What does negative dimension imply

Ask a Question

Question

I was toggling the node parameters for testing and I noticed that the ONNX checker doesn’t complain on the following model whose output has negative value for the last two dimensions. I am under the impression that the model is invalid and I wonder if negative value has special meaning.

Further information

Model gen script

from onnx import helper
from onnx import TensorProto as tp
from onnx import checker
from onnx import save
import sys
import argparse
import os

node = helper.make_node('MaxPool',
                        inputs=['input'],
                        outputs=['output'],
                        kernel_shape=[5, 5],
                        strides=[2, 2],
                        dilations=[2, 2])
graph = helper.make_graph(
    [node], 'test_graph',
    [helper.make_tensor_value_info('input', tp.FLOAT, [1, 1, 5, 5])],
    [helper.make_tensor_value_info('output', tp.FLOAT, [1, 1, -1, -1])])
model = helper.make_model(graph, producer_name='model')
checker.check_model(model, full_check=True)

Notes

Any additional information, code snippets.

[gramalingam]

Yes, in an ONNX ShapeProto, a negative value is not valid. Technically, it makes sense to have the checker catch and report an error for this case. However, some other frameworks use the special value -1 to indicate an unknown dimension. Hence, there may be some older models that incorrectly use -1 to indicate an unknown dimension. So, throwing an error could cause some compatibility issues for such models. I think it might make sense for us to have the notion of a “warning” and flag such issues with a “warning” instead of “error” to balance these requirements.

[askhade]

agree - another such violation is nodes are not sorted in topological order. I encountered 1 case where the model was valid but the nodes were not in the right order (mostly because some offline optimization script was run) … ORT allows this too