Allow the creation of ONNX initializers #965
Conversation
|
This PR does not reference an issue. Can you please file one and reference it in the PR description? |
|
|
||
| // Use implementation as in the actual conversion code | ||
| var ctx = ctxImpl as OnnxContext; | ||
| ctx.AddInitializer(9.4f, "float"); |
There was a problem hiding this comment.
ctx.AddInitializer(9.4f, "float"); [](start = 16, length = 34)
var floatScalar = ctx.AddInitializer(9.4f, "float");
There was a problem hiding this comment.
it's better than do
var floatScalar = model.Graph.Initializer[0];
If someone decide to change this test, and add one more initializer in the beginning, it will break tests.
In reply to: 220259227 [](ancestors = 220259227)
There was a problem hiding this comment.
ctx.AddInitializer returns the ONNX name of a tensor, not the actual ONNX object.
In addition, the test aims at examining the generated ONNX objects under the real scenario. Getting information before finishing the conversion is not ideal because it means we are checking an intermediate state of a conversion instead of the final state we care the most.
There was a problem hiding this comment.
Can you use that name of tensor to access initializer instead of model.Graph.Initalizer[0]?
If yes, i would suggest to use names instead of index.
If no, I find it strange, but i'm ok with current code.
In reply to: 220418126 [](ancestors = 220418126)
| @@ -349,5 +352,77 @@ public ModelArgs(string name, TensorProto.Types.DataType dataType, List<long> di | |||
|
|
|||
| return new ModelArgs(name, dataType, dimsLocal, dimsParamLocal); | |||
| } | |||
|
|
|||
| // Make long scalar in ONNX from native C# number | |||
| public static TensorProto MakeInt64(string name, long value) | |||
There was a problem hiding this comment.
(string name, long value) [](start = 43, length = 25)
Can you please add checks? See how above functions do checks. Same for below.
ghost
locked as resolved and limited conversation to collaborators
Related issue: #958.
This PR provides an ability of adding tensors into the
initializerlist of ONNX model. It may help the conversion of neural network based transforms and transforms which can be decomposed into matrix operators (some constant matrices would be saved as initializers).