Thanks, is there a way to know what operations are allowed in the context of building a loss function? I can see that T.eq would have 0 gradients everywhere except the discontinuous point at which the function equals 1, but I'm having trouble imagining what the gradient would be for something like T.set_subtensor, which also seems to have a 0 gradient.
On Monday, March 6, 2017 at 11:38:59 AM UTC-6, Jesse Livezey wrote: > > There is nothing wrong with using T.eq. But, the derivatives with respect > to the inputs will be zero, so your cost function is not useful for > training. > > On Sunday, March 5, 2017 at 8:03:12 PM UTC-8, > [email protected] wrote: >> >> Thanks Jesse, so are there operations that are "safe" to use and others >> that aren't? Where can I find this information? Also, I've used T.eq before >> in another custom loss function which works correctly and doesn't return 0 >> gradients, but my use case there is in computing array indices, such as the >> way I'm using it in this line: >> >> pred_up = T.flatten(pred_one_hot[T.eq(unchanged_col, 0).nonzero(), >> preprocess.up_index]) >> >> Is T.eq ok to use in some contexts and not others? >> >> On Sunday, March 5, 2017 at 9:14:20 PM UTC-6, Jesse Livezey wrote: >>> >>> The gradient of T.eq will be zero (almost) everywhere and you're using >>> it to compute num_win and num_lose. >>> >>> On Sunday, March 5, 2017 at 2:42:14 PM UTC-8, >>> [email protected] wrote: >>>> >>>> Also, the return values of this loss function are small compared to >>>> cross-entropy, some sample values after random initialization were around >>>> +/- 0.01. There is a LSTM layer and the input sequences are thousands of >>>> elements long, so I suspected vanishing gradients. However, I'm printing >>>> out the min, max, and mean of the gradients w.r.t each parameter, and they >>>> are all exactly equal to 0, which seems to indicate a different problem. >>>> >>>> On Sunday, March 5, 2017 at 3:59:42 PM UTC-6, >>>> [email protected] wrote: >>>>> >>>>> I have defined a custom loss function, and despite the loss function >>>>> returning correct values given the inputs, the gradients are all always 0 >>>>> w.r.t each of my parameters. I am not suppressing any theano errors >>>>> including the disconnected input error, so I can't explain what is >>>>> causing >>>>> this. I have copied the loss function below; in words, I first convert a >>>>> 3 >>>>> class softmax output into a one hot representation, then I compare a >>>>> subset >>>>> of it to the response and compute a quantity of interest. More generally, >>>>> I >>>>> was under the impression that if one could express a function using >>>>> theano >>>>> ops, it could be used as a loss function. Is this not the case? >>>>> >>>>> def calc_one_hot_loss(pred, y, mask): >>>>> mask_flat = T.flatten(mask) >>>>> length = T.sum(mask_flat, dtype='int32') >>>>> pred_unmasked = pred[mask_flat.nonzero()] >>>>> max_indices = T.argmax(pred_unmasked, axis=1) >>>>> pred_zero = T.set_subtensor(pred_unmasked[:], 0) >>>>> pred_one_hot = T.set_subtensor(pred_zero[T.arange(length), >>>>> max_indices], 1) >>>>> y_unmasked = y[mask_flat.nonzero()] >>>>> unchanged_col = pred_one_hot[:, preprocess.unchanged_index] >>>>> pred_up = T.flatten(pred_one_hot[T.eq(unchanged_col, 0).nonzero(), >>>>> preprocess.up_index]) >>>>> pred_down = T.flatten(pred_one_hot[T.eq(unchanged_col, 0).nonzero(), >>>>> preprocess.down_index]) >>>>> y_up = T.flatten(y_unmasked[T.eq(unchanged_col, 0).nonzero(), >>>>> preprocess.up_index]) >>>>> y_down = T.flatten(y_unmasked[T.eq(unchanged_col, 0).nonzero(), >>>>> preprocess.down_index]) >>>>> diff_up = T.abs_(pred_up - y_up) >>>>> diff_down = T.abs_(pred_down - y_down) >>>>> diff_sum = diff_up + diff_down >>>>> num_win = T.sum(T.eq(diff_sum, 0)) >>>>> num_lose = T.sum(T.eq(diff_sum, 2)) >>>>> loss = -1 * (num_win - num_lose) / length >>>>> return loss >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> -- --- 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.
