def validate_mffr(model, val_loader):
LOGGER.info("start running MFFR validation...")
val_loss = 0
cosine = 0
n_feat = 0
st = time()
for i, batch in enumerate(val_loader):
targets = batch['feat_targets']
pred_feat = model(batch, task='mffr', compute_loss=False)
loss = F.mse_loss(pred_feat, targets, reduction='none')
loss = torch.sqrt(loss.sum(dim=1))
val_loss += loss.sum().item()
cosine += F.cosine_similarity(pred_feat, targets, dim=-1).sum().item()
n_feat += batch['c_v_masks'].sum().item()
val_loss = sum(all_gather_list(val_loss))
cosine = sum(all_gather_list(cosine))
n_feat = sum(all_gather_list(n_feat))
tot_time = time() - st
val_loss /= n_feat
val_log = {
'loss': val_loss,
'cosine': cosine / n_feat,
'feat_per_s': n_feat / tot_time
}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"loss: {val_loss:.2f}")
return val_log
def validate(model, val_loader, split):
model.eval()
val_loss = 0
tot_score = 0
n_ex = 0
st = time()
results = []
for i, batch in enumerate(val_loader):
qids = batch['qids']
targets = batch['targets']
del batch['targets']
del batch['qids']
scores = model(batch, compute_loss=False)
loss = F.cross_entropy(scores, targets, reduction='sum')
val_loss += loss.item()
tot_score += (scores.max(dim=-1, keepdim=False)[1] == targets
).sum().item()
answers = ['True' if i == 1 else 'False'
for i in scores.max(dim=-1, keepdim=False
)[1].cpu().tolist()]
results.extend(zip(qids, answers))
n_ex += len(qids)
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time()-st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {f'valid/{split}_loss': val_loss,
f'valid/{split}_acc': val_acc,
f'valid/{split}_ex_per_s': n_ex/tot_time}
model.train()
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log, results
def validate(model, val_loader, label2ans, split='val'):
model.eval()
val_loss = 0
tot_score = 0
n_ex = 0
st = time()
results = {}
for i, batch in enumerate(val_loader):
scores = model(batch, compute_loss=False)
targets = batch['targets']
loss = F.binary_cross_entropy_with_logits(
scores, targets, reduction='sum')
val_loss += loss.item()
tot_score += compute_score_with_logits(scores, targets).sum().item()
answers = [label2ans[i]
for i in scores.max(dim=-1, keepdim=False
)[1].cpu().tolist()]
qids = batch['qids']
for qid, answer in zip(qids, answers):
results[qid] = answer
n_ex += len(qids)
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time()-st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {f'valid/{split}_loss': val_loss,
f'valid/{split}_acc': val_acc,
f'valid/{split}_ex_per_s': n_ex/tot_time}
model.train()
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log, results
def validate_mlm(model, val_loader):
LOGGER.info("start running MLM validation...")
val_loss = 0
n_correct = 0
n_word = 0
st = time()
for i, batch in enumerate(val_loader):
scores = model(batch, task='mlm', compute_loss=False)
labels = batch['txt_labels']
labels = labels[labels != -1]
loss = F.cross_entropy(scores, labels, reduction='sum')
val_loss += loss.item()
n_correct += (scores.max(dim=-1)[1] == labels).sum().item()
n_word += labels.numel()
val_loss = sum(all_gather_list(val_loss))
n_correct = sum(all_gather_list(n_correct))
n_word = sum(all_gather_list(n_word))
tot_time = time()-st
val_loss /= n_word
acc = n_correct / n_word
val_log = {'loss': val_loss,
'acc': acc,
'tok_per_s': n_word/tot_time}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"acc: {acc*100:.2f}")
return val_log
def validate_mlm_old(model, val_loader):
LOGGER.info("start running MLM validation...")
val_loss = 0
n_correct = 0
n_word = 0
st = time()
for i, batch in enumerate(val_loader):
scores = model.forward(batch, task='mlm', compute_loss=False)
loss_fct = torch.nn.CrossEntropyLoss(ignore_index=-1,
reduction='sum')
scores = scores.contiguous().view(-1, model.config.vocab_size)
labels = batch['txt_labels'].contiguous().view(-1)
loss = loss_fct(scores, labels)
val_loss += loss.item()
n_correct += accuracy_count(scores, labels)
n_word += batch['txt_labels'].numel()
val_loss = sum(all_gather_list(val_loss))
n_correct = sum(all_gather_list(n_correct))
n_word = sum(all_gather_list(n_word))
tot_time = time()-st
val_loss /= n_word
acc = n_correct / n_word
val_log = {'loss': val_loss,
'acc': acc,
'tok_per_s': n_word/tot_time}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"acc: {acc*100:.2f}")
return val_log
def validate_vmlm_soft(model, val_loader):
LOGGER.info("start running VMLM-SOFT validation...")
val_loss = 0
n_feat = 0
st = time()
tot_score = 0
#label2token_matrix = torch.float(LABEL2TOKEN_MATRIX)
for i, batch in enumerate(val_loader):
prediction_soft_label = model(
batch, task='vmlm-soft', compute_loss=False)
prediction_soft_label = F.log_softmax(
prediction_soft_label, dim=-1)
label_targets = batch['label_targets']
#convert label_targets to a new dimension
# label_targets = torch.matmul(label_targets, torch.cuda.FloatTensor(LABEL2TOKEN_MATRIX))
# label_targets = label_targets[:, VALID_XLMR_TOKEN_IDS]
# label_targets = label_targets / torch.sum(label_targets, dim=1, keepdim=True)
tot_score += compute_accuracy_for_soft_targets(
prediction_soft_label, label_targets)
loss = F.kl_div(
prediction_soft_label, label_targets, reduction='sum')
val_loss += loss.item()
n_feat += batch['tgt_masks'].sum().item()
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_feat = sum(all_gather_list(n_feat))
tot_time = time()-st
val_loss /= n_feat
val_acc = tot_score / n_feat
val_log = {'loss': val_loss,
'acc': val_acc,
'feat_per_s': n_feat/tot_time}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log
def validate_mmxlm_soft(model, val_loader):
LOGGER.info("start running MMXLM_SOFT validation...")
val_loss = 0
n_feat = 0
st = time()
tot_score = 0
for i, batch in enumerate(val_loader):
prediction_soft_label = model(
batch, task="mmxlm-soft", compute_loss=False)
#if "kl" in task:
prediction_soft_label = F.log_softmax(
prediction_soft_label, dim=-1)
label_targets = batch['label_targets']
loss = F.kl_div(
prediction_soft_label, label_targets, reduction='sum')
tot_score += compute_accuracy_for_soft_targets(
prediction_soft_label, label_targets)
val_loss += loss.item()
n_feat += batch['tgt_masks'].sum().item()
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_feat = sum(all_gather_list(n_feat))
tot_time = time()-st
val_loss /= n_feat
val_acc = tot_score / n_feat
val_log = {'loss': val_loss,
'acc': val_acc,
'feat_per_s': n_feat/tot_time}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log
def validate_videoQA(model, val_loader, split, task="tvqa", save_logits=False):
LOGGER.info(f"start running validation on {task} {split} split...")
model.eval()
val_loss = 0
n_ex = 0
tot_score = 0
results = {}
logits = {}
val_log = {}
st = time()
has_gt_target = True
for i, batch in enumerate(val_loader):
targets = batch['targets']
if has_gt_target and targets.min() < 0:
has_gt_target = False
LOGGER.info(
"No GT annotations provided, only generate predictions")
if 'qids' in batch:
qids = batch['qids']
del batch['qids']
scores = model(batch, task, compute_loss=False)
answers = [
i for i in scores.max(dim=-1, keepdim=False)[1].cpu().tolist()
]
for qid, answer in zip(qids, answers):
results[str(qid)] = answer
if save_logits:
scores = scores.cpu().tolist()
for qid, logit in zip(qids, scores):
logits[str(qid)] = logit
if has_gt_target:
loss = F.cross_entropy(scores,
targets.squeeze(-1),
reduction='sum')
val_loss += loss.item()
tot_score += compute_accuracies(scores, targets)
n_ex += len(qids)
if has_gt_target:
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {
'valid/loss': val_loss,
'valid/acc': val_acc,
'valid/ex_per_s': n_ex / tot_time
}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"loss:{val_loss:.2f}, score: {val_acc*100:.2f}")
model.train()
return val_log, results, logits
def validate_vcmr(model, val_loader, opts):
LOGGER.info(
"start running validation (easy version with loss computed)...")
val_loss = 0
val_loss_st_ed = 0
val_loss_neg_ctx = 0
val_loss_neg_q = 0
n_ex = 0
n_ex_pos = 0
st = time()
for i, batch in enumerate(val_loader):
if 'qids' in batch:
# qids = batch['qids']
del batch['qids']
n_ex += len(batch['q_vidx'])
loss_st_ed, loss_neg_ctx, loss_neg_q =\
model(batch, opts.task, compute_loss=True)
val_loss_st_ed += loss_st_ed.item()
if opts.lw_neg_ctx != 0 or opts.lw_neg_q != 0:
n_pos = len(loss_neg_ctx)
val_loss_neg_ctx += loss_neg_ctx.sum().item()
val_loss_neg_q += loss_neg_q.sum().item()
n_ex_pos += n_pos
val_loss_st_ed = sum(all_gather_list(val_loss_st_ed))
val_loss_neg_ctx = sum(all_gather_list(val_loss_neg_ctx))
val_loss_neg_q = sum(all_gather_list(val_loss_neg_q))
n_ex = sum(all_gather_list(n_ex))
n_ex_pos = sum(all_gather_list(n_ex_pos))
tot_time = time() - st
if opts.lw_st_ed:
val_loss_st_ed /= n_ex
val_loss_st_ed /= opts.lw_st_ed
if n_ex_pos > 0 and opts.lw_neg_q > 0 and\
opts.lw_neg_ctx > 0:
val_loss_neg_ctx /= n_ex_pos
val_loss_neg_q /= n_ex_pos
val_loss_neg_ctx /= opts.lw_neg_ctx
val_loss_neg_q /= opts.lw_neg_q
val_loss = opts.lw_st_ed * val_loss_st_ed +\
opts.lw_neg_ctx * val_loss_neg_ctx +\
opts.lw_neg_q * val_loss_neg_q
val_log = {
'valid/loss_overall': val_loss,
'valid/loss_st_ed': val_loss_st_ed,
'valid/loss_neg_ctx': val_loss_neg_ctx,
'valid/loss_neg_q': val_loss_neg_q,
'valid/ex_per_s': n_ex / tot_time
}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"loss: {val_loss:.2f}")
return val_log
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 compute_hard_neg(model, loader, dataset, hard_negative_num, hard_neg_dir):
txt2hardimgs, img2hardtxts = get_hard_negs(model, loader,
hard_negative_num)
with open(f'{hard_neg_dir}/'
f'txt2hardimgs_rank{hvd.rank()}.json', 'w') as f:
json.dump(txt2hardimgs, f)
if hvd.rank() == 0:
with open(f'{hard_neg_dir}/img2hardtxts.json', 'w') as f:
json.dump(img2hardtxts, f)
all_gather_list(None) # dummy sync to wait for writing
if isinstance(dataset, ConcatDataset):
for dset in dataset.datasets:
dset.reload_hard_negs(hard_neg_dir)
else:
dataset.reload_hard_negs(hard_neg_dir)
def all_gather_stats_list(stat_list, max_size=4096):
"""
Gather a `Statistics` list accross all processes/nodes
Args:
stat_list(list([`Statistics`])): list of statistics objects to
gather accross all processes/nodes
max_size(int): max buffer size to use
Returns:
our_stats(list([`Statistics`])): list of updated stats
"""
from torch.distributed import get_rank
from utils.distributed import all_gather_list
# Get a list of world_size lists with len(stat_list) Statistics objects
all_stats = all_gather_list(stat_list, max_size=max_size)
our_rank = get_rank()
our_stats = all_stats[our_rank]
for other_rank, stats in enumerate(all_stats):
if other_rank == our_rank:
continue
for i, stat in enumerate(stats):
our_stats[i].update(stat, update_n_src_words=True)
return our_stats
def validate(loader, generator, tokenizer, evaluator):
st = time()
generator.model.eval()
results = []
for batch in loader:
vids = batch['vid_names']
cids = batch['clip_ids']
all_ts = batch['all_ts']
outputs = generator.greedy_decode(batch)
for vid, cid, ts, out_ids in zip(vids, cids, all_ts, outputs):
output = tokenizer.convert_tokens_to_string(
tokenizer.convert_ids_to_tokens(out_ids))
results.append({
'vid_name': vid,
'clip_id': cid,
'ts': ts,
'descs': [{
'desc': output
}]
})
results = [r for rs in all_gather_list(results) for r in rs]
LOGGER.info(f'decoding finished in {int(time() - st)} seconds')
if hvd.rank() == 0:
val_log = evaluator(results)
LOGGER.info(f'Validation finished in {int(time() - st)} seconds')
LOGGER.info(f'CIDEr: {val_log["CIDEr"]}')
else:
val_log = {}
generator.model.train()
return val_log, results
def test(model, val_loader, label2ans):
from utils.logger import LOGGER, TB_LOGGER, RunningMeter, add_log_to_file
LOGGER.info("start running test...")
model.eval()
val_loss = 0
tot_score = 0
n_ex = 0
st = time()
results = []
for i, batch in enumerate(val_loader):
scores = model(batch, compute_loss=False)
answers = torch.nn.functional.softmax(scores)
for qid, answer in zip(batch['qids'], answers):
results.append({
'id': qid,
'proba': answer[0].cpu().item(),
'label': int(answer[0].cpu().item() > 0.5)
})
n_ex += len(batch['qids'])
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_log = {'valid/ex_per_s': n_ex / tot_time}
model.train()
LOGGER.info(f"validation finished in {int(tot_time)} seconds, ")
return val_log, results
def evaluate(model, eval_loader, label2ans, save_logits=False):
LOGGER.info("start running evaluation...")
model.eval()
n_ex = 0
st = time()
results = []
logits = {}
for i, batch in enumerate(eval_loader):
qids = batch["qids"]
scores = model(batch, compute_loss=False)
answers = [
label2ans[i]
for i in scores.max(dim=-1, keepdim=False)[1].cpu().tolist()
]
for qid, answer in zip(qids, answers):
results.append({"answer": answer, "question_id": int(qid)})
if save_logits:
scores = scores.cpu()
for i, qid in enumerate(qids):
logits[qid] = scores[i].half().numpy()
if i % 100 == 0 and hvd.rank() == 0:
n_results = len(results)
n_results *= hvd.size() # an approximation to avoid hangs
LOGGER.info(f"{n_results}/{len(eval_loader.dataset)} "
"answers predicted")
n_ex += len(qids)
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_log = {"valid/ex_per_s": n_ex / tot_time}
model.train()
LOGGER.info(f"evaluation finished in {int(tot_time)} seconds "
f"at {int(n_ex/tot_time)} examples per second")
return val_log, results, logits
def validate_violin(model, val_loader, split, save_logits=False):
LOGGER.info("start running validation on VIOLIN {split} split...")
model.eval()
val_loss = 0
n_ex = 0
tot_score = 0
results = {}
logits = {}
st = time()
for i, batch in enumerate(val_loader):
targets = batch['targets']
if 'qids' in batch:
qids = batch['qids']
del batch['qids']
scores = model(batch, "violin", compute_loss=False)
predictions = (torch.sigmoid(scores) > 0.5).long()
answers = predictions.squeeze().cpu().tolist()
for qid, answer in zip(qids, answers):
results[str(qid)] = answer
if save_logits:
scores = scores.cpu().tolist()
for qid, logit in zip(qids, scores):
logits[str(qid)] = logit
loss = F.binary_cross_entropy(torch.sigmoid(scores),
targets.to(dtype=scores.dtype),
reduction='sum')
val_loss += loss.item()
tot_score += compute_accuracies(predictions, targets)
n_ex += len(qids)
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {
f'valid/{split}_loss': val_loss,
f'valid/{split}_acc': val_acc,
f'valid/{split}_ex_per_s': n_ex / tot_time
}
LOGGER.info(f"validation of {split} split finished in {int(tot_time)}s, "
f"loss:{val_loss:.2f}, score: {val_acc*100:.2f}")
return val_log, results, logits
def validate_mrfr(model, val_loader):
LOGGER.info("start running MRFR validation...")
val_loss = 0
n_feat = 0
st = time()
for i, batch in enumerate(val_loader):
loss = model(batch, task='mrfr', compute_loss=True)
val_loss += loss.sum().item() / IMG_DIM
n_feat += batch['img_mask_tgt'].sum().item()
val_loss = sum(all_gather_list(val_loss))
n_feat = sum(all_gather_list(n_feat))
tot_time = time() - st
val_loss /= n_feat
val_log = {'loss': val_loss, 'feat_per_s': n_feat / tot_time}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"loss: {val_loss:.2f}")
return val_log
def validate_mrc(model, val_loader, task):
LOGGER.info("start running MRC validation...")
val_loss = 0
n_feat = 0
st = time()
tot_score = 0
for i, batch in enumerate(val_loader):
prediction_soft_label = model(batch, task=task, compute_loss=False)
if "kl" in task:
prediction_soft_label = F.log_softmax(prediction_soft_label,
dim=-1)
label_targets = batch["label_targets"]
loss = F.kl_div(prediction_soft_label,
label_targets,
reduction="sum")
tot_score += compute_accuracy_for_soft_targets(
prediction_soft_label, label_targets)
else:
# background class should not be the target
cls_label_targets = label_targets[:, 1:].max(dim=-1)[1] + 1
loss = F.cross_entropy(
prediction_soft_label,
cls_label_targets,
ignore_index=0,
reduction="sum",
)
tot_score += compute_accuracy_for_soft_targets(
prediction_soft_label[:, 1:], label_targets[:, 1:])
val_loss += loss.item()
n_feat += batch["img_mask_tgt"].sum().item()
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_feat = sum(all_gather_list(n_feat))
tot_time = time() - st
val_loss /= n_feat
val_acc = tot_score / n_feat
val_log = {
"loss": val_loss,
"acc": val_acc,
"feat_per_s": n_feat / tot_time
}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log
def validate_fom(model, val_loader):
LOGGER.info("start running FOM validation...")
val_loss = 0
n_ex = 0
n_valid_ex = 0
tot_score = 0
st = time()
for i, batch in enumerate(val_loader):
targets = batch['targets']
batch_size, seq_len = targets.size()
vids = batch['vids']
del batch['targets']
del batch['vids']
scores = model(batch, task='fom', compute_loss=False)
targets_valid = targets.view(scores.shape[0], )
loc = (targets_valid != -1).nonzero().squeeze()
scores_valid = scores[loc, :]
targets_valid = targets_valid[loc]
loss = F.cross_entropy(scores_valid, targets_valid, reduction='sum')
val_loss += loss.item()
tot_score += (scores_valid.max(
dim=-1, keepdim=False)[1] == targets_valid).sum().item()
n_valid_ex += len(targets_valid)
n_ex += len(vids)
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
n_valid_ex = sum(all_gather_list(n_valid_ex))
tot_time = time() - st
val_loss /= n_valid_ex
val_acc = tot_score / n_valid_ex
val_log = {
'valid/loss': val_loss,
'valid/acc': val_acc,
'valid/ex_per_s': n_ex / tot_time
}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log
def evaluate(model,
eval_loader,
eval_len,
label2ans,
save_logits=False,
task='vqa'):
LOGGER.info("start running evaluation {}...".format(task))
model.eval()
n_ex = 0
tot_score = 0
st = time()
results = []
logits = {}
pbar = tqdm(total=eval_len)
for i, batch in enumerate(eval_loader):
qids = batch['qids']
scores = model(batch, compute_loss=False, task=task)
targets = batch['targets']
tot_score += compute_score_with_logits(scores, targets).sum().item()
answers = [
label2ans[i]
for i in scores.max(dim=-1, keepdim=False)[1].cpu().tolist()
]
for qid, answer in zip(qids, answers):
results.append({'answer': answer, 'question_id': qid})
if save_logits:
scores = scores.cpu()
for i, qid in enumerate(qids):
logits[qid] = scores[i].half().numpy()
n_ex += len(qids)
pbar.update(len(qids))
# TODO: dont commit, for testing only
#if i > 4:
# break
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
acc = tot_score / n_ex
tot_time = time() - st
val_log = {'valid/ex_per_s': n_ex / tot_time, 'valid/acc': acc}
model.train()
LOGGER.info(f"evaluation finished in {int(tot_time)} seconds "
f"at {int(n_ex/tot_time)} examples per second")
return val_log, results, logits
def validate(model, val_loader):
if hvd.rank() == 0:
pbar = tqdm(total=len(val_loader))
else:
pbar = NoOp()
LOGGER.info("start running Image Retrieval validation ...")
model.eval()
n_ex = 0
st = time()
recall_at_1, recall_at_5, recall_at_10 = 0, 0, 0
for batch in val_loader:
scores = model(batch, compute_loss=False)
_, indices = scores.squeeze(1).topk(10, dim=0)
rank = (indices == 0).nonzero()
if rank.numel():
rank = rank.item()
if rank < 1:
recall_at_1 += 1
if rank < 5:
recall_at_5 += 1
if rank < 10:
recall_at_10 += 1
n_ex += 1
pbar.update(1)
n_ex = sum(all_gather_list(n_ex))
recall_at_1 = sum(all_gather_list(recall_at_1)) / n_ex
recall_at_5 = sum(all_gather_list(recall_at_5)) / n_ex
recall_at_10 = sum(all_gather_list(recall_at_10)) / n_ex
tot_time = time() - st
val_log = {
'valid/ex_per_s': n_ex / tot_time,
'valid/recall_1': recall_at_1,
'valid/recall_5': recall_at_5,
'valid/recall_10': recall_at_10
}
model.train()
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"recall_1: {recall_at_1*100:.2f}, "
f"recall_5: {recall_at_5*100:.2f}, "
f"recall_10: {recall_at_10*100:.2f}")
pbar.close()
return val_log
def evaluate(model, eval_loader):
LOGGER.info("start running evaluation...")
model.eval()
tot_score = 0
n_ex = 0
st = time()
predictions = []
for i, batch in enumerate(eval_loader):
(tgt_box_list, obj_boxes_list, sent_ids) = (
batch['tgt_box'], batch['obj_boxes'], batch['sent_ids'])
# scores (n, max_num_bb)
scores = model(batch, compute_loss=False)
ixs = torch.argmax(scores, 1).cpu().detach().numpy() # (n, )
# pred_boxes
for ix, obj_boxes, tgt_box, sent_id in \
zip(ixs, obj_boxes_list, tgt_box_list, sent_ids):
pred_box = obj_boxes[ix]
predictions.append({'sent_id': int(sent_id),
'pred_box': pred_box.tolist(),
'tgt_box': tgt_box.tolist()})
if eval_loader.loader.dataset.computeIoU(pred_box, tgt_box) > .5:
tot_score += 1
n_ex += 1
if i % 100 == 0 and hvd.rank() == 0:
n_results = len(predictions)
n_results *= hvd.size() # an approximation to avoid hangs
LOGGER.info(f'{n_results}/{len(eval_loader.dataset)} '
'answers predicted')
n_ex = sum(all_gather_list(n_ex))
tot_time = time()-st
tot_score = sum(all_gather_list(tot_score))
val_acc = tot_score / n_ex
val_log = {'valid/acc': val_acc, 'valid/ex_per_s': n_ex/tot_time}
model.train()
LOGGER.info(f"validation ({n_ex} sents) finished in"
f" {int(tot_time)} seconds"
f", accuracy: {val_acc*100:.2f}%")
# summarizae
results = {'acc': val_acc, 'predictions': predictions}
return val_log, results
def validate(model, val_loader, label2ans):
LOGGER.info("start running validation...")
model.eval()
val_loss = 0
tot_score = 0
n_ex = 0
st = time()
results = {}
for i, batch in enumerate(val_loader):
scores = model(batch, compute_loss=False, task='vqa')
targets = batch['targets']
loss = F.binary_cross_entropy_with_logits(scores,
targets,
reduction='sum')
val_loss += loss.item()
tot_score += compute_score_with_logits(scores, targets).sum().item()
answers = [
label2ans[i]
for i in scores.max(dim=-1, keepdim=False)[1].cpu().tolist()
]
for qid, answer in zip(batch['qids'], answers):
results[qid] = answer
n_ex += len(batch['qids'])
# TODO: remove grossness when actual training don't commit this
#if i > 4:
# break
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {
'valid/loss': val_loss,
'valid/acc': val_acc,
'valid/ex_per_s': n_ex / tot_time
}
model.train()
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log, results
def validate(model, val_dataloader):
LOGGER.info("start running evaluation.")
model.eval()
tot_score = 0
n_ex = 0
st = time()
predictions = {}
for i, batch in enumerate(val_dataloader):
# inputs
(tgt_box_list, obj_boxes_list,
sent_ids) = (batch['tgt_box'], batch['obj_boxes'], batch['sent_ids'])
# scores (n, max_num_bb)
scores = model(batch, compute_loss=False)
ixs = torch.argmax(scores, 1).cpu().detach().numpy() # (n, )
# pred_boxes
for ix, obj_boxes, tgt_box, sent_id in \
zip(ixs, obj_boxes_list, tgt_box_list, sent_ids):
pred_box = obj_boxes[ix]
predictions[int(sent_id)] = {
'pred_box': pred_box.tolist(),
'tgt_box': tgt_box.tolist()
}
if val_dataloader.loader.dataset.computeIoU(pred_box,
tgt_box) > .5:
tot_score += 1
n_ex += 1
tot_time = time() - st
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
val_acc = tot_score / n_ex
val_log = {'valid/acc': val_acc, 'valid/ex_per_s': n_ex / tot_time}
model.train()
LOGGER.info(
f"validation ({n_ex} sents) finished in {int(tot_time)} seconds"
f", accuracy: {val_acc*100:.2f}%")
return val_log, predictions
def validate(model, val_loader, label2ans):
LOGGER.info("start running validation...")
model.eval()
val_loss = 0
tot_score = 0
n_ex = 0
st = time()
results = {}
for i, batch in enumerate(val_loader):
scores = model(batch, compute_loss=False)
targets = batch["targets"]
loss = F.binary_cross_entropy_with_logits(scores,
targets,
reduction="sum")
val_loss += loss.item()
tot_score += compute_score_with_logits(scores, targets).sum().item()
answers = [
label2ans[i]
for i in scores.max(dim=-1, keepdim=False)[1].cpu().tolist()
]
for qid, answer in zip(batch["qids"], answers):
results[qid] = answer
n_ex += len(batch["qids"])
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {
"valid/loss": val_loss,
"valid/acc": val_acc,
"valid/ex_per_s": n_ex / tot_time,
}
model.train()
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log, results
def decode(loader, generator, tokenizer):
st = time()
results = []
for batch in tqdm(loader, desc='decoding...'):
vids = batch['vid_names']
cids = batch['clip_ids']
all_ts = batch['all_ts']
outputs = generator.greedy_decode(batch)
for vid, cid, ts, out_ids in zip(vids, cids, all_ts, outputs):
output = tokenizer.convert_tokens_to_string(
tokenizer.convert_ids_to_tokens(out_ids))
results.append({'vid_name': vid, 'clip_id': cid, 'ts': ts,
'descs': [{'desc': output}]})
results = [r for rs in all_gather_list(results) for r in rs]
print(f'decoding finished in {int(time() - st)} seconds')
return results
def validate_itm(model, val_loader):
LOGGER.info("start running ITM validation...")
val_loss = 0
tot_ot_loss = 0
tot_ot_pos = 0
tot_ot_neg = 0
tot_score = 0
n_ex = 0
st = time()
for i, batch in enumerate(val_loader):
scores, ot_loss = model(batch, task='itm', compute_loss=False)
if ot_loss is not None:
if isinstance(ot_loss, tuple):
ot_pos, ot_neg = ot_loss
ot_pos = ot_pos.sum().item()
ot_neg = ot_neg.sum().item()
tot_ot_pos += ot_pos
tot_ot_neg += ot_neg
tot_ot_loss += ot_pos - ot_neg
else:
tot_ot_loss += ot_loss.sum().item()
targets = batch['targets']
loss = F.cross_entropy(scores, targets, reduction='sum')
val_loss += loss.item()
tot_score += (scores.max(dim=-1)[1] == targets).sum().item()
n_ex += len(targets)
val_loss = sum(all_gather_list(val_loss))
tot_score = sum(all_gather_list(tot_score))
n_ex = sum(all_gather_list(n_ex))
tot_time = time() - st
val_loss /= n_ex
val_acc = tot_score / n_ex
val_log = {
'valid/loss': val_loss,
'valid/acc': val_acc,
'valid/ex_per_s': n_ex / tot_time
}
if ot_loss is not None:
tot_ot_loss = sum(all_gather_list(tot_ot_loss))
tot_ot_pos = sum(all_gather_list(tot_ot_pos))
tot_ot_neg = sum(all_gather_list(tot_ot_neg))
val_log['valid/ot_loss'] = tot_ot_loss / n_ex
val_log['valid/ot_pos'] = tot_ot_pos / n_ex
val_log['valid/ot_neg'] = tot_ot_neg / n_ex
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"score: {val_acc*100:.2f}")
return val_log
def evaluate(model, eval_loader):
model.eval()
st = time()
LOGGER.info("start running Image/Text Retrieval evaluation ...")
score_matrix = inference(model, eval_loader)
dset = eval_loader.dataset
all_score = hvd.allgather(score_matrix)
all_txt_ids = [i for ids in all_gather_list(dset.ids) for i in ids]
all_img_ids = dset.all_img_ids
assert all_score.size() == (len(all_txt_ids), len(all_img_ids))
if hvd.rank() != 0:
return {}, tuple()
# NOTE: only use rank0 to compute final scores
eval_log = itm_eval(all_score, all_txt_ids, all_img_ids, dset.txt2img,
dset.img2txts)
results = (all_score, all_txt_ids, all_img_ids)
tot_time = time() - st
LOGGER.info(f"evaluation finished in {int(tot_time)} seconds, ")
return eval_log, results
def validate_mfm_nce(model, val_loader):
LOGGER.info("start running MFM-NCE validation...")
val_loss = 0
val_l2 = 0
n_correct = 0
cosine = 0
n_feat = 0
n_neg = 0
st = time()
for i, batch in enumerate(val_loader):
feats, neg_feats = model(batch, task='mfm-nce', compute_loss=False)
pos_feats = batch['feat_targets']
logits = model.v_encoder.mfm_nce(feats,
pos_feats,
neg_feats,
compute_loss=False)
targets = torch.arange(0,
logits.size(0),
dtype=torch.long,
device=logits.device)
val_loss += F.cross_entropy(logits, targets, reduction='sum').item()
val_l2 += F.mse_loss(feats, pos_feats,
reduction='none').sum(dim=1).sqrt().sum().item()
n_correct += (logits.max(dim=-1)[1] == targets).sum().item()
cosine += F.cosine_similarity(feats, pos_feats, dim=-1).sum().item()
nf = pos_feats.size(0)
n_feat += nf
n_neg += neg_feats.size(0) * nf
val_loss = sum(all_gather_list(val_loss))
val_l2 = sum(all_gather_list(val_l2))
n_correct = sum(all_gather_list(n_correct))
cosine = sum(all_gather_list(cosine))
n_feat = sum(all_gather_list(n_feat))
n_neg = sum(all_gather_list(n_neg))
tot_time = time() - st
val_loss /= n_feat
val_acc = n_correct / n_feat
val_log = {
'loss': val_loss,
'acc': val_acc,
'l2': val_l2 / n_feat,
'cosine': cosine / n_feat,
'feat_per_s': n_feat / tot_time
}
LOGGER.info(f"validation finished in {int(tot_time)} seconds, "
f"loss: {val_loss:.2f}, acc: {val_acc*100:.2f} "
f"(average {n_neg/n_feat:.0f} negatives)")
return val_log
def train(self, train_iter_fct, valid_iter_fct, train_steps, valid_steps):
"""
The main training loops.
by iterating over training data (i.e. `train_iter_fct`)
and running validation (i.e. iterating over `valid_iter_fct`
Args:
train_iter_fct(function): a function that returns the train
iterator. e.g. something like
train_iter_fct = lambda: generator(*args, **kwargs)
valid_iter_fct(function): same as train_iter_fct, for valid data
train_steps(int):
valid_steps(int):
save_checkpoint_steps(int):
Return:
None
"""
logger.info('Start training...')
step = self.optim._step + 1
true_batchs = []
accum = 0
normalization = 0
train_iter = train_iter_fct()
total_stats = Statistics()
report_stats = Statistics()
self._start_report_manager(start_time=total_stats.start_time)
while step <= train_steps:
reduce_counter = 0
for i, batch in enumerate(train_iter):
if self.n_gpu == 0 or (i % self.n_gpu == self.gpu_rank):
if self.gpu_verbose_level > 1:
logger.info("GpuRank %d: index: %d accum: %d" %
(self.gpu_rank, i, accum))
true_batchs.append(batch)
if self.norm_method == "tokens":
num_tokens = batch.tgt[1:].ne(
self.train_loss.padding_idx).sum()
normalization += num_tokens.item()
else:
normalization += batch.batch_size
accum += 1
if accum == self.grad_accum_count:
reduce_counter += 1
if self.gpu_verbose_level > 0:
logger.info("GpuRank %d: reduce_counter: %d \
n_minibatch %d" % (self.gpu_rank, reduce_counter,
len(true_batchs)))
if self.n_gpu > 1:
normalization = sum(all_gather_list(normalization))
self._gradient_accumulation(true_batchs, normalization,
total_stats, report_stats)
report_stats = self._maybe_report_training(
step, train_steps, self.optim.learning_rate,
report_stats)
true_batchs = []
accum = 0
normalization = 0
if (step % valid_steps == 0):
if self.gpu_verbose_level > 0:
logger.info('GpuRank %d: validate step %d' %
(self.gpu_rank, step))
valid_iter = valid_iter_fct()
valid_stats = self.validate(valid_iter)
if self.gpu_verbose_level > 0:
logger.info('GpuRank %d: gather valid stat \
step %d' % (self.gpu_rank, step))
valid_stats = self._maybe_gather_stats(valid_stats)
if self.gpu_verbose_level > 0:
logger.info('GpuRank %d: report stat step %d' %
(self.gpu_rank, step))
self._report_step(self.optim.learning_rate,
step,
valid_stats=valid_stats)
if self.gpu_rank == 0:
self._maybe_save(step)
step += 1
if step > train_steps:
break
if self.gpu_verbose_level > 0:
logger.info('GpuRank %d: we completed an epoch \
at step %d' % (self.gpu_rank, step))
train_iter = train_iter_fct()
return total_stats