Summary: 76 instances, 65 unique

Text	Count
// TODO: more precise error message should be provided by backend	2
# TODO: Currently Julia does not support closure in c-callbacks, so we are making use of the	1
@inplace W .+= -η .* ∇ ./ sqrt(s .+ ϵ)  # FIXME: sqrt should be dot-call	1
// TODO: Currently we do not add place holder for NDArray	1
self.index = None # TODO: what is index?	3
// TODO: Put Symbol.api.foo --> Stable APIs	1
# TODO: is there better way to compare two symbols	1
// TODO: shutdown hook won't work on Windows	1
* TODO: Optimize for GPU; this is a baseline implementation providing	1
// TODO: remove kernel dependency on warpSize=32	1
% TODO convert from c order to matlab order...	1
// TODO: change the data/label type into IndexedSeq[(NDArray, DataDesc)]	1
# TODO the explicit exclusion of take will no longer be necessary when it is removed from Base	1
//     -use 0 for testing (TODO: remove this)	1
# TODO Backward using NDArray will cause some troubles see `https://github.com/dmlc/mxnet/issues/1720'	1
# TODO Here we can in fact play multiple gaming instances simultaneously and make actions for each	1
* TODO: write a faster kernel optimized for GPU	1
# TODO: better way to shape the anchors??	1
# TODO Profiling the speed of this part!	1
#TODO Update logging patterns in other files	1
// TODO: remove kernel dependency on warpSize=32	3
// TODO: dynamically get from mxnet env to support other serializers like Kyro	2
# TODO: support storage type (stype in Python)	1
# TODO: bug in nnvm, if do not call this, call get handle "_copyto" will fail	1
# TODO `wait_to_read()` here seems unnecessary, remove it in the future!	1
s.uₜ = _maximum(β2 * s.uₜ, abs(∇))  # FIXME abs dot-call	1
# TODO: Use logging later	1
super(LiveTimeSeries, self).__init__(None, None)  # TODO: clean up this class hierarchy	1
# TODO These operators still need to be converted (listing in order of priority):	1
// TODO: shall we dispose the replaced array here?	1
## TODO write logging	1
#TODO add SCALED_TANH, SOFTPLUS, SOFTSIGN, SIGMOID_HARD, LEAKYRELU, PRELU, ELU, PARAMETRICSOFTPLUS, THRESHOLDEDRELU, LINEAR	1
// TODO: more nature way to get the # of examples?	1
// TODO should we average over number of GPUs	1
#TODO Possibly states + inds for less memory access	1
# TODO: sparse support?	1
self.index = None  # TODO: what is index?	1
// TODO: make state a ClassTag	1
* TODO: shutdown hook won't work on Windows	1
#  TODO: User-defined differentiable function	1
TODO: remove `data_padding` and `label_padding`, and implement rollover that copies	1
#TODO Optimize the reshaping functionality!	1
// TODO: need to allow user to specify DType and Layout	2
// TODO: Implement CustomOp for NDArray	1
#TODO: pull this into default	1
// TODO: naive implementation	1
# TODO: make sure it works with gpu array	1
for l, n in len_dict.items(): # TODO: There are better heuristic ways to do this	3
@inplace W .+= -η .* ∇ ./ sqrt(x .+ ϵ)  # FIXME: sqrt dot-call	1
// TODO is it possible to do in place???	1
// TODO: imdecode	1
// TODO: check ip & port available	1
Δxₜ = ∇ .* sqrt(Δx .+ ϵ) ./ sqrt(x .+ ϵ)  # FIXME: sqrt dot-call	1
// TODO: Consider implementing a vector kernel for SpMV (similar to DotCsrDnsDns)	1
#TODO: pull this into default	1
// TODO: Seq() here allows user to place Symbols rather than normal arguments to run, need to fix if old API deprecated	1
// TODO: remove sequential search, this is a bottleneck	1
# TODO: find a better solution in case this cause issues in the future.	1
// TODO: Make DType, Layout configurable	1
// TODO: make returned state a ClassTag	1
// TODO: Add '_linalg_', '_sparse_', '_image_' support	2
# TODO Test the copy function	1
// TODO: Add Filter to the same location in case of refactor	2
// TODO: make DataIter implement Iterator	1
// TODO: need to allow user to specify DType and Layout	1