Add NameEntityRecognition and Q&A deep learning tasks. #6760
Conversation
ghost
assigned 
|
Q&A currently has a runtime error I am working on resolving so the builds will fail for now. Getting the PR up so reviews can start while I finish debugging. |
michaelgsharp
requested review from
ericstj,
tannergooding,
LittleLittleCloud,
JakeRadMSFT and
luisquintanilla
|
|
||
| namespace Microsoft.ML.TorchSharp.NasBert.Models | ||
| { | ||
| internal sealed class ModelForPrediction : NasBertModel |
There was a problem hiding this comment.
This isn't for NER. Its for SentenceSimilarity and TextClassification. How about TextModel? TextModelForPrediction? Thoughts?
|
|
||
| for (var i = 0; i < srcTokens.size(0); ++i) | ||
| { | ||
| var srcTokenArray = srcTokens[i].ToArray<int>(); |
There was a problem hiding this comment.
I wonder if we could use the TensorAccessor exposed by the data method and avoid this array?
There was a problem hiding this comment.
Looking into it it doesn't look like the TensorAccessor exposes enough to be able to do that. @NiklasGustafsson do you know if that is correct?
There was a problem hiding this comment.
@michaelgsharp , I'm not sure which 'that' you're referring to. TensorAccessor implements IEnumerable.
There was a problem hiding this comment.
I know the TensorAccessor has direct access to the underlying memory, but it doesn't expose the underlying memory directly, correct?
There was a problem hiding this comment.
Indexing into the accessor will access the underlying native memory directly, both reading and writing, while ToArray() will make a copy.
public T this[params long[] indices] {
get {
long index = 0;
if (indices.Length == 1) {
index = indices[0];
validate(index);
unsafe {
T* ptr = (T*)_tensor_data_ptr;
return ptr[TranslateIndex(index, _tensor)];
}
} else {
unsafe {
T* ptr = (T*)_tensor_data_ptr;
return ptr[TranslateIndex(indices, _tensor)];
}
}
}
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## main #6760 +/- ##
==========================================
+ Coverage 68.89% 68.99% +0.10%
==========================================
Files 1216 1237 +21
Lines 250915 252836 +1921
Branches 26259 26445 +186
==========================================
+ Hits 172857 174450 +1593
- Misses 71238 71454 +216
- Partials 6820 6932 +112
Flags with carried forward coverage won't be shown. Click here to find out more.
|
There was a problem hiding this comment.
Thanks for resolving the feedback around GetSubArray - I'd still like for @LittleLittleCloud, @zewditu, or @JakeRadMSFT to give a pass
This PR adds in 2 new deep learning scenarios, Name Entity Recognition and Q&A.
The main files to focus on are NerTrainer.cs and Roberta/QATrainer.cs. Most of the rest are either part of the deep learning model itself or internal implementations of things I had to copy over from runtime for them to work on netstandard.