def main(args):
poas = []
init(args.seed, args.device)
print("* loading data")
testdata = ChunkDataSet(
*load_data(limit=args.chunks, shuffle=args.shuffle))
dataloader = DataLoader(testdata, batch_size=args.batchsize)
for w in [int(i) for i in args.weights.split(',')]:
print("* loading model", w)
model = load_model(args.model_directory, args.device, weights=w)
print("* calling")
predictions = []
t0 = time.perf_counter()
for data, *_ in dataloader:
with torch.no_grad():
log_probs = model(data.to(args.device))
predictions.append(log_probs.exp().cpu().numpy())
duration = time.perf_counter() - t0
references = [
decode_ref(target, model.alphabet)
for target in dataloader.dataset.targets
]
sequences = [
decode_ctc(post, model.alphabet)
for post in np.concatenate(predictions)
]
accuracies = list(starmap(accuracy, zip(references, sequences)))
if args.poa: poas.append(sequences)
print("* mean %.2f%%" % np.mean(accuracies))
print("* median %.2f%%" % np.median(accuracies))
print("* time %.2f" % duration)
print("* samples/s %.2E" % (args.chunks * data.shape[2] / duration))
if args.poa:
print("* doing poa")
t0 = time.perf_counter()
# group each sequence prediction per model together
poas = [list(seq) for seq in zip(*poas)]
consensuses = poa(poas)
duration = time.perf_counter() - t0
accuracies = list(starmap(accuracy, zip(references, consensuses)))
print("* mean %.2f%%" % np.mean(accuracies))
print("* median %.2f%%" % np.median(accuracies))
print("* time %.2f" % duration)
def main(args):
poas = []
init(args.seed, args.device)
print("* loading data")
testdata = ChunkDataSet(
*load_data(
limit=args.chunks, shuffle=args.shuffle,
directory=args.directory, validation=True
)
)
dataloader = DataLoader(testdata, batch_size=args.batchsize)
accuracy_with_cov = lambda ref, seq: accuracy(ref, seq, min_coverage=args.min_coverage)
for w in [int(i) for i in args.weights.split(',')]:
seqs = []
print("* loading model", w)
model = load_model(args.model_directory, args.device, weights=w)
print("* calling")
t0 = time.perf_counter()
with torch.no_grad():
for data, *_ in dataloader:
if half_supported():
data = data.type(torch.float16).to(args.device)
else:
data = data.to(args.device)
log_probs = model(data)
if hasattr(model, 'decode_batch'):
seqs.extend(model.decode_batch(log_probs))
else:
seqs.extend([model.decode(p) for p in permute(log_probs, 'TNC', 'NTC')])
duration = time.perf_counter() - t0
refs = [decode_ref(target, model.alphabet) for target in dataloader.dataset.targets]
accuracies = [accuracy_with_cov(ref, seq) if len(seq) else 0. for ref, seq in zip(refs, seqs)]
if args.poa: poas.append(sequences)
print("* mean %.2f%%" % np.mean(accuracies))
print("* median %.2f%%" % np.median(accuracies))
print("* time %.2f" % duration)
print("* samples/s %.2E" % (args.chunks * data.shape[2] / duration))
if args.poa:
print("* doing poa")
t0 = time.perf_counter()
# group each sequence prediction per model together
poas = [list(seq) for seq in zip(*poas)]
consensuses = poa(poas)
duration = time.perf_counter() - t0
accuracies = list(starmap(accuracy_with_coverage_filter, zip(references, consensuses)))
print("* mean %.2f%%" % np.mean(accuracies))
print("* median %.2f%%" % np.median(accuracies))
print("* time %.2f" % duration)
