Source data too small. Allocating larger array instrument.ts:82

[fumiya0]

I created a model for image classification using pytorch and transformed it to onnx.
Trying a random value input [1,3,224,224] to it on background 'webgl', it seems the output is almost same, regardless of the value of the input.
Console said "Souce data too small. Allocating larger array instrument.ts:82". I'm not sure how and where to fix. I think there is something wrong with the model.
My model is at https://github.com/fumiya0/patimg-classification-test/TCGA3.onnx
and you can click "predict" at https://fumiya0.github.io//patimg-classification-test/ to see what happens when the model deals with an input. Sorry for a messy code!
When I try the model res50_8.onnx, which I downloaded at onnx demo repogitory, there is no warning message and it seems to work well.
Thanks!

[SubhiH]

I am getting the same warning message. The output of my onnx model is not the same of the original model but it is not always the same. Were you able to solve this issue?
@fumiya0

[fs-eire]

I tried to figure out the cause of that warning message. It looks like the webgl backend implementation in Conv operator is not working correctly when strides is not set to [1, 1]. This may need a fix for the operator. @NTT123 @hariharans29

[SubhiH]

@fs-eire
I am using webgl as a backend. My original model is built using MXnet framework then I converted it to onnx and called it from onnxjs. In the original model, I set all the stride values for conv layers to either (1,1) or (2,2).
When I convert the back-end to 'cpu', I get an error (I think I should open an issue for that):

Error: cwise: Arrays do not all have the same shape!

Thanks!

[SubhiH]

I would like to mention that I have built the latest version of the code and I can run the model with "wasm" back-end.
When I switch to "webgl" , I get "Source data too small. Allocating larger array".
For "wasm" back-end and after running the model tens of time, an OutOfMemory exception shows.

[Kirayue]

Hello, everyone
I am getting the same warning message. According to @fs-eire, it's due to strides of Conv operator is not set to [1, 1]. But I test my two own models, both have non-[1, 1] strides, only one model causes this warning.
My question is does this affect the inference time? I want to compare the inference time to tf.js, so I want to know whether the time measured is precise with this warning.

[photopea]

I also have this problem. I have to run it on the CPU version, as WASM version does not work either.

[local-dev-31]

Same problem here, can't use the GPU and WASM doesn't support most of the operators, the inference time drop from 250fps, with onnx on python, to 8fps with the CPU backend in the browser.
Hoping for someone to fix this issue, meanwhile thanks for the great work.

[qube13]

Anybody found a solution for that problem?

[photopea]

Guys, I think Microsoft programmers stopped working on this project, because Microsoft stopped to invest money into it. So I doubt anybody will ever fix it. The sooner we switch to some alternative, the better :/

[Whiax]

Guys, I think Microsoft programmers stopped working on this project, because Microsoft stopped to invest money into it. So I doubt anybody will ever fix it. The sooner we switch to some alternative, the better :/

Probably. Sadly we don't need much but if they can't even get simple convolutions to work we should probably try tensorflowjs. I removed all strides and used maxpooling instead and it still doesn't work. I wonder how other people managed to have it working.