-
Notifications
You must be signed in to change notification settings - Fork 262
/
evaluate.py
245 lines (210 loc) · 9.06 KB
/
evaluate.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
from __future__ import division
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import models as M
import numpy as np
from matplotlib import pyplot as plt
#Keras
from keras.models import load_model
from keras.models import model_from_json
from keras.models import Model
#scikit learn
from sklearn.metrics import roc_curve
from sklearn.metrics import roc_auc_score
from sklearn.metrics import confusion_matrix
from sklearn.metrics import precision_recall_curve
from sklearn.metrics import jaccard_similarity_score
from sklearn.metrics import f1_score
# help_functions.py
from help_functions import *
# extract_patches.py
from extract_patches import recompone
from extract_patches import recompone_overlap
from extract_patches import paint_border
from extract_patches import kill_border
from extract_patches import pred_only_FOV
from extract_patches import get_data_testing
from extract_patches import get_data_testing_overlap
# pre_processing.py
from pre_processing import my_PreProc
#========= CONFIG FILE TO READ FROM =======
#===========================================
#run the training on invariant or local
path_data = './DRIVE_datasets_training_testing/'
#original test images (for FOV selection)
DRIVE_test_imgs_original = path_data + 'DRIVE_dataset_imgs_test.hdf5'
test_imgs_orig = load_hdf5(DRIVE_test_imgs_original)
full_img_height = test_imgs_orig.shape[2]
full_img_width = test_imgs_orig.shape[3]
#the border masks provided by the DRIVE
DRIVE_test_border_masks = path_data + 'DRIVE_dataset_borderMasks_test.hdf5'
test_border_masks = load_hdf5(DRIVE_test_border_masks)
# dimension of the patches
patch_height = 64
patch_width = 64
#the stride in case output with average
stride_height = 5
stride_width = 5
assert (stride_height < patch_height and stride_width < patch_width)
#model name
name_experiment = 'test'
path_experiment = './' +name_experiment +'/'
#N full images to be predicted
Imgs_to_test = 2
#Grouping of the predicted images
N_visual = 1
#====== average mode ===========
average_mode = True
#============ Load the data and divide in patches
patches_imgs_test = None
new_height = None
new_width = None
masks_test = None
patches_masks_test = None
if average_mode == True:
patches_imgs_test, new_height, new_width, masks_test = get_data_testing_overlap(
DRIVE_test_imgs_original = DRIVE_test_imgs_original, #original
DRIVE_test_groudTruth = path_data + 'DRIVE_dataset_groundTruth_test.hdf5', #masks
Imgs_to_test = 20,
patch_height = patch_height,
patch_width = patch_width,
stride_height = stride_height,
stride_width = stride_width
)
else:
patches_imgs_test, patches_masks_test = get_data_testing(
DRIVE_test_imgs_original = DRIVE_test_imgs_original, #original
DRIVE_test_groudTruth = path_data + 'DRIVE_dataset_groundTruth_test.hdf5', #masks
Imgs_to_test = 20,
patch_height = patch_height,
patch_width = patch_width,
)
#================ Run the prediction of the patches ==================================
best_last = 'best'
patches_imgs_test = np.einsum('klij->kijl', patches_imgs_test)
model = M.BCDU_net_D3(input_size = (64,64,1))
model.summary()
model.load_weights('weight_lstm.hdf5')
predictions = model.predict(patches_imgs_test, batch_size=16, verbose=1)
predictions = np.einsum('kijl->klij', predictions)
print(patches_imgs_test.shape)
pred_patches = predictions
print ("predicted images size :")
print (predictions.shape)
#===== Convert the prediction arrays in corresponding images
#========== Elaborate and visualize the predicted images ====================
pred_imgs = None
orig_imgs = None
gtruth_masks = None
if average_mode == True:
pred_imgs = recompone_overlap(pred_patches, new_height, new_width, stride_height, stride_width)# predictions
orig_imgs = my_PreProc(test_imgs_orig[0:pred_imgs.shape[0],:,:,:]) #originals
gtruth_masks = masks_test #ground truth masks
else:
pred_imgs = recompone(pred_patches,13,12) # predictions
orig_imgs = recompone(patches_imgs_test,13,12) # originals
gtruth_masks = recompone(patches_masks_test,13,12) #masks
# apply the DRIVE masks on the repdictions #set everything outside the FOV to zero!!
print('killing border')
kill_border(pred_imgs, test_border_masks) #DRIVE MASK #only for visualization
## back to original dimensions
orig_imgs = orig_imgs[:,:,0:full_img_height,0:full_img_width]
pred_imgs = pred_imgs[:,:,0:full_img_height,0:full_img_width]
gtruth_masks = gtruth_masks[:,:,0:full_img_height,0:full_img_width]
np.save('pred_imgs',pred_imgs)
print ("Orig imgs shape: " +str(orig_imgs.shape))
print ("pred imgs shape: " +str(pred_imgs.shape))
print ("Gtruth imgs shape: " +str(gtruth_masks.shape))
np.save('resutls', pred_imgs)
np.save('origin', gtruth_masks)
assert (orig_imgs.shape[0]==pred_imgs.shape[0] and orig_imgs.shape[0]==gtruth_masks.shape[0])
N_predicted = orig_imgs.shape[0]
group = N_visual
assert (N_predicted%group==0)
#====== Evaluate the results
print ("\n\n======== Evaluate the results =======================")
#predictions only inside the FOV
y_scores, y_true = pred_only_FOV(pred_imgs,gtruth_masks, test_border_masks) #returns data only inside the FOV
print(y_scores.shape)
print ("Calculating results only inside the FOV:")
print ("y scores pixels: " +str(y_scores.shape[0]) +" (radius 270: 270*270*3.14==228906), including background around retina: " +str(pred_imgs.shape[0]*pred_imgs.shape[2]*pred_imgs.shape[3]) +" (584*565==329960)")
print ("y true pixels: " +str(y_true.shape[0]) +" (radius 270: 270*270*3.14==228906), including background around retina: " +str(gtruth_masks.shape[2]*gtruth_masks.shape[3]*gtruth_masks.shape[0])+" (584*565==329960)")
#Area under the ROC curve
fpr, tpr, thresholds = roc_curve((y_true), y_scores)
AUC_ROC = roc_auc_score(y_true, y_scores)
# test_integral = np.trapz(tpr,fpr) #trapz is numpy integration
print ("\nArea under the ROC curve: " +str(AUC_ROC))
roc_curve =plt.figure()
plt.plot(fpr,tpr,'-',label='Area Under the Curve (AUC = %0.4f)' % AUC_ROC)
plt.title('ROC curve')
plt.xlabel("FPR (False Positive Rate)")
plt.ylabel("TPR (True Positive Rate)")
plt.legend(loc="lower right")
plt.savefig(path_experiment+"ROC.png")
#Precision-recall curve
precision, recall, thresholds = precision_recall_curve(y_true, y_scores)
precision = np.fliplr([precision])[0] #so the array is increasing (you won't get negative AUC)
recall = np.fliplr([recall])[0] #so the array is increasing (you won't get negative AUC)
AUC_prec_rec = np.trapz(precision,recall)
print ("\nArea under Precision-Recall curve: " +str(AUC_prec_rec))
prec_rec_curve = plt.figure()
plt.plot(recall,precision,'-',label='Area Under the Curve (AUC = %0.4f)' % AUC_prec_rec)
plt.title('Precision - Recall curve')
plt.xlabel("Recall")
plt.ylabel("Precision")
plt.legend(loc="lower right")
plt.savefig(path_experiment+"Precision_recall.png")
#Confusion matrix
threshold_confusion = 0.5
print ("\nConfusion matrix: Custom threshold (for positive) of " +str(threshold_confusion))
y_pred = np.empty((y_scores.shape[0]))
for i in range(y_scores.shape[0]):
if y_scores[i]>=threshold_confusion:
y_pred[i]=1
else:
y_pred[i]=0
confusion = confusion_matrix(y_true, y_pred)
print (confusion)
accuracy = 0
if float(np.sum(confusion))!=0:
accuracy = float(confusion[0,0]+confusion[1,1])/float(np.sum(confusion))
print ("Global Accuracy: " +str(accuracy))
specificity = 0
if float(confusion[0,0]+confusion[0,1])!=0:
specificity = float(confusion[0,0])/float(confusion[0,0]+confusion[0,1])
print ("Specificity: " +str(specificity))
sensitivity = 0
if float(confusion[1,1]+confusion[1,0])!=0:
sensitivity = float(confusion[1,1])/float(confusion[1,1]+confusion[1,0])
print ("Sensitivity: " +str(sensitivity))
precision = 0
if float(confusion[1,1]+confusion[0,1])!=0:
precision = float(confusion[1,1])/float(confusion[1,1]+confusion[0,1])
print ("Precision: " +str(precision))
#Jaccard similarity index
jaccard_index = jaccard_similarity_score(y_true, y_pred, normalize=True)
print ("\nJaccard similarity score: " +str(jaccard_index))
#F1 score
F1_score = f1_score(y_true, y_pred, labels=None, average='binary', sample_weight=None)
print ("\nF1 score (F-measure): " +str(F1_score))
#Save the results
file_perf = open(path_experiment+'performances.txt', 'w')
file_perf.write("Area under the ROC curve: "+str(AUC_ROC)
+ "\nArea under Precision-Recall curve: " +str(AUC_prec_rec)
+ "\nJaccard similarity score: " +str(jaccard_index)
+ "\nF1 score (F-measure): " +str(F1_score)
+"\n\nConfusion matrix:"
+str(confusion)
+"\nACCURACY: " +str(accuracy)
+"\nSENSITIVITY: " +str(sensitivity)
+"\nSPECIFICITY: " +str(specificity)
+"\nPRECISION: " +str(precision)
)
file_perf.close()
# Visualize
fig,ax = plt.subplots(10,3,figsize=[15,15])
for idx in range(10):
ax[idx, 0].imshow(np.uint8(np.squeeze((orig_imgs[idx]))))
ax[idx, 1].imshow(np.squeeze(gtruth_masks[idx]), cmap='gray')
ax[idx, 2].imshow(np.squeeze(pred_imgs[idx]), cmap='gray')
plt.savefig(path_experiment+'sample_results.png')