[pytorch] where is the "CPUFloatType" class for tensor datatype to match ? and the tensor convert to Buffer is null

[mullerhai]

HI , In scala ,I want to match the tensor datatype to operate something, when I debug ,I find the Tensor type is [CPUFloatType ], But I can not find it ? how to security compare the tensor datatype with declare ,is instance of something? and If the tensor is FloatType ,after I invoke the tensor.data_ptr_byte() or tensor.data_ptr_double() or tensor.data_ptr_int(), the process will crash !!! ,how to security invoke it??

import org.bytedeco.pytorch.global.torch

val tensor = torch.randn(34,34)  //as you know  the tensor  is CPUFloatType 

if(tensor.data_ptr().isInstanceOf[BytePointer] ||tensor.data_ptr().equals(tensor.data_ptr_byte()) ){
      println(" data_ptr_byte") // here will crash first 
    }else if (tensor.data_ptr.isInstanceOf[IntPointer]||tensor.data_ptr().equals(tensor.data_ptr_int())){
      println(" data_ptr_int ") // here will crash second 
    }else if (tensor.data_ptr.isInstanceOf[FloatPointer]||tensor.data_ptr().equals(tensor.data_ptr_float())){
      println(" data_ptr_float ") // can not come here !!
    }else if (tensor.data_ptr.isInstanceOf[DoublePointer]||tensor.data_ptr().equals(tensor.data_ptr_double())){
      println(" data_ptr_double ")
    }else{
      println(" data_ptr_bool")
    }

finally  ,I want to know how to get  or convert to java  data buffer from  javacpp pytorch tensor?
I try to this ,but crash ,it is null

val tensor = torch.randn(34,34) val arr = tensor.data_ptr_float().asInstanceOf[FloatPointer].asBuffer().array() //will crash array is null

val flat = tensor.data_ptr_float() val buffer = DataBuffers.of(flat.asBuffer()) // buffer is null

[mullerhai]

I also found "class org.bytedeco.javacpp.Pointer cannot be cast to class org.bytedeco.javacpp.FloatPointer " why?

val floatTensor= tensor.data_ptr.asInstanceOf[FloatPointer]

how to convert it

[mullerhai]

only tensor.toType(torch.ScalarType.Byte) , but I need to compare the datatype will BytePointer or CPUByteType

[saudet]

Have you tried Tensor.createBuffer() and found something missing? If so, please explain more clearly what you need to do exactly that you are not able to do with that method: http://bytedeco.org/javacpp-presets/pytorch/apidocs/org/bytedeco/pytorch/AbstractTensor.html#createBuffer--

[mullerhai]

http://bytedeco.org/javacpp-presets/pytorch/apidocs/org/bytedeco/pytorch/AbstractTensor.html#createBuffer-- when I use Tensor.createBuffer() ,also crash, If you success use Tensor.createBuffer() ,please give me Tensor.createBuffer() a code example to do, thanks

var tensor2 = torch.randn(34,34)
val buffer2 = tensor2.createBuffer[FloatBuffer]()
val array = buffer2.array() // here will crash 

[saudet]

Those buffers don't have an array, so don't call array() on them.

[mullerhai]

Those buffers don't have an array, so don't call array() on them.

how to convert the buffer2 to java arrayList or Array?

[mullerhai]

Those buffers don't have an array, so don't call array() on them.

how to see and get the real float data num from those buffers

[saudet]

Everything you need to know about Buffer is documented in the JDK: https://docs.oracle.com/javase/8/docs/api/java/nio/Buffer.html

[mullerhai]

I found the tensor.createBuffer is DirectBuffer ,is out jvm buffer. But I need to read the buffer inner data to array, I can not get the suitable method .

[saudet]

We can copy the data with one of the get() methods