def fetch_hidden_vector(hparam, vocab_size, data, model_path):
task = transformer_nli_hidden(hparam, vocab_size, 0, False)
sess = init_session()
sess.run(tf.global_variables_initializer())
load_model_w_scope(sess, model_path, ["bert"])
batches = get_batches_ex(data, hparam.batch_size, 4)
def batch2feed_dict(batch):
x0, x1, x2, y = batch
feed_dict = {
task.x_list[0]: x0,
task.x_list[1]: x1,
task.x_list[2]: x2,
task.y: y,
}
return feed_dict
def pred_fn():
outputs = []
for batch in batches:
x0, x1, x2, y = batch
all_layers, emb_outputs = sess.run(
[task.all_layers, task.embedding_output],
feed_dict=batch2feed_dict(batch))
outputs.append((all_layers, emb_outputs, x0))
return outputs
return pred_fn()
def __init__(self, hparam, voca_size, start_model_path):
print("AgreePredictor")
tf.reset_default_graph()
self.task = transformer_weight(hparam, voca_size, False)
self.sess = init_session()
self.sess.run(tf.global_variables_initializer())
load_model_w_scope(self.sess, start_model_path, ['bert', 'cls_dense'])
def fetch_hidden_vector(hparam, vocab_size, run_name, data_loader, model_path):
print("fetch_hidden_vector:", run_name)
task = transformer_nli_hidden(hparam, vocab_size, 0, False)
sess = init_session()
sess.run(tf.global_variables_initializer())
load_model_w_scope(sess, model_path, ["bert"])
dev_batches = get_batches_ex(data_loader.get_dev_data(), hparam.batch_size,
4)
def batch2feed_dict(batch):
x0, x1, x2, y = batch
feed_dict = {
task.x_list[0]: x0,
task.x_list[1]: x1,
task.x_list[2]: x2,
task.y: y,
}
return feed_dict
def pred_fn():
outputs = []
for batch in dev_batches[:100]:
x0, x1, x2, y = batch
all_layers, emb_outputs = sess.run(
[vars], feed_dict=batch2feed_dict(batch))
outputs.append((all_layers, emb_outputs, x0))
return outputs
return pred_fn()
def __init__(self, hparam, voca_size, start_model_path):
print("run_ukp_ex")
tf.reset_default_graph()
self.task = transformer_nli(hparam, voca_size, 5, True)
self.sess = init_session()
self.sess.run(tf.global_variables_initializer())
load_model_w_scope(self.sess, start_model_path,
['bert', 'cls_dense', 'aux_conflict'])
def fetch_params(hparam, vocab_size, run_name, data_loader, model_path):
print("fetch_hidden_vector:", run_name)
task = transformer_nli_hidden(hparam, vocab_size, 0, False)
sess = init_session()
sess.run(tf.global_variables_initializer())
load_model_w_scope(sess, model_path, ["bert"])
vars = tf.trainable_variables()
names = list([v.name for v in vars])
vars_out, = sess.run([vars])
return names, vars_out
def init_fn_generic(sess, start_type, start_model_path):
if start_type == "cls":
load_model_with_blacklist(sess, start_model_path,
["explain", "explain_optimizer"])
elif start_type == "cls_new":
load_model_with_blacklist(
sess, start_model_path,
["explain", "explain_optimizer", "optimizer"])
elif start_type == "cls_ex":
load_model(sess, start_model_path)
elif start_type == "as_is":
load_model(sess, start_model_path)
elif start_type == "cls_ex_for_pairing":
load_model_with_blacklist(sess, start_model_path,
["match_predictor", "match_optimizer"])
elif start_type == "bert":
load_model_w_scope(sess, start_model_path, ["bert"])
elif start_type == "cold":
pass
else:
assert False
def load_fn(sess, model_path):
return load_model_w_scope(sess, model_path, "bert")
def load_fn(sess, model_path):
if not resume:
return load_model_w_scope(sess, model_path, "bert")
else:
return load_model(sess, model_path)
