def infer_speaker(env, tok):
import tqdm
from utils import load_datasets
listner = Seq2SeqAgent(env, "", tok, args.maxAction)
speaker = Speaker(env, listner, tok)
speaker.load(args.load)
dataset = load_datasets(env.splits)
key_map = {}
for i, item in enumerate(dataset):
key_map[item["path_id"]] = i
path2inst = speaker.get_insts(wrapper=tqdm.tqdm)
for path_id in path2inst.keys():
speaker_pred = tok.decode_sentence(path2inst[path_id])
dataset[key_map[path_id]]['instructions'] = [speaker_pred]
with open("tasks/R2R/data/aug_paths_unseen_infer.json", "w") as f:
json.dump(dataset, f, indent=4, sort_keys=True)
def create_augment_data():
setup()
# Create a batch training environment that will also preprocess text
vocab = read_vocab(TRAIN_VOCAB)
tok = Tokenizer(vocab=vocab, encoding_length=args.maxInput)
# Load features
feat_dict = read_img_features(features)
candidate_dict = utils.read_candidates(CANDIDATE_FEATURES)
# The datasets to be augmented
print("Start to augment the data")
aug_envs = []
# aug_envs.append(
# R2RBatch(
# feat_dict, candidate_dict, batch_size=args.batchSize, splits=['train'], tokenizer=tok
# )
# )
# aug_envs.append(
# SemiBatch(False, 'tasks/R2R/data/all_paths_46_removetrain.json',
# feat_dict, candidate_dict, batch_size=args.batchSize, splits=['train', 'val_seen'], tokenizer=tok)
# )
aug_envs.append(
SemiBatch(False,
'tasks/R2R/data/all_paths_46_removevalunseen.json',
"unseen",
feat_dict,
candidate_dict,
batch_size=args.batchSize,
splits=['val_unseen'],
tokenizer=tok))
aug_envs.append(
SemiBatch(False,
'tasks/R2R/data/all_paths_46_removetest.json',
"test",
feat_dict,
candidate_dict,
batch_size=args.batchSize,
splits=['test'],
tokenizer=tok))
# aug_envs.append(
# R2RBatch(
# feat_dict, candidate_dict, batch_size=args.batchSize, splits=['val_seen'], tokenizer=tok
# )
# )
# aug_envs.append(
# R2RBatch(
# feat_dict, candidate_dict, batch_size=args.batchSize, splits=['val_unseen'], tokenizer=tok
# )
# )
for snapshot in os.listdir(os.path.join(log_dir, 'state_dict')):
# if snapshot != "best_val_unseen_bleu": # Select a particular snapshot to process. (O/w, it will make for every snapshot)
if snapshot != "best_val_unseen_bleu":
continue
# Create Speaker
listner = Seq2SeqAgent(aug_envs[0], "", tok, args.maxAction)
speaker = Speaker(aug_envs[0], listner, tok)
# Load Weight
load_iter = speaker.load(os.path.join(log_dir, 'state_dict', snapshot))
print("Load from iter %d" % (load_iter))
# Augment the env from aug_envs
for aug_env in aug_envs:
speaker.env = aug_env
# Create the aug data
import tqdm
path2inst = speaker.get_insts(beam=args.beam, wrapper=tqdm.tqdm)
data = []
for datum in aug_env.fake_data:
datum = datum.copy()
path_id = datum['path_id']
if path_id in path2inst:
datum['instructions'] = [
tok.decode_sentence(path2inst[path_id])
]
datum.pop('instr_encoding') # Remove Redundant keys
datum.pop('instr_id')
data.append(datum)
print("Totally, %d data has been generated for snapshot %s." %
(len(data), snapshot))
print("Average Length %0.4f" % utils.average_length(path2inst))
print(datum) # Print a Sample
# Save the data
import json
os.makedirs(os.path.join(log_dir, 'aug_data'), exist_ok=True)
beam_tag = "_beam" if args.beam else ""
json.dump(data,
open(
os.path.join(
log_dir, 'aug_data', '%s_%s%s.json' %
(snapshot, aug_env.name, beam_tag)), 'w'),
sort_keys=True,
indent=4,
separators=(',', ': '))
