def eval_recall(testloader, net, gpus):
net.eval()
total_num = 0
tp = 0
prep = 0
labelp = 0
acc = 0
for i, data in enumerate(testloader):
with torch.no_grad():
if gpus>0:
img, label = data[0].cuda(), data[1].cuda()
else:
img, label = data[0], data[1]
total_num += img.size(0)
pre = net(img)
_, prelabel = torch.max(pre, 1)
acc += torch.sum(prelabel.data == label.data)
prep += torch.sum(prelabel.data==1)
labelp += torch.sum(label.data==1)
tp += torch.sum((prelabel+label).data == 2)
progress_bar(i, len(testloader), 'val acc')
test_acc = float(acc)/total_num
recall = float(tp.item())/(labelp.item() + 1e-6)
precision = float(tp.item())/(prep.item()+1e-6)
return test_acc, recall, precision
def eval_fuse(testloader, net, gpus):
fuse_matrix = np.zeros((NUMCLASS, NUMCLASS))
net.eval()
total_num = 0
acc = 0
for i, data in enumerate(testloader):
with torch.no_grad():
if gpus>0:
img, label = data[0].cuda(), data[1].cuda()
else:
img, label = data[0], data[1]
total_num += img.size(0)
pre = net(img)
_, prelabel = torch.max(pre, 1)
for i in range(NUMCLASS):
for j in range(NUMCLASS):
fuse_matrix[i][j] += torch.sum((label.data==i)&(prelabel.data==j)).item()
acc += torch.sum(prelabel.data == label.data)
progress_bar(i, len(testloader), 'val acc')
test_acc = float(acc)/total_num
recalldic = {}
precisiondic = {}
for i in range(NUMCLASS):
t = fuse_matrix[i][i]
prenum = np.sum(fuse_matrix[:, i])
num = np.sum(fuse_matrix[i,:])
recalldic[i] = t/num
precisiondic[i] = t/prenum
print(fuse_matrix)
return test_acc, recalldic, precisiondic
def eval_multi_fuse(testloader, net, gpus):
fuse_classi_matrix = np.zeros((NUMCLASS, NUMCLASS))
fuse_label_matrix = np.zeros((2, 2))
net.eval()
total_num = 0
acc = 0
label_acc = 0
for i, data in enumerate(testloader):
with torch.no_grad():
if gpus>0:
img, label = data[0].cuda(), data[1].cuda()
else:
img, label = data[0], data[1]
total_num += img.size(0)
classipre, binpre = net(img)
pre = [classipre, binpre]
_, prelabel = torch.max(pre[0], 1)
_, pretlabel = torch.max(pre[1], 1)
for i in range(NUMCLASS):
for j in range(NUMCLASS):
fuse_classi_matrix[i][j] += torch.sum((label[:, 0].data==i)&(prelabel.data==j)).item()
acc += torch.sum(prelabel.data == label[:, 0].data)
for p in range(2):
for q in range(2):
fuse_label_matrix[p][q] += torch.sum((label[:, 1].data==p)&(pretlabel==q)).item()
label_acc += torch.sum(pretlabel.data==label[:, 1].data)
progress_bar(i, len(testloader), 'val acc')
test_acc = float(acc)/total_num
test_label_acc = float(label_acc)/total_num
print(test_label_acc)
recalldic = {}
precisiondic = {}
for i in range(NUMCLASS):
t = fuse_classi_matrix[i][i]
prenum = np.sum(fuse_classi_matrix[:, i])
num = np.sum(fuse_classi_matrix[i,:])
recalldic[i] = t/num
precisiondic[i] = t/prenum
print(fuse_classi_matrix)
print(fuse_label_matrix)
return test_acc, recalldic, precisiondic
img, label = data[0].cuda(), data[1].cuda()
batch_size = img.size(0)
optimizer.zero_grad()
if MIXUP:
img, labela, labelb, lam = mixup_data(img, label)
pre = net(img)
criterion = torch.nn.CrossEntropyLoss()
loss = mixup_criterion(criterion, pre, labela, labelb, lam)
else:
pre = net(img)
loss = torch.nn.CrossEntropyLoss()(pre, label)
train_loss += loss * batch_size
total += batch_size
loss.backward()
optimizer.step()
progress_bar(i, len(trainloader), 'train')
if epoch > WARM:
scheduler.step()
else:
warmup.step()
print('[*] epoch:{} - train loss: {:.3f}'.format(epoch,
train_loss / total))
acc, recalldic, precisiondic = eval_fuse(testloader, net,
torch.cuda.device_count())
if acc > best_acc:
best_acc = acc
best_epoch = epoch
print('[*] epoch:{} - test acc: {:.3f} - best acc: {}_{}'.format(
epoch, acc, best_epoch, best_acc))
displaymetric(recalldic, precisiondic)
# save train_loss and val_acc