def predict(self,
string: str,
get_proba: bool = False,
add_neutral: bool = True):
"""
Classify a string.
Parameters
----------
string : str
get_proba: bool, optional (default=False)
If True, it will return probability of classes.
add_neutral: bool, optional (default=True)
if True, it will add neutral probability.
Returns
-------
string: result
"""
if add_neutral:
label = self._label + ['neutral']
else:
label = self._label
vectors = self._vectorize.transform([self._cleaning(string)])
result = self._multinomial.predict_proba(vectors)
if add_neutral:
result = neutral(result)
result = result[0]
if get_proba:
return {label[i]: result[i] for i in range(len(result))}
else:
return label[np.argmax(result)]
def _predict(self, strings, add_neutral=False):
results = self._classify(strings)
if add_neutral:
result = neutral(results)
label = self._label + ['neutral']
else:
label = self._label
return [label[result] for result in np.argmax(results, axis=1)]
def _predict_proba(self, strings, add_neutral=False):
results = self._classify(strings)
if add_neutral:
results = neutral(results)
label = self._label + ['neutral']
else:
label = self._label
outputs = []
for result in results:
outputs.append({label[i]: result[i] for i in range(len(result))})
return outputs
def _predict(self, strings, add_neutral):
input_ids, input_masks, segment_ids, _ = xlnet_tokenization(
self._tokenizer, strings)
result = self._sess.run(
self._softmax,
feed_dict={
self._X: input_ids,
self._segment_ids: segment_ids,
self._input_masks: input_masks,
},
)
if add_neutral:
result = neutral(result)
return result
def predict_batch(
self,
strings: List[str],
get_proba: bool = False,
add_neutral: bool = True,
):
"""
Classify a list of strings.
Parameters
----------
strings: List[str]
get_proba: bool, optional (default=False)
If True, it will return probability of classes.
add_neutral: bool, optional (default=True)
if True, it will add neutral probability.
Returns
-------
string: list of results
"""
if add_neutral:
label = self._label + ['neutral']
else:
label = self._label
strings = [self._cleaning(string) for string in strings]
vectors = self._vectorize.transform(strings)
results = self._multinomial.predict_proba(vectors)
if add_neutral:
results = neutral(results)
if get_proba:
outputs = []
for result in results:
outputs.append(
{label[i]: result[i]
for i in range(len(result))})
return outputs
else:
return [label[result] for result in np.argmax(results, axis=1)]
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, 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
input_ids, input_masks, segment_ids, s_tokens = xlnet_tokenization(
self._tokenizer, [string])
r = self._execute(
inputs=[input_ids, segment_ids, input_masks],
input_labels=['Placeholder', 'Placeholder_1', 'Placeholder_2'],
output_labels=['logits', 'attention', 'logits_seq'],
)
result = softmax(r['logits'], axis=-1)
words = softmax(r['logits_seq'], axis=-1)
attentions = r['attention']
if method == 'first':
cls_attn = attentions[0][:, :, 0, :]
if method == 'last':
cls_attn = attentions[-1][:, :, 0, :]
if method == 'mean':
cls_attn = np.mean(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])),
model='xlnet')
for i in range(words.shape[1]):
m = merge_sentencepiece_tokens(
list(zip(s_tokens[0], words[:, i])),
weighted=False,
model='xlnet',
)
_, 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 Exception(
"method not supported, only support 'last', 'first' and 'mean'"
)
label = self._label + ['neutral']
batch_x, input_masks, segment_ids, s_tokens = xlnet_tokenization(
self._tokenizer, [string])
result, attentions, words = self._sess.run(
[self._softmax, self._attns, self._softmax_seq],
feed_dict={
self._X: batch_x,
self._segment_ids: segment_ids,
self._input_masks: input_masks,
},
)
if method == 'first':
cls_attn = attentions[0][:, :, 0, :]
if method == 'last':
cls_attn = attentions[-1][:, :, 0, :]
if method == 'mean':
cls_attn = np.mean(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 = neutral(result)
result = result[0]
words = neutral(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(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_binary(dict_result)
else:
return dict_result
