def load_model(opts, device):
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = Struct(load_json(hps_file))
model_config = f'{opts.output_dir}/log/model_config.json'
# Prepare model
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f'{opts.output_dir}/ckpt/model_step_{opts.checkpoint}.pt'
checkpoint = torch.load(ckpt_file)
img_pos_embed_weight_key = (
"v_encoder.f_encoder.img_embeddings.position_embeddings.weight")
assert img_pos_embed_weight_key in checkpoint
max_frm_seq_len = len(checkpoint[img_pos_embed_weight_key])
model = HeroForVcmr.from_pretrained(
model_config,
state_dict=checkpoint,
vfeat_dim=VFEAT_DIM,
max_frm_seq_len=max_frm_seq_len,
lw_neg_ctx=model_opts.lw_neg_ctx,
lw_neg_q=model_opts.lw_neg_q,
lw_st_ed=0,
ranking_loss_type=model_opts.ranking_loss_type,
use_hard_negative=False,
hard_pool_size=model_opts.hard_pool_size,
margin=model_opts.margin,
use_all_neg=model_opts.use_all_neg,
drop_svmr_prob=model_opts.drop_svmr_prob)
model.to(device)
if opts.fp16:
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
return model, model_opts
def main(opts):
hvd.init()
device = torch.device("cuda") # support single GPU only
train_opts = Struct(json.load(open(f'{opts.train_dir}/log/hps.json')))
if 'paired' in train_opts.model:
EvalDatasetCls = Nlvr2PairedEvalDataset
eval_collate_fn = nlvr2_paired_eval_collate
if train_opts.model == 'paired':
ModelCls = UniterForNlvr2Paired
elif train_opts.model == 'paired-attn':
ModelCls = UniterForNlvr2PairedAttn
else:
raise ValueError('unrecognized model type')
elif train_opts.model == 'triplet':
EvalDatasetCls = Nlvr2TripletEvalDataset
ModelCls = UniterForNlvr2Triplet
eval_collate_fn = nlvr2_triplet_eval_collate
else:
raise ValueError('unrecognized model type')
img_db = DetectFeatLmdb(opts.img_db, train_opts.conf_th, train_opts.max_bb,
train_opts.min_bb, train_opts.num_bb,
opts.compressed_db)
txt_db = TxtTokLmdb(opts.txt_db, -1)
dset = EvalDatasetCls(txt_db, img_db, train_opts.use_img_type)
batch_size = (train_opts.val_batch_size
if opts.batch_size is None else opts.batch_size)
sampler = TokenBucketSampler(dset.lens,
bucket_size=BUCKET_SIZE,
batch_size=batch_size,
droplast=False)
eval_dataloader = DataLoader(dset,
batch_sampler=sampler,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=eval_collate_fn)
eval_dataloader = PrefetchLoader(eval_dataloader)
# Prepare model
ckpt_file = f'{opts.train_dir}/ckpt/model_step_{opts.ckpt}.pt'
checkpoint = torch.load(ckpt_file)
model_config = UniterConfig.from_json_file(
f'{opts.train_dir}/log/model.json')
model = ModelCls(model_config, img_dim=IMG_DIM)
model.init_type_embedding()
model.load_state_dict(checkpoint, strict=False)
model.to(device)
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
results = evaluate(model, eval_dataloader, device)
# write results
if not exists(opts.output_dir):
os.makedirs(opts.output_dir)
with open(f'{opts.output_dir}/results.csv', 'w') as f:
for id_, ans in results:
f.write(f'{id_},{ans}\n')
print(f'all results written')
def main(opts):
hvd.init()
if hvd.rank() == 0:
toker = RobertaTokenizer.from_pretrained('roberta-base')
all_gather_list(None)
else:
all_gather_list(None)
toker = RobertaTokenizer.from_pretrained('roberta-base')
model_opts = Struct(json.load(open(f"{opts.model_dir}/log/hps.json")))
model_config = f"{opts.model_dir}/log/model_config.json"
video_db = load_video_sub_dataset(model_opts.vfeat_db,
model_opts.sub_txt_db,
model_opts.vfeat_interval,
model_opts)
dset = TvcEvalDataset(video_db, opts.target_clip)
loader = build_dataloader(dset, opts.batch_size,
TvcEvalDataset.collate, False, opts)
checkpoint = torch.load(f"{opts.model_dir}/ckpt/"
f"model_step_{opts.ckpt_step}.pt")
img_pos_embed_weight_key = "v_encoder.f_encoder.img_embeddings" +\
".position_embeddings.weight"
if img_pos_embed_weight_key in checkpoint:
max_frm_seq_len = len(checkpoint[img_pos_embed_weight_key])
else:
max_frm_seq_len = MAX_FRM_SEQ_LEN
model = HeroForTvc.from_pretrained(model_config,
state_dict=checkpoint,
vfeat_dim=VFEAT_DIM,
max_frm_seq_len=max_frm_seq_len,
lsr=model_opts.lsr)
model.cuda()
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
bos = toker.convert_tokens_to_ids(['<s>'])[0]
eos = toker.convert_tokens_to_ids(['</s>'])[0]
model.eval()
generator = TvcGenerator(model, opts.max_gen_step, bos, eos, opts.fp16)
results = decode(loader, generator, toker)
save_jsonl(results, opts.output)
# evaluate score if possible
if (hvd.rank() == 0
and 'descs' in json.loads(next(iter(open(opts.target_clip))))):
evaluator = TVCEval(opts.target_clip)
score = evaluator(results)
print(score)
def main(opts):
hvd.init()
n_gpu = hvd.size()
print('fasfafs: ', n_gpu)
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(device, n_gpu, hvd.rank(),
opts.fp16))
if opts.train_config is not None:
train_opts = Struct(json.load(open(opts.train_config)))
opts.conf_th = train_opts.conf_th
opts.max_bb = train_opts.max_bb
opts.min_bb = train_opts.min_bb
opts.num_bb = train_opts.num_bb
# load DBs and image dirs
eval_img_db = DetectFeatLmdb(opts.img_db, opts.conf_th, opts.max_bb,
opts.min_bb, opts.num_bb, opts.compressed_db)
eval_txt_db = TxtTokLmdb(opts.txt_db, -1)
eval_dataset = ItmEvalDataset(eval_txt_db, eval_img_db, opts.batch_size)
# Prepare model
checkpoint = torch.load(opts.checkpoint)
model = UniterForImageTextRetrieval.from_pretrained(opts.model_config,
checkpoint,
img_dim=IMG_DIM)
if 'rank_output' not in checkpoint:
model.init_output() # zero shot setting
model.to(device)
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
eval_dataloader = DataLoader(eval_dataset,
batch_size=1,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=itm_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
eval_log, results = evaluate(model, eval_dataloader)
def build_dataloader(opts):
# Load ground truth, query db and video db
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = Struct(load_json(hps_file))
video_ids = get_video_ids(opts.query_txt_db)
video_db = load_video_sub_dataset(opts.vfeat_db, opts.sub_txt_db,
model_opts.vfeat_interval, model_opts)
assert opts.split in opts.query_txt_db, (opts.split, opts.query_txt_db)
q_txt_db = QueryTokLmdb(opts.query_txt_db, -1)
eval_dataset = VcmrFullEvalDataset(video_ids,
video_db,
q_txt_db,
distributed=model_opts.distributed_eval)
eval_dataloader = DataLoader(eval_dataset,
batch_size=opts.batch_size,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=vcmr_full_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
return eval_dataloader
def main(opts):
hvd.init()
n_gpu = hvd.size()
print('fasfafs: ', n_gpu)
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(device, n_gpu, hvd.rank(),
opts.fp16))
if opts.train_config is not None:
train_opts = Struct(json.load(open(opts.train_config)))
opts.conf_th = train_opts.conf_th
opts.max_bb = train_opts.max_bb
opts.min_bb = train_opts.min_bb
opts.num_bb = train_opts.num_bb
# Prepare model
checkpoint = torch.load(opts.checkpoint)
model = UniterForImageTextRetrieval.from_pretrained(opts.model_config,
checkpoint,
img_dim=IMG_DIM)
if 'rank_output' not in checkpoint:
model.init_output() # zero shot setting
# model.to(device)
# model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
# load DBs and image dirs
# npz_path = os.listdir('/root/output_meme_butd')
# npz_path = ['/root/output_meme_butd/' + i for i in npz_path]
json_files = open("/root/meme/train.json", "r")
json_files = json_files.read().split('\n')
json_files = [json.loads(i) for i in json_files]
LOGGER.info('load {} file '.format(len(json_files)))
eval_log, results = evaluate(model, json_files)
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(device, n_gpu, hvd.rank(),
opts.fp16))
if opts.train_config is not None:
train_opts = Struct(json.load(open(opts.train_config)))
opts.conf_th = train_opts.conf_th
opts.max_bb = train_opts.max_bb
opts.min_bb = train_opts.min_bb
opts.num_bb = train_opts.num_bb
# load DBs and image dirs
eval_img_db = DetectFeatLmdb(opts.img_db, opts.conf_th, opts.max_bb,
opts.min_bb, opts.num_bb, opts.compressed_db)
eval_txt_db = TxtTokLmdb(opts.txt_db, -1)
eval_dataset = ItmEvalDataset(eval_txt_db, eval_img_db, opts.batch_size)
# Prepare model
checkpoint = torch.load(opts.checkpoint)
model = UniterForImageTextRetrieval.from_pretrained(opts.model_config,
checkpoint,
img_dim=IMG_DIM)
if 'rank_output' not in checkpoint:
model.init_output() # zero shot setting
model.to(device)
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
eval_dataloader = DataLoader(eval_dataset,
batch_size=1,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=itm_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
eval_log, results = evaluate(model, eval_dataloader)
if hvd.rank() == 0:
if not exists(opts.output_dir) and rank == 0:
os.makedirs(opts.output_dir)
with open(f'{opts.output_dir}/config.json', 'w') as f:
json.dump(vars(opts), f)
with open(f'{opts.output_dir}/results.bin', 'wb') as f:
pickle.dump(results, f)
with open(f'{opts.output_dir}/scores.json', 'w') as f:
json.dump(eval_log, f)
LOGGER.info(f'evaluation finished')
LOGGER.info(
f"======================== Results =========================\n"
f"image retrieval R1: {eval_log['img_r1']*100:.2f},\n"
f"image retrieval R5: {eval_log['img_r5']*100:.2f},\n"
f"image retrieval R10: {eval_log['img_r10']*100:.2f}\n"
f"text retrieval R1: {eval_log['txt_r1']*100:.2f},\n"
f"text retrieval R5: {eval_log['txt_r5']*100:.2f},\n"
f"text retrieval R10: {eval_log['txt_r10']*100:.2f}")
LOGGER.info("========================================================")
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(device, n_gpu, hvd.rank(),
opts.fp16))
hps_file = f"{opts.output_dir}/log/hps.json"
model_opts = Struct(json.load(open(hps_file)))
# train_examples = None
ans2label_file = f"{opts.output_dir}/ckpt/ans2label.json"
ans2label = json.load(open(ans2label_file))
label2ans = {label: ans for ans, label in ans2label.items()}
# load DBs and image dirs
eval_img_db = DetectFeatLmdb(
opts.img_db,
model_opts.conf_th,
model_opts.max_bb,
model_opts.min_bb,
model_opts.num_bb,
opts.compressed_db,
)
eval_txt_db = TxtTokLmdb(opts.txt_db, -1)
eval_dataset = VqaEvalDataset(len(ans2label), eval_txt_db, eval_img_db)
# Prepare model
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f"{opts.output_dir}/ckpt/model_step_{opts.checkpoint}.pt"
checkpoint = torch.load(ckpt_file)
model = UniterForVisualQuestionAnswering.from_pretrained(
f"{opts.output_dir}/log/model.json",
checkpoint,
img_dim=IMG_DIM,
num_answer=len(ans2label),
)
model.to(device)
if opts.fp16:
model = amp.initialize(model, enabled=True, opt_level="O2")
sampler = TokenBucketSampler(
eval_dataset.lens,
bucket_size=BUCKET_SIZE,
batch_size=opts.batch_size,
droplast=False,
)
eval_dataloader = DataLoader(
eval_dataset,
batch_sampler=sampler,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=vqa_eval_collate,
)
eval_dataloader = PrefetchLoader(eval_dataloader)
val_log, results, logits = evaluate(model, eval_dataloader, label2ans,
opts.save_logits)
result_dir = f"{opts.output_dir}/results_test"
if not exists(result_dir) and rank == 0:
os.makedirs(result_dir)
all_results = list(concat(all_gather_list(results)))
if opts.save_logits:
all_logits = {}
for id2logit in all_gather_list(logits):
all_logits.update(id2logit)
if hvd.rank() == 0:
with open(f"{result_dir}/"
f"results_{opts.checkpoint}_all.json", "w") as f:
json.dump(all_results, f)
if opts.save_logits:
np.savez(f"{result_dir}/logits_{opts.checkpoint}_all.npz",
**all_logits)
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(device, n_gpu, hvd.rank(),
opts.fp16))
if hvd.rank() != 0:
LOGGER.disabled = True
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = Struct(json.load(open(hps_file)))
model_config = f'{opts.output_dir}/log/model_config.json'
# load DBs and image dirs
video_ids = get_video_ids(opts.query_txt_db)
video_db = load_video_sub_dataset(opts.vfeat_db, opts.sub_txt_db,
model_opts.vfeat_interval, model_opts)
assert opts.split in opts.query_txt_db
q_txt_db = QaQueryTokLmdb(opts.query_txt_db, -1)
eval_dataset = ViolinEvalDataset(video_ids,
video_db,
q_txt_db,
sampled_by_q=model_opts.sampled_by_q)
collate_fn = violin_eval_collate
# Prepare model
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f'{opts.output_dir}/ckpt/model_step_{opts.checkpoint}.pt'
checkpoint = torch.load(ckpt_file)
img_pos_embed_weight_key = "v_encoder.f_encoder.img_embeddings" +\
".position_embeddings.weight"
assert img_pos_embed_weight_key in checkpoint
max_frm_seq_len = len(checkpoint[img_pos_embed_weight_key])
model = HeroForViolin.from_pretrained(model_config,
state_dict=checkpoint,
vfeat_dim=VFEAT_DIM,
max_frm_seq_len=max_frm_seq_len)
model.to(device)
if opts.fp16:
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
eval_dataloader = DataLoader(eval_dataset,
batch_size=opts.batch_size,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=collate_fn)
eval_dataloader = PrefetchLoader(eval_dataloader)
_, results, logits = validate_violin(model, eval_dataloader, opts.split,
opts.save_logits)
result_dir = f'{opts.output_dir}/results_{opts.split}'
if opts.save_logits:
result_dir += '_w_logit'
if not exists(result_dir) and hvd.rank() == 0:
os.makedirs(result_dir)
all_results = {}
for id2res in all_gather_list(results):
all_results.update(id2res)
if opts.save_logits:
all_logits = {}
for id2logit in all_gather_list(logits):
all_logits.update(id2logit)
if hvd.rank() == 0:
save_json(all_results,
f'{result_dir}/results_{opts.checkpoint}_all.json')
LOGGER.info('All results written......')
if opts.save_logits:
save_pickle(all_logits,
f'{result_dir}/logits_{opts.checkpoint}_all.pkl')
LOGGER.info('All logits written......')
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(
device, n_gpu, hvd.rank(), opts.fp16))
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = json.load(open(hps_file))
if 'mlp' not in model_opts:
model_opts['mlp'] = 1
model_opts = Struct(model_opts)
# Prepare model
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f'{opts.output_dir}/ckpt/model_epoch_{opts.checkpoint}.pt'
checkpoint = torch.load(ckpt_file)
model = UniterForReferringExpressionComprehension.from_pretrained(
f'{opts.output_dir}/log/model.json', checkpoint,
img_dim=IMG_DIM, mlp=model_opts.mlp)
model.to(device)
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
if opts.fp16:
model = amp.initialize(model, enabled=True, opt_level='O2')
# load DBs and image dirs
img_db_type = "gt" if "coco_gt" in opts.img_db else "det"
conf_th = -1 if img_db_type == "gt" else model_opts.conf_th
num_bb = 100 if img_db_type == "gt" else model_opts.num_bb
eval_img_db = DetectFeatLmdb(opts.img_db,
conf_th, model_opts.max_bb,
model_opts.min_bb, num_bb,
opts.compressed_db)
# Prepro txt_dbs
txt_dbs = opts.txt_db.split(':')
for txt_db in txt_dbs:
print(f'Evaluating {txt_db}')
eval_txt_db = ReTxtTokLmdb(txt_db, -1)
eval_dataset = ReEvalDataset(
eval_txt_db, eval_img_db, use_gt_feat=img_db_type == "gt")
sampler = DistributedSampler(eval_dataset, num_replicas=n_gpu,
rank=rank, shuffle=False)
eval_dataloader = DataLoader(eval_dataset,
sampler=sampler,
batch_size=opts.batch_size,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=re_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
# evaluate
val_log, results = evaluate(model, eval_dataloader)
result_dir = f'{opts.output_dir}/results_test'
if not exists(result_dir) and rank == 0:
os.makedirs(result_dir)
write_to_tmp(
f"{txt_db.split('_')[1].split('.')[0]}-acc({img_db_type}): {results['acc']*100:.2f}% ",
args.tmp_file)
all_results = list(concat(all_gather_list(results)))
if hvd.rank() == 0:
db_split = txt_db.split('/')[-1].split('.')[0] # refcoco+_val
img_dir = opts.img_db.split('/')[-1] # re_coco_gt
with open(f'{result_dir}/'
f'results_{opts.checkpoint}_{db_split}_on_{img_dir}_all.json', 'w') as f:
json.dump(all_results, f)
# print
print(f'{opts.output_dir}/results_test')
write_to_tmp(f'\n', args.tmp_file)
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(
device, n_gpu, hvd.rank(), opts.fp16))
if rank != 0:
LOGGER.disabled = True
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = Struct(json.load(open(hps_file)))
assert opts.split in opts.img_db and opts.split in opts.txt_db
# load DBs and image dirs
eval_img_db, eval_img_db_gt = load_img_feat(opts.img_db, model_opts)
eval_txt_db = VcrTxtTokLmdb(opts.txt_db, -1)
eval_dataset = VcrEvalDataset(
"test", eval_txt_db, img_db=eval_img_db,
img_db_gt=eval_img_db_gt)
# Prepare model
model = UniterForVisualCommonsenseReasoning.from_pretrained(
f'{opts.output_dir}/log/model.json', state_dict={},
img_dim=IMG_DIM)
model.init_type_embedding()
model.init_word_embedding(NUM_SPECIAL_TOKENS)
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f'{opts.output_dir}/ckpt/model_step_{opts.checkpoint}.pt'
checkpoint = torch.load(ckpt_file)
state_dict = checkpoint.get('model_state', checkpoint)
matched_state_dict = {}
unexpected_keys = set()
missing_keys = set()
for name, param in model.named_parameters():
missing_keys.add(name)
for key, data in state_dict.items():
if key in missing_keys:
matched_state_dict[key] = data
missing_keys.remove(key)
else:
unexpected_keys.add(key)
LOGGER.info(f"Unexpected_keys: {list(unexpected_keys)}")
LOGGER.info(f"Missing_keys: {list(missing_keys)}")
model.load_state_dict(matched_state_dict, strict=False)
model.to(device)
if opts.fp16:
model = amp.initialize(model, enabled=True, opt_level='O2')
eval_dataloader = DataLoader(eval_dataset,
batch_size=opts.batch_size,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
shuffle=False,
collate_fn=vcr_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
_, results = evaluate(model, eval_dataloader)
result_dir = f'{opts.output_dir}/results_{opts.split}'
if not exists(result_dir) and rank == 0:
os.makedirs(result_dir)
all_results = {}
for id2res in all_gather_list(results):
all_results.update(id2res)
if hvd.rank() == 0:
with open(f'{result_dir}/'
f'results_{opts.checkpoint}_all.json', 'w') as f:
json.dump(all_results, f)
probs_df = save_for_submission(
f'{result_dir}/results_{opts.checkpoint}_all.json')
probs_df.to_csv(f'{result_dir}/results_{opts.checkpoint}_all.csv')
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, 16-bits training: {}".format(device, n_gpu, hvd.rank(), opts.fp16))
if hvd.rank() != 0:
LOGGER.disabled = True
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = Struct(load_json(hps_file))
model_config = f'{opts.output_dir}/log/model_config.json'
# load DBs and image dirs
video_ids = get_video_ids(opts.query_txt_db)
if opts.task != "didemo_video_only":
video_db = load_video_sub_dataset(opts.vfeat_db, opts.sub_txt_db, model_opts.vfeat_interval, model_opts)
else:
txt_meta = load_json(os.path.join(opts.query_txt_db, "meta.json"))
video_db = load_video_only_dataset(opts.vfeat_db, txt_meta, model_opts.vfeat_interval, model_opts)
assert opts.split in opts.query_txt_db
q_txt_db = QueryTokLmdb(opts.query_txt_db, -1)
if opts.task != "didemo_video_only":
inf_dataset = VcmrFullEvalDataset
else:
inf_dataset = VcmrVideoOnlyFullEvalDataset
eval_dataset = inf_dataset(video_ids, video_db, q_txt_db, distributed=model_opts.distributed_eval)
# Prepare model
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f'{opts.output_dir}/ckpt/model_step_{opts.checkpoint}.pt'
checkpoint = torch.load(ckpt_file)
img_pos_embed_weight_key = ("v_encoder.f_encoder.img_embeddings.position_embeddings.weight")
assert img_pos_embed_weight_key in checkpoint
max_frm_seq_len = len(checkpoint[img_pos_embed_weight_key])
model = HeroForVcmr.from_pretrained(
model_config,
state_dict=checkpoint,
vfeat_dim=VFEAT_DIM,
max_frm_seq_len=max_frm_seq_len,
lw_neg_ctx=model_opts.lw_neg_ctx,
lw_neg_q=model_opts.lw_neg_q, lw_st_ed=0,
ranking_loss_type=model_opts.ranking_loss_type,
use_hard_negative=False,
hard_pool_size=model_opts.hard_pool_size,
margin=model_opts.margin,
use_all_neg=model_opts.use_all_neg,
drop_svmr_prob=model_opts.drop_svmr_prob)
model.to(device)
if opts.fp16:
model = amp.initialize(model, enabled=opts.fp16, opt_level='O2')
eval_dataloader = DataLoader(eval_dataset, batch_size=opts.batch_size,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
collate_fn=vcmr_full_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
_, results = validate_full_vcmr(model, eval_dataloader, opts.split, opts, model_opts)
result_dir = f'{opts.output_dir}/results_{opts.split}'
if not exists(result_dir) and rank == 0:
os.makedirs(result_dir)
all_results_list = all_gather_list(results)
if hvd.rank() == 0: # save for only one time
all_results = {"video2idx": all_results_list[0]["video2idx"]}
for rank_id in range(hvd.size()):
for key, val in all_results_list[rank_id].items():
if key == "video2idx":
continue
if key not in all_results:
all_results[key] = []
all_results[key].extend(all_results_list[rank_id][key])
LOGGER.info('All results joined......')
# save_vr(all_results, f'{result_dir}/results_{opts.checkpoint}_{opts.split}_vr.json')
# save_vcmr_base_on_vr(all_results, f'{result_dir}/results_{opts.checkpoint}_{opts.split}_vcmr_base_on_vr.json')
save_vcmr(all_results, f'{result_dir}/results_{opts.checkpoint}_{opts.split}_vcmr.json')
def main(opts):
hvd.init()
n_gpu = hvd.size()
device = torch.device("cuda", hvd.local_rank())
torch.cuda.set_device(hvd.local_rank())
rank = hvd.rank()
LOGGER.info("device: {} n_gpu: {}, rank: {}, "
"16-bits training: {}".format(device, n_gpu, hvd.rank(),
opts.fp16))
if rank != 0:
LOGGER.disabled = True
hps_file = f'{opts.output_dir}/log/hps.json'
model_opts = Struct(json.load(open(hps_file)))
assert opts.split in opts.img_db and opts.split in opts.txt_db
# load DBs and image dirs
eval_img_db, eval_img_db_gt = load_img_feat(opts.img_db, model_opts)
eval_txt_db = VcrTxtTokLmdb(opts.txt_db, -1)
eval_dataset = VcrEvalDataset("val",
eval_txt_db,
img_db=eval_img_db,
img_db_gt=eval_img_db_gt)
# Prepare model
model = UniterForVisualCommonsenseReasoning.from_pretrained(
f'{opts.output_dir}/log/model.json', state_dict={}, img_dim=IMG_DIM)
model.init_type_embedding()
model.init_type_embedding_know()
model.init_word_embedding(NUM_SPECIAL_TOKENS)
if exists(opts.checkpoint):
ckpt_file = opts.checkpoint
else:
ckpt_file = f'{opts.output_dir}/ckpt/model_step_{opts.checkpoint}.pt'
checkpoint = torch.load(ckpt_file)
state_dict = checkpoint.get('model_state', checkpoint)
matched_state_dict = {}
unexpected_keys = set()
missing_keys = set()
for name, param in model.named_parameters():
missing_keys.add(name)
for key, data in state_dict.items():
if key in missing_keys:
matched_state_dict[key] = data
missing_keys.remove(key)
else:
unexpected_keys.add(key)
LOGGER.info(f"Unexpected_keys: {list(unexpected_keys)}")
LOGGER.info(f"Missing_keys: {list(missing_keys)}")
model.load_state_dict(matched_state_dict, strict=False)
model.to(device)
if opts.fp16:
model = amp.initialize(model, enabled=True, opt_level='O2')
eval_dataloader = DataLoader(eval_dataset,
batch_size=opts.batch_size,
num_workers=opts.n_workers,
pin_memory=opts.pin_mem,
shuffle=False,
collate_fn=vcr_eval_collate)
eval_dataloader = PrefetchLoader(eval_dataloader)
results = evaluate(model, eval_dataloader)
output = '/src/vlkaf.json'
before_json = ""
for i, item in enumerate(results):
jstring = json.dumps(item)
before_json += jstring + '\n'
f = open(output, "w")
f.write(before_json)
f.close()
'''
