def predict(self, data_loader):
n = len(data_loader) * data_loader.batch_size
Yhat = np.zeros((n, self.num_output_channels))
Ids = np.zeros((n, 1))
k = 0
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
end = time.time()
for i, (Id, inputs) in enumerate(tqdm(data_loader)):
x = inputs.cuda() if self.cuda else inputs
# fit (forward)
yhat = self.net(x)
yhat = F.softmax(yhat, dim=1)
yhat = pytutils.to_np(yhat)
for j in range(yhat.shape[0]):
Yhat[k, :] = yhat[j]
Ids[k] = Id[j]
k += 1
Yhat = Yhat[:k, :]
Ids = Ids[:k]
return Ids, Yhat
def test(self, data_loader):
masks = []
ids = []
k = 0
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
end = time.time()
for i, sample in enumerate(tqdm(data_loader)):
# get data (image, label)
inputs, meta = sample['image'], sample['metadata']
idd = meta[:, 0]
x = inputs.cuda() if self.cuda else inputs
# fit (forward)
yhat = self.net(x)
yhat = F.softmax(yhat, dim=1)
yhat = pytutils.to_np(yhat)
masks.append(yhat)
ids.append(idd)
return ids, masks
def test(self, data_loader):
n = len(data_loader) * data_loader.batch_size
Yhat = np.zeros((n, self.num_output_channels))
Y = np.zeros((n, 1))
k = 0
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
end = time.time()
for i, sample in enumerate(tqdm(data_loader)):
# get data (image, label)
x, y = sample['image'], sample['label'].argmax(1).long()
x = x.cuda() if self.cuda else x
# fit (forward)
yhat = self.net(x)
yhat = F.softmax(yhat, dim=1)
yhat = pytutils.to_np(yhat)
for j in range(yhat.shape[0]):
Y[k] = y[j]
Yhat[k, :] = yhat[j]
k += 1
Yhat = Yhat[:k, :]
Y = Y[:k]
return Yhat, Y
def predict(self, data_loader):
n = len(data_loader) * data_loader.batch_size
Yhat = np.zeros((n, self.num_output_channels))
Ids = np.zeros((n, 1))
k = 0
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
end = time.time()
for i, (Id, inputs) in enumerate(tqdm(data_loader)):
# get data (image, label)
#inputs = sample['image']
#Id = sample['id']
x = inputs.cuda() if self.cuda else inputs
x = Variable(x, requires_grad=False, volatile=True)
# fit (forward)
yhat = self.net(x)
yhat = F.softmax(yhat, dim=1)
yhat = pytutils.to_np(yhat)
for j in range(yhat.shape[0]):
Yhat[k, :] = yhat[j]
Ids[k] = Id[j]
k += 1
Yhat = Yhat[:k, :]
Ids = Ids[:k]
return Ids, Yhat
def __call__(self, image):
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
x = image.cuda() if self.cuda else image
msoft = nn.Softmax()
yhat = msoft(self.net(x))
yhat = pytutils.to_np(yhat)
return yhat
def __call__(self, image):
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
x = image.cuda() if self.cuda else image
yhat = F.softmax(self.net(x), dim=1)
yhat = pytutils.to_np(yhat).transpose(2, 3, 1, 0)[..., 0]
return yhat
def __call__(self, image):
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
x = image.cuda() if self.cuda else image
x = Variable(x, requires_grad=False, volatile=True)
msoft = nn.Softmax()
yhat = msoft(self.net(x))
yhat = pytutils.to_np(yhat)
return yhat
def representation(self, data_loader):
""""
Representation
-data_loader: simple data loader for image
"""
# switch to evaluate mode
self.net.eval()
n = len(data_loader) * data_loader.batch_size
k = 0
# embebed features
embX = np.zeros([n, self.net.dim])
embY = np.zeros([n, 1])
batch_time = AverageMeter()
end = time.time()
for i, sample in enumerate(data_loader):
# get data (image, label)
x, y = sample['image'], sample['label'].argmax(1).long()
x = x.cuda() if self.cuda else x
# representation
emb = self.net.representation(x)
emb = pytutils.to_np(emb)
for j in range(emb.shape[0]):
embX[k, :] = emb[j, :]
embY[k] = y[j]
k += 1
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
print(
'Representation: |{:06d}/{:06d}||{batch_time.val:.3f} ({batch_time.avg:.3f})|'
.format(i, len(data_loader), batch_time=batch_time))
embX = embX[:k, :]
embY = embY[:k]
return embX, embY
def test(self, data_loader):
n = len(data_loader) * data_loader.batch_size
Yhat = np.zeros((n, self.num_output_channels))
Y = np.zeros((n, 1))
k = 0
# switch to evaluate mode
self.net.eval()
with torch.no_grad():
end = time.time()
for i, sample in enumerate(tqdm(data_loader)):
# get data (image, label)
inputs = sample['image']
targets = pytutils.argmax(sample['label'])
x = inputs.cuda() if self.cuda else inputs
x = Variable(x, requires_grad=False, volatile=True)
# fit (forward)
yhat = self.net(x)
yhat = F.softmax(yhat, dim=1)
yhat = pytutils.to_np(yhat)
for j in range(yhat.shape[0]):
Y[k] = targets[j]
Yhat[k, :] = yhat[j]
k += 1
#print( 'Test:', i , flush=True )
Yhat = Yhat[:k, :]
Y = Y[:k]
return Yhat, Y
