def make_predict(img_obj, train_obj, test_img):
# just makes the predicted image
print "Loading trained model..."
test_model = os.path.join(train_obj.model_path,
train_obj.model_name + ".hdf5")
mean_t, std_t = np.load(
os.path.join(img_obj.out_path, img_obj.out_name + "_meanstd.npy"))
if train_obj.model_arch == "vanilla_unet":
model = vanilla_unet()
model.load_weights(test_model)
elif train_obj.model_arch == "rob_unet":
model = rob_unet()
model.load_weights(test_model)
elif train_obj.model_arch == "dil_unet":
model = dil_unet()
model.load_weights(test_model)
else:
print "you didn't chose one of the possible model architecutres: "
print "- vanilla_unet \n -rob_unet \n dil_unet"
print "Loading test image..."
img = np.load(test_img)
print "Trying to segment..."
predict = seg_img(model, img, mean_t, std_t)
#predicted_imgs.append(predict)
print "Saving Segmentation Result..."
np.save(os.path.join(train_obj.model_path, train_obj.model_name + ".npy"),
predict)
return
def train_model(img_obj, train_obj, log=1):
''' Train a model for the configuration provided'''
# get training imgs, labels and val imgs, labels
train_img, train_label, val_img, val_label = load_patch(
patch_path=img_obj.out_path, patch_name=img_obj.out_name)
# get arguments for the data augumentation generator
gen_arg = train_obj.get_data_aug()
# create the data augumentation generator .. bs is batch size
combined = aug_gen(train_img, train_label, bs=32, **gen_arg)
psize = img_obj.patch_size * img_obj.patch_scale
if train_obj.model_arch == "vanilla_unet":
model = vanilla_unet(patch_size=(psize, psize),
dropout=train_obj.model_dropout)
elif train_obj.model_arch == "rob_unet":
model = rob_unet(patch_size=(psize, psize),
dropout=train_obj.model_dropout)
elif train_obj.model_arch == "dil_unet":
model = dil_unet(patch_size=(psize, psize),
dropout=train_obj.model_dropout)
else:
print "you didn't chose one of the possible model architecutres: "
print "- vanilla_unet \n -rob_unet \n dil_unet"
# this looks more scary than it really is , I made the property names same as
# the arugument names for a reason!
run_dict = train_seg(train_patch=train_img,
train_label=train_label,
val_patch=val_img,
val_label=val_label,
img_gen=combined,
model=model,
model_path=train_obj.model_path,
model_name=train_obj.model_name,
epochs=train_obj.model_epoch,
monitor=train_obj.model_monitor,
early_stop=train_obj.early_stopping,
patience=train_obj.model_patience)
print run_dict
if (log):
make_log(img_obj, train_obj, **run_dict)
return
def make_predict(img_obj,train_obj,test_img):
print "Loading trained model..."
test_model=os.path.join(train_obj.model_path,train_obj.model_name+".hdf5")
mean_t, std_t = np.load(os.path.join(img_obj.out_path,img_obj.out_name+"_meanstd.npy"))
if train_obj.model_arch=="vanilla_unet":
model = vanilla_unet()
model.load_weights(test_model)
elif train_obj.model_arch=="rob_unet":
model = rob_unet()
model.load_weights(test_model)
elif train_obj.model_arch=="dil_unet":
model=dil_unet()
model.load_weights(test_model)
else:
print "you didn't chose one of the possible model architecutres: "
print "- vanilla_unet \n -rob_unet \n dil_unet"
print "Loading test image..."
img = np.load(test_img)
img_rgb = rgb_images(img,period)
img_rgb = 100.0*(img_rgb-img_rgb.min(axis=(1,2),keepdims = True))/(img_rgb.max(axis=(1,2),keepdims = True)
-img_rgb.min(axis=(1,2),keepdims = True))
img_rgb-=mean_t
img_rgb/=std_t
img_rgb,nw,nz = pad_images(img_rgb)
print "Segmenting test image..."
predict = model.predict(img_rgb)
predict = predict.transpose(1,2,0,3)[:-nw,:-nz,:,0]
predict = predict[:,:-nw,:-nz,0]
#predicted_imgs.append(predict)
print "Saving Segmentation Result..."
np.save(os.path.join(train_obj.model_path,train_obj.model_name+".npy"),predict)
return
def make_predict(img_obj,train_obj,img_test):
test_model=os.path.join(train_obj.model_path,train_obj.model_name+".hdf5")
mean_t, std_t = np.load(os.path.join(img_obj.out_path,img_obj.out_name+"_meanstd.npy"))
if train_obj.model_arch=="vanilla_unet":
model = vanilla_unet()
model.load_weights(test_model)
elif train_obj.model_arch=="rob_unet":
model = rob_unet()
model.load_weights(test_model)
elif train_obj.model_arch=="dil_unet":
model=dil_unet()
model.load_weights(test_model)
else:
print "you didn't chose one of the possible model architecutres: "
print "- vanilla_unet \n -rob_unet \n dil_unet"
#unsure how i want to load in the imgs...
#predict = unet.predict(img_rgb)
#predict = predict.transpose(1,2,0,3)[:-nw,:-nz,:,0]
return