我正在尝试在ImageNet预训练权重上实现ResNet50。以下是完整的代码：
from tensorflow import keras
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense
from tensorflow.keras import optimizers
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.preprocessing import image
#Data preparation below
train_datagen = ImageDataGenerator(rescale = 1./255,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True)
test_datagen = ImageDataGenerator(rescale = 1./255)
traindata = train_datagen.flow_from_directory('dataset/training_set',
target_size = (50, 50),
batch_size = 32,
class_mode = 'categorical')
testdata = test_datagen.flow_from_directory('dataset/test_set',
target_size = (50, 50),
batch_size = 32,
class_mode = 'categorical')
''' Building RESNET50 by freezing convolutional layers, clipping FC layers (include_top=False) and # manually including them by Dense layers. If include_top is set to True, a custom input shape of 50x50x3 cannot be provided which is imperative due to the size of the images of dataset'''
from tensorflow.keras.applications import ResNet50
resnet50model = ResNet50(weights='imagenet', include_top=False, input_shape=(50, 50, 3))
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Conv2D, MaxPool2D , Flatten
resnet50_1 = Sequential()
resnet50_1.add(Flatten())
resnet50_1.add(Dense(units=1000,activation="relu"))
resnet50_1.add(Dense(units=2, activation="softmax"))
resnet50model_new = Model(inputs = resnet50model.input, outputs = resnet50_1(resnet50model.output))
#freezing convolutional layers below
for layers in (resnet50model_new.layers)[:-2]:
print(layers)
layers.trainable = False
#Using keras-buoy for resumable training
resnet50model_new.compile(loss = "categorical_crossentropy", optimizer = optimizers.SGD(lr=0.0001, momentum=0.9), metrics=["accuracy"])
from keras_buoy.models import ResumableModel
resumable_model_resnet50 = ResumableModel(resnet50model_new, save_every_epochs = 1, custom_objects=None, to_path='dataset/saved_models/resnet50.h5')
history_resnet50 = resumable_model_resnet50.fit_generator(traindata, steps_per_epoch=157106/32, epochs=15, validation_data=testdata, validation_steps=55504/32)
#Now, to evaluate
import glob
normal_cases_dir = xxx
idc_cases_dir = xxx
normal_cases = glob.glob(normal_cases_dir+'/*.png')
idc_cases = glob.glob(idc_cases_dir+'/*.png')
test_data = []
test_labels = []
import cv2
import keras
for img in normal_cases:
img = cv2.imread(str(img))
img = cv2.resize(img, (50,50))
if img.shape[2] ==1:
img = np.dstack([img, img, img])
else:
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)/255.
label = keras.utils.to_categorical(0, num_classes=2)
test_data.append(img)
test_labels.append(label)
for img in idc_cases:
img = cv2.imread(str(img))
img = cv2.resize(img, (50,50))
if img.shape[2] ==1:
img = np.dstack([img, img, img])
else:
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32)/255.
label = keras.utils.to_categorical(1, num_classes=2)
test_data.append(img)
test_labels.append(label)
test_data = np.array(test_data)
test_labels = np.array(test_labels)
test_loss_resnet50, test_score_resnet50 = resnet50model_new.evaluate(test_data, test_labels, batch_size=32)
print("Loss on test set: ", test_loss_resnet50)
print("Accuracy on test set: ", test_score_resnet50)
'''The above 2 print statement gives:
Loss on test set: 0.5971299170691742
Accuracy on test set: 0.7161286
'''
preds_resnet50 = resnet50model_new.predict(test_data, batch_size = 32)
preds_resnet50 = np.argmax(preds_resnet50, axis=-1)
orig_test_labels = np.argmax(test_labels, axis=-1)
cm = confusion_matrix(orig_test_labels, preds_resnet50)
print(cm)
'''The above print statements give a confusion matrix:
array([[39748, 0],
[15756, 0]], dtype=int64)
'''
#The above confusion matrix has zero values for false positive (FP) and true positive (TP)
#preds_resnet50 returns:
# array([0, 0, 0, ..., 0, 0, 0], dtype=int64)
# all zeroes ^
问题是所有的预测都是负数(零(。我发现这个问题很相似：Keras model.predict总是0然而，并没有正确的答案，他们所建议的只是即使在测试集上进行评估时也要重新缩放值。我已经这样做了，但问题仍然存在。
model.evaluate给出了71.6%的准确率，因此测试数据和测试标签似乎是正确的。在使用argmax之前，请尝试打印出preds_resnet50。这应该显示每个样本的类为标签0的概率和类为标签1的概率。可能会让我们对发生的事情有一些了解。试着从argmax语句中删除axis-1，我不明白为什么需要它。