listwise_cost
Closed this issue · 4 comments
I want to train ListNet to re-rank retrieved document so I got this error
Traceback (most recent call last):
File "ranking/rank/train.py", line 88, in
model.fit(X_train, y_train, X_test, y_test, Query, Query2, batchsize, n_epoch, n_hidden1, n_hidden2)
File "/home/ama/mhidy/nn_ranking/RankNet_chainer/ranking/rank/ListNet.py", line 193, in fit
self.trainModel(train_X, train_y, validate_X, validate_y, query, query_validate, n_epoch, batchsize)
File "/home/ama/mhidy/nn_ranking/RankNet_chainer/ranking/rank/ListNet.py", line 160, in trainModel
self.optimizer.update(self.model, x, t)
File "/home/ama/mhidy/.local/lib/python2.7/site-packages/chainer/optimizer.py", line 392, in update
loss = lossfun(*args, **kwds)
File "/home/ama/mhidy/nn_ranking/RankNet_chainer/ranking/rank/ListNet.py", line 61, in call
self.loss = self.listwise_cost(y, t)
File "/home/ama/mhidy/nn_ranking/RankNet_chainer/ranking/rank/ListNet.py", line 91, in listwise_cost
return - np.sum(self.topkprob(list_ans) * np.log(self.topkprob(list_pred)))
File "/home/ama/mhidy/nn_ranking/RankNet_chainer/ranking/rank/ListNet.py", line 85, in topkprob
vec_sort = np.sort(vec)[-1::-1]
File "/home/ama/mhidy/.local/lib/python2.7/site-packages/numpy/core/fromnumeric.py", line 824, in sort
a = asanyarray(a).copy(order="K")
File "/home/ama/mhidy/.local/lib/python2.7/site-packages/numpy/core/numeric.py", line 533, in asanyarray
return array(a, dtype, copy=False, order=order, subok=True)
File "/home/ama/mhidy/.local/lib/python2.7/site-packages/chainer/functions/array/get_item.py", line 71, in get_item
return GetItem(slices)(x)
File "/home/ama/mhidy/.local/lib/python2.7/site-packages/chainer/function.py", line 189, in call
self._check_data_type_forward(in_data)
File "/home/ama/mhidy/.local/lib/python2.7/site-packages/chainer/function.py", line 271, in _check_data_type_forward
six.raise_from(
AttributeError: 'module' object has no attribute 'raise_from'
Hi @mhidy thank you for the report.
Actually listwise_cost seems to be not implemented properly yet, although the functions are inside the file. In the current version, the cost function is only jensen-shannon divergence. (please see https://github.com/shiba24/learning2rank/blob/master/rank/ListNet.py#L29 )
In order to make listwise_cost available, we need to impement listwise_cost function with not numpy, but chainer standard functions ( http://docs.chainer.org/en/stable/reference/functions.html ).
I will implement it later on, however, i will appreciate it if you commit this repo!
Hi @shiba24 and @mhidy I uploaded a chainer implementation of ListNet with listwise cost based on top-1 probability. FYI: https://github.com/fullflu/learning-to-rank