def predict():
test_data = cancerloader.load_cancer_data()
torch.nn.Module.dump_patches = True
model_name = "ubploss/transformer_res18-128_size0_autoloss_alpha16"
model_dir = os.path.join("./modeldir/%s" % model_name)
model_pth = os.path.join(model_dir, "model.pth")
model = torch.load(model_pth).cuda()
model.eval()
with torch.no_grad():
all_gene = []
all_gt = []
all_pd = []
for item in cancerloader.batch_fv(test_data, len(test_data)):
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
pred = model(inputs)
test_pd = torch_util.threshold_tensor_batch(pred)
all_gene.extend(genes)
all_gt.extend(labels.astype(np.int))
all_pd.extend(np.array(test_pd))
df = pd.DataFrame({'gene': all_gene, 'gt': all_gt, 'pd': all_pd})
df.to_csv("cancer.txt", header=True, index=False, sep=',')
def run_val(rnn, val_data, writer, val_step, criterion):
print("------run val-----------", val_step)
rnn.eval()
with torch.no_grad():
st = time.time()
# for item in fvloader.batch_fv(val_data, len(val_data)):
for item in matloader.batch_fv(val_data, len(val_data)):
hidden = None
genes, nimgs, labels, timesteps = item
idx = np.argsort(np.array(-timesteps))
s_nimgs = torch.from_numpy(np.stack(nimgs[idx])).type(
torch.cuda.FloatTensor)
s_labels = torch.from_numpy(labels[idx]).type(
torch.cuda.FloatTensor)
s_timesteps = timesteps[idx]
out_pack, hidden = rnn(s_nimgs, s_timesteps)
loss = criterion(out_pack, s_labels)
writer.add_scalar("val loss", loss.item(), val_step)
val_pd = torch_util.threshold_tensor_batch(out_pack)
np_pd = val_pd.data.cpu().numpy()
lab_f1_macro = torch_util.torch_metrics(labels[idx], np_pd, writer,
val_step)
et = time.time()
writer.add_scalar("val time", et - st, val_step)
return loss.item(), lab_f1_macro
def run_val(model, val_dataset, writer, val_step, criterion, dcfg):
model.eval()
with torch.no_grad():
val_loader = DataLoader(val_dataset,
batch_size=dcfg['bsize'],
shuffle=False,
num_workers=dcfg['nworker'],
collate_fn=dcfg['collate'])
tot_loss = 0.0
np_label = []
np_pd = []
for i_batch, sample_batched in enumerate(val_loader):
(img, label) = sample_batched
inputs = img.type(torch.cuda.FloatTensor)
gt = label.type(torch.cuda.FloatTensor)
pd = model(inputs)
loss = criterion(pd, gt)
tot_loss += loss
val_pd = torch_util.threshold_tensor_batch(pd)
np_pd.append(val_pd.data.cpu().numpy())
np_label.append(gt.data.cpu().numpy())
np_label = np.concatenate(np_label)
np_pd = np.concatenate(np_pd)
tot_loss = tot_loss / len(val_loader)
writer.add_scalar("val loss", tot_loss.item(), val_step)
torch_util.torch_metrics(np_label, np_pd, writer, val_step)
return tot_loss.item()
def run_val(model, dloader, val_data, writer, val_step, criterion):
print("------run val-----------", val_step)
model.eval()
with torch.no_grad():
st = time.time()
for item in dloader.batch_fv(val_data, len(val_data)):
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
gt = torch.from_numpy(labels).type(torch.cuda.FloatTensor)
pd = model(inputs)
# loss = criterion(pd, gt)
# criterion = torch.nn.BCELoss(reduction='none')
all_loss = criterion(pd, gt)
label_loss = torch.mean(all_loss, dim=0)
loss = torch.mean(label_loss)
for i in range(6):
writer.add_scalar("val sl_%d_loss" % i,
label_loss[i].item(), val_step)
writer.add_scalar("val loss", loss.item(), val_step)
val_pd = torch_util.threshold_tensor_batch(pd)
# val_pd = torch.ge(pd, 0.5)
np_pd = val_pd.data.cpu().numpy()
lab_f1_macro = torch_util.torch_metrics(
labels, np_pd, writer, val_step)
et = time.time()
writer.add_scalar("val time", et - st, val_step)
return loss.item(), lab_f1_macro
def bag_mislocation():
genes = [
'ENSG00000128342', 'ENSG00000023734', 'ENSG00000105705',
'ENSG00000111011', 'ENSG00000116863', 'ENSG00000197265',
'ENSG00000160208', 'ENSG00000159352', 'ENSG00000115539',
'ENSG00000163946', 'ENSG00000116161'
]
torch.nn.Module.dump_patches = True
model_name = "ubploss/transformer_res18-128_size0_autoloss_alpha16"
model_dir = os.path.join("./modeldir/%s" % model_name)
model_pth = os.path.join(model_dir, "model.pth")
model = torch.load(model_pth).cuda()
cfv_dir = '/ndata/longwei/hpa/cancerfv_4tissue'
nfv_dir = '/ndata/longwei/hpa/tissuefv/res18_128'
model.eval()
with torch.no_grad():
for g in genes:
ngp = os.path.join(nfv_dir, '%s.npy' % g)
cgp = os.path.join(cfv_dir, '%s.npy' % g)
if not os.path.exists(ngp):
print("normal fv for %s not exists" % g)
continue
if not os.path.exists(cgp):
print("cancer fv for %s not exists" % g)
continue
npy = np.expand_dims(np.load(ngp), axis=0)
nin = torch.from_numpy(npy).type(torch.cuda.FloatTensor)
npd = model(nin)
test_npd = torch_util.threshold_tensor_batch(npd)
time.sleep(1)
cpy = np.expand_dims(np.load(cgp), axis=0)
cin = torch.from_numpy(cpy).type(torch.cuda.FloatTensor)
cpd = model(cin)
test_cpd = torch_util.threshold_tensor_batch(cpd)
print("gene", g)
print("normal pd", test_npd.data.cpu().numpy())
print("cancer pd", test_cpd.data.cpu().numpy())
def run_test(model, dloader, test_data, result):
model.eval()
with torch.no_grad():
for item in dloader.batch_fv(test_data, len(test_data)):
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
pd = model(inputs)
test_pd = torch_util.threshold_tensor_batch(pd)
np_pd = test_pd.data.cpu().numpy()
npmetrics.write_metrics(labels, np_pd, result)
def run_val(model, dloader, val_data, writer, val_step, criterion):
print("------run val-----------", val_step)
model.eval()
with torch.no_grad():
st = time.time()
# loss = 0
# lab_f1_macro = 0
num = 0
gt = torch.from_numpy(np.array([[0 for _ in range(10)]
])).type(torch.cuda.FloatTensor)
pd = torch.from_numpy(np.array([[0 for _ in range(10)]
])).type(torch.cuda.FloatTensor)
for item in dloader.batch_fv(val_data, batch=1):
num += 1
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
gt = torch.cat(
(gt, torch.from_numpy(labels).type(torch.cuda.FloatTensor)))
# print(gt)
pd = torch.cat((pd, model(inputs)))
# print(pd)
# loss = criterion(pd, gt)
# criterion = torch.nn.BCELoss(reduction='none')
gt = gt[1:, :]
pd = pd[1:, :]
all_loss = criterion(pd, gt)
label_loss = torch.mean(all_loss, dim=0)
loss = torch.mean(label_loss)
for i in range(num_class):
writer.add_scalar("val sl_%d_loss" % i, label_loss[i].item(),
val_step)
writer.add_scalar("val loss", loss.item(), val_step)
val_pd = torch_util.threshold_tensor_batch(pd)
# val_pd = torch.ge(pd, 0.5)
np_pd = val_pd.data.cpu().numpy() # hard predicts
lab_f1_macro = torch_util.torch_metrics(gt.cpu().numpy(), np_pd,
writer, val_step)
# print(labels, pd.cpu().numpy())
auc = evaluate.auc(gt.cpu().numpy(), pd.cpu().numpy())
mif1 = evaluate.micro_f1(gt.cpu().numpy(), np_pd)
maf1 = evaluate.macro_f1(gt.cpu().numpy(), np_pd)
et = time.time()
writer.add_scalar("val time", et - st, val_step)
return loss.item(), lab_f1_macro, auc, mif1, maf1
def run_test(rnn, test_data, result):
rnn.eval()
with torch.no_grad():
# for item in fvloader.batch_fv(test_data, len(test_data)):
for item in matloader.batch_fv(test_data, len(test_data)):
genes, nimgs, labels, timesteps = item
idx = np.argsort(np.array(-timesteps))
s_nimgs = torch.from_numpy(np.stack(nimgs[idx])).type(
torch.cuda.FloatTensor)
s_timesteps = timesteps[idx]
out_pack, hidden = rnn(s_nimgs, s_timesteps)
test_pd = torch_util.threshold_tensor_batch(out_pack)
npmetrics.write_metrics(labels[idx], np.array(test_pd), result)
def predict_cancer_gene(model, gene):
cfv_dir = '/ndata/longwei/hpa/cancerfv_4tissue'
cgp = os.path.join(cfv_dir, '%s.npy' % gene)
if not os.path.exists(cgp):
print("cancer fv for %s not exists" % gene)
return [0] * 6
model.eval()
with torch.no_grad():
cpy = np.expand_dims(np.load(cgp), axis=0)
cin = torch.from_numpy(cpy).type(torch.cuda.FloatTensor)
cpd = model(cin)
test_cpd = torch_util.threshold_tensor_batch(cpd)
return test_cpd.data.cpu().numpy()
def predict_normal_gene(model, gene):
nfv_dir = '/ndata/longwei/hpa/tissuefv/res18_128'
ngp = os.path.join(nfv_dir, '%s.npy' % gene)
if not os.path.exists(ngp):
print("normal fv for %s not exists" % gene)
return [0] * 6
model.eval()
with torch.no_grad():
npy = np.expand_dims(np.load(ngp), axis=0)
nin = torch.from_numpy(npy).type(torch.cuda.FloatTensor)
npd = model(nin)
test_npd = torch_util.threshold_tensor_batch(npd)
time.sleep(1)
return test_npd.data.cpu().numpy()
def run_test(model, test_dataset, result, dcfg):
model.eval()
with torch.no_grad():
test_loader = DataLoader(test_dataset,
batch_size=dcfg['bsize'],
shuffle=False,
num_workers=dcfg['nworker'],
collate_fn=dcfg['collate'])
np_label = []
np_pd = []
for i_batch, sample_batched in enumerate(test_loader):
(img, label) = sample_batched
inputs = img.type(torch.cuda.FloatTensor)
gt = label.type(torch.cuda.FloatTensor)
pd = model(inputs)
test_pd = torch_util.threshold_tensor_batch(pd)
np_pd.append(test_pd.data.cpu().numpy())
np_label.append(gt.data.cpu().numpy())
np_label = np.concatenate(np_label)
np_pd = np.concatenate(np_pd)
npmetrics.write_metrics(np_label, np_pd, result)
def run_test(model, dloader, test_data, result):
model.eval()
with torch.no_grad():
gt = np.array([[0 for _ in range(10)]])
pd = np.array([[0 for _ in range(10)]])
for item in dloader.batch_fv(test_data, batch=1):
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
# print(gt)
apd = model(inputs)
test_pd = torch_util.threshold_tensor_batch(apd)
np_pd = test_pd.data.cpu().numpy()
gt = np.concatenate((gt, labels))
pd = np.concatenate((pd, np_pd))
gt = gt[1:, :]
pd = pd[1:, :]
npmetrics.write_metrics(gt, pd, result)
def run_origin_train(model, dloader, imbtrain_data, writer, step, criterion):
print("------run origin imblance train data-----------", step)
model.eval()
with torch.no_grad():
st = time.time()
for item in dloader.batch_fv(imbtrain_data, len(imbtrain_data)):
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
gt = torch.from_numpy(labels).type(torch.cuda.FloatTensor)
pd = model(inputs)
# loss = criterion(pd, gt)
# criterion = torch.nn.BCELoss(reduction='none')
all_loss = criterion(pd, gt)
label_loss = torch.mean(all_loss, dim=0)
loss = torch.mean(label_loss)
for i in range(num_class):
writer.add_scalar("origin sl_%d_loss" % i,
label_loss[i].item(), step)
writer.add_scalar("origin loss", loss.item(), step)
val_pd = torch_util.threshold_tensor_batch(pd)
# val_pd = torch.ge(pd, 0.5)
np_pd = val_pd.data.cpu().numpy()
torch_util.torch_metrics(labels,
np_pd,
writer,
step,
mode="origin")
et = time.time()
writer.add_scalar("origin time", et - st, step)
return loss.item()
def train(fv, model_name, criterion, balance=False, batchsize=64, size=0):
if fv == "matlab":
dloader = matloader
else:
dloader = fvloader
train_data = dloader.load_train_data(size=size, balance=balance, fv=fv)
val_data = dloader.load_val_data(size=size, fv=fv)
test_data = dloader.load_test_data(size=size, fv=fv)
# model_name = "transformer_%s_size%d_bce" % (fv, size)
model_dir = os.path.join("./modeldir/%s" % model_name)
model_pth = os.path.join(model_dir, "model.pth")
writer = tensorboardX.SummaryWriter(model_dir)
if os.path.exists(model_pth):
print("------load model--------")
model = torch.load(model_pth)
else:
# model = Transformer(fv, NUM_HEADS=4, NUM_LAYERS=3).cuda()
model = Transformer(fv).cuda()
model = nn.DataParallel(model)
optimizer = torch.optim.Adam(model.parameters(),
lr=0.0001, weight_decay=0.001)
# scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
# optimizer, factor=0.5,
# patience=30, min_lr=1e-4)
epochs = 2000
step = 1
val_step = 1
max_f1 = 0.0
for e in range(epochs):
model.train()
print("------epoch--------", e)
st = time.time()
train_shuffle = fvloader.shuffle(train_data)
for item in fvloader.batch_fv(train_shuffle, batch=batchsize):
# for name, param in model.named_parameters():
# writer.add_histogram(
# name, param.clone().cpu().data.numpy(), step)
# writer.add_histogram(
# "grad/"+name, param.grad.clone().cpu().data.numpy(), step)
model.zero_grad()
genes, nimgs, labels, timesteps = item
inputs = torch.from_numpy(nimgs).type(torch.cuda.FloatTensor)
gt = torch.from_numpy(labels).type(torch.cuda.FloatTensor)
pd = model(inputs)
# loss = criterion(pd, gt)
all_loss = criterion(pd, gt)
label_loss = torch.mean(all_loss, dim=0)
loss = torch.mean(label_loss)
# for i in range(6):
# writer.add_scalar("train sl_%d_loss" % i,
# label_loss[i].item(), step)
train_pd = torch_util.threshold_tensor_batch(pd)
np_pd = train_pd.data.cpu().numpy()
torch_util.torch_metrics(
labels, np_pd, writer, step, mode="train")
writer.add_scalar("train loss", loss, step)
loss.backward()
optimizer.step()
step += 1
et = time.time()
writer.add_scalar("train time", et - st, e)
for param_group in optimizer.param_groups:
writer.add_scalar("lr", param_group['lr'], e)
# run_origin_train(model, imbtrain_data, writer, e, criterion)
if e % 1 == 0:
val_loss, val_f1 = run_val(
model, dloader, val_data, writer, val_step, criterion)
# scheduler.step(val_loss)
val_step += 1
if e == 0:
start_loss = val_loss
min_loss = start_loss
# if val_loss > 2 * min_loss:
# print("early stopping at %d" % e)
# break
# if e % 50 == 0:
# pt = os.path.join(model_dir, "%d.pt" % e)
# torch.save(model.state_dict(), pt)
# result = os.path.join(model_dir, "result_epoch%d.txt" % e)
# run_test(model, test_data, result)
if min_loss > val_loss or max_f1 < val_f1:
if min_loss > val_loss:
print("---------save best----------", "loss", val_loss)
min_loss = val_loss
if max_f1 < val_f1:
print("---------save best----------", "f1", val_f1)
max_f1 = val_f1
torch.save(model, model_pth)
result = os.path.join(model_dir, "result_epoch%d.txt" % e)
run_test(model, dloader, test_data, result)
