def evaluate(model=None,
inp_images=None,
annotations=None,
checkpoints_path=None,
epoch=None):
#Finished implementation of the evaluate function in keras_segmentation.predict
#Input: array of paths or nd arrays
if model is None and (not checkpoints_path is None):
model = model_from_checkpoint_path(checkpoints_path, epoch)
ious = []
pred_time = []
for inp, ann in zip(inp_images, annotations):
t_start = time.time()
pr = predict_fast(model, inp)
t_end = time.time()
pred_time.append(t_end - t_start)
print('Prediction time: ', t_end - t_start)
gt = get_segmentation_arr(ann,
model.n_classes,
model.output_width,
model.output_height,
no_reshape=True)
gt = gt.argmax(-1)
iou = metrics.get_iou(gt, pr, model.n_classes)
ious.append(iou)
print("Class wise IoU ", np.mean(ious, axis=0))
print("Total IoU ", np.mean(ious))
print("Median prediction time:", np.median(pred_time))
def evaluate(model=None,
inp_images=None,
annotations=None,
checkpoints_path=None,
epoch=None,
visualize=False,
output_folder=''):
#Finished implementation of the evaluate function in keras_segmentation.predict
#Input: array of paths or nd arrays
if model is None and (not checkpoints_path is None):
model = model_from_checkpoint_path(checkpoints_path, epoch)
ious = []
pred_time = []
for inp, ann in tqdm(zip(inp_images, annotations)):
t_start = time.time()
pr = predict_fast(model, inp)
t_end = time.time()
pred_time.append(t_end - t_start)
#print('Prediction time: ', pred_time)
gt = get_segmentation_arr(ann,
model.n_classes,
model.output_width,
model.output_height,
no_reshape=True)
gt = gt.argmax(-1)
iou = metrics.get_iou(gt, pr, model.n_classes)
ious.append(iou)
if visualize:
fig = vis_pred_vs_gt_separate(inp, pr, gt)
plt.title("Predicted mask and errors. "
"IOU (bg, crop, lane):" + str(iou))
if not output_folder:
if epoch is None:
epoch = ''
print(checkpoints_path, epoch, os.path.basename(inp))
fig.savefig(checkpoints_path + epoch + '_IOU_' +
os.path.basename(inp))
print(
'Saving to: ',
checkpoints_path + epoch + '_IOU_' + os.path.basename(inp))
else:
fig.savefig(
os.path.join(
output_folder,
os.path.basename(checkpoints_path) + '_IOU_' +
os.path.basename(inp)))
print ious
ious = np.array(ious)
print("Class wise IoU ", np.mean(ious, axis=0))
print("Total IoU ", np.mean(ious))
print("Mean prediction time:", np.mean(pred_time))
import time
#from scipy.misc import imresize
from PIL import Image
#Temporary code with one image, will be added back to evaluate function when done
#TODO: should be based on colormaps for error images and masks separately, instead of RGB encoding
labels = [0, 1, 2]
gt = np.load('gt_array.npy')
pr = np.load('pr_array.npy')
inp = r'Frogn_Dataset/images_prepped_test/20190913_LR1_S_0420.png'
image_mode = 'F'
input_image = Image.open(inp)
iou = metrics.get_iou(gt, pr, 3)
ious = []
ious.append(iou)
#visualization
fig = plt.figure()
plt.title("IOU:" + str(iou))
ax1 = fig.add_subplot(2, 2, 1)
ax1.imshow(gt - pr)
#ax1.colorbar()
ax1.title.set_text("Difference GT-pred")
ax2 = fig.add_subplot(2, 2, 2)
ax2.imshow(gt)
ax2.title.set_text('GT')
def evaluate_camvid():
# get seg (ground truth) and compare with pr (prediction)
# Get seg
segs_path = "./dataset3/label/test/"
segs = glob.glob( os.path.join(segs_path,"*.png") )
#print(len(segs))
imglabels = np.ndarray((len(segs),2710, 3384), dtype=np.uint8)
i=0
for x in range(len(segs)):
imgpath = segs[x]
pic_name = imgpath.split('/')[-1]
labelpath = "dataset3/label/" + pic_name.split('.')[0] + '.png'
label = load_img(labelpath, grayscale=True, target_size=[2710, 3384]) # grayscale = False
label = img_to_array(label)
imglabels[i] = label[:,:,0]
if i % 100 == 0:
print('Creating testing images: {0}/{1} images'.format(i, len(segs)))
i += 1
#np.save('./segs_test.npy', imglabels)
i=0
orininal_w = 3384
orininal_h = 2710
n_classes = 12
ious=[]
precisions=[]
recalls=[]
for pr in tqdm(all_prs):
gt = imglabels[i]
pr = cv2.resize(pr, (orininal_w , orininal_h), interpolation=cv2.INTER_NEAREST)
#IoU
iou = metrics.get_iou( gt , pr , n_classes )
ious.append( iou )
# precision
precision = metrics.get_precision( gt , pr , n_classes )
precisions.append( precision )
# recall
recall = metrics.get_recall( gt , pr , n_classes )
recalls.append( recall )
i+=1
ious = np.array(ious)
precisions = np.array( precisions )
recalls = np.array(recalls)
#print("Class wise IoU Class:{:d} / IoU:{:.2f}\n".format(11, np.mean(ious , axis=0 )[11]))
#print("Class wise Precision Class:{:d} / Precision:{:.2f}\n".format(11, np.mean(precisions , axis=0 )[11]))
#print("Class wise Recall Class:{:d} / Recall:{:.2f}\n".format(11, np.mean(recalls , axis=0 )[11]))
print("Class wise IoU " , np.mean(ious , axis=0 ))
print("Total IoU " , np.mean(ious ))
#Save one at a time as image
for ind, fname in enumerate(inp_names):
#Load images
inp = os.path.join(image_folder, fname)
ann = os.path.join(annotations_folder, fname)
#Run prediction
pr = predict_fast(model, inp)
gt = get_segmentation_arr(ann,
model.n_classes,
model.output_width,
model.output_height,
no_reshape=True)
gt = gt.argmax(-1)
iou = metrics.get_iou(gt, pr, model.n_classes)
#--Make overlay image
im_overlay = vis_pred_vs_gt_overlay(
inp, pr, gt, figure_width_mm=85) #check column widht for CASE
plt.imsave(output_name + '_' + os.path.basename(inp), im_overlay)
np.savetxt('IOU_' + os.path.basename(inp) + '.txt', iou)
#All in one figure
'''
fig,ax = plt.subplots(nrows=len(inp_names),ncols=2,figsize=(figure_width_mm/25.4,figure_height_mm/25.4))
for ind,fname in enumerate(inp_names):
#Load images
inp = os.path.join(image_folder,fname)
ann = os.path.join(annotations_folder,fname)
#Run prediction