Slice op incorrectly parses 'ends' attribute

[kernhanda]

I was trying to run onnx2c on a small model of mine (coco_3M.zip), but I ran into the following error:

$ ./onnx2c ~/code/inferred_person_coco_3M.onnx
2022-03-05 15-18-39.725 [Fatal] (resolveOutput) Concat's input tensors must have the same shape, except for the dimension size of the axis to concatenate on.
onnx2c: ~/code/onnx2c/src/nodes/concat.h:103: virtual void toC::Concat::resolveOutput(const std::vector<const toC::Tensor*>&, std::vector<toC::Tensor*>&): Assertion `false' failed.
[1]    1633 abort      ./onnx2c ~/code/inferred_person_coco_3M.onnx

The same model passes ONNX's model checker, renders fine in netron.app, and executes correctly through ONNXRuntime. I traced the differences to the Slice nodes in this model, which has an ends attribute of 9223372036854775807, aka 2**63 - 1, aka std::numeric_limits<int64_t>::max(). Protobuf for some reason parses this number as -1, even though the value is being stored in an int64_t. Not sure why that is, but I thought you could take a look.

This issue can be seen starting at node Slice_33. onnx2c calculates the output as 1x16x14x16 because the ends is read as [-1] instead of [2**63 - 1].

The following python script can be used to update models for a workaround:

import onnx

MODEL_IN_FILENAME = 'foo.onnx'
MODEL_OUT_FILENAME = 'foo_adjusted.onnx'

m = onnx.load_model(MODEL_IN_FILENAME)

def handle_slice_node(slice_node):                     
    for a in slice_node.attribute:                     
        if a.name == 'ends':                           
            for i, v in enumerate(a.ints):             
                if v == (2**63 - 1):                   
                    print("Found a value to adjust")   
                    a.ints[i] = 2**31 - 1              

for n in m.graph.node:        
    if n.op_type == 'Slice':  
        handle_slice_node(n)  

onnx.save_model(m, MODEL_OUT_FILENAME)

[kraiskil]

Thanks for reporting. With such a through report, it was easy to spot this - this is a clear cut bug in onnx2c. The branch mentioned above fixes this. I want to see if I can make a single-node regression test for this before merging into master.

Slightly related: in order to be able to replicate this issue, I had to fix another bug in the Concat node (patch in the branch above) that caused onnx2c to error out before hitting this. Do you have local changes to onnx2c?

[kernhanda]

Wow, keen eye! I totally missed the obvious bug fix :)

Good point about the other changes in concat. Here they are

diff --git a/src/nodes/concat.h b/src/nodes/concat.h
index d158bf1..eea1ad6 100644
--- a/src/nodes/concat.h
+++ b/src/nodes/concat.h
@@ -85,7 +85,7 @@ namespace toC {

            void resolveOutput(const std::vector<const Tensor *> &inputs, std::vector<Tensor *> &outputs) override {
                    node_inputs = inputs;
-                 if (inputs.size() < 2)
+                 if (inputs.size() < 1)
                            ERROR("Wrong number of inputs to Concat");

                    if (axis < 0)
@@ -98,7 +98,7 @@ namespace toC {
                    size_t i, j;
                    std::vector<int> dims = inputs[0]->data_dim;
                    for (i = 0; i < input_count; i++) {
-                         for (j = 1; j < dims.size(); j++) {
+                         for (j = 0; j < dims.size(); j++) {
                                    if (dims[j] != inputs[i]->data_dim[j] && (int) j != axis)
                                            ERROR("Concat's input tensors must have the same shape, except for the "
                                                      "dimension size of the axis to concatenate on.");

Rationate:

• Min number of inputs to concat is actually 1
• If axis == 0, it didn't look like the dims were being checked, but I didn't check if there was another code path for the verification

[kraiskil]

Thanks for the diff. I didn't find an easy way to add a minimized test for this, so I merged this branch issue_15 to master without a test. Looking at the onnx2c test suite there already was a slice_INT64_MAX test, but that was for ONNX opset 10 or later, where the start/end/axis had moved from attributes to input tensors, so this problem did not repeat... I could not find with quick searching an api call to set the ONNX opset version with the script using scailable onnx which I use to make the unit tests, so I let this one go without a regression test.

About the changes to concat:

• the first point is the bug I had to fix too
• the second: that looks like a bug too - I don't see why dimension 0 should not be checked at this point. It looks like a case of "the programmer stops working when the code starts working"... (this was a hobby project, and that shows at times :) ). I'll commit that fix, thanks.