# Combine layer, task and layer, attention maps
# layer_task_config = {}
# layer_attention_config = {}
# for task_or_attn_name, layer in layer_config.items():
# if task_or_attn_name in attention_config:
# layer_attention_config[layer] = attention_config[task_or_attn_name]
# elif task_or_attn_name in task_config:
# if layer not in layer_task_config:
# layer_task_config[layer] = {}
# layer_task_config[layer][task_or_attn_name] = task_config[task_or_attn_name]
# else:
# util.fatal_error('No task or attention config "%s"' % task_or_attn_name)
layer_task_config, layer_attention_config = util.combine_attn_maps(
layer_config, attention_config, task_config)
hparams = train_utils.load_hparams(args, model_config)
dev_filenames = args.dev_files.split(',')
test_filenames = args.test_files.split(',') if args.test_files else []
vocab = Vocab(data_config, args.save_dir)
vocab.update(test_filenames)
embedding_files = [
embeddings_map['pretrained_embeddings']
for embeddings_map in model_config['embeddings'].values()
if 'pretrained_embeddings' in embeddings_map
]
# Generate mappings from feature/label names to indices in the model_fn inputs
# feature_idx_map = {}
model_config = train_utils.load_json_configs(args.model_configs)
task_config = train_utils.load_json_configs(args.task_configs, args)
# print("debug <task_config>: ", task_config)
layer_config = train_utils.load_json_configs(args.layer_configs)
attention_config = train_utils.load_json_configs(args.attention_configs)
# attention_config = {}
# if args.attention_configs and args.attention_configs != '':
# attention_config = train_utils.load_json_configs(args.attention_configs)
# Combine layer, task and layer, attention maps
# todo save these maps in save_dir
layer_task_config, layer_attention_config = util.combine_attn_maps(layer_config, attention_config, task_config)
hparams = train_utils.load_hparams(args, model_config, neptune_handler)
## NEED TO REMOVE
if args.attn_debug:
hparams.attn_debug = True
# Set the random seed. This defaults to int(time.time()) if not otherwise set.
np.random.seed(hparams.random_seed)
tf.set_random_seed(hparams.random_seed)
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
train_filenames = args.train_files.split(',')
dev_filenames = args.dev_files.split(',')
vocab = Vocab(data_config, args.save_dir, train_filenames)