self._execute(
inputs=[input_ids],
input_labels=['Placeholder'],
output_labels=['logits'],
)
return softmax(r['logits'], axis=-1)
def _base(self, strings_left, strings_right):
input_ids_left, input_masks_left, _, _ = bert_tokenization(
self._tokenizer, strings_left)
input_ids_right, input_masks_right, _, _ = bert_tokenization(
self._tokenizer, strings_right)
r = self._execute(
inputs=[
input_ids_left,
input_masks_left,
input_ids_right,
input_masks_right,
],
input_labels=[
'Placeholder',
'Placeholder_1',
'Placeholder_2',
'Placeholder_3',
],
output_labels=['logits'],
)
return softmax(r['logits'], axis=-1)
def _classify(self, strings):
input_ids, input_masks, _, _ = bert_tokenization(
self._tokenizer, strings, socialmedia=self._socialmedia
)
r = self._execute(
inputs=[input_ids, input_masks],
input_labels=['Placeholder'],
output_labels=['logits'],
)
if self._multilabels:
return sigmoid(r['logits'])
else:
return softmax(r['logits'], axis=-1)
def vectorize(self, strings: List[str], method: str = 'first'):
"""
vectorize list of strings.
Parameters
----------
strings: List[str]
method : str, optional (default='first')
Vectorization layer supported. Allowed values:
* ``'last'`` - vector from last sequence.
* ``'first'`` - vector from first sequence.
* ``'mean'`` - average vectors from all sequences.
* ``'word'`` - average vectors based on tokens.
Returns
-------
result: np.array
"""
method = method.lower()
if method not in ['first', 'last', 'mean', 'word']:
raise ValueError(
"method not supported, only support 'first', 'last', 'mean' and 'word'"
)
input_ids, input_masks, _, s_tokens = bert_tokenization(
self._tokenizer, strings
)
r = self._execute(
inputs=[input_ids, input_masks],
input_labels=['Placeholder', 'Placeholder_1'],
output_labels=['vectorizer'],
)
v = r['vectorizer']
if method == 'first':
v = v[:, 0]
elif method == 'last':
v = v[:, -1]
elif method == 'mean':
v = np.mean(v, axis=1)
else:
v = [
merge_sentencepiece_tokens(
list(zip(s_tokens[i], v[i][: len(s_tokens[i])])),
weighted=False,
vectorize=True,
)
for i in range(len(v))
]
return v
def _paraphrase(self, strings):
batch_x, input_masks, input_segments, _ = bert_tokenization(
self._tokenizer, strings)
outputs = self._sess.run(
self._logits,
feed_dict={
self._X: batch_x,
self._segment_ids: input_segments,
self._input_masks: input_masks,
},
)[:, 0, :].tolist()
results = []
for output in outputs:
output = [i for i in output if i > 0]
output = self._tokenizer.convert_ids_to_tokens(output)
output = [(t, 1) for t in output]
output = merge_sentencepiece_tokens(output)
output = [t[0] for t in output]
results.append(' '.join(output))
return results
def vectorize(self, strings: List[str]):
"""
Vectorize list of strings.
Parameters
----------
strings : List[str]
Returns
-------
result: np.array
"""
input_ids, input_masks, _, _ = bert_tokenization(
self._tokenizer, strings)
r = self._execute(
inputs=[input_ids, input_masks],
input_labels=['Placeholder', 'Placeholder_1'],
output_labels=['bert/summary'],
)
return r['bert/summary']
def vectorize(self, strings: List[str]):
"""
Vectorize list of strings.
Parameters
----------
strings : List[str]
Returns
-------
result: np.array
"""
input_ids, input_masks, segment_ids, _ = bert_tokenization(
self._tokenizer, strings)
segment_ids = np.array(segment_ids) + 1
return self._sess.run(
self._vectorizer,
feed_dict={
self._X: input_ids,
self._segment_ids: segment_ids,
self._input_masks: input_masks,
},
)
def _predict_words(self, string, method, visualization, add_neutral=False):
method = method.lower()
if method not in ['last', 'first', 'mean']:
raise ValueError(
"method not supported, only support 'last', 'first' and 'mean'"
)
if add_neutral:
label = self._label + ['neutral']
else:
label = self._label
batch_x, batch_mask, _, s_tokens = bert_tokenization(
self._tokenizer, [string])
result, attentions, words = self._sess.run(
[self._softmax, self._attns, self._softmax_seq],
feed_dict={
self._X: batch_x,
self._input_masks: batch_mask
},
)
if method == 'first':
cls_attn = list(attentions[0].values())[0][:, :, 0, :]
if method == 'last':
cls_attn = list(attentions[-1].values())[0][:, :, 0, :]
if method == 'mean':
combined_attentions = []
for a in attentions:
combined_attentions.append(list(a.values())[0])
cls_attn = np.mean(combined_attentions, axis=0).mean(axis=2)
cls_attn = np.mean(cls_attn, axis=1)
total_weights = np.sum(cls_attn, axis=-1, keepdims=True)
attn = cls_attn / total_weights
words = words[0]
if add_neutral:
result = neutral(result)
words = neutral(words)
result = result[0]
weights = []
merged = merge_sentencepiece_tokens(list(zip(s_tokens[0], attn[0])))
for i in range(words.shape[1]):
m = merge_sentencepiece_tokens(list(zip(s_tokens[0], words[:, i])),
weighted=False)
_, weight = zip(*m)
weights.append(weight)
w, a = zip(*merged)
words = np.array(weights).T
distribution_words = words[:, np.argmax(words.sum(axis=0))]
y_histogram, x_histogram = np.histogram(distribution_words,
bins=np.arange(0, 1, 0.05))
y_histogram = y_histogram / y_histogram.sum()
x_attention = np.arange(len(w))
left, right = np.unique(np.argmax(words, axis=1), return_counts=True)
left = left.tolist()
y_barplot = []
for i in range(len(label)):
if i not in left:
y_barplot.append(i)
else:
y_barplot.append(right[left.index(i)])
dict_result = {label[i]: result[i] for i in range(len(result))}
dict_result['alphas'] = {w: a[no] for no, w in enumerate(w)}
dict_result['word'] = {w: words[no] for no, w in enumerate(w)}
dict_result['histogram'] = {'x': x_histogram, 'y': y_histogram}
dict_result['attention'] = {'x': x_attention, 'y': np.array(a)}
dict_result['barplot'] = {'x': label, 'y': y_barplot}
dict_result['class_name'] = self._class_name
if visualization:
render_dict[self._class_name](dict_result)
else:
return dict_result
def predict_words(self,
string: str,
method: str = 'last',
visualization: bool = True):
"""
classify words.
Parameters
----------
string : str
method : str, optional (default='last')
Attention layer supported. Allowed values:
* ``'last'`` - attention from last layer.
* ``'first'`` - attention from first layer.
* ``'mean'`` - average attentions from all layers.
visualization: bool, optional (default=True)
If True, it will open the visualization dashboard.
Returns
-------
dictionary: results
"""
method = method.lower()
if method not in ['last', 'first', 'mean']:
raise ValueError(
"method not supported, only support 'last', 'first' and 'mean'"
)
batch_x, input_masks, _, s_tokens = bert_tokenization(
self._tokenizer, [string])
result, attentions, words = self._sess.run(
[self._sigmoid, self._attns, self._sigmoid_seq],
feed_dict={
self._X: batch_x,
self._input_masks: input_masks
},
)
if method == 'first':
cls_attn = list(attentions[0].values())[0][:, :, 0, :]
if method == 'last':
cls_attn = list(attentions[-1].values())[0][:, :, 0, :]
if method == 'mean':
combined_attentions = []
for a in attentions:
combined_attentions.append(list(a.values())[0])
cls_attn = np.mean(combined_attentions, axis=0).mean(axis=2)
cls_attn = np.mean(cls_attn, axis=1)
total_weights = np.sum(cls_attn, axis=-1, keepdims=True)
attn = cls_attn / total_weights
result = result[0]
words = words[0]
weights = []
merged = merge_sentencepiece_tokens(list(zip(s_tokens[0], attn[0])))
for i in range(words.shape[1]):
m = merge_sentencepiece_tokens(list(zip(s_tokens[0], words[:, i])),
weighted=False)
_, weight = zip(*m)
weights.append(weight)
w, a = zip(*merged)
words = np.array(weights).T
distribution_words = words[:, np.argmax(words.sum(axis=0))]
y_histogram, x_histogram = np.histogram(distribution_words,
bins=np.arange(0, 1, 0.05))
y_histogram = y_histogram / y_histogram.sum()
x_attention = np.arange(len(w))
left, right = np.unique(np.argmax(words, axis=1), return_counts=True)
left = left.tolist()
y_barplot = []
for i in range(len(self._label)):
if i not in left:
y_barplot.append(i)
else:
y_barplot.append(right[left.index(i)])
dict_result = {self._label[i]: result[i] for i in range(len(result))}
dict_result['alphas'] = {w: a[no] for no, w in enumerate(w)}
dict_result['word'] = {w: words[no] for no, w in enumerate(w)}
dict_result['histogram'] = {'x': x_histogram, 'y': y_histogram}
dict_result['attention'] = {'x': x_attention, 'y': np.array(a)}
dict_result['barplot'] = {'x': self._label, 'y': y_barplot}
dict_result['class_name'] = self._class_name
if visualization:
_render_toxic(dict_result)
else:
return dict_result
def _predict_words(
self,
string,
method,
visualization,
add_neutral=False,
bins_size=0.05,
**kwargs,
):
method = method.lower()
if method not in ['last', 'first', 'mean']:
raise ValueError(
"method not supported, only support 'last', 'first' and 'mean'"
)
if add_neutral and not self._multilabels:
label = self._label + ['neutral']
else:
label = self._label
input_ids, input_masks, _, s_tokens = bert_tokenization(
self._tokenizer, [string]
)
r = self._execute(
inputs=[input_ids, input_masks],
input_labels=['Placeholder', 'Placeholder_1'],
output_labels=['logits', 'attention', 'logits_seq'],
)
if self._multilabels:
result = sigmoid(r['logits'])
words = sigmoid(r['logits_seq'])
else:
result = softmax(r['logits'], axis=-1)
words = softmax(r['logits_seq'], axis=-1)
attentions = r['attention']
if method == 'first':
cls_attn = list(attentions[0].values())[0][:, :, 0, :]
if method == 'last':
cls_attn = list(attentions[-1].values())[0][:, :, 0, :]
if method == 'mean':
combined_attentions = []
for a in attentions:
combined_attentions.append(list(a.values())[0])
cls_attn = np.mean(combined_attentions, axis=0).mean(axis=2)
cls_attn = np.mean(cls_attn, axis=1)
total_weights = np.sum(cls_attn, axis=-1, keepdims=True)
attn = cls_attn / total_weights
words = words[0]
if add_neutral and not self._multilabels:
result = neutral(result)
words = neutral(words)
result = result[0]
weights = []
merged = merge_sentencepiece_tokens(list(zip(s_tokens[0], attn[0])))
for i in range(words.shape[1]):
m = merge_sentencepiece_tokens(
list(zip(s_tokens[0], words[:, i])), weighted=False
)
_, weight = zip(*m)
weights.append(weight)
w, a = zip(*merged)
words = np.array(weights).T
distribution_words = words[:, np.argmax(words.sum(axis=0))]
y_histogram, x_histogram = np.histogram(
distribution_words, bins=np.arange(0, 1 + bins_size, bins_size)
)
y_histogram = y_histogram / y_histogram.sum()
x_attention = np.arange(len(w))
left, right = np.unique(
np.argmax(words, axis=1), return_counts=True
)
left = left.tolist()
y_barplot = []
for i in range(len(label)):
if i not in left:
y_barplot.append(i)
else:
y_barplot.append(right[left.index(i)])
dict_result = {label[i]: result[i] for i in range(len(result))}
dict_result['alphas'] = {w: a[no] for no, w in enumerate(w)}
dict_result['word'] = {w: words[no] for no, w in enumerate(w)}
dict_result['histogram'] = {'x': x_histogram, 'y': y_histogram}
dict_result['attention'] = {'x': x_attention, 'y': np.array(a)}
dict_result['barplot'] = {'x': label, 'y': y_barplot}
dict_result['module'] = self._module
if visualization:
render_dict[self._module](dict_result, **kwargs)
else:
return dict_result