❓ [Question] How to set three input tensor shape in input_shape?

[wxyhv]

❓ Question

I have three input tensor:src_tokens, dummy_embeded_x, dummy_encoder_embedding
In this case, I don't konw how to set input_shape in compile_settings's "input_shape"
Who can help me? Thank you!

encoder_out = model.forward_encoder([src_tokens, dummy_embeded_x, dummy_encoder_embedding])
...
script_encoder = torch.jit.script(encoder)...
compile_settings = { "input_shapes": [[2, 16]], "op_precision": torch.float32 }

[narendasan]

You set the input shapes as a list of lists or a list of dicts if you are using dynamic shape.

So for your 3 inputs you would do something like this:

"input_shapes": [[2,16], [1,10], [1, 15]]

or something similar where the entries correspond to src_tokens, dummy_embeded_x, dummy_encoder_embedding in the order they will appear in the function signature for your forward i.e. in this case maybe something like forward(src_tokens, dummy_embeded_x, dummy_encoder_embedding).

To do dynamic shape it would look like this:

"input_shapes": [
  {
    "min": [2, 16],
    "opt": [2, 20],
    "max: [2, 25],
   }, {
    "min": [1, 10],
    "opt": [1, 20],
    "max: [1, 25],
   }, {
    "min": [1, 20],
    "opt": [1, 25],
    "max: [1, 35],
   }
]