CNN performance worse when loading data with tf.Data

Question

I have a trained EfficientNetB2 neural network that I'm using for image classification. When I'm loading the images with PIL like this:

image = Image.open(item)
image = image.convert('RGB').resize((120, 120))
image = np.array(image)

if image.ndim == 3:
    image = np.expand_dims(image, axis = 0)


predictions.append(model.predict(image))

I get accuracy at around 90%. This is, however, extremely slow so I tried using tf.data to load my dataset. This looks something like this:

ds = tf.data.Dataset.list_files(str(test_dir / '*' / '*'))
ds = (ds
      .map(load_data, num_parallel_calls=tf.data.experimental.AUTOTUNE)
      .cache()
      .batch(32)
        
      .prefetch(tf.data.experimental.AUTOTUNE)
    )

And this is the load_data function:

def load_image_and_label(file_path):
   label = tf.strings.split(file_path, os.sep)[-2]
   image = tf.io.read_file(file_path)
   image = tf.io.decode_jpeg(image, channels=3, dct_method='INTEGER_ACCURATE')
   image = tf.image.resize(image, target_size)
   

   return image, label

This is, as expected, much much faster but the accuracy drops to around ~70%. I've tried moving stuff around but I just can't figure out why this happens. If anyone has any suggestions it would be much appreciated.

P.S. I'm aware that there is an almost identical question already asked on stack overflow but the answer to that question doesn't change anything for my situation, this is why I'm posting this as a separate question. Thank you.

Edit: I tried not using tf.Dataset but still using my load_image_and_label function, the results were again ~90% accuracy, meaning that there is a problem somewhere with tf.Dataset pipeline, anyone got any experience with this kind of a problem?

Why are you shuffling your images during inference (predicting) ? — AloneTogether, Nov 04 '22 at 11:12
I'm not, I've edited that out, I've put that in the question by accident — WholesomeGhost, Nov 04 '22 at 11:15
Can you set dct_method='INTEGER_ACCURATE' for tf and check- https://stackoverflow.com/questions/44514897/difference-between-the-return-values-of-pil-image-open-and-tf-image-decode-jpeg — Vijay Mariappan, Nov 04 '22 at 11:22
Can you try loading your images in your first approach like you are doing in `load_image_and_label`, to see if it is a problem with `PIL` and `TF` or the `tf.data.Dataset`. The image resizing methods could also be different. — AloneTogether, Nov 04 '22 at 11:31
I wanted to do that before but I get stuck because `file_path` variable is a tensor and I don't know how to extract the String value from it, do you have any suggestions on that front ? — WholesomeGhost, Nov 04 '22 at 11:33
Do you set a batch size when you don't use `tf.data.Dataset`? — Djinn, Nov 04 '22 at 22:33

CNN performance worse when loading data with tf.Data

0 Answers0