-
|
I saved a model to torchscript from a publicly available pytorch model. Example here: https://aiqm.github.io/torchani-test-docs/examples/jit.html Here is what my code looks like: Path modelDir = Paths.get("./models");
Model model = Model.newInstance("ANI-1ccx");
model.load(modelDir, "compiled_model.pt"); |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 9 replies
-
|
Beta Was this translation helpful? Give feedback.
-
|
I was actually able to successfully load the model after switching to pytorch 1.12.1 when saving the model. However, I am still running into an issue when trying to run predictions: It looks like there is still an issue with a tuple being the model input. Is it possible for the torchscript model in DJL to accept a tuple as input? |
Beta Was this translation helpful? Give feedback.
-
|
@mhermon You just provide a name for each NDArray with |
Beta Was this translation helpful? Give feedback.
-
|
Thanks @frankfliu @siddvenk, I am now running into another error regarding pytorch expecting a tuple but getting a string similar to the initial error. My code: Path modelDir = Paths.get("./models");
Model model = Model.newInstance("ANI-1ccx");
try {
model.load(modelDir, "compiled_model.pt");
} catch (MalformedModelException e) {
e.printStackTrace();
} catch (IOException e) {
e.printStackTrace();
}
Translator<NDList, Float> translator = new Translator<NDList, Float>() {
@Override
public NDList processInput(TranslatorContext ctx, NDList input) {
return input;
}
@Override
public Float processOutput(TranslatorContext ctx, NDList list) {
NDArray array = list.singletonOrThrow();
return array.getFloat();
}
@Override
public Batchifier getBatchifier() {
return Batchifier.STACK;
}
};
Predictor<NDList, Float> predictor = model.newPredictor(translator);
try(NDManager manager = NDManager.newBaseManager()){
NDArray species = manager.randomInteger(1, 6, new Shape(1, 4), DataType.INT32);
species.setName("input1()");
NDArray coordinates = manager.randomUniform(-1, 1, new Shape(1, 4, 3));
coordinates.setName("input1()");
NDList input = new NDList(species, coordinates);
input.attach(manager);
Float energy = predictor.predict(input);
} catch (Exception e) {
e.printStackTrace();
}I saved the torchscript using: I am using djl version 0.18.0 |
Beta Was this translation helpful? Give feedback.
-
|
I want to reproduce your issue faithfully, so can you provide the following?
|
Beta Was this translation helpful? Give feedback.
-
|
@siddvenk Thanks for your help. I was actually able to resolve this issue. The problem was an incompatibility in the pytorch version I was using to save the torchscript and the pytorch version used by djl. I ended up having to use the snapshot version of djl, 0.19.0, to support pytorch version 1.12.1. |
Beta Was this translation helpful? Give feedback.
-
|
If you use python 1.12.1 to trace the model, you cannot use python 1.11.0 to load it. jit trace model is not forward compatible. |
Beta Was this translation helpful? Give feedback.
@mhermon
jit.trace()instead ofjit.script()?