def confusion_matrix(output, labels, num_labels):
conf_meter = ConfusionMeter(num_labels)
auc_meter = AUCMeter()
preds = output.max(1)[1].type_as(labels)
conf_meter.add(preds.data.squeeze(), labels.type(torch.LongTensor).data)
auc_meter.add(preds.data.squeeze(), labels.data.squeeze())
return conf_meter, auc_meter
def __init__(self, labels, logger=None):
# self._labels = labels
self._logger = logger
self._loss = []
self._acc = []
self._labels = labels
self._conf = ConfusionMeter(len(self._labels))
# self._auc = {label: AUCMeter() for label in range(len(labels))}
self._auc = AUCMeter()
self._c_inp, self._c_tar = Counter(), Counter()
class ModelMeter:
def __init__(self, labels, logger=None):
# self._labels = labels
self._logger = logger
self._loss = []
self._acc = []
self._labels = labels
self._conf = ConfusionMeter(len(self._labels))
# self._auc = {label: AUCMeter() for label in range(len(labels))}
self._auc = AUCMeter()
self._c_inp, self._c_tar = Counter(), Counter()
def update(self, loss, output, targets):
self._loss.append(loss)
self._acc.append(accuracy(output, targets))
preds = output.max(1)[1].type_as(targets)
p = preds.cpu().data.numpy()
t = targets.cpu().data.numpy()
cur_indexes = np.logical_and(p != 2, t != 2)
self._auc.add(p[cur_indexes], t[cur_indexes])
self._conf.add(preds.data.squeeze(), targets.data.squeeze())
self._c_inp.update(p)
self._c_tar.update(t)
# preds = preds[preds != 2 and (targets != 2)]
# for label, meter in self._auc.items():
# cur_pred = preds.eq(label)
# cur_label = targets.eq(label)
# meter.add(cur_pred.data.squeeze(), cur_label.data.squeeze())
def log(self, msg, level=logging.DEBUG, func=None):
if func is None:
loss, acc = self._loss[-1].data[0], self._acc[-1].data[0]
else:
loss, acc = func(self._loss).data[0], func(self._acc).data[0]
self._logger.log(level, "%s: loss: %3.4f, acc: %3.4f", msg, loss, acc)
@property
def last_acc(self):
return self._acc[-1].data[0]
@property
def last_loss(self):
return self._loss[-1].data[0]
def test(segment_size):
test_list_pub = pickle.load(
open(os.path.join(MTAT_SPLIT_FOLDER, 'test_list_pub.cP'), 'rb'))
total_test_size = len(test_list_pub)
model = local_model(segment_size).cuda()
model.load_state_dict(
torch.load(
os.path.join(ENCODER_FOLDER,
'local_model_' + str(segment_size) + '.pt')))
model.eval()
auc = AUCMeter()
for start in range(0, total_test_size, n_songs):
print("Loading dataset...", start)
test_features = np.concatenate([
np.load(
os.path.join(MTAT_NPY_FOLDER, 'testing/' + test_list_pub[i]))
for i in range(start, min(start + n_songs, total_test_size))
])
test_labels = np.load(os.path.join(
MTAT_SPLIT_FOLDER,
'y_test_pub.npy'))[start:min(start + n_songs, total_test_size)]
if normalization:
mean = np.mean(test_features, axis=0)
var = np.var(test_features, axis=0)
test_features = (test_features - mean) / np.sqrt(var)
test_data = CustomDataset(test_features, test_labels)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=batch_size)
print("Dataset loaded")
for data, labels in test_loader:
X = Variable(data).cuda()
out, _ = model(X)
auc_out = np.reshape(out.data.cpu().numpy(), -1)
auc_target = np.reshape(labels, -1)
auc.add(auc_out, auc_target)
auc_tuple = auc.value()
print("AUC = ", auc_tuple[0])
plt.plot(auc_tuple[2], auc_tuple[1])
plt.plot([0, 1])
plt.show()
def evaluate(model, eval_iter, opt):
model.eval()
accuracy = []
threshold = 0.5
AUC_list = [AUCMeter() for _ in range(opt.label_size)]
for index, batch in enumerate(eval_iter):
text = batch.comment_text.data
label = torch.stack([
batch.toxic, batch.severe_toxic, batch.obscene,
batch.threat, batch.insult, batch.identity_hate
], dim=1)
label = label.float()
# for label (batch_size, classes_size)
# for text (batch_size, max_seq_len)
pred = model(text)
is_class = pred > threshold # is it is greater than the threshold
is_class = is_class.float() # (batch_size, classes_size)
# for AUC_meter
pred = torch.nn.functional.sigmoid(pred)
print(pred)
print(label)
for i in range(opt.label_size):
if opt.use_cuda:
AUC_list[i].add(
output=pred.data.cpu().numpy()[:, i],
target=label.data.cpu().numpy()[:, i]
)
else:
AUC_list[i].add(
output=pred.data.numpy()[:, i],
target=label.data.numpy()[:, i]
)
percision = is_class == label # (batch_size, classes_size)
percision = percision.float()
percision = percision.mean(dim=0) # (classes_size)
if opt.use_cuda:
accuracy.append(percision.data.cpu().numpy())
else:
accuracy.append(percision.data.numpy())
# accuracy ()
model.train()
# return (classes_size)
# AUC
AUC_scores = [AUC_list[i].value()[0] for i in range(opt.label_size)]
return np.mean(accuracy, axis=0), AUC_scores # return the mean data, the accuracy for all six classes
def __init__(self,
input_key: str = "targets",
output_key: str = "logits",
prefix: str = "auc",
class_names: List[str] = None,
num_classes: int = 1):
self.prefix = prefix
self.input_key = input_key
self.output_key = output_key
self.class_names = class_names
self.num_classes = num_classes \
if class_names is None \
else len(class_names)
assert self.num_classes is not None
self.auc_meters = [AUCMeter() for _ in range(self.num_classes)]
def __init__(self, r, transform, mode, pred=[], probability=[], log=''):
self.r = r # noise ratio
self.transform = transform
self.mode = mode
train_loader, val_loader = get_chexpert_loaders(r, batch_size=32)
if self.mode == 'test':
self.test_data = val_loader.get_all_samples()
self.test_label = val_loader.get_all_real_ground_truth()
else:
train_label = train_loader.get_all_real_ground_truth()
train_data = train_loader.get_all_samples()
noise_label = train_loader.get_all_labels()
if self.mode == 'all':
self.train_data = train_data
self.noise_label = noise_label
elif self.mode == 'labeled':
pred_idx = pred.nonzero()[0]
self.probability = [probability[i] for i in pred_idx]
clean = (np.array(noise_label) == np.array(train_label))
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability, clean)
auc, _, _ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n' %
(pred.sum(), auc))
log.flush()
self.train_data = train_data[pred_idx]
self.noise_label = noise_label[pred_idx]
print("%s data has a size of %d" %
(self.mode, len(self.noise_label)))
elif self.mode == "unlabeled":
pred_idx = (1 - pred).nonzero()[0]
self.train_data = train_data[pred_idx]
self.noise_label = noise_label[pred_idx]
print("%s data has a size of %d" %
(self.mode, len(self.noise_label)))
def __init__(self, dataset, r, noise_mode, root_dir, transform, mode, noise_file='', pred=[], probability=[], log=''):
self.r = r # noise ratio
self.transform = transform
self.mode = mode
self.transition = {0:0,2:0,4:7,7:7,1:1,9:1,3:5,5:3,6:6,8:8} # class transition for asymmetric noise
if self.mode=='test' or self.mode=='test_average':
if dataset=='cifar10':
test_dic = unpickle('%s/test_batch'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
elif dataset=='cifar100':
test_dic = unpickle('%s/test'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
else:
train_data = []
train_label = []
if dataset=='cifar10':
for n in range(1,6):
dpath = '%s/data_batch_%d'%(root_dir,n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label+data_dic['labels']
train_data = np.concatenate(train_data)
elif dataset=='cifar100':
train_dic = unpickle('%s/train'%root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
if self.mode == 'eval' or self.mode == 'eval_average':
self.eval_data = train_data[45000:]
self.eval_label = train_label[45000:]
else:
if os.path.exists(noise_file):
noise_label = json.load(open(noise_file,"r"))
else: #inject noise
noise_label = []
if self.mode in ['all', 'benchmark_all', 'benchmark_all_average']:
size = 50000
elif self.mode in ['train', 'benchmark', 'benchmark_average']:
size = 45000
idx = list(range(size))
random.shuffle(idx)
num_noise = int(self.r*size)
noise_idx = idx[:num_noise]
for i in range(size):
if i in noise_idx:
if noise_mode=='sym':
if dataset=='cifar10':
noiselabel = random.randint(0,9)
elif dataset=='cifar100':
noiselabel = random.randint(0,99)
noise_label.append(noiselabel)
elif noise_mode=='asym':
noiselabel = self.transition[train_label[i]]
noise_label.append(noiselabel)
else:
noise_label.append(train_label[i])
print("save noisy labels to %s ..."%noise_file)
json.dump(noise_label,open(noise_file,"w"))
if self.mode in ['all', 'benchmark_all', 'benchmark_all_average']:
self.train_data = train_data
self.noise_label = noise_label
self.clean_label = train_label
elif self.mode in ['train', 'benchmark', 'benchmark_average']:
self.train_data = train_data[:45000]
self.noise_label = noise_label[:45000]
self.clean_label = train_label[:45000]
else:
if self.mode == "labeled":
pred_idx = pred.nonzero()[0]
self.probability = [probability[i] for i in pred_idx]
clean = (np.array(noise_label)==np.array(train_label))
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability,clean)
auc,_,_ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n'%(pred.sum(),auc))
log.flush()
elif self.mode == "unlabeled":
pred_idx = (1-pred).nonzero()[0]
self.train_data = train_data[pred_idx]
self.noise_label = [noise_label[i] for i in pred_idx]
print("%s data has a size of %d"%(self.mode,len(self.noise_label)))
background_image = Image.fromarray(background)
foreground = np.uint8(cm.jet(heatmap) * 255)
heatmap_opacity = foreground[:, :, 3]
heatmap_opacity[:] = 64
threshold_prob = min(0.3, heatmap.max() - 0.05)
heatmap_opacity[heatmap < threshold_prob] = 0
foreground_image = Image.fromarray(foreground)
image = Image.alpha_composite(background_image, foreground_image)
image.load() # needed for split()
background = Image.new('RGB', image.size, color)
background.paste(image, mask=image.split()[3])
image_array = np.array(background, dtype=np.uint8)
return image_array
auc_meter = AUCMeter()
conf_meter = ConfusionMeter(options.n_classes)
iterator = get_train_val_iterators(options)
bar = progressbar.ProgressBar()
for batch_idx, data in bar(enumerate(iterator['val']())):
output = model(Variable(data['input'].cuda(), volatile=True))
target = data['target'].cpu().numpy()
prob_tensor = F.softmax(output['classification']).data
prob = prob_tensor.cpu().numpy()
heatmap = F.softmax(output['segmentation']).data.cpu().numpy()
auc_meter.add(prob[:, 1], target)
conf_meter.add(prob_tensor, data['target'])
input_images = data['input'].cpu().numpy()
for i in range(input_images.shape[0]):
image = np.repeat(input_images[i], 3, axis=0)
def __init__(self, dataset, r, noise_mode, root_dir, transform, mode, noise_file='', pred=[], probability=[], log='', teacher_idx=None, truncate_mode=None, refinement=None):
self.r = r # noise ratio
self.transform = transform
self.mode = mode
self.transition = {0:0,2:0,4:7,7:7,1:1,9:1,3:5,5:3,6:6,8:8} # class transition for asymmetric noise
# For distill test
self.teacher_idx = teacher_idx
self.truncate_mode = truncate_mode
self.train_label = None
self.refinement = refinement
if self.mode=='test':
if dataset=='cifar10':
test_dic = unpickle('%s/test_batch'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
elif dataset=='cifar100':
test_dic = unpickle('%s/test'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
else:
train_data=[]
train_label=[]
if dataset=='cifar10':
for n in range(1,6):
dpath = '%s/data_batch_%d'%(root_dir,n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label+data_dic['labels']
train_data = np.concatenate(train_data)
elif dataset=='cifar100':
train_dic = unpickle('%s/train'%root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
self.train_label = train_label
if os.path.exists(noise_file):
noise_label = json.load(open(noise_file,"r"))
else: #inject noise
fix_seed()
noise_label = []
idx = list(range(50000))
random.shuffle(idx)
num_noise = int(self.r*50000)
noise_idx = idx[:num_noise]
for i in range(50000):
if i in noise_idx:
if noise_mode=='sym':
if dataset=='cifar10':
noiselabel = random.randint(0,9)
elif dataset=='cifar100':
noiselabel = random.randint(0,99)
noise_label.append(noiselabel)
elif noise_mode=='asym':
noiselabel = self.transition[train_label[i]]
noise_label.append(noiselabel)
else:
noise_label.append(train_label[i])
print("save noisy labels to %s ..."%noise_file)
json.dump(noise_label,open(noise_file,"w"))
if self.mode == 'all':
self.train_data = train_data
self.noise_label = noise_label
if self.truncate_mode == 'initial':
self.train_data = self.train_data[teacher_idx]
self.noise_label = [noise_label[i] for i in teacher_idx]
else:
if self.mode == "labeled":
pred_idx = pred.nonzero()[0]
if self.truncate_mode == 'initial':
pred_idx = pred_idx.tolist()
teacher_idx = teacher_idx.tolist()
pred_idx = list(set(pred_idx) & set(teacher_idx))
pred_idx = torch.tensor(pred_idx)
self.probability = [probability[i] for i in pred_idx]
clean = (np.array(noise_label)==np.array(train_label))
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability,clean)
auc,_,_ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n'%(pred.sum(),auc))
log.flush()
elif self.mode == "unlabeled":
pred_idx = (1-pred).nonzero()[0]
if self.truncate_mode == 'initial':
whole_idx = list(range(50000))
pred_idx = pred_idx.tolist()
teacher_idx = teacher_idx.tolist()
tmp_set = set(whole_idx) - set(teacher_idx)
tmp_set = tmp_set | set(pred_idx)
pred_idx = torch.tensor(list(tmp_set))
elif self.mode == "labeled_svd":
if self.refinement:
pred_idx = pred.nonzero()[0]
pred_idx_set = set(pred_idx.tolist())
teacher_idx_set = set(teacher_idx.tolist())
pred_idx = torch.tensor(list(pred_idx_set & teacher_idx_set))
self.probability = [probability[i] for i in pred_idx]
clean = (np.array(noise_label)==np.array(train_label))
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability,clean)
auc,_,_ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n'%(pred.sum(),auc))
log.flush()
else:
pred_idx = teacher_idx
probability = torch.ones(50000,)
self.probability = [probability[i] for i in pred_idx]
log.write('Number of labeled samples (by svd) : %d' % teacher_idx.shape[0])
elif self.mode == "unlabeled_svd":
if self.refinement:
clean_pred_idx = pred.nonzero()[0]
clean_pred_idx_set = set(clean_pred_idx.tolist())
teacher_idx_set = set(teacher_idx.tolist())
all_idx_set = set(range(50000))
pred_idx = torch.tensor(list(all_idx_set - (clean_pred_idx_set & teacher_idx_set)))
else:
pred_idx = torch.arange(0, 50000)
pred_idx_set = set(pred_idx.tolist()) - set(teacher_idx.tolist())
pred_idx = torch.tensor(list(pred_idx_set))
self.train_data = train_data[pred_idx]
self.noise_label = [noise_label[i] for i in pred_idx]
print("%s data has a size of %d"%(self.mode,len(self.noise_label)))
model = model_data["model"]
loss_fn = model_data["criterion"]
optim = model_data["optim"]
nb_data_test = 300
nb_canaux = 3
data_root = "./data/"
data = load_preprocess_data(data_root, 0, nb_data_test)
signals = data["signals"]
target = data["targets"]
res = np.zeros(target.shape)
auc_c0 = AUCMeter()
auc_c1 = AUCMeter()
auc_c2 = AUCMeter()
print("Testing model...")
model.eval()
for i in tqdm(range(nb_data_test)):
out = model(th.Tensor(signals[None, i, :, :])).detach().numpy()
auc_c0.add(out[None, 0, 0], target[None, i, 0])
auc_c1.add(out[None, 0, 1], target[None, i, 1])
auc_c2.add(out[None, 0, 2], target[None, i, 2])
res[i, 0] = 1 if out[0, 0] > 0.5 else -1
res[i, 1] = 1 if out[0, 1] > 0.5 else -1
res[i, 2] = 1 if out[0, 2] > 0.5 else -1
def __init__(self,
dataset,
noisy_dataset,
r,
on,
noise_mode,
root_dir,
noise_data_dir,
transform,
mode,
noise_file='',
pred=[],
probability=[],
log='',
targets=None):
self.r = r # total noise ratio
self.on = on # proportion of open noise
self.transform = transform
self.mode = mode
self.transition = {
0: 0,
2: 0,
4: 7,
7: 7,
1: 1,
9: 1,
3: 5,
5: 3,
6: 6,
8: 8
} # class transition for asymmetric noise
self.open_noise = None
self.closed_noise = None
if self.mode == 'test':
if dataset == 'cifar10':
test_dic = unpickle('%s/test_batch' % root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
elif dataset == 'cifar100':
test_dic = unpickle('%s/test' % root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
elif self.mode == 'clean':
if not os.path.exists(noise_file):
print('Noise not defined')
return
if self.open_noise is None or self.closed_noise is not None:
noise = json.load(open(noise_file, "r"))
noise_labels = noise['noise_labels']
self.open_noise = noise['open_noise']
self.closed_noise = noise['closed_noise']
train_data = []
train_label = []
noise_data = []
if dataset == 'cifar10':
for n in range(1, 6):
dpath = '%s/data_batch_%d' % (root_dir, n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label + data_dic['labels']
train_data = np.concatenate(train_data)
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
open_noise = [item[0] for item in self.open_noise]
clean_indices = list(
set(range(50000)) - set(open_noise) - set(self.closed_noise))
self.clean_data = train_data[clean_indices]
self.clean_label = np.asarray(train_label)[clean_indices]
else:
train_data = []
train_label = []
noise_data = []
if dataset == 'cifar10':
for n in range(1, 6):
dpath = '%s/data_batch_%d' % (root_dir, n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label + data_dic['labels']
train_data = np.concatenate(train_data)
elif dataset == 'cifar100':
train_dic = unpickle('%s/train' % root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
if noisy_dataset == 'imagenet32':
noise_data = None
else:
noise_data = unpickle(
'%s/train' % noise_data_dir)['data'].reshape(
(50000, 3, 32, 32)).transpose((0, 2, 3, 1))
if os.path.exists(noise_file):
noise = json.load(open(noise_file, "r"))
noise_labels = noise['noise_labels']
self.open_noise = noise['open_noise']
self.closed_noise = noise['closed_noise']
for cleanIdx, noisyIdx in noise['open_noise']:
if noisy_dataset == 'imagenet32':
train_data[cleanIdx] = np.asarray(
Image.open('{}/{}.png'.format(
noise_data_dir,
str(noisyIdx + 1).zfill(7)))).reshape(
(32, 32, 3))
else:
train_data[cleanIdx] = noise_data[noisyIdx]
else:
#inject noise
noise_labels = [] # all labels (some noisy, some clean)
idx = list(range(50000)) # indices of cifar dataset
random.shuffle(idx)
num_total_noise = int(self.r * 50000) # total amount of noise
num_open_noise = int(
self.on *
num_total_noise) # total amount of noisy/openset images
if noisy_dataset == 'imagenet32': # indices of openset source images
target_noise_idx = list(range(1281149))
else:
target_noise_idx = list(range(50000))
random.shuffle(target_noise_idx)
self.open_noise = list(
zip(idx[:num_open_noise], target_noise_idx[:num_open_noise]
)) # clean sample -> openset sample mapping
self.closed_noise = idx[
num_open_noise:num_total_noise] # closed set noise indices
# populate noise_labels
for i in range(50000):
if i in self.closed_noise:
if noise_mode == 'sym':
if dataset == 'cifar10':
noiselabel = random.randint(0, 9)
elif dataset == 'cifar100':
noiselabel = random.randint(0, 99)
noise_labels.append(noiselabel)
elif noise_mode == 'asym':
noiselabel = self.transition[train_label[i]]
noise_labels.append(noiselabel)
else:
noise_labels.append(train_label[i])
# populate openset noise images
for cleanIdx, noisyIdx in self.open_noise:
if noisy_dataset == 'imagenet32':
train_data[cleanIdx] = np.asarray(
Image.open('{}/{}.png'.format(
noise_data_dir,
str(noisyIdx + 1).zfill(7)))).reshape(
(32, 32, 3))
else:
train_data[cleanIdx] = noise_data[noisyIdx]
# write noise to a file, to re-use
noise = {
'noise_labels': noise_labels,
'open_noise': self.open_noise,
'closed_noise': self.closed_noise
}
print("save noise to %s ..." % noise_file)
json.dump(noise, open(noise_file, "w"))
if self.mode == 'all':
self.train_data = train_data
if targets is None:
self.noise_labels = noise_labels
else:
self.noise_labels = targets
else:
if self.mode == "labeled":
pred_idx = pred.nonzero()[0]
self.probability = [probability[i] for i in pred_idx]
clean = (np.array(noise_labels) == np.array(train_label))
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability, clean)
# note: If all the labels are clean, the following will return NaN
auc, _, _ = auc_meter.value()
elif self.mode == "unlabeled":
pred_idx = pred.nonzero()[0]
self.train_data = train_data[pred_idx]
self.noise_labels = [noise_labels[i] for i in pred_idx]
print("%s data has a size of %d" %
(self.mode, len(self.noise_labels)))
def __init__(self, num_class):
super().__init__()
self.num_class = num_class
self.meters: List[AUCMeter] = [
AUCMeter() for k in range(self.num_class)
]
def __init__(self, dataset, r, noise_mode, root_dir, transform, mode, noise_file='', pred=[], probability=[], log=''):
# mode
# Test : Test
# All : All
# Labeled : Labeled
# UnLabeled : UnLabeled
self.r = r # noise ratio
self.transform = transform
self.mode = mode
self.transition = {0:0,2:0,4:4,7:7,1:1,9:1,3:5,5:3,6:6,8:8} # class transition for asymmetric noise
if self.mode=='test':
if dataset=='cifar10':
test_dic = unpickle('%s/test_batch'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
elif dataset=='cifar100':
test_dic = unpickle('%s/test'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
else:
train_data=[]
train_label=[]
if dataset=='cifar10':
for n in range(1,6):
dpath = '%s/data_batch_%d'%(root_dir,n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label+data_dic['labels']
train_data = np.concatenate(train_data)
elif dataset=='cifar100':
train_dic = unpickle('%s/train'%root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
# Noise Label 생성
if os.path.exists(noise_file):
noise_label = json.load(open(noise_file,"r"))
else: #inject noise
noise_label = []
idx = list(range(50000))
random.shuffle(idx)
# num_noise = int(self.r*50000) -> 순 사기꾼이여ㅡㅡ
# noise_idx = idx[:num_noise]
noise_idx = idx[:]
num_classes = 10 if dataset == 'cifar10' else 100
if noise_mode == 'sym':
C = uniform_mix_C(self.r, num_classes)
# if dataset=='cifar10':
# noiselabel = random.randint(0,9)
# elif dataset=='cifar100':
# noiselabel = random.randint(0,99)
# noise_label.append(noiselabel)
elif noise_mode == 'asym':
C = flip_labels_C(self.r, num_classes)
for i in range(50000):
if i in noise_idx:
noiselabel = np.random.choice(num_classes, p=C[train_label[i]])
noise_label.append(noiselabel)
else:
noise_label.append(train_label[i])
print("save noisy labels to %s ..."%noise_file)
json.dump(noise_label,open(noise_file,"w"))
# 전체 부분
if self.mode == 'all':
self.train_data = train_data
self.noise_label = noise_label
else:
if self.mode == "labeled":
pred_idx = pred.nonzero()[0] # 4770
self.probability = [probability[i] for i in pred_idx] # 4770
clean = (np.array(noise_label)==np.array(train_label)) # 39981
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability,clean)
auc,_,_ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n'%(pred.sum(),auc))
log.flush()
elif self.mode == "unlabeled":
pred_idx = (1-pred).nonzero()[0] # 45230
self.train_data = train_data[pred_idx]
self.noise_label = [noise_label[i] for i in pred_idx]
print("%s data has a size of %d"%(self.mode,len(self.noise_label)))
def test_vgg16():
parser = argparse.ArgumentParser("Test VGG16 Main")
parser.add_argument("-d",
"--data-path",
type=str,
required=True,
dest="data_path")
parser.add_argument("-l",
"--label-path",
type=str,
required=True,
dest="label_path")
parser.add_argument("-m",
"--model-path",
type=str,
required=True,
dest="model_path")
args = parser.parse_args()
data_path = args.data_path
label_path = args.label_path
model_path = args.model_path
# Test if numpy data and labels exist
if not exists(data_path):
raise FileNotFoundError(
"Numpy data file doesn't exist ({}) !".format(data_path))
if not exists(label_path):
raise FileNotFoundError(
"Numpy label file doesn't exist ({}) !".format(label_path))
# Test if model save file exist
if not exists(model_path):
raise FileNotFoundError(
"Model state dict file doesn't exist ({}) !".format(model_path))
print("Load model...")
# Load model
vgg16 = get_vgg16_modified()
vgg16.load_state_dict(th.load(model_path))
vgg16.cuda()
vgg16.eval()
# Create AUC Meter
auc_meter = AUCMeter()
print("Load data...")
# Load data
data = np.load(data_path)
labels = np.load(label_path)
# Split eval
nb_split = int(data.shape[0] * train_ratio)
data = data[nb_split:]
labels = labels[nb_split:]
batch_size = 32
nb_batch = ceil(data.shape[0] / batch_size)
# Loop on eval data
for i_b in tqdm(range(nb_batch)):
# Get batch indexes
i_min = i_b * batch_size
i_max = (i_b + 1) * batch_size
i_max = i_max if i_max < data.shape[0] else data.shape[0]
# Slice data to get batch
batch = data[i_min:i_max, :, :, :]
batch = batch.transpose(0, 3, 1, 2)
batch = th.tensor(batch).cuda().float() / 255.
# And labels
batch_label = th.tensor(labels[i_min:i_max]).cuda().float()
# Forward - Inférence
out = vgg16(batch).squeeze(1)
# Update metric
auc_meter.add(out.cpu().detach(), batch_label.cpu().detach())
print("AUC value = {}".format(auc_meter.value()[0]))
simulator.train()
for epoch in range(4):
print("---epoch {}---".format(epoch))
for step, batch in enumerate(train_loader):
feats, labels = batch
logits = simulator(**feats)
loss = criterion(logits, labels)
opt.zero_grad()
loss.backward()
opt.step()
if (step + 1) % 500 == 0:
with torch.no_grad():
simulator.eval()
auc = AUCMeter()
for feats, labels in val_loader:
outputs = torch.sigmoid(simulator(**feats))
auc.add(outputs, labels)
print(step, auc.value()[0])
if auc.value()[0] > 0.735:
break
simulator.train()
simulator.to("cpu")
torch.save(simulator.state_dict(), simulator_path)
# create a torch dataset class that adopt the simulator and generate the synthetic dataset
synthetic_data_path = os.path.join(filepath, "full_impression_feats.pt")
syn = SyntheticMovieLensDataset(filepath,
simulator_path,
def __init__(self, dataset, r, noise_mode, root_dir, transform, mode, noise_file='', pred=[], probability=[], log=''):
self.r = r # noise ratio
self.transform = transform
self.mode = mode
#self.transition = {0:0,2:0,4:7,7:7,1:1,9:1,3:5,5:3,6:6,8:8} # class transition for asymmetric noise
#self.transition = {0: 1, 1: 2, 2: 3, 3: 4, 4: 5, 5: 6, 6: 7, 7: 8, 9: 9, 9: 0} # 十类
self.transition = {0: 1, 1: 0} # 两类
if self.mode=='test':
if dataset=='cifar10':
test_dic = unpickle('%s/test_batch'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
if noise_mode == 'asym_two_unbalanced_classes':
for i in range(len(self.test_label)):
if self.test_label[i] != 1:
self.test_label[i] = 0
#print("self.test_label=",self.test_label)
elif dataset=='cifar100':
test_dic = unpickle('%s/test'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
else:
train_data=[]
train_label=[]
if dataset=='cifar10':
for n in range(1,6):
dpath = '%s/data_batch_%d'%(root_dir,n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label+data_dic['labels']
train_data = np.concatenate(train_data)
if noise_mode == 'asym_two_unbalanced_classes':
for i in range(len(train_label)):
if train_label[i] != 1:
train_label[i] = 0
#print("train_label=",train_label)
elif dataset=='cifar100':
train_dic = unpickle('%s/train'%root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
#每次重新获得第一次初始化的数据
if os.path.exists(noise_file):
noise_label = json.load(open(noise_file,"r"))
else: #inject noise
noise_label = []
idx = list(range(50000))
random.shuffle(idx)
#num_noise = int(self.r*50000)
if noise_mode == 'sym':
num_noise = int((self.r / 9) / ( 1-self.r + self.r / 9 ) * 50000)
else:
num_noise = int(self.r * 50000)
noise_idx = idx[:num_noise]
for i in range(50000):
if i in noise_idx:
if noise_mode=='sym':
if dataset=='cifar10':
noiselabel = random.randint(0, 9)
#print("noiselabel=",noiselabel)
#print("train_label[i]=",train_label[i])
while noiselabel == train_label[i]:
noiselabel = random.randint(0,9)
elif dataset=='cifar100':
noiselabel = random.randint(0,99)
noise_label.append(noiselabel)
elif noise_mode=='asym_two_unbalanced_classes':
noiselabel = self.transition[train_label[i]]
noise_label.append(noiselabel)
else:
noise_label.append(train_label[i])
print("save noisy labels to %s ..."%noise_file)
json.dump(noise_label,open(noise_file,"w"))
if self.mode == 'all':
self.train_data = train_data
self.noise_label = noise_label
else:
if self.mode == "labeled":
pred_idx = pred.nonzero()[0]#二维数组，返回第几个samples，[0]表示第几行
self.probability = [probability[i] for i in pred_idx]
clean = (np.array(noise_label)==np.array(train_label))
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability,clean)
auc,_,_ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n'%(pred.sum(),auc))
log.flush()
elif self.mode == "unlabeled":
pred_idx = (1-pred).nonzero()[0]
self.train_data = train_data[pred_idx]# 每次初始化的时候 这里面是有label的数据 或者 无label的数据
self.noise_label = [noise_label[i] for i in pred_idx] #每次初始化的时候 这里面是有label的数据 或者 无label的数据
print("%s data has a size of %d"%(self.mode,len(self.noise_label)))
import torchvision import torchvision.transforms as transforms from torch.utils.data import DataLoader from asteroid_dataset import * import torch.optim as optim from classifier import * import visdom from torchnet.meter import ConfusionMeter from torchnet.meter import AUCMeter from sklearn.metrics import matthews_corrcoef confusion_matrix = ConfusionMeter(2) # temp_confusion_matrix = ConfusionMeter(2) auc_meter = AUCMeter() # confusion_matrix_validation = ConfusionMeter(2) vis = visdom.Visdom() draw_graph = None draw_accuracy = None draw_roc_curve = None csv_file = "classifications.csv" root_dir = "data/" # hyperparameters batch_size = 159 learning_rate = 0.001 epoch_num = 50 # experiment parameters real_exp = True
def __init__(self, device):
self.device = device
# super(AllInOneMeter, self).__init__()
self.out1auc1 = AUCMeter()
self.out1auc2 = AUCMeter()
self.out1auc3 = AUCMeter()
self.out1auc4 = AUCMeter()
self.out1auc5 = AUCMeter()
self.out2auc1 = AUCMeter()
self.out2auc2 = AUCMeter()
self.out2auc3 = AUCMeter()
self.out2auc4 = AUCMeter()
self.out2auc5 = AUCMeter()
self.loss1 = []
self.loss2 = []
self.loss3 = []
self.loss = []
self.jaccard = []
# self.nbatch = 0
self.intersection = torch.zeros([5],
dtype=torch.float,
device=self.device)
self.union = torch.zeros([5], dtype=torch.float, device=self.device)
self.reset()
class AllInOneMeter(object):
"""
All in one meter: AUC
"""
def __init__(self, device):
self.device = device
# super(AllInOneMeter, self).__init__()
self.out1auc1 = AUCMeter()
self.out1auc2 = AUCMeter()
self.out1auc3 = AUCMeter()
self.out1auc4 = AUCMeter()
self.out1auc5 = AUCMeter()
self.out2auc1 = AUCMeter()
self.out2auc2 = AUCMeter()
self.out2auc3 = AUCMeter()
self.out2auc4 = AUCMeter()
self.out2auc5 = AUCMeter()
self.loss1 = []
self.loss2 = []
self.loss3 = []
self.loss = []
self.jaccard = []
# self.nbatch = 0
self.intersection = torch.zeros([5],
dtype=torch.float,
device=self.device)
self.union = torch.zeros([5], dtype=torch.float, device=self.device)
self.reset()
def reset(self):
# self.scores = torch.DoubleTensor(torch.DoubleStorage()).numpy()
# self.targets = torch.LongTensor(torch.LongStorage()).numpy()
self.out1auc1.reset()
self.out1auc2.reset()
self.out1auc3.reset()
self.out1auc4.reset()
self.out1auc5.reset()
self.out2auc1.reset()
self.out2auc2.reset()
self.out2auc3.reset()
self.out2auc4.reset()
self.out2auc5.reset()
self.loss1 = []
self.loss2 = []
self.loss3 = []
self.loss = []
self.jaccard = []
self.intersection = torch.zeros([5],
dtype=torch.float,
device=self.device)
self.union = torch.zeros([5], dtype=torch.float, device=self.device)
# self.nbatch = 0
def add(self, mask_prob, true_mask, mask_ind_prob1, mask_ind_prob2,
true_mask_ind, loss1, loss2, loss3, loss):
self.out1auc1.add(mask_ind_prob1[:, 0].data, true_mask_ind[:, 0].data)
self.out1auc2.add(mask_ind_prob1[:, 1].data, true_mask_ind[:, 1].data)
self.out1auc3.add(mask_ind_prob1[:, 2].data, true_mask_ind[:, 2].data)
self.out1auc4.add(mask_ind_prob1[:, 3].data, true_mask_ind[:, 3].data)
self.out1auc5.add(mask_ind_prob1[:, 4].data, true_mask_ind[:, 4].data)
self.out2auc1.add(mask_ind_prob2[:, 0].data, true_mask_ind[:, 0].data)
self.out2auc2.add(mask_ind_prob2[:, 1].data, true_mask_ind[:, 1].data)
self.out2auc3.add(mask_ind_prob2[:, 2].data, true_mask_ind[:, 2].data)
self.out2auc4.add(mask_ind_prob2[:, 3].data, true_mask_ind[:, 3].data)
self.out2auc5.add(mask_ind_prob2[:, 4].data, true_mask_ind[:, 4].data)
self.loss1.append(loss1)
self.loss2.append(loss2)
self.loss3.append(loss3)
self.loss.append(loss)
# self.nbatch += true_mask.shape[0]
y_pred = (mask_prob > 0.3).type(true_mask.dtype)
y_true = true_mask
self.intersection += (y_pred *
y_true).sum(dim=-2).sum(dim=-1).sum(dim=0)
self.union += y_true.sum(dim=-2).sum(dim=-1).sum(dim=0) + y_pred.sum(
dim=-2).sum(dim=-1).sum(dim=0)
def value(self):
jaccard_array = (self.intersection / (self.union - self.intersection))
# jaccard_array = jaccard_array.data.cpu().numpy()
jaccard = jaccard_array.mean()
metrics = {
'out1auc1': self.out1auc1.value()[0],
'out1auc2': self.out1auc2.value()[0],
'out1auc3': self.out1auc3.value()[0],
'out1auc4': self.out1auc4.value()[0],
'out1auc5': self.out1auc5.value()[0],
'out2auc1': self.out2auc1.value()[0],
'out2auc2': self.out2auc2.value()[0],
'out2auc3': self.out2auc3.value()[0],
'out2auc4': self.out2auc4.value()[0],
'out2auc5': self.out2auc5.value()[0],
'loss1': np.mean(self.loss1),
'loss2': np.mean(self.loss2),
'loss3': np.mean(self.loss3),
'loss': np.mean(self.loss),
'jaccard': jaccard.item(),
'jaccard1': jaccard_array[0].item(),
'jaccard2': jaccard_array[1].item(),
'jaccard3': jaccard_array[2].item(),
'jaccard4': jaccard_array[3].item(),
'jaccard5': jaccard_array[4].item(),
}
return metrics
def test(segment_size_list):
if test_dataset == 'MTAT':
test_list_pub = pickle.load(open(os.path.join(MTAT_SPLIT_FOLDER, 'test_list_pub.cP'), 'rb'))
if test_dataset == 'MSD':
id7d_to_path = pickle.load(open(os.path.join(MSD_SPLIT_FOLDER, '7D_id_to_path.pkl'), 'rb'))
idmsd_to_id7d = pickle.load(
open(os.path.join(MSD_SPLIT_FOLDER, 'MSD_id_to_7D_id.pkl'), 'rb'))
test_list_pub_id = pickle.load(
open(os.path.join(MSD_SPLIT_FOLDER, 'filtered_list_test.cP'), 'rb'))
test_list_pub = [id7d_to_path[idmsd_to_id7d[song]][:-9] + '.npy' for song in test_list_pub_id]
del id7d_to_path, idmsd_to_id7d
total_test_size = len(test_list_pub)
n_inputs = 0
for segment_size in segment_size_list:
if segment_size == 18:
n_inputs += 512
if segment_size == 27:
n_inputs += 512
if segment_size == 54:
n_inputs += 768
if segment_size == 108:
n_inputs += 1024
if segment_size == 216:
n_inputs += 1280
local_models = []
for segment_size in segment_size_list:
loc_model = local_model(segment_size).cuda()
loc_model.load_state_dict(os.path.join(ENCODER_FOLDER, torch.load('local_model_' + str(segment_size) + '.pt')))
loc_model.eval()
local_models.append(loc_model)
model = global_model(n_inputs, 512).cuda()
model.load_state_dict(torch.load(os.path.join(ENCODER_FOLDER, 'global_model_18_27_54_9051_123.pt')))
model.eval()
auc = AUCMeter()
for start in range(0, total_test_size, n_songs):
print("Loading dataset...", start)
if test_dataset == 'MTAT':
test_features = np.concatenate(
[np.load(os.path.join(MTAT_NPY_FOLDER, 'testing/' + test_list_pub[i])) for i in
range(start, min(start + n_songs, total_test_size))])
test_labels = np.load(
os.path.join(MTAT_SPLIT_FOLDER, 'y_test_pub.npy'))[start:min(start + n_songs, total_test_size)]
if test_dataset == 'MSD':
test_features = np.concatenate(
[np.expand_dims(np.load(os.path.join(MSD_NPY_FOLDER, 'testing/' + test_list_pub[i]))[:, :1255], axis=0)
for i in range(start, min(start + n_songs, total_test_size))])
idmsd_to_tag = pickle.load(
open(os.path.join(MSD_SPLIT_FOLDER, 'msd_id_to_tag_vector.cP'), 'rb'))
test_labels = np.concatenate(
[idmsd_to_tag[idmsd] for idmsd in test_list_pub_id[start:min(start + n_songs, total_test_size)]],
axis=1)
if normalization:
mean = np.mean(test_features, axis=0)
var = np.var(test_features, axis=0)
test_features = (test_features - mean) / np.sqrt(var)
test_data = CustomDataset(test_features, test_labels)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=batch_size)
print("Dataset loaded")
for data, labels in test_loader:
X = Variable(data).cuda()
X = torch.cat([loc_model(X)[1] for loc_model in local_models], dim=1)
out, _ = model(X)
auc_out = np.reshape(out.data.cpu().numpy(), -1)
auc_target = np.reshape(labels, -1)
auc.add(auc_out, auc_target)
del test_features, test_labels, test_data, test_loader
auc_tuple = auc.value()
print("AUC = ", auc_tuple[0])
def __init__(self,
dataset,
r,
noise_mode,
root_dir,
transform,
mode,
noise_file='',
clean_file='',
pred=[],
probability=[],
log=''):
self.r = r # noise ratio
self.transform = transform
self.noise_mode = noise_mode
self.mode = mode
self.transition = {
0: 0,
2: 0,
4: 7,
7: 7,
1: 1,
9: 1,
3: 5,
5: 3,
6: 6,
8: 8
} # class transition for asymmetric noise
if self.mode == 'test':
if dataset == 'cifar10':
test_dic = unpickle('%s/data/cifar-10-batches-py/test_batch' %
root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
elif dataset == 'cifar100':
test_dic = unpickle('%s/data/cifar-100-python/test' % root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
else:
train_data = []
train_label = []
if dataset == 'cifar10':
#print("current path is {}".format(sys.path[0]))
for n in range(1, 6):
dpath = '%s/data/cifar-10-batches-py/data_batch_%d' % (
root_dir, n)
#print("path is {}".format(dpath))
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label + data_dic['labels']
train_data = np.concatenate(train_data)
elif dataset == 'cifar100':
train_dic = unpickle('%s/data/cifar-100-python/train' %
root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
train_label = np.array(train_label)
noise_label = train_label.copy()
if dataset == 'cifar10':
nb_classes = 10
elif dataset == 'cifar100':
nb_classes = 100
clean_per_class = int(5000 / nb_classes) # cifar10: 100 else: 10
noise_per_class = int(50000 / nb_classes * r)
#select clean_per_class numbers of data in each class as clean data
#leave the other data to add noise
#the 0th data processing is at the outer loop
#0th add noise (for index)
all_index = np.arange(50000).reshape(-1)
clean_indices = all_index[np.where(
train_label == 0)[0]][-clean_per_class:]
noise_idx = [
all_index[np.where(train_label == 0)[0]][:-clean_per_class]
]
#from 1th to 9th to add noise (for index)
for i in range(nb_classes - 1):
indices1 = all_index[np.where(train_label == i +
1)[0]][-clean_per_class:]
noisy_indices1 = all_index[np.where(train_label == i +
1)[0]][:-clean_per_class]
clean_indices = np.concatenate((clean_indices, indices1))
noise_idx.append(noisy_indices1)
#add noise
for t, i in enumerate(noise_idx):
# randomly selected one image as the center
image_center = train_data[i[10]]
norm_loss = np.zeros(len(i))
for j, k in enumerate(i):
images = train_data[k]
norm_loss[j] = np.linalg.norm(image_center - images)
noisy_indices = i[norm_loss.argsort()[:noise_per_class]]
noise_label[noisy_indices] = (t + 1) % nb_classes
if self.mode == 'all':
self.train_data = train_data
self.noise_label = noise_label
elif self.mode == 'small':
self.train_data = train_data[::100]
self.noise_label = noise_label[::100]
else:
if self.mode == "labeled":
pred_idx = pred.nonzero()[0]
self.probability = [probability[i] for i in pred_idx]
#clean = (np.array(noise_label)==np.array(train_label))
clean = (noise_label == train_label)
auc_meter = AUCMeter()
auc_meter.reset()
auc_meter.add(probability, clean)
auc, _, _ = auc_meter.value()
log.write('Numer of labeled samples:%d AUC:%.3f\n' %
(pred.sum(), auc))
log.flush()
elif self.mode == "unlabeled":
pred_idx = (1 - pred).nonzero()[0]
self.train_data = train_data[pred_idx]
self.noise_label = noise_label[pred_idx]
print("%s data has a size of %d" %
(self.mode, len(self.noise_label)))
