proper way to construct a tensor on the fly?

[vincentme]

My goal is to pad the image to the power of two. The following code works. But the way to construct pads pads = op.ConcatFromSequence(op.SequenceConstruct(zero, zero, padded_row, padded_col), axis = 0, new_axis=1) seems a bit cumbersome. Is this the proper way to construct a tensor on the fly? I tried the literal way # pads = [zero, zero, padded_row, padded_col], it does not work.

Thanks!

@script()
def pad_image(img_in):
    shape = op.Shape(img_in)
    next_row = find_next_power_of_two(shape[0])
    next_col = find_next_power_of_two(shape[1])

    zero = op.Constant(value=helper.make_tensor("int_zero", TensorProto.INT64, (), [0]))
    padded_row = next_row - shape[0]
    padded_col = next_col - shape[1]
    pads = op.ConcatFromSequence(op.SequenceConstruct(zero, zero, padded_row, padded_col), axis = 0, new_axis=1)
    # pads = [zero, zero, padded_row, padded_col]
    print(pads)

    img_padded = op.Pad(img_in, pads = pads, mode = 'reflect')
    return img_padded

[justinchuby]

Does op.Concat(zero, zero, padded_row, padded_col) work for you?

[vincentme]

Does op.Concat(zero, zero, padded_row, padded_col) work for you?

Thanks! Because these tensors are indeed scalars, so I have to unsqueeze them to 1D then concat, the following works. Not sure if this is a better way...

pads = op.Concat(op.Unsqueeze(zero, axes=[0]), op.Unsqueeze(zero, axes=[0]), op.Unsqueeze(pad_row, axes=[0]), op.Unsqueeze(pad_col, axes=[0]), axis = 0)

[gramalingam]

A naive question: how is this being run? I am a bit puzzled because this mixed-mode construction (interleaving script-time computation with onnx-construction) is doable in trace-mode, but not in script-mode.

Otherwise: nothing much to add to the existing discussion. If the shape is statically known, there may be other ways of doing this. But I assume you want to handle dynamic shapes here? So, Concat, as above, seems reasonable.

[vincentme]

A naive question: how is this being run? I am a bit puzzled because this mixed-mode construction (interleaving script-time computation with onnx-construction) is doable in trace-mode, but not in script-mode.

Otherwise: nothing much to add to the existing discussion. If the shape is statically known, there may be other ways of doing this. But I assume you want to handle dynamic shapes here? So, Concat, as above, seems reasonable.

Sorry, i am not quite understand your comments... The script or trace modes are the ones in pytorch export?

If so, my context is different. The goal is not convert from pytorch, but directly write algorithm(not necessary deep learning model) in onnxscript. As shown inthis example, one step is to pad the image to the power of two. The shape of the input image is dynamic, but the shape of the pads is of course fixed as (4,).