DALI
DALI copied to clipboard
Speed up Dino with DALI
Describe the question.
Hello,
I'm trying to speed up the image augmentations in DINO with DALI. I'm trying to start with a naive solution which is just calling the augmentations from DALI. I'll optimize it once I get this running, but I can't get this to build. I'm getting this error:
TypeError: Illegal pipeline output type. The output 0 contains a nested `DataNode`. Missing list[/tuple](http://localhost:8888/tuple) expansion (*) is the likely cause.
Can you please help me find the issue with this code?
from torchvision import transforms
import nvidia.dali.plugin.pytorch as dalitorch
from nvidia.dali import pipeline_def
import nvidia.dali.fn as fn
import nvidia.dali.types as types
import matplotlib.gridspec as gridspec
import matplotlib.pyplot as plt
from nvidia.dali.backend import TensorListGPU
import numpy as np
from PIL import Image
flip_and_color_jitter = transforms.Compose([
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomApply(
[transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)],
p=0.8
),
transforms.RandomGrayscale(p=0.2),
])
normalize = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
def global_transform1(images):
global_crops_scale=1.0
func = transforms.Compose([
transforms.RandomResizedCrop(224, scale=global_crops_scale, interpolation=Image.BICUBIC),
flip_and_color_jitter,
utils.GaussianBlur(1.0),
normalize,
])
return func(images)
def global_transform2(images):
global_crops_scale=1.0
func = transforms.Compose([
transforms.RandomResizedCrop(224, scale=global_crops_scale, interpolation=Image.BICUBIC),
flip_and_color_jitter,
utils.GaussianBlur(0.1),
utils.Solarization(0.2),
normalize,
])
return func(images)
def local_transform(images):
local_crops_scale=0.5
func = transforms.Compose([
transforms.RandomResizedCrop(96, scale=local_crops_scale, interpolation=Image.BICUBIC),
flip_and_color_jitter,
utils.GaussianBlur(p=0.5),
normalize,
])
return func(images)
@pipeline_def(seed=seed, enable_conditionals=True)
def dino_dali_pipeline(image_dir, local_crops_number=8, device="mixed"):
jpegs, labels = fn.readers.file(file_root=image_dir)
decoded_jpegs = fn.decoders.image(jpegs, device=device, output_type=types.RGB)
crops = []
crops.append(dalitorch.fn.torch_python_function(decoded_jpegs, function=global_transform1, num_outputs=1))
crops.append(dalitorch.fn.torch_python_function(decoded_jpegs, function=global_transform2, num_outputs=1))
for _ in range(local_crops_number):
crops.append(dalitorch.fn.torch_python_function(decoded_jpegs, function=local_transform, num_outputs=1))
return crops, labels
Check for duplicates
- [X] I have searched the open bugs/issues and have found no duplicates for this bug report
Hi @jpfeil,
Thank you for reaching out.
In your code crops
is a list while DALI expects outputs to be plain types. So in your case, you need to either unpack the list:
return *crops, labels
or concatenate all into one tensor using cat
or stack
operators (as long as the tensors have uniform shapes samplewise).
Thank you, @JanuszL!
I've implemented the pipeline with your suggestions, but now I'm having issues with the iterator.
class Solarize:
def __init__(self, threshold: int = 128) -> None:
self._threshold = threshold
def __call__(self, img):
inverted_img = 255 - img
mask = img >= self._threshold
return mask * inverted_img + (True ^ mask) * img
solarize = Solarize()
@pipeline_def(seed=seed, batch_size=1, enable_conditionals=True)
def dino_dali_pipeline(image_dir, local_crops_number=8, device="mixed"):
jpegs, _ = fn.readers.file(file_root=image_dir, random_shuffle=True)
decoded_jpegs = fn.decoders.image(jpegs, device=device)
cropped_jpegs = fn.crop(decoded_jpegs, crop=(16384, 16384))
#
# Global Transform 1
#
gt1 = fn.random_resized_crop(cropped_jpegs, size=224, random_aspect_ratio=(1.0, 1.0))
## Random Horizontal Flip
coin = fn.random.coin_flip()
if coin:
gt1 = fn.flip(gt1, horizontal=1)
## Color Jitter
coin = fn.random.coin_flip(probability=0.8)
if coin:
gt1 = fn.color_twist(gt1, brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)
## Random Grayscale
coin = fn.random.coin_flip(probability=0.2)
if coin:
gt1 = fn.color_space_conversion(gt1, image_type=types.RGB, output_type=types.GRAY)
## Gaussian Blur
gt1 = fn.gaussian_blur(gt1, sigma=(0.1, 2.0))
## Normalize
gt1 = fn.normalize(gt1)
#
# Global Transform 2
#
gt2 = fn.random_resized_crop(cropped_jpegs, size=224, random_aspect_ratio=(1.0, 1.0))
## Random Horizontal Flip
coin = fn.random.coin_flip()
if coin:
gt2 = fn.flip(gt2, horizontal=1)
## Color Jitter
coin = fn.random.coin_flip(probability=0.8)
if coin:
gt2 = fn.color_twist(gt2, brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)
## Random Grayscale
coin = fn.random.coin_flip(probability=0.2)
if coin:
gt2 = fn.color_space_conversion(gt2, image_type=types.RGB, output_type=types.GRAY)
## Gaussian Blur
coin = fn.random.coin_flip(probability=0.1)
if coin:
gt2 = fn.gaussian_blur(gt2, sigma=(0.1, 2.0))
## Solarize
coin = fn.random.coin_flip(probability=0.1)
if coin:
gt2 = fn.cast(solarize(gt2), dtype=types.UINT8)
gt2 = fn.normalize(gt2)
#
# Local Transformations
#
crops = [gt1, gt2]
for _ in range(local_crops_number):
lt = fn.random_resized_crop(cropped_jpegs, size=96, random_aspect_ratio=(0.5, 0.5))
## Random Horizontal Flip
coin = fn.random.coin_flip()
if coin:
lt = fn.flip(lt, horizontal=1)
## Color Jitter
coin = fn.random.coin_flip(probability=0.8)
if coin:
lt = fn.color_twist(lt, brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)
## Random Grayscale
coin = fn.random.coin_flip(probability=0.2)
if coin:
gt1 = fn.color_space_conversion(lt, image_type=types.RGB, output_type=types.GRAY)
## Gaussian Blur
coin = fn.random.coin_flip(probability=0.5)
if coin:
lt = fn.gaussian_blur(lt, sigma=(0.1, 2.0))
## Normalize
lt = fn.normalize(lt)
return *crops,
This works and I can get augmented images out of it. The only issue is that I'm used to the pytorch representation using floats whereas in DALI it usually represents the image as uint8, but hopefully that doesn't influence training that much.
The problem I have now is passing the pipeline to the iterator:
pipe = dino_dali_pipeline(image_dir, batch_size=4, num_threads=4, device_id=0)
pipe.build()
iterator = DALIGenericIterator(
pipelines=pipe,
output_map=["gt1", "gt2", "lt1", "lt2", "lt3", "lt4", "lt5", "lt6", "lt7", "lt8"],
)
for i, (batch,) in enumerate(iterator):
print(batch)
break
This will run for a little while and then it throws this error:
RuntimeError: [[/opt/dali/dali/pipeline/data/tensor_list.cc:1012](http://localhost:8888/opt/dali/dali/pipeline/data/tensor_list.cc#line=1011)] Assert on "IsDenseTensor()" failed: The batch must be representable as a tensor - it must have uniform shape and be allocated in contiguous memory.
Stacktrace (88 entries):
This seems to be related to the global and local crops being different sizes. Is there a way to support this kind of data in DALI?
Thanks!
Hi @jpfeil,
The only issue is that I'm used to the pytorch representation using floats whereas in DALI it usually represents the image as uint8
You can use the crop_mirror_normalize
operator and pass float as the output type, and 255
as the std
to scale from uin8 to 0-1 float.
This seems to be related to the global and local crops being different sizes. Is there a way to support this kind of data in DALI?
Yes, the iterator expects that the batch of samples can be represented as the tensor where one of the dimensions is the batch size. In this case, you can either pad samples to have them uniform or try out PyTorch DALIRaggedIterator.