層“max_pooling2d_30”的輸入0與層不兼容：預期ndim=4，發現ndim=5

Question

我剛剛開始接近 ml，我正在嘗試在使用 tf.keras.utils.image_dataset_from_directory 從圖像目錄獲得的圖像數據集上訓練模型（然后對其進行酸洗），以便它可以預測哪個字母是手指拼寫的在圖像中。

所以，這就是我解開我的數據集的方式：

#PICKLE LOAD

#TRAIN
  #images
with open('x_train.pkl', 'rb') as x_train_pickle:
  x_train_data = pickle.load(x_train_pickle)
  #labels
with open('y_train.pkl', 'rb') as y_train_pickle:
  y_train_data = pickle.load(y_train_pickle)

#VALIDATION
with open('x_val.pkl', 'rb') as x_val_pickle:
  x_val_data = pickle.load(x_val_pickle)

with open('y_val.pkl', 'rb') as y_val_pickle:
  y_val_data = pickle.load(y_val_pickle)

#TEST
with open('x_test.pkl', 'rb') as x_test_pickle:
  x_test_data = pickle.load(x_test_pickle)

with open('y_test.pkl', 'rb') as y_test_pickle:
  y_test_data = pickle.load(y_test_pickle)

Answer 1

您將 5D 數據提供給 Conv2d 而不是 4D。因此，您的數據應為 (batch_size*32, 180, 180, 3) 形狀，或者您可以使用 TimeDistributed 層包裝器在所有圖像上應用相同的卷積層在 5D 張量中。 例如：

model = Sequential()
model.add(tf.keras.layers.TimeDistributed(tf.keras.layers.Conv2D(32, 3, activation='relu')))

model.summary()