TensorFlow 在训练或验证时给出错误“InvalidArgumentError: Input is empty”

Question

I am trying to run the Resnet model on custom images (transfer learning).

My directory tree looks like this:

|-train
|  |-class1
|  |    |-image1
|  |    |-image2
|  |    |-....
|  |-class2
|       |-image1
|       |-image2
|       |-....
|-val
   |-class1
   |    |-image1
   |    |-image2
   |    |-....
   |-class2
        |-image1
        |-image2
        |-....

And I created the tensorflow datasets like this:

train_ds = tf.keras.preprocessing.image_dataset_from_directory( "train", labels='inferred', label_mode='int', image_size=(img_height, img_width), batch_size=batch_size)

val_ds = tf.keras.preprocessing.image_dataset_from_directory( "val", labels='inferred', label_mode='int', image_size=(img_height, img_width), batch_size=batch_size)

But when I train or test the dataset, after a few images it gives me an error:

InvalidArgumentError: Input is empty [[{{node decode_image/DecodeImage}}]] [Op:IteratorGetNext]

<Figure size 720x720 with 0 Axes>.

The dataset I am using is here - https://github.com/xuequanlu/I-Nema - and I have converted all the .tif images to .jpg . What could be causing this?

Thanks in advance!

EDIT: here is the error log: https://pastecode.io/s/82hk68ar

Answer 1

in your code you have

train_ds = tf.keras.preprocessing.image_dataset_from_directory( "train", labels='inferred', label_mode='int', image_size=(img_height, img_width), batch_size=batch_size)

"train" should be the full path to the directory with the training images not a string. for example train_dir=os.path.join(my_dir, 'train') where my_dir holds the train and val sub directories

train_dir=os.path.join(my_dir, 'train')
val_dir=os.path.join(my_dir, 'val')
train_ds = tf.keras.preprocessing.image_dataset_from_directory( train_dir, labels='inferred', label_mode='int', image_size=(img_height, img_width), batch_size=batch_size)
val_ds = tf.keras.preprocessing.image_dataset_from_directory( val_dir, labels='inferred', label_mode='int', image_size=(img_height, img_width), batch_size=batch_size)

Also note since you are using label_mode='int' then in model.compile specify the loss as sparse_categorical_crossentropy