def _run(self, sess, model_file, embeddings_set, use_preproc=True):
indices, vocabs = self._create_vocabs(model_file)
extra_features_required = [
x for x in vocabs.keys()
if x not in TensorFlowExporter.DEFAULT_VOCABS
]
self.assign_char_lookup()
labels = self.load_labels(model_file)
mxlen, mxwlen = self._get_max_lens(model_file)
model_params = self.task.config_params["model"]
if use_preproc:
serialized_tf_example, tf_example, raw_posts, lengths = self._run_preproc(
model_params, vocabs, model_file, indices,
extra_features_required)
model_params["pkeep"] = 1
model_params["sess"] = sess
model_params["maxs"] = mxlen
model_params["maxw"] = mxwlen
print(model_params)
embeddings = self._initialize_embeddings_map(vocabs, embeddings_set)
model = baseline.tf.classify.create_model(embeddings, labels,
**model_params)
softmax_output = tf.nn.softmax(model.logits)
values, indices = tf.nn.top_k(softmax_output, len(labels))
class_tensor = tf.constant(model.labels)
table = tf.contrib.lookup.index_to_string_table_from_tensor(
class_tensor)
classes = table.lookup(tf.to_int64(indices))
self.restore_model(sess, model_file)
if use_preproc:
sig_input = SignatureInput(serialized_tf_example, tf_example,
extra_features_required)
else:
sig_input = SignatureInput(None,
None,
extra_features_required,
model=model)
sig_output = SignatureOutput(classes, values)
return sig_input, sig_output, 'predict_text'
def _run(self, sess, model_file, embeddings_set, use_preproc=True):
mxlen, mxwlen = self._get_max_lens(model_file)
indices, vocabs = self._create_vocabs(model_file)
self.assign_char_lookup()
labels = self.load_labels(model_file)
extra_features_required = [
x for x in vocabs.keys()
if x not in TensorFlowExporter.DEFAULT_VOCABS
]
model_params = self.task.config_params["model"]
lengths = []
if use_preproc:
serialized_tf_example, tf_example, raw_posts, lengths = self._run_preproc(
model_params, vocabs, model_file, indices,
extra_features_required)
model_params["lengths"] = lengths
model_params["pkeep"] = 1
model_params["sess"] = sess
model_params["maxs"] = mxlen
model_params["maxw"] = mxwlen
model_params['span_type'] = self.task.config_params['train'].get(
'span_type')
print(model_params)
classes, values, model = self._create_model(vocabs, labels,
embeddings_set, mxlen,
model_params)
self.restore_model(sess, model_file)
if use_preproc:
sig_input = SignatureInput(serialized_tf_example, tf_example,
extra_features_required)
else:
sig_input = SignatureInput(None,
None,
extra_features_required + ['lengths'],
model=model)
sig_output = SignatureOutput(classes, values)
return sig_input, sig_output, 'tag_text'
def _run(self, sess, model_file, embeddings_set):
self.word2input, vocab1 = Seq2SeqTensorFlowExporter.read_input_vocab(
model_file)
self.output2word, vocab2 = Seq2SeqTensorFlowExporter.read_output_vocab(
model_file)
# Make the TF example, network input
serialized_tf_example = tf.placeholder(tf.string, name='tf_example')
feature_configs = {
FIELD_NAME: tf.FixedLenFeature(shape=[], dtype=tf.string),
}
tf_example = tf.parse_example(serialized_tf_example, feature_configs)
raw_posts = tf_example[FIELD_NAME]
# Run for each post
dense, length = tf.map_fn(self._preproc_post_creator(),
raw_posts,
dtype=(tf.int32, tf.int32))
model_params = self.task.config_params["model"]
model_params["dsz"] = self.get_dsz(embeddings_set)
model_params["src"] = dense
model_params["src_len"] = length
model_params["mx_tgt_len"] = self.task.config_params["preproc"][
"mxlen"]
model_params["tgt_len"] = 1
model_params["pkeep"] = 1
model_params["sess"] = sess
model_params["predict"] = True
print(model_params)
model = baseline.tf.seq2seq.create_model(vocab1, vocab2,
**model_params)
output = self.output2word.lookup(tf.cast(model.best, dtype=tf.int64))
self.restore_model(sess, model_file)
sig_input = SignatureInput(serialized_tf_example, raw_posts)
sig_output = SignatureOutput(classes, None)
return sig_input, sig_output, 'suggest_text'
def _run(self, sess, model_file, embeddings_set, output_dir, model_version, use_preproc=True):
indices, vocabs = self._create_vocabs(model_file)
self.assign_char_lookup()
labels = self.load_labels(model_file)
extra_features_required = []
# Make the TF example, network input
serialized_tf_example, tf_example = self._create_example(extra_features_required)
raw_posts = tf_example[FIELD_NAME]
mxlen, mxwlen = self._get_max_lens(model_file)
preprocessor = ElmoPreprocessorCreator(
indices, self.lchars, self.upchars_lut,
self.task, FIELD_NAME, extra_features_required, mxlen, mxwlen
)
types = {k: tf.int64 for k in indices.keys()}
types.update({'mixed_case_word': tf.int64})
# Run for each post
preprocessed, lengths = tf.map_fn(preprocessor.preproc_post, tf_example,
dtype=(types, tf.int32),
back_prop=False)
embeddings = self._initialize_embeddings_map(vocabs, embeddings_set)
model_params = self.task.config_params["model"]
model_params["x"] = preprocessed['mixed_case_word']
model_params["xch"] = preprocessed['char']
model_params["x_lc"] = preprocessed['word']
model_params["lengths"] = lengths
model_params["pkeep"] = 1
model_params["sess"] = sess
model_params["maxs"] = mxlen
model_params["maxw"] = mxwlen
model_params['span_type'] = self.task.config_params['train'].get('span_type')
print(model_params)
model = baseline.tf.tagger.create_model(labels, embeddings, **model_params)
model.create_loss()
softmax_output = tf.nn.softmax(model.probs)
values, indices = tf.nn.top_k(softmax_output, 1)
start_np = np.full((1, 1, len(labels)), -1e4, dtype=np.float32)
start_np[:, 0, labels['<GO>']] = 0
start = tf.constant(start_np)
model.probs = tf.concat([start, model.probs], 1)
if model.crf is True:
indices, _ = tf.contrib.crf.crf_decode(model.probs, model.A, tf.constant([mxlen + 1]))## We are assuming the batchsz is 1 here
indices = indices[:, 1:]
list_of_labels = [''] * len(labels)
for label, idval in labels.items():
list_of_labels[idval] = label
class_tensor = tf.constant(list_of_labels)
table = tf.contrib.lookup.index_to_string_table_from_tensor(class_tensor)
classes = table.lookup(tf.to_int64(indices))
self.restore_checkpoint(sess, model_file)
sig_input = SignatureInput(None, raw_posts)
sig_output = SignatureOutput(classes, values)
return sig_input, sig_output, "tag_text"