General models of Neural Networks compute the error using its output, compared with the groundtruth by some custom function. I'm trying build a model in witch the output vector *O* of a CNN is used to extract a feature vector *F* from the original input images *I*. I mean I want to compute error and gradients of the model using *F*, not directly *O.* My problem is that Theano doesn't "understand" *F* vector as part of computational graph cost and raises the following error:
*File "/usr/local/lib/python2.7/dist-packages/theano/gradient.py", line 532, in handle_disconnected raise DisconnectedInputError(message)theano.gradient.DisconnectedInputError: grad method was asked to compute the gradient with respect to a variable that is not part of the computational graph of the cost, or is used only by a non-differentiable operator: convolution2d_1_W* I tried to pass *disconnected_inputs='ignore'* to the *T.grad()* function and my program could run, but it means that the partial derivative of my loss function with respect to *F* will be 0 (zero) and it doesn't solve my problem. How can I make Theano understand *F* as part of computational cost? -- --- 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.
