print("datasetSize:", len(train_dataset))
for i_epoach in range(nepoach):
for i_batch, (img, label, name) in enumerate(train_loader):
img = img.cuda()
img = img.type(torch.cuda.FloatTensor)
label = label.cuda()
output = model(img)
loss = [crossEntropyLoss(x, label) for x in output]
nploss = [x.cpu().data.numpy() for x in loss]
losss.append(nploss)
if len(losss) == 10:
print(i_epoach, i_batch, np.average(losss))
losss.clear()
loss_total = None
for i, l in enumerate(loss):
if i == 0:
loss_total = l
else:
loss_total = loss_total + l
loss_total.backward()
optimizer.step()
optimizer.zero_grad()
pth = './model_resnet/net_%d.pth' % (i_epoach)
print('save to', pth)
torch.save(model.state_dict(), pth)
benchmark.eval(model, "/home/afan/tumor/data/train_refine_v5",
"val.csv")
#train_dataset.csvData = train_dataset.csvData[:100]
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=8,
num_workers=8,
drop_last=False,
shuffle=True)
frate, trate = train_dataset.distribution
print("data distribution frate %f trate %f" % (frate, trate))
optimizer = torch.optim.SGD(model.parameters(), lr=3e-4, momentum=0.9)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 10, 0.3)
nepoach = 30
losss = []
print("datasetSize:", len(train_dataset))
benchmark.eval(model,
"../data/train_img_refine",
"val.csv",
dataTransform=utils.getTemplateData)
for i_epoach in range(nepoach):
if i_epoach == 5:
for p in model.featureExtractor.parameters():
p.requires_grad = True
for i_batch, (img, label) in enumerate(train_loader):
img = utils.getTemplateData(img)
img = img.cuda()
label = label.cuda()
output, _ = model(img)
loss = [crossEntropyLoss(x, label, frate, trate) for x in output]
nploss = [x.cpu().data.numpy() for x in loss]
losss.append(nploss)
# quit()
#train_dataset.csvData = train_dataset.csvData[:100]
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=8,
num_workers=8,
drop_last=False,
shuffle=True)
frate, trate = train_dataset.distribution
print("data distribution frate %f trate %f" % (frate, trate))
optimizer = torch.optim.SGD(model.parameters(), lr=3e-4, momentum=0.9)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 10, 0.3)
nepoach = 30
losss = []
print("datasetSize:", len(train_dataset))
benchmark.eval(model, "../data/train_seg", "val.csv")
for i_epoach in range(nepoach):
for i_batch, (img, label) in enumerate(train_loader):
#continue
img = img.cuda()
label = label.cuda()
output, _ = model(img)
loss = [crossEntropyLoss(x, label) for x in output]
nploss = [x.cpu().data.numpy() for x in loss]
losss.append(nploss)
if len(losss) == 10:
print(i_epoach, i_batch, np.average(losss))
losss.clear()
# if i_epoach == 0 and i_batch < 100:
# quit()
#train_dataset.csvData = train_dataset.csvData[:100]
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=8,
num_workers=10,
drop_last=False,
shuffle=True)
frate, trate = train_dataset.distribution
print("data distribution frate %f trate %f" % (frate, trate))
optimizer = torch.optim.SGD(model.parameters(), lr=3e-5, momentum=0.9)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 10, 0.3)
nepoach = 30
losss = []
print("datasetSize:", len(train_dataset))
benchmark.eval(model, "../data/train_img_refine", "val.csv")
for i_epoach in range(nepoach):
for i_batch, (img, label) in enumerate(train_loader):
img = img.cuda()
label = label.cuda()
output, outmap = model(img)
loss = [crossEntropyLoss(x, label) for x in output]
nploss = [x.cpu().data.numpy() for x in loss]
#outmap = [x.cpu().data.numpy() for x in outmap]
# for i in range(len(outmap[0])):
# m = [cv2.resize(outmap[j][i][0],(224,224)) for j in range(len(outmap))]
# cv2.imwrite("./output/%5d.jpg"%(random.randrange(0,1000)),(np.concatenate(m,axis=1)*255).astype(np.uint8))
losss.append(nploss)
if len(losss) == 10:
print(i_epoach, i_batch, np.average(losss))
img = img.cuda()
label = label.cuda()
output, _ = model(img)
loss = [crossEntropyLoss(x, label) for x in output]
nploss = [x.cpu().data.numpy() for x in loss]
losss.append(nploss)
if len(losss) == 10:
print(i_epoach, i_batch, np.average(losss))
losss.clear()
# if i_epoach == 0 and i_batch < 100:
# loss = loss*0.1
loss_total = None
for i, l in enumerate(loss):
if i == 0:
loss_total = l
else:
loss_total = loss_total + l
loss_total.backward()
optimizer.step()
optimizer.zero_grad()
pth = './model/net_%d.pth' % (i_epoach)
print('save to', pth)
torch.save(model.state_dict(), pth)
benchmark.eval(model,
"../data/train_seg",
"val.csv",
dataTransform=utils.toCubeData)
scheduler.step()