How to create an ndArray using the byte[] obtained by ndArray.data().asBytes()?

ND4J
1

Why is it that for the byte array dataBuffer created by Nd4j.createTypedBuffer(bytes, DataType.FLOAT) , when using Nd4j.create(dataBuffer, shape); to create an ndarray , the values of the ndarray are float numbers that are the same as the integer values in the bytes , instead of converting the byte array into floats?

2

@cqiaoYc I would recommend staying away from that in general. At least in M2.1. Before I rewrote it in the newer code (no the rewrite’s not done yet..) it was really ambiguous as to what the byte array is. I would manually convert it like I did later:
@Override
public void put(byte element) {
if (released.get())
throw new IllegalStateException(“You can’t use DataBuffer once it was released”);

    switch (dataType()) {
        case BOOL:
            ((BooleanIndexer) indexer).put(0, ArrayUtil.toBooleanArray(element));
            break;
        case BYTE:
            ((ByteIndexer) indexer).put(0,element);
            break;
        case UBYTE:
            ((UByteIndexer) indexer).put(0,ArrayUtil.toIntArraySimple(element));
            break;
        case UINT16:
            ((UShortIndexer) indexer).put(0,ArrayUtil.toIntArraySimple(element));
            break;
        case SHORT:
            ((ShortIndexer) indexer).put(0,ArrayUtil.toShorts(element));
            break;
        case UINT32:
            ((UIntIndexer) indexer).put(0,ArrayUtil.toLongArray(element));
            break;
        case INT:
            ((IntIndexer) indexer).put(0,ArrayUtil.toIntArraySimple(element));
            break;
        case UINT64:
            ((ULongIndexer) indexer).put(0,ArrayUtil.toBigInteger(element));
            break;
        case LONG:
            ((LongIndexer) indexer).put(0,ArrayUtil.toLongArray(element));
            break;
        case BFLOAT16:
            ((Bfloat16Indexer) indexer).put(0,ArrayUtil.toFloatArraySimple(element));
            break;
        case HALF:
            ((HalfIndexer) indexer).put(0,ArrayUtil.toFloatArraySimple(element));
            break;
        case FLOAT:
            ((FloatIndexer) indexer).put(0,ArrayUtil.toFloatArraySimple(element));
            break;
        case DOUBLE:
            ((DoubleIndexer) indexer).put(0,ArrayUtil.toDoubleArraySimple(element));
            break;
        default:
            throw new UnsupportedOperationException("Unsupported data type: " + dataType());
    }
}

It’s just an extra method call and should be enough for your use case.

3

@agibsonccc Maybe I didn’t make my requirement clear. First, save the bytes obtained from ndArray.data().asBytes() into a binary data file. Then, load the bytes from the binary data file into memory. Finally, create an ndArray based on the bytes .

4

@cqiaoYc yes and I’m saying either create a wrapper method using what I specified. I’m also mentioning to you that in M2.1 the asBytes() needed a bit of rewriting due to what “bytes” can mean. The way I did it in the latest master should work for you.