Model Saving

[djsaunde]

I am using the file SdA.py. I am wondering if there is a way to save the model which has been generated and trained; I would like to run vectors through the network and get back the "encoding" from the trained network.

[amit4111989]

Use pickle or cpickle (Python modules for serializing objects ) to store the generated model

[djsaunde]

This allows me to save the SdA object, but I would like to input vectors and retrieve their encoding, or "h(x)" in the literature, where "x" is the input.

[amarczew]

You can try something like:

def new_representation_X_train(hidden_layer):
    new_representation_train = theano.function(
         inputs = [index],
        outputs = self.sigmoid_layers[hidden_layer].output,
        givens = {
            self.x: train_set_x[index: index + 1 ]
        }
    )
    train_lines = train_set_x.get_value(borrow=True).shape[0]
    return [new_representation_train(i) for i in xrange(train_lines)]

[ohuole233]

Thanks for your help, another question is that what does [index] mean in your code?^-^

At 2016-09-22 08:22:24, "Alison Marczewski" [email protected] wrote:

You can try something like:

def new_representation_X_train(hidden_layer):
new_representation_train = theano.function(
inputs = [index],
outputs = self.sigmoid_layers[hidden_layer].output,
givens = {
self.x: train_set_x[
index: index + 1
]
}
)
train_lines = train_set_x.get_value(borrow=True).shape[0]
return [new_representation_train(i) for i in xrange(train_lines)]

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub, or mute the thread.

[ohuole233]

Oh, I get it and I have added it to the DBN class. Thank you again.

At 2016-09-22 08:22:24, "Alison Marczewski" [email protected] wrote:

You can try something like:

def new_representation_X_train(hidden_layer):
new_representation_train = theano.function(
inputs = [index],
outputs = self.sigmoid_layers[hidden_layer].output,
givens = {
self.x: train_set_x[
index: index + 1
]
}
)
train_lines = train_set_x.get_value(borrow=True).shape[0]
return [new_representation_train(i) for i in xrange(train_lines)]

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub, or mute the thread.

[amarczew]

@ohuole233 remember this value comes from a sigmoid function. If you want the linear output, you need to create a linear_output = T.dot(input, self.W) + self.b and get its value inside the function I wrote above.

[ohuole233]

@amarczew Hi.professor~ After the method new_representation_X_train added in the class DBN,and I run the code like this:

import theano
import DBN
import mlp
import rbm
import numpy

pretraining_epochs=2
pretrain_lr=0.1
k=1
batch_size=2

train_set=[[1,2,3,1],[2,3,1,4],[3,6,4,2],[3,5,4,5],[4,2,5,2],[4,3,2,6]]
train_set_x = theano.shared(numpy.asarray(train_set,
dtype=theano.config.floatX),
borrow=True)
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
numpy_rng=numpy.random.RandomState(123)

dbn=DBN.DBN(numpy_rng=numpy_rng,n_ins=4,hidden_layers_sizes=[3,3],n_outs=2)
pretraining_fns=dbn.pretraining_functions(train_set_x=train_set_x,batch_size=batch_size,k=k)

print dbn.new_representation_X_train(2)

An unexpected bug occured to remind me that "TypeError: new_representation_X_train() takes exactly 1 argument (2 given)".
Is there any idea cope with this?
Much thanks and Waiting for your reply soon.

[amarczew]

@ohuole233 probably the problem is related to self parameter that is passed implicitly, because you are calling a function from an object. Change to def new_representation_X_train(self, hidden_layer): may fix it.

[ohuole233]

Grateful for your reply. but it seems 'global name 'index' is not defined. I haven't got how to define it and my objective is to derive new representation for each input.Waiting for your assistance sincerely .
😀

发自网易邮箱大师
On 09/28/2016 03:41, Alison Marczewski wrote:

@ohuole233 probably the problem is related to self parameter that is passed implicitly, because you are calling a function from an object. Change to def new_representation_X_train(self, hidden_layer): may fix it.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.

[amarczew]

@ohuole233 you need to declare this tensor index = T.lscalar('index'), before this function

[ohuole233]

@amarczew Hi, Professor. I have abandoned training on GUPs with some special settlements. However, after the above repair including declaring the tensor and adding self, a new error named '' NameError: global name 'train_set_x' is not defined '' occurred. Is there anything to fix the bug?
It would be very kind of you to send me the DBN class code including the intermediate representation method and my email is [email protected].
Waiting for your reply. Thanks~^-^

[amarczew]

@ohuole233 Unfortunately I don't have this implementation. This function is working on SdA.py implementation. You can take a look on this file to solve the problem in your DBN implementation. I have declared this function inside build_finetune_functions(...)