prediction difference between batch=1 and batch=16

[alanbekker]

Any ideas why I'm receiving different prediction values when running with batch_size=1,16?
find code below:
Thanks!

def predict(self,speech_input):
labels = np.empty(0, int)
labels = np.append(labels, range(speech_input.shape[0]), axis=0)
feature,logits,_ = self.session.run(
[self.features,self.logits,self.end_points_speech],
feed_dict={self.is_training: False, self.batch_dynamic: labels.shape[0],
self.margin_imp_tensor: 50,
self.batch_speech: speech_input})
#self.batch_labels: labels.reshape([labels.shape[0], 1])})

    # Extracting the associated numpy array.
    #print (feature[0])

    return  feature,logits

[astorfi]

Are you running on your own dataset?

[alanbekker]

Yes

…

On Sun, May 13, 2018, 20:00 Amirsina Torfi ***@***.***> wrote: Are you running on your own dataset? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#16 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AE_DSc68BlIm6FqR0HSmh25ZDkJCOEB9ks5tyGa3gaJpZM4T02CP> .

[astorfi]

I do not get your question actually. Why are you expecting to get the same prediction values?

[alanbekker]

Because I'm in inference stage after I finished the training stage look this pseudo code: for i in batch16: yi=model.predict(xi) and I compare it to: y=model.predict(batch16) and this 16 length vector y look different... and the results look different

…

On Sun, May 13, 2018, 20:09 Amirsina Torfi ***@***.***> wrote: I do not get your question actually. Why are you expecting to get the same prediction values? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#16 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AE_DSUtaa0xtwPCiiTDfms0D4OrUTynJks5tyGjcgaJpZM4T02CP> .

[astorfi]

What do you mean by batch16?

[alanbekker]

Batch16 is a (16,80,40,20) tensor a opposed to yi which is (1,80,40,20) signle example

…

On Mon, May 14, 2018, 04:23 Amirsina Torfi ***@***.***> wrote: What do you mean by batch16? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#16 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AE_DSTwBRBfnPg8gVq0csGjtDKI5tuUyks5tyNybgaJpZM4T02CP> .

[astorfi]

I think you should debug the values more carefully. It's weird. It's hard for me to have an idea about the details and the values that you are getting. Please see if the averaging is performing correctly. This might be due to the drastic changes happened in new TensorFlow versions.

[astorfi]

Please open this issue if the problem has not been resolved!
Thanks