def get_dataloaders(args):
train_set = GTZAN(phase='train',
min_segments=args['train_segments'],
randomized=True,
overlap=args['overlap'])
val_set = GTZAN(phase='val',
min_segments=args['inference_segments'],
overlap=args['overlap'])
test_set = GTZAN(phase='test',
min_segments=args['inference_segments'],
overlap=args['overlap'])
train_loader = DataLoader(train_set,
batch_size=args['batch_size'],
shuffle=True,
num_workers=args['num_workers'],
pin_memory=True)
val_loader = DataLoader(val_set,
batch_size=args['batch_size'],
num_workers=args['num_workers'],
pin_memory=True)
test_loader = DataLoader(test_set,
batch_size=args['batch_size'],
num_workers=args['num_workers'],
pin_memory=True)
dataloaders = {
'train': train_loader,
'val': val_loader,
'test': test_loader,
}
return dataloaders
def get_datasets(args):
dataset = GTZAN(phase='all', min_segments=args['train_segments'])
train_set = GTZAN(phase='all',
min_segments=args['train_segments'],
randomized=True,
overlap=args['overlap'])
val_set = GTZAN(phase='all',
min_segments=args['inference_segments'],
overlap=args['overlap'])
test_set = GTZAN(phase='all',
min_segments=args['inference_segments'],
overlap=args['overlap'])
return {
'all': dataset,
'train': train_set,
'val': val_set,
'test': test_set,
}
NUM_EARLY_STOPPING_PATIENCE = 50
DEVICE = torch.device('cuda:0')
# In[4]:
# CNN pretrain
criterion = nn.NLLLoss()
net = DenseInception(1, 32, 3, cnn_pretrain=True).to(DEVICE)
best_net = DenseInception(1, 32, 3, cnn_pretrain=True).to(DEVICE)
if not osp.exists(PRETRAIN_CHECKPOINT):
optimizer = optim.Adam(net.parameters(), lr=LR)
train_set = GTZAN(phase='train',
min_segments=PRETRAIN_SEGMENTS,
randomized=True,
overlap=OVERLAP,
noise_rate=1e-3)
val_set = GTZAN(phase='val', min_segments=10, overlap=OVERLAP)
test_set = GTZAN(phase='test', min_segments=10, overlap=OVERLAP)
manual_seed()
train_loader = DataLoader(train_set,
batch_size=PRETRAIN_BATCH_SIZE,
shuffle=True,
num_workers=9,
pin_memory=True,
worker_init_fn=lambda x: manual_seed(x))
val_loader = DataLoader(val_set,
batch_size=PRETRAIN_BATCH_SIZE,
num_workers=9,
pin_memory=True)
loss = running_loss / running_segments
acc = running_corrects / running_samples
seg_acc = running_seg_corrects / running_segments
progress.set_postfix(OrderedDict(
phase=phase,
loss='{:.4f}'.format(loss),
acc='{:.2%}'.format(acc),
seg_acc='{:.2%}'.format(seg_acc)))
return loss, acc, seg_acc
criterion = nn.CrossEntropyLoss()
cv_results = []
dataset = GTZAN(phase='all', min_segments=SEGMENTS)
train_set = GTZAN(phase='all', min_segments=SEGMENTS)
test_set = GTZAN(phase='all', min_segments=SEGMENTS)
skf = StratifiedKFold(NUM_KFOLD, shuffle=True, random_state=1234)
for kfold, (train_index, test_index) in enumerate(skf.split(dataset.X, dataset.Y)):
train_set.X, train_set.Y = dataset.X[train_index], dataset.Y[train_index]
test_set.X, test_set.Y = dataset.X[test_index], dataset.Y[test_index]
train_loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True, num_workers=9, pin_memory=True)
test_loader = DataLoader(test_set, batch_size=BATCH_SIZE, num_workers=9, pin_memory=True)
dataloaders = {
'train': train_loader,
'test': test_loader,
}
loss = running_loss / running_segments
acc = running_corrects / running_samples
seg_acc = running_seg_corrects / running_segments
progress.set_postfix(OrderedDict(
phase=phase,
loss='{:.4f}'.format(loss),
acc='{:.2%}'.format(acc),
seg_acc='{:.2%}'.format(seg_acc)))
return loss, acc, seg_acc
criterion = nn.CrossEntropyLoss()
cv_results = []
dataset = GTZAN(phase='all', min_segments=SEGMENTS)
train_set = GTZAN(phase='all', min_segments=SEGMENTS, randomized=True, overlap=OVERLAP)
test_set = GTZAN(phase='all', min_segments=SEGMENTS, overlap=OVERLAP)
skf = StratifiedKFold(NUM_KFOLD, shuffle=True, random_state=1234)
for kfold, (train_index, test_index) in enumerate(skf.split(dataset.X, dataset.Y)):
if kfold not in (6, 9):
continue
train_set.X, train_set.Y = dataset.X[train_index], dataset.Y[train_index]
test_set.X, test_set.Y = dataset.X[test_index], dataset.Y[test_index]
train_loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True, num_workers=9, pin_memory=True)
test_loader = DataLoader(test_set, batch_size=BATCH_SIZE, num_workers=9, pin_memory=True)
dataloaders = {
'train': train_loader,
'test': test_loader,
loss = running_loss / running_segments
acc = running_corrects / running_samples
seg_acc = running_seg_corrects / running_segments
progress.set_postfix(
OrderedDict(phase=phase,
loss='{:.4f}'.format(loss),
acc='{:.2%}'.format(acc),
seg_acc='{:.2%}'.format(seg_acc)))
return loss, acc, seg_acc
criterion = nn.CrossEntropyLoss()
cv_results = []
dataset = GTZAN(phase='all', min_segments=SEGMENTS)
train_set = GTZAN(phase='all', min_segments=SEGMENTS, overlap=0.5)
test_set = GTZAN(phase='all', min_segments=SEGMENTS, overlap=0.5)
skf = StratifiedKFold(NUM_KFOLD, shuffle=True, random_state=1234)
for kfold, (train_index,
test_index) in enumerate(skf.split(dataset.X, dataset.Y)):
train_set.X, train_set.Y = dataset.X[train_index], dataset.Y[train_index]
test_set.X, test_set.Y = dataset.X[test_index], dataset.Y[test_index]
train_loader = DataLoader(train_set,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=9,
pin_memory=True)
test_loader = DataLoader(test_set,
seg_acc = running_seg_corrects / running_segments
progress.set_postfix(
OrderedDict(phase=phase,
loss='{:.4f}'.format(loss),
acc='{:.2%}'.format(acc),
seg_acc='{:.2%}'.format(seg_acc)))
return loss, acc, seg_acc
# In[5]:
criterion = nn.CrossEntropyLoss()
cv_results = []
dataset = GTZAN(phase='all', min_segments=SEGMENTS)
train_set = GTZAN(phase='all', min_segments=SEGMENTS, randomized=True)
test_set = GTZAN(phase='all', min_segments=SEGMENTS)
skf = StratifiedKFold(NUM_KFOLD, shuffle=True, random_state=1234)
for kfold, (train_index,
test_index) in enumerate(skf.split(dataset.X, dataset.Y)):
train_set.X, train_set.Y = dataset.X[train_index], dataset.Y[train_index]
test_set.X, test_set.Y = dataset.X[test_index], dataset.Y[test_index]
train_loader = DataLoader(train_set,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=9)
test_loader = DataLoader(test_set, batch_size=BATCH_SIZE, num_workers=9)
dataloaders = {