def RunModel(saved_model_dir, signature_def_key, tag, text, ngrams_list=None):
saved_model = reader.read_saved_model(saved_model_dir)
meta_graph = None
for meta_graph_def in saved_model.meta_graphs:
if tag in meta_graph_def.meta_info_def.tags:
meta_graph = meta_graph_def
break
if meta_graph_def is None:
raise ValueError("Cannot find saved_model with tag" + tag)
signature_def = signature_def_utils.get_signature_def_by_key(
meta_graph, signature_def_key)
text = text_utils.TokenizeText(text)
ngrams = None
if ngrams_list is not None:
ngrams_list = text_utils.ParseNgramsOpts(ngrams_list)
ngrams = text_utils.GenerateNgrams(text, ngrams_list)
example = inputs.BuildTextExample(text, ngrams=ngrams)
example = example.SerializeToString()
inputs_feed_dict = {
signature_def.inputs["inputs"].name: [example],
}
if signature_def_key == "proba":
output_key = "scores"
elif signature_def_key == "embedding":
output_key = "outputs"
else:
raise ValueError("Unrecognised signature_def %s" % (signature_def_key))
output_tensor = signature_def.outputs[output_key].name
with tf.Session() as sess:
loader.load(sess, [tag], saved_model_dir)
outputs = sess.run(output_tensor, feed_dict=inputs_feed_dict)
return outputs
def Request(text):
example = inputs.BuildTextExample(text_utils.TokenizeText(text))
request = classification_pb2.ClassificationRequest()
request.model_spec.name = 'default'
request.model_spec.signature_name = FLAGS.signature_def
request.input.example_list.examples.extend([example])
return request
def ParseText_and_label_Input(textfile, labelsmap, stopwords_file, ngrams):
"""从两个文件中解析数据，一个是训练集。一行是一条数据，用"""
idx_to_textlabel = {}
with open(labelsmap, encoding='utf-8') as f:
lines = f.readlines()
for idx, line in enumerate(lines):
idx_to_textlabel[idx] = line.strip()
examples = []
with open(stopwords_file, encoding="utf-8") as sf:
lines = sf.readlines()
stop_words = set([line.strip() for line in lines])
stop_words.add(' ')
stop_words.add('\n')
stop_words.add('')
with open(textfile, encoding='utf-8') as f1:
for record in f1:
text, label = record.strip().split(' ')
text = text.strip()
label = idx_to_textlabel[int(label.strip())]
words = text_utils.TokenizeText(text, stop_words)
examples.append({
"text": words,
"label": label,
})
if ngrams:
examples[-1]["ngrams"] = text_utils.GenerateNgrams(
words, ngrams)
return examples
def Request(text, ngrams):
text = text_utils.TokenizeText(text)
ngrams = None
if ngrams is not None:
ngrams_list = text_utils.ParseNgramsOpts(ngrams)
ngrams = text_utils.GenerateNgrams(text, ngrams_list)
example = inputs.BuildTextExample(text, ngrams=ngrams)
request = classification_pb2.ClassificationRequest()
request.model_spec.name = 'default'
request.model_spec.signature_name = 'proba'
request.input.example_list.examples.extend([example])
return request
def ParseTextInput(textfile, labelsfile):
"""Parse input from two text files: text and labels.
labels are specified 0-offset one per line.
"""
examples = []
with open(textfile) as f1, open(labelsfile) as f2:
for text, label in zip(f1, f2):
words = text_utils.TokenizeText(text)
examples.append({
"text": words,
"label": label,
})
return examples
def ParseFacebookInput(inputfile):
"""Parse input in the format used by facebook FastText.
labels are formatted as __label__1
where the label values start at 0.
"""
examples = []
for line in open(inputfile):
words = line.split(' ', 1)
# label is first field with __label__ removed
match = re.match(r'__label__(.+)', words[0])
label = match.group(1) if match else None
words = text_utils.TokenizeText(words[1])
examples.append({
"text": words,
"label": label
})
return examples