def test(args, model, noise_std=25, device=torch.device('cpu')):
loader = data.getDataLoaders(**args['train']['loaders'])['test']
model.eval()
t = tqdm(iter(loader), desc=f"TEST", dynamic_ncols=True)
psnr = 0
for itern, batch in enumerate(t):
batch = batch.to(device)
noisy_batch = utils.awgn(batch, noise_std)
with torch.no_grad():
output = model(noisy_batch)
mse = torch.mean((batch - output)**2).item()
psnr = psnr - 10 * np.log10(mse)
psnr = psnr / (itern + 1)
print(f"Test PSNR = {psnr:.2f} dB")
with open(os.path.join(args['paths']['save'], f'test.psnr'),
'a') as psnr_file:
psnr_file.write(f'{psnr} ')
def main():
args = set_args()
setattr(args, 'model_time', strftime('%H:%M:%S', gmtime()))
setattr(args, 'class_size', 4)
# loading EmoContext data
print("loading data")
train_dataloader, valid_dataloader, num_train_examples = getDataLoaders(
args)
best_model, max_dev_f1 = train(args, train_dataloader, valid_dataloader,
num_train_examples)
if not os.path.exists('saved_models'):
os.makedirs('saved_models')
torch.save(best_model,
f'saved_models/BERT_{args.model_time}_{max_dev_f1}.pt')
print('training finished!')
def test_getDataLoaders(self):
train_ds, val_ds = getDatasets(
'images/labels.csv', 'images/12_images', 'fname', 'is_mono',
train_size=1.0, train_transform=transforms.ToTensor()
)
train_dl, val_dl = getDataLoaders(
train_ds, val_ds, batch_size=1, num_workers=2,
use_weighted_sampling=True
)
# Verify that weighted sampling is working.
counts = [0, 0]
for i in range(100):
for samp in train_dl:
label = int(samp[1][0])
counts[label] += 1
counts[0] /= 100 * 12
counts[1] /= 100 * 12
print(counts)
self.assertLess(abs(0.5 - counts[0]), 0.03)
fout.write('{0}: {1}\n'.format(key, argsdict[key]))
n_gpus = args.n_gpus
if n_gpus < 0:
n_gpus = torch.cuda.device_count()
train_data, val_data = getDatasets(args.labels_csv,
args.images,
args.fnames_col,
args.labels_col,
train_size=args.train_split,
train_transform=train_transform,
val_transform=val_transform,
rng=rng)
trainloader, valloader = getDataLoaders(train_data,
val_data,
batch_size=args.batch_size,
use_weighted_sampling=True)
if args.model_wts != '':
print(f'Loading model weights from {args.model_wts}...')
model = ENModel.load_from_checkpoint(args.model_wts,
lr=args.learning_rate,
n_classes=len(
train_data.dataset.classes))
else:
model = ENModel(args.learning_rate, len(train_data.dataset.classes))
if args.top_only:
model.setTrainTopOnly()
tb_logger = pl_loggers.TensorBoardLogger(outputdir, exp_name)
#print(len(train_idx), len(valid_idx))
fold_folder = outpath / ('fold_' + str(loop_count))
fold_folder.mkdir()
print(
f'\n#\n# Cross-validation fold {loop_count}; saving results to '
f'{fold_folder}.\n#'
)
train_data, val_data = getDatasets(
args.labels_csv, args.images, args.fnames_col, args.labels_col,
train_idx=train_idx, valid_idx=valid_idx,
train_transform=train_transform, val_transform=val_transform, rng=rng
)
trainloader, valloader = getDataLoaders(
train_data, val_data, batch_size=args.batch_size
)
if args.model_wts != '':
print(f'Loading model weights from {args.model_wts}...')
model = ENModel.load_from_checkpoint(
args.model_wts, lr=args.learning_rate,
n_classes=len(train_data.dataset.classes)
)
else:
model = ENModel(args.learning_rate, len(train_data.dataset.classes))
if args.top_only:
model.setTrainTopOnly()
tb_logger = pl_loggers.TensorBoardLogger(