訓練損失減少但准確率始終為 0？

Question

我嘗試訓練一個 model，輸入是 (3000,1) 向量，主要由負數組成，非規范化輸入。 Output 是二進制圖像，表示為矢量 (2500,1)。

我的 model 是這樣的：

model = Sequential()
model.add(Dense(3000, input_shape=(x_train.shape[1:]), activation='linear'))
model.add(Dense(2500, activation='relu'))
model.add(Dense(2500, activation='relu'))
model.add(Dense(2500, activation='relu'))
model.add(Dense(2500, activation='relu'))
model.add(Dense(y_train.shape[1], activation='sigmoid'))
model.compile(optimizer=Adam(learning_rate=0.0001), loss='binary_crossentropy', metrics=['accuracy'])

結果是這樣的：

Epoch 1/300
1/1 - 0s - loss: 0.6999 - accuracy: 0.0000e+00 - val_loss: 0.6930 - val_accuracy: 0.0000e+00
Epoch 2/300
1/1 - 0s - loss: 0.6843 - accuracy: 0.0000e+00 - val_loss: 0.6911 - val_accuracy: 0.0000e+00
Epoch 3/300
1/1 - 0s - loss: 0.6700 - accuracy: 0.0000e+00 - val_loss: 0.6944 - val_accuracy: 0.0000e+00
Epoch 4/300
1/1 - 0s - loss: 0.6515 - accuracy: 0.0000e+00 - val_loss: 0.7081 - val_accuracy: 0.0000e+00
Epoch 5/300
1/1 - 0s - loss: 0.6314 - accuracy: 0.0000e+00 - val_loss: 0.7349 - val_accuracy: 0.0000e+00
Epoch 6/300
1/1 - 0s - loss: 0.6147 - accuracy: 0.0000e+00 - val_loss: 0.7568 - val_accuracy: 0.0000e+00
Epoch 7/300
1/1 - 0s - loss: 0.6006 - accuracy: 0.0000e+00 - val_loss: 0.7615 - val_accuracy: 0.0000e+00
Epoch 8/300
1/1 - 0s - loss: 0.5865 - accuracy: 0.0000e+00 - val_loss: 0.7560 - val_accuracy: 0.0000e+00
Epoch 9/300
1/1 - 0s - loss: 0.5738 - accuracy: 0.0000e+00 - val_loss: 0.7515 - val_accuracy: 0.0000e+00
Epoch 10/300
1/1 - 0s - loss: 0.5637 - accuracy: 0.0000e+00 - val_loss: 0.7533 - val_accuracy: 0.0000e+00
Epoch 11/300
1/1 - 0s - loss: 0.5555 - accuracy: 0.0000e+00 - val_loss: 0.7629 - val_accuracy: 0.0000e+00
Epoch 12/300
1/1 - 0s - loss: 0.5490 - accuracy: 0.0000e+00 - val_loss: 0.7766 - val_accuracy: 0.0000e+00
Epoch 13/300
1/1 - 0s - loss: 0.5441 - accuracy: 0.0000e+00 - val_loss: 0.7877 - val_accuracy: 0.0000e+00
Epoch 14/300
1/1 - 0s - loss: 0.5402 - accuracy: 0.0000e+00 - val_loss: 0.7937 - val_accuracy: 0.0000e+00
Epoch 15/300
1/1 - 0s - loss: 0.5370 - accuracy: 0.0000e+00 - val_loss: 0.7966 - val_accuracy: 0.0000e+00
Epoch 16/300
1/1 - 0s - loss: 0.5346 - accuracy: 0.0000e+00 - val_loss: 0.8001 - val_accuracy: 0.0000e+00
Epoch 17/300
1/1 - 0s - loss: 0.5329 - accuracy: 0.0000e+00 - val_loss: 0.8065 - val_accuracy: 0.0000e+00
Epoch 18/300
1/1 - 0s - loss: 0.5315 - accuracy: 0.0000e+00 - val_loss: 0.8152 - val_accuracy: 0.0000e+00
Epoch 19/300
1/1 - 0s - loss: 0.5305 - accuracy: 0.0000e+00 - val_loss: 0.8253 - val_accuracy: 0.0000e+00
Epoch 20/300
1/1 - 0s - loss: 0.5294 - accuracy: 0.0000e+00 - val_loss: 0.8337 - val_accuracy: 0.0000e+00
Epoch 21/300
1/1 - 0s - loss: 0.5283 - accuracy: 0.0000e+00 - val_loss: 0.8408 - val_accuracy: 0.0000e+00
Epoch 22/300
1/1 - 0s - loss: 0.5271 - accuracy: 0.0000e+00 - val_loss: 0.8476 - val_accuracy: 0.0000e+00
Epoch 23/300
1/1 - 0s - loss: 0.5259 - accuracy: 0.0000e+00 - val_loss: 0.8550 - val_accuracy: 0.0000e+00
Epoch 24/300
1/1 - 0s - loss: 0.5247 - accuracy: 0.0000e+00 - val_loss: 0.8625 - val_accuracy: 0.0000e+00
Epoch 25/300
1/1 - 0s - loss: 0.5235 - accuracy: 0.0000e+00 - val_loss: 0.8705 - val_accuracy: 0.0000e+00
Epoch 26/300
1/1 - 0s - loss: 0.5223 - accuracy: 0.0000e+00 - val_loss: 0.8794 - val_accuracy: 0.0000e+00
Epoch 27/300
1/1 - 0s - loss: 0.5211 - accuracy: 0.0000e+00 - val_loss: 0.8872 - val_accuracy: 0.0000e+00
Epoch 28/300
1/1 - 0s - loss: 0.5200 - accuracy: 0.0000e+00 - val_loss: 0.8940 - val_accuracy: 0.0000e+00
Epoch 29/300
1/1 - 0s - loss: 0.5188 - accuracy: 0.0000e+00 - val_loss: 0.8982 - val_accuracy: 0.0000e+00

准確性和驗證沒有增加。 驗證損失在某個時間點后開始增加。

即使我嘗試這個網絡非常小的數據集（17 daatset），它也不會順利收斂。

然后我嘗試決策樹回歸器，決策樹的得分為負數。 我檢查了數據集，但我找不到錯誤。 有什么問題，你能幫幫我嗎？

Answer 1

model 的任務很難獲得良好的准確性。

您在一個 output 中有 2500 個值，如果其中一個值錯誤，則 output 將是每個數據樣本的零精度。 您的任務不需要您計算准確性，您可以在這里只關注損失。

或者您可以手動定義您的特定 output 的准確度，這 2500 個值中有多少必須被正確預測才能稱為正確預測。 例如，這些值中有 50% 正確分類，每個值的誤差小於 0.5。