With ONNX export, does it mean it will avoid the problem when not using jit.script in torch.trace?

JonathanSum · October 23, 2020, 2:15pm

In the torch.trace tutorial, it says if we don’t do jit.script in some if-else logic statement in the model, the exported trace model won’t capture it. However, in ONNX, does it mean we don’t need to do jit.script in those logic statements in the model?

Thank you for answering.

ebarsoum · October 23, 2020, 11:11pm

ONNX export leverage TorchIR to convert PyTorch model to ONNX graph. TorchIR is generated by trace or script, so you will still have the same limitation if you use trace to export to ONNX.

What the problem of using script to capture control flow and loop?

JonathanSum · October 25, 2020, 11:33pm

github.com/microsoft/onnxjs

Error loading onnxjs from webpack

opened 05:52AM - 16 Jun 19 UTC

csaroff

I've been trying to add onnxjs support in [VoTT](https://github.com/microsoft/Vo…TT/issues/794), but I've been hitting some issues. I have a model that I was able to successfully run in node.js, but when I try to run it on the frontend, I get ```javascript TypeError: cannot resolve operator 'Shape' with opsets: ai.onnx v9 ``` I tried [converting the model](https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md#converting-version-of-an-onnx-model-within-default-domain-aionnx) to use a lower opset, but I faced similar errors with the converted models. So I thought that I'd be clever and try a workaround and run the model on the backend using onnxjs-node, sending the results back to the frontend using IPC. After adding a little code into src/electron/main.ts ```javascript import { Tensor, InferenceSession } from "onnxjs-node"; const inferenceSession = new InferenceSession(); ``` and updating my webpack config, I'm seeing a couple of new errors. ``` σ csaroff:~/fiddle/VoTT$ node_modules/.bin/webpack-cli --config config/webpack.dev.js ... WARNING in ./node_modules/onnxjs-node/bin/napi-v3 sync ^\.\/.*\.node$ Module not found: Error: Can't resolve 'node-loader' in '/Users/csaroff/fiddle/VoTT' @ ./node_modules/onnxjs-node/bin/napi-v3 sync ^\.\/.*\.node$ @ ./node_modules/onnxjs-node/lib/binding.js @ ./node_modules/onnxjs-node/lib/inference-session-override.js @ ./node_modules/onnxjs-node/lib/index.js @ ./src/electron/main.ts ERROR in ./node_modules/onnxjs/lib/wasm-binding.js Module not found: Error: Can't resolve 'worker-loader' in '/Users/csaroff/fiddle/VoTT/node_modules/onnxjs/lib' @ ./node_modules/onnxjs/lib/wasm-binding.js 101:29-96 @ ./node_modules/onnxjs/lib/backends/backend-wasm.js @ ./node_modules/onnxjs/lib/api/onnx-impl.js @ ./node_modules/onnxjs/lib/api/index.js @ ./node_modules/onnxjs-node/lib/index.js @ ./src/electron/main.ts ``` So then I ran `npm install --save-dev node-loader worker-loader` and tried rerunning webpack. ``` σ csaroff:~/fiddle/VoTT$node_modules/.bin/webpack-cli --config config/webpack.dev.js ... ERROR in ./node_modules/onnxjs/lib/wasm-binding.js Module not found: Error: Can't resolve './worker/worker-main' in '/Users/csaroff/fiddle/VoTT/node_modules/onnxjs/lib' @ ./node_modules/onnxjs/lib/wasm-binding.js 101:29-96 @ ./node_modules/onnxjs/lib/backends/backend-wasm.js @ ./node_modules/onnxjs/lib/api/onnx-impl.js @ ./node_modules/onnxjs/lib/api/index.js @ ./node_modules/onnxjs-node/lib/index.js @ ./src/electron/main.ts ``` And this is where I'm stuck. I see that you chose not to include lib/worker in your [tsconfig file](https://github.com/microsoft/onnxjs/blob/master/tsconfig.json#L3), but I couldn't figure out why. So I tried rebuilding the project, removing the ` "exclude": ["lib/worker"],` line from the tsconfig.json, but that just gives me a whole mess of typescript errors. ``` σ csaroff:~/fiddle/onnxjs$npm run build > onnxjs@0.1.7 build /Users/csaroff/fiddle/onnxjs > tsc && node tools/build --build-wasm --build-bundle node_modules/typescript/lib/lib.dom.d.ts:25:1 - error TS6200: Definitions of the following identifiers conflict with those in another file: EventListenerOrEventListenerObject, ImportExportKind, TableKind, BlobPart, HeadersInit, BodyInit, RequestInfo, DOMHighResTimeStamp, CanvasImageSource, OffscreenRenderingContext, MessageEventSource, ImageBitmapSource, TimerHandler, PerformanceEntryList, VibratePattern, AlgorithmIdentifier, HashAlgorithmIdentifier, BigInteger, NamedCurve, GLenum, GLboolean, GLbitfield, GLint, GLsizei, GLintptr, GLsizeiptr, GLuint, GLfloat, GLclampf, TexImageSource, Float32List, Int32List, BufferSource, DOMTimeStamp, FormDataEntryValue, IDBValidKey, Transferable, BinaryType, CanvasDirection, CanvasFillRule, CanvasLineCap, CanvasLineJoin, CanvasTextAlign, CanvasTextBaseline, ClientTypes, EndingType, IDBCursorDirection, IDBRequestReadyState, IDBTransactionMode, ImageSmoothingQuality, KeyFormat, KeyType, KeyUsage, NotificationDirection, NotificationPermission, OffscreenRenderingContextId, PermissionName, PermissionState, PushEncryptionKeyName, PushPermissionState, ReferrerPolicy, RequestCache, RequestCredentials, RequestDestination, RequestMode, RequestRedirect, ResponseType, ServiceWorkerState, ServiceWorkerUpdateViaCache, VisibilityState, WebGLPowerPreference, WorkerType, XMLHttpRequestResponseType 25 interface Account { ~~~~~~~~~ node_modules/typescript/lib/lib.webworker.d.ts:25:1 25 interface AddEventListenerOptions extends EventListenerOptions { ~~~~~~~~~ Conflicts are in this file. ... ``` I'm pretty much out of ideas. I checked in a minimally reproducible example on [this branch](https://github.com/csaroff/VoTT/tree/onnxjs-backend-example). Here's a [diff of the changes](https://github.com/csaroff/VoTT/commit/a90769d2009c154ff47c412d63a305b7e5943c83). Any help/insight here would be great!

I guess it is kind of like underdevelopment.