def train(run, model_name, data_path, epochs, batch_size, mlflow_custom_log,
log_as_onnx):
x_train, y_train, x_test, y_test = utils.get_train_data(data_path)
model = build_model()
model.compile(optimizer="rmsprop",
loss="categorical_crossentropy",
metrics=["accuracy"])
model.summary()
model.fit(x_train,
y_train,
epochs=epochs,
batch_size=batch_size,
verbose=0)
test_loss, test_acc = model.evaluate(x_test, y_test)
print("test_acc:", test_acc)
print("test_loss:", test_loss)
if mlflow_custom_log:
mlflow.log_param("epochs", epochs)
mlflow.log_param("batch_size", batch_size)
mlflow.log_metric("test_acc", test_acc)
mlflow.log_metric("test_loss", test_loss)
mlflow.keras.log_model(model,
"keras-model",
registered_model_name=model_name)
# write model summary
summary = []
model.summary(print_fn=summary.append)
summary = "\n".join(summary)
with open("model_summary.txt", "w") as f:
f.write(summary)
mlflow.log_artifact("model_summary.txt")
# Save as TensorFlow SavedModel format
path = "tensorflow-model"
tf.keras.models.save_model(model,
path,
overwrite=True,
include_optimizer=True)
mlflow.log_artifact(path)
else:
utils.register_model(run, model_name)
# write model as yaml file
with open("model.yaml", "w") as f:
f.write(model.to_yaml())
mlflow.log_artifact("model.yaml")
# MLflow - log onnx model
if log_as_onnx:
import onnx_utils
mname = f"{model_name}_onnx" if model_name else None
onnx_utils.log_model(model, "onnx-model", mname)
predictions = model.predict_classes(x_test)
print("predictions:", predictions)
def train(run, model_name, epochs, batch_size, mlflow_custom_log, log_as_onnx):
x_train, y_train, x_test, y_test = utils.get_train_data()
model = build_model()
model.compile(
optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
model.fit(x_train, y_train, epochs=epochs, batch_size=batch_size, verbose=0)
test_loss, test_acc = model.evaluate(x_test, y_test)
print("test_acc:", test_acc)
print("test_loss:", test_loss)
if mlflow_custom_log:
mlflow.log_param("epochs", epochs)
mlflow.log_param("batch_size", batch_size)
mlflow.log_metric("test_acc", test_acc)
mlflow.log_metric("test_loss", test_loss)
mlflow.keras.log_model(model, "keras-model", registered_model_name=model_name)
# write model summary
summary = []
model.summary(print_fn=summary.append)
summary = '\n'.join(summary)
with open("model_summary.txt", "w") as f:
f.write(summary)
mlflow.log_artifact("model_summary.txt")
else:
utils.register_model(run, model_name)
# write model as yaml file
with open("model.yaml", "w") as f:
f.write(model.to_yaml())
mlflow.log_artifact("model.yaml")
# MLflow - log onnx model
if log_as_onnx:
import onnx_utils
mname = f"{model_name}_onnx" if model_name else None
onnx_utils.log_model(model, "onnx-model", mname)
predictions = model.predict_classes(x_test)
print("predictions:", predictions)
def main():
batch_size = 2
train_dir = 'D:/DATA/JHUBrain/Train/'
val_dir = 'D:/DATA/JHUBrain/Val/'
save_dir = 'ViTVNet_reg0.02_mse_diff/'
lr = 0.0001
epoch_start = 0
max_epoch = 500
cont_training = False
config_vit = CONFIGS_ViT_seg['ViT-V-Net']
reg_model = utils.register_model((160, 192, 224), 'nearest')
reg_model.cuda()
model = models.ViTVNet(config_vit, img_size=(160, 192, 224))
if cont_training:
epoch_start = 335
model_dir = 'experiments/'+save_dir
updated_lr = round(lr * np.power(1 - (epoch_start) / max_epoch,0.9),8)
best_model = torch.load(model_dir + natsorted(os.listdir(model_dir))[0])['state_dict']
model.load_state_dict(best_model)
else:
updated_lr = lr
model.cuda()
train_composed = transforms.Compose([trans.RandomFlip(0),
trans.NumpyType((np.float32, np.float32)),
])
val_composed = transforms.Compose([trans.Seg_norm(), #rearrange segmentation label to 1 to 46
trans.NumpyType((np.float32, np.int16)),
])
train_set = datasets.JHUBrainDataset(glob.glob(train_dir + '*.pkl'), transforms=train_composed)
val_set = datasets.JHUBrainInferDataset(glob.glob(val_dir + '*.pkl'), transforms=val_composed)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True)
val_loader = DataLoader(val_set, batch_size=1, shuffle=False, num_workers=4, pin_memory=True, drop_last=True)
optimizer = optim.Adam(model.parameters(), lr=updated_lr, weight_decay=0, amsgrad=True)
criterion = nn.MSELoss()
criterions = [criterion]
weights = [1]
# prepare deformation loss
criterions += [losses.Grad3d(penalty='l2')]
weights += [0.02]
best_mse = 0
writer = SummaryWriter(log_dir='ViTVNet_log')
for epoch in range(epoch_start, max_epoch):
print('Training Starts')
'''
Training
'''
loss_all = AverageMeter()
idx = 0
for data in train_loader:
idx += 1
model.train()
adjust_learning_rate(optimizer, epoch, max_epoch, lr)
data = [t.cuda() for t in data]
x = data[0]
y = data[1]
x_in = torch.cat((x,y), dim=1)
output = model(x_in)
loss = 0
loss_vals = []
for n, loss_function in enumerate(criterions):
curr_loss = loss_function(output[n], y) * weights[n]
loss_vals.append(curr_loss)
loss += curr_loss
loss_all.update(loss.item(), y.numel())
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
del x_in
del output
# flip fixed and moving images
loss = 0
x_in = torch.cat((y, x), dim=1)
output = model(x_in)
for n, loss_function in enumerate(criterions):
curr_loss = loss_function(output[n], x) * weights[n]
loss_vals[n] += curr_loss
loss += curr_loss
loss_all.update(loss.item(), y.numel())
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('Iter {} of {} loss {:.4f}, Img Sim: {:.6f}, Reg: {:.6f}'.format(idx, len(train_loader), loss.item(), loss_vals[0].item()/2, loss_vals[1].item()/2))
writer.add_scalar('Loss/train', loss_all.avg, epoch)
print('Epoch {} loss {:.4f}'.format(epoch, loss_all.avg))
'''
Validation
'''
eval_dsc = AverageMeter()
with torch.no_grad():
for data in val_loader:
model.eval()
data = [t.cuda() for t in data]
x = data[0]
y = data[1]
x_seg = data[2]
y_seg = data[3]
# x = x.squeeze(0).permute(1, 0, 2, 3)
# y = y.squeeze(0).permute(1, 0, 2, 3)
x_in = torch.cat((x, y), dim=1)
output = model(x_in)
def_out = reg_model([x_seg.cuda().float(), output[1].cuda()])
dsc = utils.dice_val(def_out.long(), y_seg.long(), 46)
eval_dsc.update(dsc.item(), x.size(0))
print(eval_dsc.avg)
best_mse = max(eval_dsc.avg, best_mse)
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_mse': best_mse,
'optimizer': optimizer.state_dict(),
}, save_dir='experiments/'+save_dir, filename='dsc{:.3f}.pth.tar'.format(eval_dsc.avg))
writer.add_scalar('MSE/validate', eval_dsc.avg, epoch)
plt.switch_backend('agg')
pred_fig = comput_fig(def_out)
x_fig = comput_fig(x_seg)
tar_fig = comput_fig(y_seg)
writer.add_figure('input', x_fig, epoch)
plt.close(x_fig)
writer.add_figure('ground truth', tar_fig, epoch)
plt.close(tar_fig)
writer.add_figure('prediction', pred_fig, epoch)
plt.close(pred_fig)
loss_all.reset()
writer.close()
def train(run, model_name, data_path, epochs, batch_size, mlflow_custom_log,
log_as_onnx, log_as_tensorflow_lite, log_as_tensorflow_js):
print("mlflow_custom_log:", mlflow_custom_log)
x_train, _, y_train, _ = utils.build_data(data_path)
ncols = x_train.shape[1]
def baseline_model():
model = Sequential()
model.add(
Dense(ncols,
input_dim=ncols,
kernel_initializer='normal',
activation='relu'))
model.add(Dense(1, kernel_initializer='normal'))
model.compile(loss='mean_squared_error', optimizer='adam')
return model
model = baseline_model()
if mlflow_custom_log:
print("Logging with mlflow.log")
mlflow.log_param("epochs", epochs)
mlflow.log_param("batch_size", batch_size)
mlflow.keras.log_model(model,
"tensorflow-model",
registered_model_name=model_name)
else:
utils.register_model(run, model_name)
# MLflow - log as ONNX model
if log_as_onnx:
import onnx_utils
mname = f"{model_name}_onnx" if model_name else None
onnx_utils.log_model(model, "onnx-model", model_name=mname)
# Save as TensorFlow Lite format
if log_as_tensorflow_lite:
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
path = "model.tflite"
with open(path, "wb") as f:
f.write(tflite_model)
mlflow.log_artifact(path, "tensorflow-lite-model")
# Save as TensorFlow.js format
if log_as_tensorflow_js:
import tensorflowjs as tfjs
path = "model.tfjs"
tfjs.converters.save_keras_model(model, path)
mlflow.log_artifact(path, "tensorflow-js-model")
# Evaluate model
estimator = KerasRegressor(build_fn=baseline_model,
epochs=epochs,
batch_size=batch_size,
verbose=0)
kfold = KFold(n_splits=10)
results = cross_val_score(estimator, x_train, y_train, cv=kfold)
print(
f"Baseline MSE: mean: {round(results.mean(),2)} std: {round(results.std(),2)}"
)
if mlflow_custom_log:
mlflow.log_metric("mse_mean", results.mean())
mlflow.log_metric("mse_std", results.std())
# Score
data = x_train
predictions = model.predict(data)
predictions = pd.DataFrame(data=predictions, columns=["prediction"])
print("predictions.shape:", predictions.shape)
print("predictions:", predictions)
def main():
test_dir = 'D:/DATA/JHUBrain/Test/'
model_idx = -1
model_folder = 'ViTVNet_reg0.02_mse_diff/'
model_dir = 'experiments/' + model_folder
config_vit = CONFIGS_ViT_seg['ViT-V-Net']
dict = utils.process_label()
if os.path.exists('experiments/' + model_folder[:-1] + '.csv'):
os.remove('experiments/' + model_folder[:-1] + '.csv')
csv_writter(model_folder[:-1], 'experiments/' + model_folder[:-1])
line = ''
for i in range(46):
line = line + ',' + dict[i]
csv_writter(line, 'experiments/' + model_folder[:-1])
model = models.ViTVNet(config_vit, img_size=(160, 192, 224))
best_model = torch.load(
model_dir + natsorted(os.listdir(model_dir))[model_idx])['state_dict']
print('Best model: {}'.format(natsorted(os.listdir(model_dir))[model_idx]))
model.load_state_dict(best_model)
model.cuda()
reg_model = utils.register_model((160, 192, 224), 'nearest')
reg_model.cuda()
test_composed = transforms.Compose([
trans.Seg_norm(),
trans.NumpyType((np.float32, np.int16)),
])
test_set = datasets.JHUBrainInferDataset(glob.glob(test_dir + '*.pkl'),
transforms=test_composed)
test_loader = DataLoader(test_set,
batch_size=1,
shuffle=False,
num_workers=1,
pin_memory=True,
drop_last=True)
eval_dsc_def = AverageMeter()
eval_dsc_raw = AverageMeter()
eval_det = AverageMeter()
with torch.no_grad():
stdy_idx = 0
for data in test_loader:
model.eval()
data = [t.cuda() for t in data]
x = data[0]
y = data[1]
x_seg = data[2]
y_seg = data[3]
x_in = torch.cat((x, y), dim=1)
x_def, flow = model(x_in)
def_out = reg_model([x_seg.cuda().float(), flow.cuda()])
tar = y.detach().cpu().numpy()[0, 0, :, :, :]
#jac_det = utils.jacobian_determinant(flow.detach().cpu().numpy()[0, :, :, :, :])
line = utils.dice_val_substruct(def_out.long(), y_seg.long(),
stdy_idx)
line = line #+','+str(np.sum(jac_det <= 0)/np.prod(tar.shape))
csv_writter(line, 'experiments/' + model_folder[:-1])
#eval_det.update(np.sum(jac_det <= 0) / np.prod(tar.shape), x.size(0))
dsc_trans = utils.dice_val(def_out.long(), y_seg.long(), 46)
dsc_raw = utils.dice_val(x_seg.long(), y_seg.long(), 46)
print('Trans diff: {:.4f}, Raw diff: {:.4f}'.format(
dsc_trans.item(), dsc_raw.item()))
eval_dsc_def.update(dsc_trans.item(), x.size(0))
eval_dsc_raw.update(dsc_raw.item(), x.size(0))
stdy_idx += 1
# flip moving and fixed images
y_in = torch.cat((y, x), dim=1)
y_def, flow = model(y_in)
def_out = reg_model([y_seg.cuda().float(), flow.cuda()])
tar = x.detach().cpu().numpy()[0, 0, :, :, :]
#jac_det = utils.jacobian_determinant(flow.detach().cpu().numpy()[0, :, :, :, :])
line = utils.dice_val_substruct(def_out.long(), x_seg.long(),
stdy_idx)
line = line #+ ',' + str(np.sum(jac_det < 0) / np.prod(tar.shape))
out = def_out.detach().cpu().numpy()[0, 0, :, :, :]
#print('det < 0: {}'.format(np.sum(jac_det <= 0)/np.prod(tar.shape)))
csv_writter(line, 'experiments/' + model_folder[:-1])
#eval_det.update(np.sum(jac_det <= 0) / np.prod(tar.shape), x.size(0))
dsc_trans = utils.dice_val(def_out.long(), x_seg.long(), 46)
dsc_raw = utils.dice_val(y_seg.long(), x_seg.long(), 46)
print('Trans diff: {:.4f}, Raw diff: {:.4f}'.format(
dsc_trans.item(), dsc_raw.item()))
eval_dsc_def.update(dsc_trans.item(), x.size(0))
eval_dsc_raw.update(dsc_raw.item(), x.size(0))
stdy_idx += 1
print('Deformed DSC: {:.3f} +- {:.3f}, Affine DSC: {:.3f} +- {:.3f}'.
format(eval_dsc_def.avg, eval_dsc_def.std, eval_dsc_raw.avg,
eval_dsc_raw.std))
print('deformed det: {}, std: {}'.format(eval_det.avg, eval_det.std))