Totally different results when using the same model on PC and android mobilephone?

[btdan]

Hello, I have trained a mobilenetv2 model with the help of pytorch(version: 1.9.0) and convert to .pt format， deploy the .pt model on android mobilephone. However, I get very different result from the result on PC. Can anyone help me to solve this problem? The details are as follows: First，train the new model(named result.ckpt) from mobilenetv2(the same problem with mobilenetv3 ) with the help of pytorch. Secondly, convert the model result.ckpt to result.pt the key source codes is as follows:

import torch
import torchvision
from torch.utils.mobile_optimizer import optimize_for_mobile
import torchvision.models as models

model= torch.load('result.ckpt', map_location=torch.device('cpu'))
print(model)
model.eval()
example = torch.rand(1, 3, 224, 224)
traced_script_module = torch.jit.trace(model, example)
traced_script_module.save("result.pt")

Thirdly, deploy the result.pt model inside app source code. And then compile to get the apk the key source codes are as follows:

module = LiteModuleLoader.load(assetFilePath(this, "largeMbv3.pt"));
bitmap = decodeFile(imgPath);
newbitmap = compressImageFromBitmap(bitmap,224,224);
final Tensor inputTensor = TensorImageUtils.bitmapToFloat32Tensor(newbitmap,
                TensorImageUtils.TORCHVISION_NORM_MEAN_RGB, TensorImageUtils.TORCHVISION_NORM_STD_RGB, MemoryFormat.CHANNELS_LAST);
      
        // running the model
        final Tensor outputTensor = module.forward(IValue.from(inputTensor)).toTensor();
        final float[] scores = outputTensor.getDataAsFloatArray();

And the function compressImageFromBitmap is as follows:

    public static Bitmap compressImageFromBitmap(Bitmap bitmap, int pixelW, int pixelH) {
        ByteArrayOutputStream os = new ByteArrayOutputStream();
        bitmap.compress(Bitmap.CompressFormat.JPEG, 100, os);
        if( os.toByteArray().length / 1024>512) {
            os.reset();
            bitmap.compress(Bitmap.CompressFormat.JPEG, 50, os);
        }
        ByteArrayInputStream is = new ByteArrayInputStream(os.toByteArray());
        BitmapFactory.Options options = new BitmapFactory.Options();
        options.inJustDecodeBounds = true;
        options.inPreferredConfig = Bitmap.Config.RGB_565;
        BitmapFactory.decodeStream(is, null, options);
        options.inJustDecodeBounds = false;
        options.inSampleSize = computeSampleSize(options , pixelH > pixelW ? pixelW : pixelH ,pixelW * pixelH );
        is = new ByteArrayInputStream(os.toByteArray());
        Bitmap newBitmap = BitmapFactory.decodeStream(is, null, options);
        return newBitmap;
    }

However, I get quiet different result for same image to identify on PC and android Can anyone help me to solve this problem? Thans very much.

By the way, I have set random to a fix number when training the model.Just as follows:

seed=42
random.seed(seed)
numpy.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)  # all gpus
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True
torch.use_deterministic_algorithms(True)
torch.set_default_tensor_type(torch.float32)

But it does not have any change at all. Thans very much for your help . Best wishes to you.