Onnx in Javascript about dtln HOT 6 CLOSED

Sadly I found this microsoft/onnxjs#167

from dtln.

Ahh sigh !!! Tfjs it is then :)

from dtln.

I read into the documentation and it should support lstms. I converted the model with hopfully the right opset. Try this model:
models.zip

Probably the keras2onnx converter uses by default a newer opset for converting the models as used by onnx.js. This model is converted with opset 7.

from dtln.

Tried the test model. No luck same error.
I debugged the onnx model graph loading code. The issue seems to be an input to LSTM. The object it throws the exception on seems to be this:

attribute: (3) [t, t, t]
domain: ""
input: (8) ["lstm_4_X", "lstm_4_W", "lstm_4_R", "lstm_4_B", "", "lstm_4_initial_h", "lstm_4_initial_c", ""]
name: "lstm_4"
opType: "LSTM"
output: (3) ["lstm_4_Y", "lstm_4_Y_h", "lstm_4_Y_c"]
proto: Object

Notice empty "" in 5th and last position. This is what its complaining about.

By the way in looking through the docs for Onnx Javascript LSTM op is not supported in JavaScript ?
https://github.com/microsoft/onnxjs/blob/master/docs/operators.md

In the LSTM column of the doc I don't see it mentioned for any opset.

from dtln.

Oh yeah, it looks like I skipped that when reading on my phone. Hopefully they integrate the support for lstms soon.

from dtln.

Ahh sigh !!! Tfjs it is then :)

did you get this working with JS? Any pointers?

from dtln.