0

I see example https://docs.djl.ai/jupyter/load_pytorch_model.html I made my own torchscript witch named qooModel.pt file and synset.txt

I tried to imitate the example. Here is my code.

val translator = ImageClassificationTranslator.builder()
    .setPipeline(Pipeline().apply {
        add(Resize(128))
        add(ToTensor())
    })
    .optSynsetArtifactName("synset.txt")
    .build()


val criteria = Criteria.builder()
    .setTypes(Image::class.java, Classifications::class.java)
    .optApplication(Application.CV.IMAGE_CLASSIFICATION)
    .optModelPath(Paths.get("models/qooModel.pt"))
    .optTranslator(translator)
    .build()

val imgFactory = ImageFactory.getInstance()
val imgPath    = Paths.get(args.getOrElse(0){"./images/qoo_learning.png"})
val image      = imgFactory.fromFile(imgPath)

val model = ModelZoo.loadModel(criteria)


val predict = model.newPredictor()

val result = predict.predict(image)
print(result)

model.close()

My code does not makes error, but forwarded value in result looks strange. I expect

[
    {"class": "qoo", "probability": 0.9597}
]

but

[
    {"class": "qoo", "probability": 0.00463}
]

printed.

I exepect 0.9597 because when I running under Python code in Google Colab

import torch
from PIL import Image
import torchvision.transforms as T

tensor_transform = T.Compose([T.PILToTensor()])

image = Image.open("qoo_learning.png").convert("RGB").resize((128, 128))
img_tensor = tensor_transform(image).type(torch.FloatTensor)
img_tensor = torch.unsqueeze(img_tensor, 0)
model(img_tensor)

the result of code show me tensor([[0.9597]], grad_fn=<SigmoidBackward0>)

Questions.

Q1. Why I cannot get correct result (likes 0.9597) in my Kotlin code?

Q2. What should I do for get correct result?

My guess

G1. there some issue between pytorch version between 2.0.1(I write torchscript) and 1.13.1(Kotlin use)

G2. I should use pytorch 2.0.1 in my kotlin code ....unfortunately I tried to use torch 2.0.1 but I couldn't understand document

Explicitly specify pytorch-native-xxx package version to override the version in the BOM.

and I could not find example about that

or

G1. My code wrong. It load model but it does not load pre-trained weights.

G2. I should fix my code

....example https://docs.djl.ai/jupyter/load_pytorch_model.html looks like using pre-trained model

but I think my code is almost same.

My Develop Environment.

I make my torchscript file(qooModel.pt) in Window 11, pytorch 2.0.1

I run my Kotlin code in macOS, I did not install torch by pip

... and it looks like running pytorch 1.13.1, look under debug log

[main] DEBUG ai.djl.pytorch.jni.LibUtils - Using cache dir: /Users/romeo/.djl.ai/pytorch/1.13.1-cpu-osx-aarch64
[main] DEBUG ai.djl.pytorch.jni.LibUtils - Loading native library: /Users/romeo/.djl.ai/pytorch/1.13.1-cpu-osx-aarch64/libc10.dylib
[main] DEBUG ai.djl.pytorch.jni.LibUtils - Loading native library: /Users/romeo/.djl.ai/pytorch/1.13.1-cpu-osx-aarch64/libtorch_cpu.dylib
[main] DEBUG ai.djl.pytorch.jni.LibUtils - Loading native library: /Users/romeo/.djl.ai/pytorch/1.13.1-cpu-osx-aarch64/libtorch.dylib
[main] DEBUG ai.djl.pytorch.jni.LibUtils - Loading native library: /Users/romeo/.djl.ai/pytorch/1.13.1-cpu-osx-aarch64/0.23.0-libdjl_torch.dylib

sorry about my poor english.

Romeo RKPK
  • 11
  • 1

1 Answers1

1

Summary
Image preprocessing incorrect

look at this kotlin code, it is part of code from question 
val translator = ImageClassificationTranslator.builder()
.setPipeline(Pipeline().apply {
    add(Resize(128))
    add(ToTensor())
})
.optSynsetArtifactName("synset.txt")
.build()

and python code from question

tensor_transform = T.Compose([T.PILToTensor()])

Unfortunately, that are does not same. Python code wich do preprocessing should be

preprocess = transforms.Compose([transforms.Resize(128), transforms.ToTensor()])

I fix this problem...

  1. fix my python code, about image preprocessing part
  2. train my model again
  3. export to torchscript
  4. load torchscript model by kotlin code
  5. It works well :)
Romeo RKPK
  • 11
  • 1