def get_result(val_base_dir, json_sav_dir):
'''
val_base_dir : patch saved dir
json_sav_dir : json file saved dir
'''
val_base_pth = val_base_dir
sample_name = os.listdir(val_base_dir)
x = tf.placeholder(tf.float32, [None, height, width, channel], name="inputs")
is_training = tf.placeholder(tf.bool, name="is_train")
resnet = ResNet()
last_feature = resnet.model(x, is_training)
before_prob = resnet.fc_layer(last_feature, 1, is_training)
prob = tf.nn.sigmoid(before_prob)
my_dict = dict()
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_saved_path)
cnt = 0
for name in sample_name:
pth_name = val_base_dir + "/" + name
prob_list = []
t1 = time()
for fname in os.listdir(pth_name):
img = cv.imread(pth_name + '/' + fname)
img = dataAugmentation(img)
img = (img - [MEAN_B, MEAN_G, MEAN_R]) / [VAR_B, VAR_G, VAR_R]
if img.shape[0] != 224 : img = cv.resize(img, (224, 224))
img = np.expand_dims(img, axis=0)
p = sess.run(prob, {x: img, is_training: is_train})
p = np.squeeze(p)
prob_list.append(float(p))
t2 = time()
print("%d : predict all patch spend %.5f second" % (cnt, (t2 - t1)))
print(len(prob_list))
print(type(prob_list))
my_dict[name] = prob_list
cnt += 1
with open(json_sav_dir + '/prob_TCGA.json', "w") as f:
json.dump(my_dict, f)
return
def get_result():
sample_name = os.listdir(val_base_dir)
x = tf.placeholder(tf.float32, [None, height, width, channel],
name="inputs")
is_training = tf.placeholder(tf.bool, name="is_train")
resnet = ResNet()
last_feature = resnet.model(x, is_training)
before_prob = resnet.fc_layer(last_feature, 1, is_training)
# prob = tf.nn.softmax(before_prob, axis=1, name="prob")
prob = tf.nn.sigmoid(before_prob)
grad = tf.gradients(before_prob, last_feature)
print(grad.get_shape().as_list())
# y_pred = tf.argmax(prob, axis=1)
y_pred = prob >= 0.5
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_saved_path)
t1 = time()
for name in sample_name:
pth_name = val_base_dir + "/" + name
print(pth_name)
pos_patch_num = 0
patch_num = len(os.listdir(pth_name))
for fname in os.listdir(pth_name):
# print(pth_name + '/' + fname)
img = cv.imread(pth_name + '/' + fname)
img = dataAugmentation(img)
img = (img - [MEAN_B, MEAN_G, MEAN_R]) / [VAR_B, VAR_G, VAR_R]
if img.shape[0] != 224: img = cv.resize(img, (224, 224))
img = np.expand_dims(img, axis=0)
# print(img.shape)
predict, p, g = sess.run([y_pred, prob, grad], {
x: img,
is_training: is_train
})
# print("predict : ", predict, "probability : ", p)
# print("labels : ", (1 if fname[:16] in highTMB else 0))
break
break
def get_result():
x = tf.placeholder(tf.float32, [None, height, width, channel],
name="inputs")
is_training = tf.placeholder(tf.bool, name="is_train")
vgg = VGG()
out = vgg.model(x, is_training)
prob = tf.nn.sigmoid(out)
y_pred = tf.cast(tf.greater(prob, 0.5), tf.float32)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_saved_path)
for dir_name in os.listdir(base_dir):
test_dir = base_dir + '/' + dir_name
file_name = os.listdir(test_dir)
cur_positive_sample_count = 0
cur_sample_count = len(file_name)
for name in file_name:
img = cv.imread(test_dir + '/' + name)
img = dataAugmentation(img)
img = (img - [MEAN_B, MEAN_G, MEAN_R]) / 255.0
if img.shape[0] != 224:
img = cv.resize(img, (224, 224))
img = np.expand_dims(img, axis=0)
probability, predict = sess.run([prob, y_pred], {
x: img,
is_training: False
})
print("predict is : ", predict, "labels is : ",
[1 if name[:16] in highTMB else 0])
cur_positive_sample_count += predict
print("current positive sample count from predict : ",
cur_positive_sample_count)
print("current sample count : ", cur_sample_count)
def get_result():
sample_name = os.listdir(val_base_dir)
x = tf.placeholder(tf.float32, [None, height, width, channel], name="inputs")
is_training = tf.placeholder(tf.bool, name="is_train")
resnet = ResNet()
last_feature = resnet.model(x, is_training)
before_prob = resnet.fc_layer(last_feature, 1, is_training)
# prob = tf.nn.softmax(before_prob, axis=1, name="prob")
prob = tf.nn.sigmoid(before_prob)
# y_pred = tf.argmax(prob, axis=1)
y_pred = prob >= 0.5
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_saved_path)
ratio = []
ans = []
lab = []
avt = []
t1 = time()
pc = 0
for name in sample_name:
pth_name = val_base_dir + "/" + name
print(pth_name)
pos_patch_num = 0
patch_num = len(os.listdir(pth_name))
for fname in os.listdir(pth_name):
# print(pth_name + '/' + fname)
img = cv.imread(pth_name + '/' + fname)
img = dataAugmentation(img)
img = (img - [MEAN_B, MEAN_G, MEAN_R]) / [VAR_B, VAR_G, VAR_R]
if img.shape[0] != 224 : img = cv.resize(img, (224, 224))
img = np.expand_dims(img, axis=0)
# print(img.shape)
predict, p = sess.run([y_pred, prob], {x: img, is_training: is_train})
# print("predict : ", predict, "probability : ", p)
# print("labels : ", (1 if fname[:16] in highTMB else 0))
pos_patch_num += int(predict)
ratio.append(pos_patch_num / patch_num)
ans.append((pos_patch_num / patch_num) >= 0.5)
lab.append((1 if name[:16] in highTMB else 0))
print("postive num : %d, patch num : %d, ratio : %.5f" % (pos_patch_num, patch_num, pos_patch_num / patch_num))
print("predict : %d, labels : %d" % (int((pos_patch_num / patch_num) >= 0.5), int(1 if name[:16] in highTMB else 0)))
print("Done")
acc = np.mean(np.array(ans) == np.array(lab))
print("acc : ", acc)
t2 = time()
avt.append(t2 - t1)
t = np.mean(np.array(avt))
print("Average predict time : ", t)
def get_result():
# sample_name = os.listdir(val_base_dir)
x = tf.placeholder(tf.float32, [None, height, width, channel], name="inputs")
is_training = tf.placeholder(tf.bool, name="is_train")
vgg = VGG()
out = vgg.model(x, is_training)
prob = tf.nn.sigmoid(out)
y_pred = tf.cast(tf.greater(prob, 0.5), tf.float32)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_saved_path)
ava = []
wc = 0 # whole count
wp = 0 # whole positive count
wn = 0 # whole negative count
wpp = 0 # whole predict positive
wpn = 0 # whole negative predict
pssc = 0
for sample_name in os.listdir(sample_pth):
t1 = time()
pc = 0 # 每个病例下的被预测为TMB高的癌灶区patch个数
pos_patch_num = 0
patch_num = len(os.listdir(sample_pth + '/' + sample_name))
for name in os.listdir(sample_pth + '/' + sample_name):
img_name = sample_pth + '/' + sample_name + "/" + name
# print(pth_name)
# print(pth_name + '/' + fname)
img = cv.imread(img_name)
img = dataAugmentation(img)
img = (img - [B_MEAN, G_MEAN, R_MEAN]) / 255.0
if img.shape[0] != 224 :
img = cv.resize(img, (224, 224))
img = np.expand_dims(img, axis=0)
# print(img.shape)
predict, p = sess.run([y_pred, prob], {x: img, is_training: is_train})
# print("predict : ", predict, "probability : ", p)
# print("labels : ", (1 if fname[:16] in highTMB else 0))
label = (1 if name[:16] in highTMB else 0)
wc += 1
wp += label
wn += (1 - label)
wpp += int((predict == 1) and (label == 1))
wpn += int((predict == 0) and (label == 0))
pos_patch_num += predict
# a = input()
# print("postive num : %d, patch num : %d, ratio : %.5f" % (pos_patch_num, patch_num, pos_patch_num / patch_num))
# print("predict : %d, labels : %d" % (int((pos_patch_num / patch_num) >= 0.5), int(1 if name[:16] in highTMB else 0)))
# print("Done")
pc = int((pos_patch_num / patch_num) >= 0.5)
t2 = time()
case_predict_result = pc
print("case predict : ", case_predict_result, "label is : ", int(1 if sample_name[:16] in highTMB else 0))
ava.append(t2 - t1)
t = np.mean(np.array(ava))
print("average time is : ", t)
print("recall : ", wpp / wp)
print("acc for patch : ", (wpp + wpn) / wc)
def get_result():
sample_name = os.listdir(val_base_dir)
x = tf.placeholder(tf.float32, [None, height, width, channel], name="inputs")
is_training = tf.placeholder(tf.bool, name="is_train")
vgg = VGG()
out = vgg.model(x, is_training)
prob = tf.nn.sigmoid(out)
y_pred = tf.greater(prob, 0.5)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_saved_path)
ratio = []
ans = []
lab = []
avt = []
t1 = time()
pc = 0
for name in sample_name:
pth_name = val_base_dir + "/" + name
print(pth_name)
pos_patch_num = 0
patch_num = len(os.listdir(pth_name))
for fname in os.listdir(pth_name):
# print(pth_name + '/' + fname)
img = cv.imread(pth_name + '/' + fname)
img = dataAugmentation(img)
img = (img - [B_MEAN, G_MEAN, R_MEAN]) / 255.0
img = cv.resize(img, (224, 224))
img = np.expand_dims(img, axis=0)
# print(img.shape)
predict, p = sess.run([y_pred, prob], {x: img, is_training: is_train})
# print("predict : ", predict, "probability : ", p)
# print("labels : ", (1 if fname[:16] in highTMB else 0))
pos_patch_num += int(predict)
ratio.append(pos_patch_num / patch_num)
ans.append((pos_patch_num / patch_num) >= 0.5)
lab.append((1 if name[:16] in highTMB else 0))
print("postive num : %d, patch num : %d, ratio : %.5f" % (pos_patch_num, patch_num, pos_patch_num / patch_num))
print("predict : %d, labels : %d" % (int((pos_patch_num / patch_num) >= 0.5), int(1 if name[:16] in highTMB else 0)))
print("Done")
pc += int((pos_patch_num / patch_num) >= 0.5)
if pc * 2 >= len(sample_name) :
print("TMB value is High !")
else:
print("TMB value is Low !")
acc = np.mean(np.array(ans) == np.array(lab))
print("acc : ", acc)
t2 = time()
avt.append(t2 - t1)
t = np.mean(np.array(avt))
print("Average predict time : ", t)