How to provide mean & stddev to dali.fn.normalize

[Vincouux]

Hi all, I am super new to Dali Backend, but the benchmark I recently ran are incredible. I am trying to preprocess the data with Dali backend, but struggle with the "normalize" operation. I would like to specify a "per channel" mean & stddev ([c1, c2, c3] for example), but keep getting errors.

Here is what I've tried so far:

...
mean=[0.485, 0.456, 0.406],
stddev=[0.229, 0.224, 0.225]
...

...
mean=(0.485, 0.456, 0.406),
stddev=(0.229, 0.224, 0.225)
...

...
mean=dali.types.Constant([0.485, 0.456, 0.406]),
stddev=dali.types.Constant([0.229, 0.224, 0.225])
...

Here is the full code of my pipeline:

import nvidia.dali as dali
from nvidia.dali.plugin.triton import autoserialize


@autoserialize
@dali.pipeline_def(batch_size=8, num_threads=8, device_id=0)
def pipe():

    # Getting the images.
    images = dali.fn.external_source(device="gpu", name="input_0")

    # Resizing.
    images = dali.fn.resize(images, resize_x=224, resize_y=224)

    # Permutting NHWC to NCHW.
    images = dali.fn.transpose(images, perm=[2, 0, 1])

    # Rescaling 0-255 to 0-1.
    images = images / 255.

    # Normalizing (images - mean) / std.
    images = dali.fn.normalize(
        images,
        batch=False,
        mean=dali.types.Constant([0.485, 0.456, 0.406]),
        stddev=dali.types.Constant([0.229, 0.224, 0.225])
    )

    return images

Many thanks !!

[szalpal]

Hello @Vincouux ,

Thank you for reaching out.

First thing I see is that you don't need to wrap the mean and stddev in the types.Constant:

    # Normalizing (images - mean) / std.
    images = dali.fn.normalize(
        images,
        batch=False,
        mean=[0.485, 0.456, 0.406],
        stddev=[0.229, 0.224, 0.225]
    )

If this doesn't help, could you please post the error message, so I'd know more what's going on?

[Vincouux]

Here is the error I am facing:

Traceback (most recent call last):
  File "<string>", line 5, in <module>
  File "<frozen importlib._bootstrap>", line 553, in module_from_spec
AttributeError: 'NoneType' object has no attribute 'loader'
Traceback (most recent call last):
  File "<string>", line 8, in <module>
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/_utils/autoserialize.py", line 72, in invoke_autoserialize
    dali_pipeline().serialize(filename=filename)
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/pipeline.py", line 1471, in create_pipeline
    pipe_outputs = func(*args, **fn_kwargs)
  File "/models/detection_exp114_20220523_preprocessing/1/dali.py", line 21, in pipe
    images = dali.fn.normalize(
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/fn.py", line 88, in fn_wrapper
    return op_wrapper(*inputs, **kwargs)
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/fn.py", line 80, in op_wrapper
    return op_class(**init_args)(*inputs, **call_args)
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/ops.py", line 621, in __init__
    _add_spec_args(self._schema, self._spec, kwargs)
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/ops.py", line 398, in _add_spec_args
    converted_value = _type_convert_value(dtype, value)
  File "/opt/tritonserver/backends/dali/conda/envs/dalienv/lib/python3.8/site-packages/nvidia/dali/types.py", line 110, in _type_convert_value
    return _known_types[dtype][1](val)
TypeError: float() argument must be a string or a number, not 'list'

To give a bit more context, the input image is of shape [-1, -1, 3]. After the resize and the permutation, it should be [3, 224, 224] based on the above pipeline I provided.

Many thanks @szalpal for you help !!!

[szalpal]

@Vincouux

Generally, for image normalization it'd be better if you'd go for fn.crop_mirror_normalize instead of fn.normalize. For two reasons, one is that the former fuses 3 possible operations in one and the second is that the API would be more user friendly.

Therefore I'd advice you to use this instead of fn.normalize:

    images = dali.fn.crop_mirror_normalize(images,
                                           dtype=types.FLOAT,
                                           output_layout="CHW",
                                           crop=(224, 224),
                                           mean=[0.485, 0.456, 0.406],
                                           std=[0.229, 0.224, 0.225])

Should you have any more problems, don't hesitate to reach us out. Good luck in using DALI ;)

---

Furthermore, if you'd like to squeeze even more perf from this snippet, I'd suggest, that you plug the Rescaling 0-255 to 0-1. into the normalization by multiplying your normalization parameters by 255:

                                           mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
                                           std=[0.229 * 255, 0.224 * 255, 0.225 * 255])