config = util.get_config("experiments.conf")[name]
config["log_dir"] = util.mkdirs(os.path.join(config["log_root"], name))
# Dynamic batch size.
config["batch_size"] = -1
config["max_tokens_per_batch"] = -1
# Use dev lm, if provided.
if config["lm_path"] and "lm_path_dev" in config and config["lm_path_dev"]:
config["lm_path"] = config["lm_path_dev"]
util.print_config(config)
data = LSGNData(config)
model = SRLModel(data, config)
evaluator = LSGNEvaluator(config)
variables_to_restore = []
for var in tf.global_variables():
if "module/" not in var.name:
variables_to_restore.append(var)
else:
print("Not restoring from checkpoint:", var.name)
saver = tf.train.Saver(variables_to_restore)
log_dir = config["log_dir"]
assert not ("final" in name) # Make sure we don't override a finalized checkpoint.
writer = tf.summary.FileWriter(log_dir, flush_secs=20)
evaluated_checkpoints = set()
max_f1 = 0
config["batch_size"] = -1
config["max_tokens_per_batch"] = -1
# Use dev lm, if provided.
if config["lm_path"] and "lm_path_dev" in config and config["lm_path_dev"]:
config["lm_path"] = config["lm_path_dev"]
util.print_config(config)
if len(sys.argv) > 3 and sys.argv[2] == '-gpu':
util.set_gpus(sys.argv[3])
data = LSGNData(config)
model = SRLModel(data, config)
evaluator = LSGNEvaluator(config)
variables_to_restore = []
for var in tf.global_variables():
print var.name
if "module/" not in var.name:
variables_to_restore.append(var)
saver = tf.train.Saver(variables_to_restore)
log_dir = config["log_dir"]
with tf.Session() as session:
checkpoint_path = os.path.join(log_dir, "model.max.ckpt")
print "Evaluating {}".format(checkpoint_path)
tf.global_variables_initializer().run()
saver.restore(session, checkpoint_path)
def predict():
# util.set_gpus()
path = "./tempDir/"
path2 = "./tempDir2/"
os.mkdir("./tempDir2/")
data_list = os.listdir(path)
for data_file in data_list[1:]:
name = "final"
input_filename = path + data_file
output_filename = path2 + data_file
print("Running experiment: {}.".format(name))
config = util.get_config("experiments.conf")[name]
config["log_dir"] = util.mkdirs(os.path.join(config["log_root"], name))
print("Loading data from: {}.".format(input_filename))
config["eval_path"] = input_filename
config["batch_size"] = -1
config["max_tokens_per_batch"] = -1
# Use dev lm, if provided.
if config["lm_path"] and "lm_path_dev" in config and config[
"lm_path_dev"]:
config["lm_path"] = config["lm_path_dev"]
util.print_config(config)
data = LSGNData(config)
model = SRLModel(data, config)
evaluator = LSGNEvaluator(config)
# Load data and model.
eval_data, eval_tensors, doc_level_eval_data = data.load_eval_data()
variables_to_restore = []
for var in tf.global_variables():
#print var.name
if "module/" not in var.name:
variables_to_restore.append(var)
saver = tf.train.Saver(variables_to_restore)
log_dir = config["log_dir"]
with tf.Session() as session:
checkpoint_path = os.path.join(log_dir, "model.max.ckpt")
tf.global_variables_initializer().run()
saver.restore(session, checkpoint_path)
with open(output_filename, "w") as f:
#for example_num, (tensorized_example, example) in enumerate(model.eval_data):
for i, doc_tensors in enumerate(eval_tensors):
feed_dict = dict(
zip(data.input_tensors, [
input_utils.pad_batch_tensors(doc_tensors, tn)
for tn in data.input_names + data.label_names
]))
predict_names = []
for tn in data.predict_names:
if tn in model.predictions:
predict_names.append(tn)
predict_tensors = [
model.predictions[tn] for tn in predict_names
] + [model.loss]
predict_tensors = session.run(predict_tensors,
feed_dict=feed_dict)
predict_dict = dict(
zip(predict_names + ["loss"], predict_tensors))
doc_example = doc_level_eval_data[i]
sentences = doc_example["sentences"]
predictions = inference_utils.srl_decode(
sentences, predict_dict, data.srl_labels_inv, config)
doc_example["predicted_srl"] = []
word_offset = 0
for j, sentence in enumerate(sentences):
for pred, args in predictions["srl"][j].items():
doc_example["predicted_srl"].extend([[
int(pred + word_offset),
int(a[0] + word_offset),
int(a[1] + word_offset), a[2]
] for a in args])
word_offset += len(sentence)
f.write(json.dumps(doc_example))
f.write("\n")
if (i + 1) % 10 == 0:
print("Decoded {} documents.".format(i + 1))
coref_config['coref_depth'] = 0
# Use dev lm, if provided.
if config["lm_path"] and "lm_path_dev" in config and config["lm_path_dev"]:
config["lm_path"] = config["lm_path_dev"]
if coref_config[
"lm_path"] and "lm_path_dev" in coref_config and coref_config[
"lm_path_dev"]:
coref_config["lm_path"] = coref_config["lm_path_dev"]
util.print_config(config)
data = LSGNData(config)
coref_data = LSGNData(coref_config)
model = SRLModel(data, config)
coref_model = SRLModel(coref_data, coref_config)
evaluator = LSGNEvaluator(config)
coref_evaluator = LSGNEvaluator(coref_config)
variables_to_restore = []
for var in tf.global_variables():
if "module/" not in var.name:
variables_to_restore.append(var)
else:
print "Not restoring from checkpoint:", var.name
saver = tf.train.Saver(variables_to_restore)
log_dir = config["log_dir"]
assert not ("final" in name
) # Make sure we don't override a finalized checkpoint.
writer = tf.summary.FileWriter(log_dir, flush_secs=20)
evaluated_checkpoints = set()
print("Running experiment: {}.".format(name))
config = util.get_config("experiments.conf")[name]
config["log_dir"] = util.mkdirs(os.path.join(config["log_root"], name))
print("Loading data from: {}.".format(input_filename))
config["eval_path"] = input_filename
config["batch_size"] = -1
config["max_tokens_per_batch"] = -1
# Use dev lm, if provided.
if config["lm_path"] and "lm_path_dev" in config and config["lm_path_dev"]:
config["lm_path"] = config["lm_path_dev"]
util.print_config(config)
data = LSGNData(config)
model = SRLModel(data, config)
evaluator = LSGNEvaluator(config)
# Load data and model.
eval_data, eval_tensors, doc_level_eval_data = data.load_eval_data()
variables_to_restore = []
for var in tf.global_variables():
#print var.name
if "module/" not in var.name:
variables_to_restore.append(var)
saver = tf.train.Saver(variables_to_restore)
log_dir = config["log_dir"]
with tf.Session() as session:
checkpoint_path = os.path.join(log_dir, "model.max.ckpt")
tf.global_variables_initializer().run()
saver.restore(session, checkpoint_path)