def unet_predict(self):
d = divide_image(test_image=self.test_image)
data = separate_channels(d)
preds = self.model.predict(data,verbose=1)
data = combine_channels(data)
predicted_image = preds.reshape((-1,data[0].shape[0],data[0].shape[1]))
predicted_image = stitch(predicted_image,1024,4)[0]
print "heatmap of unet",predicted_image.shape
return predicted_image
def preprocess(self,test_image):
if isinstance(test_image,str):
test_image = plt.imread(test_image)
test_image = imresize(test_image,(1024,1024))
d = divide_image(test_image=test_image)
print "d.shape is ",np.shape(d)
data = separate_channels(d)
self.orig_shape = (-1,d[0].shape[0],d[0].shape[1])
data = combine_channels(data)
return data
def _save_debug_img(self, x, y, idx):
# x: (w, h, d, 1)
# y: (w, h, d, c)
if self.categorical:
y = combine_channels(y)
x = nib.Nifti1Image(x, np.eye(4))
y = nib.Nifti1Image(y, np.eye(4))
debug_dir = "debug"
if not os.path.exists(debug_dir):
os.mkdir(debug_dir)
nib.save(x, os.path.join(debug_dir, "%d.nii" % idx))
nib.save(y, os.path.join(debug_dir, "%d_seg.nii" % idx))
def take_out(self, layer_name, image, average=True):
if isinstance(image, str):
image = plt.imread(image)
images = self.preprocess(image)
if not self.func:
self.func = self.make_func(layer_name)
feature_maps = self.func([images])[0]
if average:
feature_maps = np.mean(feature_maps, axis=1)
else:
feature_maps = combine_channels(feature_maps)
feature_map = self.custom_stitch(feature_maps)
if not average:
feature_map = separate_channels(np.array([feature_map]))[0]
return feature_map
# batches = 0 # for X_batch, Y_batch in izip(image_datagen.flow(data,shuffle=False,batch_size=1,seed=seed),mask_datagen.flow(labels,shuffle=False,batch_size=1,seed=seed)): # print X_batch.shape, Y_batch.shape # x = combine_channels([X_batch[0]])[0] # y = Y_batch[0].reshape((img_rows,img_cols)) # print x.shape, y.shape # # show_images([x,y]) # # # # raise # loss = model.train(X_batch, Y_batch) # batches += 1 # if batches >= len(X_train) / 32: # # we need to break the loop by hand because # # the generator loops indefinitely # break save_model(model,"aug_dunet") from sklearn.utils import shuffle inds = shuffle(range(len(data)))[:5] preds = model.predict(data[inds],verbose=1) print "preds orig:\n",preds print "orig shape",preds.shape data_dash = combine_channels(data[inds]) preds = preds.reshape((-1,img_rows,img_cols)) show_images(data_dash,preds/255)
data, labels = load_bce_data()
shape = data[0].shape
print data.shape, labels.shape
data = separate_channels(
data) # make data compatible with theano dimension ordering
print data.shape, labels.shape
labels = labels.reshape(
(-1, img_rows * img_cols)
) # Flatten the ground truths from 2D to 1D for evaluation of binary cross entropy at the other end of model
data_train, data_test, labels_train, labels_test = train_test_split(
data, labels, test_size=0.3) # split the data into 70-30 train-test
model.fit(data_train, labels_train, batch_size=32,
nb_epoch=3000) # Train the model on the train-data portion
save_model(model, "dunet7") # Save the model
from sklearn.utils import shuffle
inds = shuffle(range(len(data_test)))[:5]
preds = model.predict(
data_test[inds], verbose=1
) # Test the model on randomly picked 5 images from the test-portion
print "preds orig:\n", preds
print "orig shape", preds.shape
data_dash = combine_channels(
data_test[inds]) # Make data compatible for visualizing
preds = preds.reshape(
(-1, img_rows,
img_cols)) # Reshape back the predictions from flattened array to 2D
show_images(data_dash, preds / 255)
to_continue = mkdir_p(result_dir)
if not to_continue:
print("Overwrite declined, program interrupted!")
sys.exit(0)
print "loading data.."
data, labels = load_bce_data()
print "done"
# rand_inds = shuffle(range(len(data)))
# data = data[rand_inds]
# labels = labels[rand_inds]
data = separate_channels(data)
print "predicting.."
preds = model.predict(data,batch_size=1,verbose=1)
print "done"
data = combine_channels(data)
predicted_images = preds.reshape((-1,data[0].shape[0],data[0].shape[1]))
#data = stitch(data,1024,4)
#predicted_images = stitch(predicted_images,1024,4)
#labels = stitch(labels,1024,4)
print "saving results to "+result_dir
count = 0
results = {}
for orig_image, label_image, heat_map in tqdm(zip(data,labels,predicted_images)):
count = count +1
result = prepare_result(orig_image,label_image,heat_map,heat_map_metrics=[Result.BRIER_SCORE],mask_metrics=[Result.CONFUSION_MATRIX,Result.F1_SCORE,Result.ACCURACY, Result.PRECISION_SCORE, Result.RECALL_SCORE])
result.show()
result.save(result_dir+"/tr_"+model_name+"_"+str(count)+".jpg")