inference problem

[misaghkord]

hi @potterhsu i have a very odd problem i have trained my model and got an very good accuracy. but when i try to inference my images if i use batch sizes lower than 16 (i.e 1) i get rangom numbers mostly 9. can you please help me?

[potterhsu]

Could you please provide a minimal reproduction of your issue? Also have you followed our sample guide: inference_sample.ipynb and inference_outside_sample.ipynb?

[misaghkord]

tnx for your quick response: i used this code from your inference_outside_sample.ipynb:

import tensorflow as tf
from model import Model


path_to_image_files = ['data/images/18.jpg', 'data/images/31.jpg', 'data/images/62.jpg',
                       'data/images/522.jpg', 'data/images/952.jpg', 'data/images/888.jpg']

images = []
for path_to_image_file in path_to_image_files:
    image = tf.image.decode_jpeg(tf.read_file(path_to_image_file), channels=3)
    image = tf.reshape(image, tf.shape(image))
    image = tf.image.convert_image_dtype(image, dtype=tf.float32)
    image = tf.multiply(tf.subtract(image, 0.5), 2)
    image = tf.image.resize_images(image, [65, 290])
    images.append(image)
images = tf.stack(images)


digits_logits = Model.inference(images)
digits_predictions = tf.argmax(digits_logits, axis=2)
digits_predictions_string = tf.reduce_join(tf.as_string(digits_predictions), axis=1)

sess = tf.InteractiveSession()
restorer = tf.train.Saver()
checkpoint_path = tf.train.latest_checkpoint('logs/train')
restorer.restore(sess, checkpoint_path)
digits_predictions_string_val, images_val = sess.run([digits_predictions_string, images])
images_val = (images_val / 2.0) + 0.5


idx = 5
image_val = images_val[idx]
digits_prediction_string_val = digits_predictions_string_val[idx]

print ('digits: %s' % digits_prediction_string_val)

sess.close()

you will notice that my data has different dimensions. i get correct results when using 6 images in batch for example but when using 2 images i get completely different result.

[potterhsu]

It seems weird, the number of images should not affect the result. What are the shapes of images and digits_logits in your case?

[misaghkord]

image shape = [65 290 3] digits_logits = [16 11] i got around this problem by fine tuning my network with batch size 1 after the main training was done. now i get accurate results by feeding single image to the network.

[jewes]

change image = tf.reshape(image, [64, 64, 3]) to image = tf.image.resize_images(image, (64, 64), method=0)