def constrcut_predictor(path):
# warm_start_from=tf.estimator.WarmStartSettings(
# ckpt_to_initialize_from=str(path),
# vars_to_warm_start=".*" # everything in TRAINABLE_VARIABLES - excluding optimiser params
# # vars_to_warm_start=[".*"], # everything in GLOBAL_VARIABLES - including optimiser params
# )
model = LISAModel(hparams, model_config, layer_task_config, layer_attention_config, feature_idx_map, label_idx_map,
vocab)
# ws = WarmStartSettings(ckpt_to_initialize_from=path,
# vars_to_warm_start=".*")
# print("debug <loading from>:", path)
estimator = tf.estimator.Estimator(model_fn=model.model_fn, model_dir=path)
return estimator
# ('feature' in data_config[d] and data_config[d]['feature']) or
# ('label' in data_config[d] and data_config[d]['label'])]):
# if 'feature' in data_config[f] and data_config[f]['feature']:
# feature_idx_map[f] = i
# if 'label' in data_config[f] and data_config[f]['label']:
# if 'type' in data_config[f] and data_config[f]['type'] == 'range':
# idx = data_config[f]['conll_idx']
# j = i + idx[1] if idx[1] != -1 else -1
# label_idx_map[f] = (i, j)
# else:
# label_idx_map[f] = (i, i+1)
feature_idx_map, label_idx_map = util.load_feat_label_idx_maps(data_config)
# Initialize the model
model = LISAModel(hparams, model_config, layer_task_config,
layer_attention_config, feature_idx_map, label_idx_map,
vocab)
tf.logging.log(
tf.logging.INFO, "Created model with %d trainable parameters" %
tf_utils.get_num_trainable_parameters())
# Set up the Estimator
estimator = tf.estimator.Estimator(model_fn=model.model_fn,
model_dir=args.save_dir)
def dev_input_fn():
return train_utils.get_input_fn(vocab,
data_config,
dev_filenames,
hparams.batch_size,