Error when using decoder from NVIDIA DALI

Deep Learning
1

Hello,

@pipeline_def
def dali_pipeline(image_paths, mask_paths):
    images_jpg = fn.readers.webdataset(paths=image_paths, shard_id=0, num_shards=1, dtypes=types.UINT8, random_shuffle=True)
    masks_jpg = fn.readers.webdataset(paths=mask_paths, shard_id=0, num_shards=1, dtypes=types.UINT8, random_shuffle=True)
    
    images = fn.decoders.image(images_jpg, device='mixed', output_type=types.RGB)
    masks = fn.decoders.image(masks_jpg, device='mixed', output_type=types.RGB)
    
    images = fn.resize(images, device='gpu', resize_x=256, resize_y=140, interp_type=dali.types.INTERP_NN)
    masks = fn.resize(masks, device='gpu', resize_x=256, resize_y=140, interp_type=dali.types.INTERP_NN)
    
    images = fn.normalize(images, dtype=types.FLOAT)
    
    return images, masks

pipe = dali_pipeline(image_paths=image_paths, mask_paths=mask_paths, batch_size=64, num_threads=2, device_id=0)
pipe.build()
dali_dl = DALIGenericIterator(pipe, ['image', 'mask'])

I have this code here and I get the following error that I couldn’t solve even if I used another type of reader (instead of readers.file) or adding dtype=types.UINT8:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[16], line 18
     16 pipe = dali_pipeline(image_paths=image_paths, mask_paths=mask_paths, batch_size=64, num_threads=2, device_id=0)
     17 pipe.build()
---> 18 dali_dl = DALIGenericIterator(pipe, ['image', 'mask'])

File /opt/conda/lib/python3.10/site-packages/nvidia/dali/plugin/pytorch.py:194, in DALIGenericIterator.__init__(self, pipelines, output_map, size, reader_name, auto_reset, fill_last_batch, dynamic_shape, last_batch_padded, last_batch_policy, prepare_first_batch)
    192 if self._prepare_first_batch:
    193     try:
--> 194         self._first_batch = DALIGenericIterator.__next__(self)
    195         # call to `next` sets _ever_consumed to True but if we are just calling it from
    196         # here we should set if to False again
    197         self._ever_consumed = False

File /opt/conda/lib/python3.10/site-packages/nvidia/dali/plugin/pytorch.py:211, in DALIGenericIterator.__next__(self)
    208     return batch
    210 # Gather outputs
--> 211 outputs = self._get_outputs()
    213 data_batches = [None for i in range(self._num_gpus)]
    214 for i in range(self._num_gpus):

File /opt/conda/lib/python3.10/site-packages/nvidia/dali/plugin/base_iterator.py:298, in _DaliBaseIterator._get_outputs(self)
    296     for p in self._pipes:
    297         with p._check_api_type_scope(types.PipelineAPIType.ITERATOR):
--> 298             outputs.append(p.share_outputs())
    299 except StopIteration as e:
    300     # in case ExternalSource returns StopIteration
    301     if self._size < 0 and self._auto_reset == "yes":

File /opt/conda/lib/python3.10/site-packages/nvidia/dali/pipeline.py:1003, in Pipeline.share_outputs(self)
   1001 self._batches_to_consume -= 1
   1002 self._gpu_batches_to_consume -= 1
-> 1003 return self._pipe.ShareOutputs()

RuntimeError: Critical error in pipeline:
Error when executing CPU operator decoders__Image, instance name: "__Image_53", encountered:
Error in thread 0: [/opt/dali/dali/operators/decoder/host/host_decoder.cc:31] Assert on "IsType<uint8>(input.type())" failed: Input must be stored as uint8 data.
Stacktrace (7 entries):
[frame 0]: /opt/conda/lib/python3.10/site-packages/nvidia/dali/libdali_operators.so(+0x686e3e) [0x7b2235038e3e]
[frame 1]: /opt/conda/lib/python3.10/site-packages/nvidia/dali/libdali_operators.so(+0xa87601) [0x7b2235439601]
[frame 2]: /opt/conda/lib/python3.10/site-packages/nvidia/dali/libdali_operators.so(+0x6880e0) [0x7b223503a0e0]
[frame 3]: /opt/conda/lib/python3.10/site-packages/nvidia/dali/libdali.so(dali::ThreadPool::ThreadMain(int, int, bool, std::string const&)+0x1e6) [0x7b227e9e9e86]
[frame 4]: /opt/conda/lib/python3.10/site-packages/nvidia/dali/libdali.so(+0x769dd0) [0x7b227ef99dd0]
[frame 5]: /lib/x86_64-linux-gnu/libc.so.6(+0x94b43) [0x7b237cde6b43]
[frame 6]: /lib/x86_64-linux-gnu/libc.so.6(clone+0x44) [0x7b237ce77bb4]

Current pipeline object is no longer valid.

My dataset consists of a tar file that holds jpg images that are RGB.
Does anyone know how to solve this?

2

Solved the problem. It was as simple as this :man_facepalming:

@pipeline_def(batch_size=64, num_threads=2, device_id=0)
def dali_pipeline():
    images_encoded, _ = fn.readers.file(file_root="", files=image_paths, shard_id=0, num_shards=1)
    masks_encoded, _ = fn.readers.file(file_root="", files=mask_paths, shard_id=0, num_shards=1)
    
    images_decoded = fn.decoders.image(images_encoded, device='mixed', output_type=types.RGB)
    masks_decoded = fn.decoders.image(masks_encoded, device='mixed', output_type=types.RGB)
    
    images = fn.resize(images_decoded, device='gpu', resize_x=256, resize_y=140, interp_type=types.INTERP_NN)
    masks = fn.resize(masks_decoded, device='gpu', resize_x=256, resize_y=140, interp_type=types.INTERP_NN)
    
    images_normalized = fn.normalize(images)
    
    return images_normalized, masks

clean_mem()
pipe = dali_pipeline()
pipe.build()
dali_dl = DALIGenericIterator(pipe, ['image', 'mask'])
3

The error message you are getting is telling you that the input data to the decoders__Image operator is not of type uint8. This is because the webdataset reader, by default, reads the data as uint8. However, you are specifying the output type of the decoders__Image operator to be types.RGB, which is not a valid type for the webdataset reader.

To solve this, you need to either change the output type of the decoders__Image operator to uint8 or change the way you are reading the data. For example, you could use the readers.file reader instead of the webdataset reader.

Here is an example of how you could change the code to use the readers.file reader:

def dali_pipeline(image_paths, mask_paths):
    images_file = fn.readers.file(paths=image_paths, shard_id=0, num_shards=1, dtypes=types.UINT8)
    masks_file = fn.readers.file(paths=mask_paths, shard_id=0, num_shards=1, dtypes=types.UINT8)

    images = fn.decoders.image(images_file, device='mixed', output_type=types.RGB)
    masks = fn.decoders.image(masks_file, device='mixed', output_type=types.RGB)

    images = fn.resize(images, device='gpu', resize_x=256, resize_y=140, interp_type=dali.types.INTERP_NN)
    masks = fn.resize(masks, device='gpu', resize_x=256, resize_y=140, interp_type=dali.types.INTERP_NN)

    images = fn.normalize(images, dtype=types.FLOAT)

    return images, masks

This code will read the data from the files as uint8 and then decode it as RGB .

4

I’m afraid your solution won’t work either. Look at my reply and see how I solved it.

The fn.readers.file returns 2 outputs and since I don’t have labels (numbers or other stuff like that) to return, only the images, then I have to ignore the second output. If I didn’t do that then I would get another error, something with INT if I remember well.

Regardless, thank you for your reply.