def inference(epoch, vocab, frames, lang):
model = stochastic_cslr.load_model(1, epoch=epoch, lang=lang)
#model = stochastic_cslr.load_model()
model.to("cpu")
model.eval()
lpis = model([torch.tensor(frames).to("cpu")])
prob = []
prob += [lpi.exp().detach().cpu().numpy() for lpi in lpis]
hyp = model.decode(prob=prob, beam_width=10, prune=0.01, nj=4)
print([" ".join([vocab[i] for i in hi]) for hi in hyp])
hyp = [sup(h) for h in hyp]
hyp = [" ".join([vocab[i] for i in hi]) for hi in hyp]
return hyp
p_drop=0.5,
random_drop=False,
random_crop=False,
crop_size=[224, 224],
base_size=[256, 256],
)
data_loader = DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=False, # result should be strictly ordered for evaluation.
num_workers=args.nj,
collate_fn=dataset.collate_fn,
)
model = stochastic_cslr.load_model(args.model == "sfl", epoch=args.epoch)
model.to(args.device)
model.eval()
prob = []
ref = []
j = 0
with torch.no_grad():
for batch in tqdm(data_loader):
ref += batch["label"]
video = list(map(lambda v: v.to(args.device), batch["video"]))
prob += [lpi.exp().cpu().numpy() for lpi in model(video)]
hyp = model.decode(prob=prob,
beam_width=args.beam_width,
p_drop=0.5,
random_drop=False,
random_crop=False,
crop_size=[224, 224],
base_size=[256, 256],
)
data_loader = DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=False, # result should be strictly ordered for evaluation.
num_workers=args.nj,
collate_fn=dataset.collate_fn,
)
model = stochastic_cslr.load_model(args.model == "sfl")
model.to(args.device)
model.eval()
result_dir = Path("results", args.model, args.split)
prob_path = result_dir / "prob.npz"
if args.use_lm:
lm = dataset.corpus.create_lm()
else:
lm = None
if prob_path.exists():
prob = np.load(prob_path, allow_pickle=True)["prob"]
else:
prob = []
video_path = "video.avi"
#frames = get_frames(video_path="video.avi")
frames = get_frames()
indices = sample_indices(n=len(frames), p_drop=p_drop, random_drop=random_drop)
frames = [frames[i] for i in indices]
frames = map(Image.open, frames)
frames = map(transform, frames)
frames = np.stack(list(frames))
#indices = sample_indices(n=len(frames), p_drop=p_drop, random_drop=random_drop)
#frames = [Image.fromarray(frames[i].asnumpy(), 'RGB') for i in indices]
#frames = map(transform, frames)
#frames = np.stack(list(frames))
model = stochastic_cslr.load_model(args.model == "sfl",
epoch=args.epoch,
lang="German")
model.to(args.device)
model.eval()
lpis = model([torch.tensor(frames).to(args.device)])
prob = []
prob += [lpi.exp().detach().cpu().numpy() for lpi in lpis]
hyp = model.decode(prob=prob,
beam_width=args.beam_width,
prune=args.prune,
nj=args.nj)
def sup(preds):