pytorch-dnc-Question

[SoroorM]

Hi Dear Thank you for sharing code of NTM-DNC on github & thanks for the your explanations, I have a problem when trying to test the code, can you help me? the error is: NameError: name 'raw_input' is not defined how can i fixed this? your kind attention is highly appreciated in advance. best regards

[jingweiz]

Hey, So a quick guess would be: are you using python3 instead of python2? Cos according to https://stackoverflow.com/questions/954834/how-do-i-use-raw-input-in-python-3 raw_input() is renamed to input() in python3. Also this repos is only tested under python2, and I never tried running it under python3 :)

[SoroorM]

Hi Dear Jacob

Thanks again for your guide, that error was solved, but continues of testing after one epoch, stopped and not any action, is that usual? but after close the cmd for milisecond show the error and i can't see clearly... and in visdom not showing anything... please help me to solve these problems

best regards

[jingweiz]

Hey, so in testing mode, the raw_input() is used there to pause the code every step, so as to get a step by step visualization. So you can just keep pressing Enter during testing then visdom will show the visualizations after each time you pressed.

[SoroorM]

thank you very much...

but still haven't any visualizations in visdom :(

should I change the code?

i open this address : "http://localhost:8097/env/MACHINE_TIMESTAMP_test" to see the test result and for train result open this : "http://localhost:8097/env/MACHINE_TIMESTAMP" but don't show anything, and I can't see anything on the basic address, too.

your kind attention is highly appreciated in advance. best regards

[jingweiz]

Hey, so just to make sure, you replaced the MACHINE and TIMESTAMP with the values in your code right?

[SoroorM]

can you tell me what's wrong of my runs? these are very different with your results.

unfortunately I still have any errors...., but first I'll try more to solve them, if I couldn't solve, ask you again, sorry... very very thanks for your helping you are the best 👍 :)

[jingweiz]

Super wierd, I tested multiple times and the curves look quite similar ... did u change anything in the hyperparameters? and is this ntm or dnc?

[jingweiz]

also, can u show me this part of the file: utils/options.py:

class Params(object):   # NOTE: shared across all modules
    def __init__(self):
        self.verbose     = 0            # 0(warning) | 1(info) | 2(debug)

        # training signature
        self.machine     = "daim"       # "machine_id"
        self.timestamp   = "17053100"   # "yymmdd##"

[SoroorM]

no, i didn't change anything in hyperparameters ...

sure, here you are:

[SoroorM]

Hi Dear Jacob

I tried once again, and after ending training my result was little better than previous training but still different with your result.can you guide me, why?

and also result of testing was different; i think that's very bad result, and I do not know why... :(

Best Regards

[jingweiz]

Hey, sorry I'm quite busy these several days, I will try to solve this issue next week.

[SoroorM]

Hi okay, I'm waiting for your response and very very thank you for your kindness. :)

[jingweiz]

Hey, So I ran it multiple times, the training curves differ a bit but they all converge and none diverge. Are you sure that you did not change any hyperparameter settings?

[SoroorM]

Hi Dear Jacob first, thank you for your Follow up oh sure, I didn't change any hyperparameter settings. I was just want to experiment this code on my computer and after it, change it.

[jingweiz]

OK, I'm running more tests, in the meanwhile, you can try to comment off those two lines 28-29 in core/controllers/lstm_controllers.py and run it again:

        self.lstm_hidden_vb = [self.lstm_hidden_vb[0].clamp(min=-self.clip_value, max=self.clip_value),
                               self.lstm_hidden_vb[1]]

[SoroorM]

okay thanks I'll try it, but the results seem to be getting worse...

[jingweiz]

Can u tell me which commit are you in?

[SoroorM]

I see that address and two lines 28-29 in core/controllers/lstm_controllers.py, were not comments, so results of previous run the code, with these two lines; but I commented these two line codes, to see the what happen? when I run the 'main.py' , in "command prompt", I have this:

and in "visdom", already have:

[SoroorM]

final result of training is 👍

[jingweiz]

Hey, sorry for not answering earlier, but I'm moving these days, and hopefully will be done in the next day or so, then I will take a closer look. In the meanwhile, can you tell me which commit your are in? Cos for all the runs that I tested, I never saw anything like the ones you showed as they never diverge. Also a result of git diff would also be helpful! Thanks!

[SoroorM]

Hi Dear Jacob I created it.

Best Regards

[jingweiz]

What do you mean? Also one important thing I realized is that for the "ntm", the batch_size should be 16 instead 1, that should solve all your problems. I will update the repo after I run more tests, but you can test by changing the batch_size.

[SoroorM]

ok, i will change it but, what's the batch_size ?

[SoroorM]

thank you that problem was fixed.

you are the best 👍 :)

what is the batch_size?
Is that length of inputs?

[SoroorM]

but result the test mode seems not very good... 🤔

[jingweiz]

Hey, I already tested it, and now the results are consistent. Where did you change, for me, the different runs are like this:

So before when the batch_size is 1, I got lucky that the different runs for me always converge, then after you pointed it out, I did more runs and found some diverge, a reasonable guess would be that when the batch size is 1 then the sgd would be too stochastic. Now when I change it to 16, it is much more consistent and stable :) I will update the repo very soon today, thank you so much for the input :)

[SoroorM]

Thank you very much Dear Jacob 🌺

I'm just a little confused about self.seed = 1 , why changing it? can you guide me about this?

[jingweiz]

Hey, the seed is for the random number generator, so if the seed is fixed, you would basically get more or less the same learning curve, if an algorithm is stable, the curve should be somewhat consistent across different seeds.