Hello I am new to theano. Can you help me how to use get_value() function in order to save weights and biases after pertaining phase in both DBN and SDA? I am really appreciate you
On Wednesday, October 22, 2014 at 11:56:00 PM UTC+3:30, Kin wrote: > > *Thank you very much for your help, Pascal!* I am now able to save the > weights and biases by simply using the get_value function as suggested to > acquire the weights and biases of each RBM layers at the end of the > pretraining phase. If loading the pretrained model is desired, I skipped > the pretraining portion of the code entirely and the values (W, hbias, > vbias) are reassigned using set_value() prior to finetuning. > > - Cage > > On Tuesday, October 21, 2014 11:51:44 AM UTC-5, Pascal Lamblin wrote: >> >> On Mon, Oct 20, 2014, Cage Lawrence wrote: >> > I have used my own dataset and it almost took 8 hours just to pretrain >> the >> > model (without finetuning). I would love to save the weights and biases >> for >> > each layer to a file, then at the next time to use it I could simply >> load >> > the model from the file and skipping the pretraining process. What >> other >> > considerations should be noted if I were to save the weights and biases >> > *after* finetuning as well? >> >> One easy way, but maybe not the most flexible, would be to have a `save` >> method in DBN, that would get all the (vbias, W, hbias) parameter >> for all the RBMs, and for the final softmax layer (W and b). For >> each of these shared variables, you can then call numpy.save on the >> value returned by that_var.get_value(), so each of them is saved in a >> different .npy file. >> >> Then, in a `load` method, you could load them all and put them in shared >> variables, then go through the logic of `__init__` again, but specifying >> all the parameters for all of the components, instead of leaving some >> of them with the default value. For instance, you will have to provide >> vbias to the constructor of `RBM`. >> >> After fine-tuning is essentially the same thing, except that the value >> for the visible biases would not be used (and not change) during >> fine-tuning. >> >> If you want to try different hyper-parameters (sizes of layers, and so >> on), you can also create a file containing those values (maybe as a >> dictionary) in DBN.save, and use it in DBN.load. >> >> > I have read the Theano documentation on saving/loading models, and also >> > searched this group but the people who encountered the same problem did >> not >> > obtain a particularly detailed reply. Either that, or they have found >> > solutions very specific to topics outside of DBNs. What solutions would >> you >> > recommend? Should I be looking into editing the codes within RBM.py >> instead >> > of DBN.py? >> >> > >> > *Below are the threads/topics that I've read (but they didn't help my >> > case...):* >> > Theano Official Documentation: >> > >> http://deeplearning.net/tutorial/gettingstarted.html#loading-and-saving-models >> >> > Saving Weights and Biases for *MLP*: >> > >> https://groups.google.com/forum/#!searchin/theano-users/save/theano-users/gWbjkIpSjC4/mL-747Yz-P0J >> >> > Saving Weights and Biases for *Convolutional Networks: * >> > >> https://groups.google.com/forum/#!searchin/theano-users/saving/theano-users/tdF_j_JyD24/2iNU8ezProoJ >> >> > Saving & Resuming Model Training: >> > >> https://groups.google.com/forum/#!searchin/theano-users/saving$20parameters$20dbn/theano-users/2VwXqQgnBwA/-Xyu-uR4KXwJ >> >> > Save and Load Parameters of *DBN*: >> > >> https://groups.google.com/forum/#!searchin/theano-users/save$20parameters$20dbn/theano-users/GLcmlLoONic/u8KFFuG9JEMJ >> >> > >> > Thank you very much! >> > >> > - Cage >> > >> > -- >> > >> > --- >> > You received this message because you are subscribed to the Google >> Groups "theano-users" group. >> > To unsubscribe from this group and stop receiving emails from it, send >> an email to [email protected]. >> > For more options, visit https://groups.google.com/d/optout. >> >> -- >> Pascal >> > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
