我的 ResNet 遷移學習模型在訓練時始終保持 0.5 准確度 - 怎么了？

Question

我是一名開始學習 ML 的學生。 我正在做的是使用 ResNet50 遷移學習與 Keras 進行簡單分類。 我想要的結果輸出只是“臉與否（像這樣）”。 我真的不在乎模型的表現。 但訓練時我的准確率始終保持在 0.5。 我的代碼有什么問題？ 請幫忙。

數據是這樣的：

X = face_images + animal_images # 10 face_images and 10 animal_images
# if image is face :1, else: 0

y = [[1]]* len(face_images) + [[0]] * len(animal_images)
X = np.array(X)
y = np.array(y)
print(X.shape, y.shape)

(30, 75, 75, 3) (30, 1)

我的預訓練模型是：

# pre-trained model
input_layer = layers.Input(shape=(75,75,3))
base_model = ResNet50(weights='imagenet',
                      input_tensor=input_layer,
                      include_top=False)
base_model.summary()

和轉移模型：

# transfer
# flatten last layer, and added dense layer
last_layer = base_model.output
flatten = layers.Flatten()(last_layer)
my_layer = layers.Dense(256, activation='relu')(flatten)
my_layer = layers.Dense(256, activation='relu')(my_layer)
output_layer = layers.Dense(1,activation='sigmoid')(my_layer)

# input_layer = layers.Input(shape="75,75,3")
model = Model(inputs=input_layer, outputs=output_layer)
model.summary()

並編譯：

model.compile(loss='mse', optimizer='adam',metrics=['accuracy'])
print("model compile completed.")

歷史結果history = model.fit(X_p, y, epochs=100, shuffle=True)

Epoch 1/100
1/1 [==============================] - 4s 4s/step - loss: 0.2589 - accuracy: 0.5000
Epoch 2/100
1/1 [==============================] - 1s 613ms/step - loss: 0.2570 - accuracy: 0.5000
Epoch 3/100
1/1 [==============================] - 1s 611ms/step - loss: 0.2553 - accuracy: 0.5000
Epoch 4/100
1/1 [==============================] - 1s 612ms/step - loss: 0.2538 - accuracy: 0.5000
Epoch 5/100
1/1 [==============================] - 1s 620ms/step - loss: 0.2526 - accuracy: 0.5000
Epoch 6/100
1/1 [==============================] - 1s 610ms/step - loss: 0.2516 - accuracy: 0.5000
.
.
.
Epoch 96/100
1/1 [==============================] - 1s 608ms/step - loss: 0.2500 - accuracy: 0.5000
Epoch 97/100
1/1 [==============================] - 1s 607ms/step - loss: 0.2500 - accuracy: 0.5000
Epoch 98/100
1/1 [==============================] - 1s 925ms/step - loss: 0.2500 - accuracy: 0.5000
Epoch 99/100
1/1 [==============================] - 1s 916ms/step - loss: 0.2500 - accuracy: 0.5000
Epoch 100/100
1/1 [==============================] - 1s 913ms/step - loss: 0.2500 - accuracy: 0.3333

Answer 1

我解決了它：

1. 火車圖像不是 RGB
2. 'preprocess_input' 是由不同的 cnn 模型而不是 resnet 導入的