stu1130
commented
3 years ago Here is the sample translator
// I don't use input, but you should use the input to generate 4 tensors you mentioned above
public static final class MyTranslator implements Translator<NDList, NDList> {
@Override
public NDList processInput(TranslatorContext ctx, NDList input) {
NDManager manager = ctx.getNDManager();
NDArray input1 = manager.ones(new Shape(32, 96, 6));
NDArray input2 = manager.ones(new Shape(32, 96, 6));
NDArray input3 = manager.ones(new Shape(32, 72, 6));
NDArray input4 = manager.ones(new Shape(32, 72, 6));
return new NDList(input1, input2, input3, input4);
}
@Override
public NDList processOutput(TranslatorContext ctx, NDList list) {
// here you will get 32 tensors in list
// list.size() == 32
// list.get(0).getShape() == new Shape()
// take stock price prediction for instance, you can return a float representing the stock price got from 32 tensors instead of the original output list
return list;
}
@Override
public Batchifier getBatchifier() {
return null;
}
}Here is the pytorch prediction jupyter notebook. Let me know if you need more details
MaleicAcid
frankfliu
Hello, I want to use djl for time series forecasting. I have trained a pytorch model(Transformer) and converted it to torchscript format I have consulted the official documents and examples, but I still don’t know how to write Translator and use DJl to call the pt model for time series forecasting.
The input of my model is four tensors, and the shape is as follows: torch.Size([32, 96, 6]) torch.Size([32, 96, 6]) torch.Size([32, 72, 6]) torch.Size([32, 72, 6])
The model will predict 24 steps backwards, the shape of output which is as follows: torch.Size([24, 6]) torch.Size([24, 6]) ...... torch.Size([24, 6]) # A total of 32
Does DJl currently support calling such a time series model for forecasting? How should I write Translator and make predictions? I hope to get some help, thank you very much!
你好，我想用djl来进行时间序列的预测，我已经训练了一个pytorch模型(Transformer模型)，并将其转换为torchscript格式 我查阅了官方文档和示例，但我仍然不知道如何编写Translator以及使用DJl调用pt模型进行时间序列预测
我的模型输入为四个tensor, 其形状如下: torch.Size([32, 96, 6]) torch.Size([32, 96, 6]) torch.Size([32, 72, 6]) torch.Size([32, 72, 6])
模型会向后预测24步，输出32个tensor, 其形状如下： torch.Size([24, 6]) torch.Size([24, 6]) ...... torch.Size([24, 6]) # 共32个
DJl目前支持调用这样的时间序列模型进行预测嘛？我应该如何编写Translator以及进行预测？恳请您能为我提供一些帮助，万分感谢！