show_and_tell.tensorflow icon indicating copy to clipboard operation
show_and_tell.tensorflow copied to clipboard

Published 20 hours ago • verified publisher icon jazzsaxmafia

Question about different results for the same picture

Open xingwen78 opened this issue 8 years ago • 6 comments

Hi

I downloaded the code and it worked very well, nice stuff. Just one question, when testing a same picture I got different results from time to time. Is that just because of the weights' float-point rounding ?

Thank you.

xingwen78 avatar Aug 23 '16 01:08 xingwen78

I have the same problem while running test_tf(). The output of VGG is different from time to time. Any suggestion?

chingyaoc avatar Sep 20 '16 12:09 chingyaoc

Same here. I first suspected the problem is due to random initialized parameters of textual and image embeddings, however as they are loaded from the model-72, the only reason for different results are due to different VGG activations.

kilickaya avatar Nov 18 '16 14:11 kilickaya

I'm wondering if this is because of the dropout layers? Is it still working at testing stage?

YuDeng avatar Nov 22 '16 02:11 YuDeng

Yes, this is probably due to dropout dropping some random connections at test time, which is connected to fc-7 layer (previous layers do not produce random activations) Is there any way to shut it down at test time?

kilickaya avatar Nov 22 '16 08:11 kilickaya

Hi ImportError: This module is deprecated. Use tf.nn.rnn_cell instead. i use from tensorflow.python.ops import rnn_cell instead from tensorflow.models.rnn import rnn_cell but ValueError: setting an array element with a sequence.

ilovezhenzhen avatar Dec 15 '16 01:12 ilovezhenzhen

Hi, I think the result is different, that is normal. It pick the top scores result by random, not only 1 result

nqkhai1706 avatar Jan 23 '17 04:01 nqkhai1706