DataNode to integer

[Sumered]

I'm trying to implement somehow more complicated pipeline using Dali and Triton, I want for it to look like this:

• image goes to preprocessing Dali pipeline, which results in preprocessed image
• preprocessed image goes to neural net hosted on Triton, this neural net produces output in such format [number_of_bboxes,x,y,width,height,score,label,x,y,width,height...] lets call it predictions. In config.pbtxt it has size of 6001,1,1 (1000 is max number of detections)
• this output and also preprocessed image should both go to another Dali pipe which should prepare batch of cropped bounding boxes from this image (using output from neural network), and those crops in one batch should go to another neural network
• then whole pipeline should return both predictions of first neural network on whole image and all predictions on batched crops from second neural network.

Is it possible to do this in such way? Right now I'm stuck at trying to get number of bboxes from first neural network predictions so I can iterate over those bboxes. It seems that it is impossible to get DataNode as int, and use it in for loop. Code for the second Dali element looks like this:

@dali.pipeline_def(batch_size=1, num_threads=64, device_id=0)
def pipe():
    detections = dali.fn.external_source(
        device="gpu", name="DETECTIONS", batch=False)
    image = dali.fn.external_source(
        device="gpu", name="preprocess_output", batch=False)
    batch = []
    number = dali.fn.cast(detections[0], dtype=dali.types.INT64)
    for i in range(0, number, 6):
        x = detections[i+1]
        y = detections[i+2]
        width = detections[i+3]
        height = detections[i+4]
        batch.append(dali.fn.resize(
            image[y:y+height, x:x+width], resize_x=640,
            resize_y=640, dtype=dali.types.FLOAT))
    batch_result = dali.fn.expand_dims(batch[0], axes=[0])
    for i in range(1, len(batch)):
        dali.fn.cat(batch_result, batch[i], axis=0)
    return batch_result

pipe().serialize(filename="/home/somewhere/model.dali")

Error that I'm getting is:

Traceback (most recent call last):
  File "1_test.py", line 26, in <module>
    pipe().serialize(filename="/home/somewhere/model.dali")
  File "/home/user/.local/lib/python3.8/site-packages/nvidia/dali/pipeline.py", line 1346, in create_pipeline
    pipe_outputs = func(*args, **fn_kwargs)
  File "1_test.py", line 12, in pipe
    for i in range(0, number, 6):
TypeError: 'DataNode' object cannot be interpreted as an integer

[JanuszL]

Hi @Sumered,

Batch in DALI is implicit and you cannot iterate over samples inside the pipeline definition and simply create a batch appending samples. DALI doesn't support changing the batch size inside the pipeline, however, long term this could be a case for you:

• accept N x samples at the input
• crop each of it M times
• output N x M items

Now what I could advise you is to try out python TRITON backend.