(beta) Channels Last Memory Format in PyTorch

Author: Vitaly Fedyunin

What is Channels Last

Channels last memory format is an alternative way of ordering NCHW tensors in memory preserving dimensions ordering. Channels last tensors ordered in such a way that channels become the densest dimension (aka storing images pixel-per-pixel).

For example, classic (contiguous) storage of NCHW tensor (in our case it is two 4x4 images with 3 color channels) look like this:

Channels last memory format orders data differently:

Pytorch supports memory formats (and provides back compatibility with existing models including eager, JIT, and TorchScript) by utilizing existing strides structure. For example, 10x3x16x16 batch in Channels last format will have strides equal to (768, 1, 48, 3).

Channels last memory format is implemented for 4D NCHW Tensors only.

Memory Format API

Here is how to convert tensors between contiguous and channels last memory formats.

Classic PyTorch contiguous tensor

import torch

N, C, H, W = 10, 3, 32, 32
x = torch.empty(N, C, H, W)
print(x.stride())  # Outputs: (3072, 1024, 32, 1)

Conversion operator

x = x.to(memory_format=torch.channels_last)
print(x.shape)  # Outputs: (10, 3, 32, 32) as dimensions order preserved
print(x.stride())  # Outputs: (3072, 1, 96, 3)

Back to contiguous

x = x.to(memory_format=torch.contiguous_format)
print(x.stride())  # Outputs: (3072, 1024, 32, 1)

Alternative option

x = x.contiguous(memory_format=torch.channels_last)
print(x.stride())  # Outputs: (3072, 1, 96, 3)

Format checks

print(x.is_contiguous(memory_format=torch.channels_last))  # Outputs: True

There are minor difference between the two APIs to and contiguous. We suggest to stick with to when explicitly converting memory format of tensor.

For general cases the two APIs behave the same. However in special cases for a 4D tensor with size NCHW when either: C==1 or H==1 && W==1, only to would generate a proper stride to represent channels last memory format.

This is because in either of the two cases above, the memory format of a tensor is ambiguous, i.e. a contiguous tensor with size N1HW is both contiguous and channels last in memory storage. Therefore, they are already considered as is_contiguous for the given memory format and hence contiguous call becomes a no-op and would not update the stride. On the contrary, to would restride tensor with a meaningful stride on dimensions whose sizes are 1 in order to properly represent the intended memory format

special_x = torch.empty(4, 1, 4, 4)
print(special_x.is_contiguous(memory_format=torch.channels_last))  # Outputs: True
print(special_x.is_contiguous(memory_format=torch.contiguous_format))  # Outputs: True

Same thing applies to explicit permutation API permute. In special case where ambiguity could occur, permute does not guarantee to produce a stride that properly carry the intended memory format. We suggest to use to with explicit memory format to avoid unintended behavior.

And a side note that in the extreme case, where three non-batch dimensions are all equal to 1 (C==1 && H==1 && W==1), current implementation cannot mark a tensor as channels last memory format.

Create as channels last

x = torch.empty(N, C, H, W, memory_format=torch.channels_last)
print(x.stride())  # Outputs: (3072, 1, 96, 3)

clone preserves memory format

y = x.clone()
print(y.stride())  # Outputs: (3072, 1, 96, 3)

to, cuda, float … preserves memory format

if torch.cuda.is_available():
    y = x.cuda()
    print(y.stride())  # Outputs: (3072, 1, 96, 3)

empty_like, *_like operators preserves memory format

y = torch.empty_like(x)
print(y.stride())  # Outputs: (3072, 1, 96, 3)

Pointwise operators preserves memory format

z = x + y
print(z.stride())  # Outputs: (3072, 1, 96, 3)

Conv, Batchnorm modules using cudnn backends support channels last (only works for cuDNN >= 7.6). Convolution modules, unlike binary p-wise operator, have channels last as the dominating memory format. If all inputs are in contiguous memory format, the operator produces output in contiguous memory format. Otherwise, output will be in channels last memory format.

if torch.backends.cudnn.is_available() and torch.backends.cudnn.version() >= 7603:
    model = torch.nn.Conv2d(8, 4, 3).cuda().half()
    model = model.to(memory_format=torch.channels_last)  # Module parameters need to be channels last

    input = torch.randint(1, 10, (2, 8, 4, 4), dtype=torch.float32, requires_grad=True)
    input = input.to(device="cuda", memory_format=torch.channels_last, dtype=torch.float16)

    out = model(input)
    print(out.is_contiguous(memory_format=torch.channels_last))  # Outputs: True

When input tensor reaches a operator without channels last support, a permutation should automatically apply in the kernel to restore contiguous on input tensor. This introduces overhead and stops the channels last memory format propagation. Nevertheless, it guarantees correct output.

Performance Gains

Channels last memory format optimizations are available on both GPU and CPU. On GPU, the most significant performance gains are observed on NVIDIA’s hardware with Tensor Cores support running on reduced precision (torch.float16). We were able to archive over 22% performance gains with channels last comparing to contiguous format, both while utilizing ‘AMP (Automated Mixed Precision)’ training scripts. Our scripts uses AMP supplied by NVIDIA https://github.com/NVIDIA/apex.

python main_amp.py -a resnet50 --b 200 --workers 16 --opt-level O2  ./data

# opt_level = O2
# keep_batchnorm_fp32 = None <class 'NoneType'>
# loss_scale = None <class 'NoneType'>
# CUDNN VERSION: 7603
# => creating model 'resnet50'
# Selected optimization level O2:  FP16 training with FP32 batchnorm and FP32 master weights.
# Defaults for this optimization level are:
# enabled                : True
# opt_level              : O2
# cast_model_type        : torch.float16
# patch_torch_functions  : False
# keep_batchnorm_fp32    : True
# master_weights         : True
# loss_scale             : dynamic
# Processing user overrides (additional kwargs that are not None)...
# After processing overrides, optimization options are:
# enabled                : True
# opt_level              : O2
# cast_model_type        : torch.float16
# patch_torch_functions  : False
# keep_batchnorm_fp32    : True
# master_weights         : True
# loss_scale             : dynamic
# Epoch: [0][10/125] Time 0.866 (0.866) Speed 230.949 (230.949) Loss 0.6735125184 (0.6735) Prec@1 61.000 (61.000) Prec@5 100.000 (100.000)
# Epoch: [0][20/125] Time 0.259 (0.562) Speed 773.481 (355.693) Loss 0.6968704462 (0.6852) Prec@1 55.000 (58.000) Prec@5 100.000 (100.000)
# Epoch: [0][30/125] Time 0.258 (0.461) Speed 775.089 (433.965) Loss 0.7877287269 (0.7194) Prec@1 51.500 (55.833) Prec@5 100.000 (100.000)
# Epoch: [0][40/125] Time 0.259 (0.410) Speed 771.710 (487.281) Loss 0.8285319805 (0.7467) Prec@1 48.500 (54.000) Prec@5 100.000 (100.000)
# Epoch: [0][50/125] Time 0.260 (0.380) Speed 770.090 (525.908) Loss 0.7370464802 (0.7447) Prec@1 56.500 (54.500) Prec@5 100.000 (100.000)
# Epoch: [0][60/125] Time 0.258 (0.360) Speed 775.623 (555.728) Loss 0.7592862844 (0.7472) Prec@1 51.000 (53.917) Prec@5 100.000 (100.000)
# Epoch: [0][70/125] Time 0.258 (0.345) Speed 774.746 (579.115) Loss 1.9698858261 (0.9218) Prec@1 49.500 (53.286) Prec@5 100.000 (100.000)
# Epoch: [0][80/125] Time 0.260 (0.335) Speed 770.324 (597.659) Loss 2.2505953312 (1.0879) Prec@1 50.500 (52.938) Prec@5 100.000 (100.000)

Passing --channels-last true allows running a model in Channels last format with observed 22% performance gain.

python main_amp.py -a resnet50 --b 200 --workers 16 --opt-level O2 --channels-last true ./data

# opt_level = O2
# keep_batchnorm_fp32 = None <class 'NoneType'>
# loss_scale = None <class 'NoneType'>
#
# CUDNN VERSION: 7603
#
# => creating model 'resnet50'
# Selected optimization level O2:  FP16 training with FP32 batchnorm and FP32 master weights.
#
# Defaults for this optimization level are:
# enabled                : True
# opt_level              : O2
# cast_model_type        : torch.float16
# patch_torch_functions  : False
# keep_batchnorm_fp32    : True
# master_weights         : True
# loss_scale             : dynamic
# Processing user overrides (additional kwargs that are not None)...
# After processing overrides, optimization options are:
# enabled                : True
# opt_level              : O2
# cast_model_type        : torch.float16
# patch_torch_functions  : False
# keep_batchnorm_fp32    : True
# master_weights         : True
# loss_scale             : dynamic
#
# Epoch: [0][10/125] Time 0.767 (0.767) Speed 260.785 (260.785) Loss 0.7579724789 (0.7580) Prec@1 53.500 (53.500) Prec@5 100.000 (100.000)
# Epoch: [0][20/125] Time 0.198 (0.482) Speed 1012.135 (414.716) Loss 0.7007197738 (0.7293) Prec@1 49.000 (51.250) Prec@5 100.000 (100.000)
# Epoch: [0][30/125] Time 0.198 (0.387) Speed 1010.977 (516.198) Loss 0.7113101482 (0.7233) Prec@1 55.500 (52.667) Prec@5 100.000 (100.000)
# Epoch: [0][40/125] Time 0.197 (0.340) Speed 1013.023 (588.333) Loss 0.8943189979 (0.7661) Prec@1 54.000 (53.000) Prec@5 100.000 (100.000)
# Epoch: [0][50/125] Time 0.198 (0.312) Speed 1010.541 (641.977) Loss 1.7113249302 (0.9551) Prec@1 51.000 (52.600) Prec@5 100.000 (100.000)
# Epoch: [0][60/125] Time 0.198 (0.293) Speed 1011.163 (683.574) Loss 5.8537774086 (1.7716) Prec@1 50.500 (52.250) Prec@5 100.000 (100.000)
# Epoch: [0][70/125] Time 0.198 (0.279) Speed 1011.453 (716.767) Loss 5.7595844269 (2.3413) Prec@1 46.500 (51.429) Prec@5 100.000 (100.000)
# Epoch: [0][80/125] Time 0.198 (0.269) Speed 1011.827 (743.883) Loss 2.8196096420 (2.4011) Prec@1 47.500 (50.938) Prec@5 100.000 (100.000)

The following list of models has the full support of Channels last and showing 8%-35% performance gains on Volta devices: alexnet, mnasnet0_5, mnasnet0_75, mnasnet1_0, mnasnet1_3, mobilenet_v2, resnet101, resnet152, resnet18, resnet34, resnet50, resnext50_32x4d, shufflenet_v2_x0_5, shufflenet_v2_x1_0, shufflenet_v2_x1_5, shufflenet_v2_x2_0, squeezenet1_0, squeezenet1_1, vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn, wide_resnet101_2, wide_resnet50_2

The following list of models has the full support of Channels last and showing 26%-76% performance gains on Intel(R) Xeon(R) Ice Lake (or newer) CPUs: alexnet, densenet121, densenet161, densenet169, googlenet, inception_v3, mnasnet0_5, mnasnet1_0, resnet101, resnet152, resnet18, resnet34, resnet50, resnext101_32x8d, resnext50_32x4d, shufflenet_v2_x0_5, shufflenet_v2_x1_0, squeezenet1_0, squeezenet1_1, vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn, wide_resnet101_2, wide_resnet50_2

Converting existing models

Channels last support is not limited by existing models, as any model can be converted to channels last and propagate format through the graph as soon as input (or certain weight) is formatted correctly.

# Need to be done once, after model initialization (or load)
model = model.to(memory_format=torch.channels_last)  # Replace with your model

# Need to be done for every input
input = input.to(memory_format=torch.channels_last)  # Replace with your input
output = model(input)

However, not all operators fully converted to support channels last (usually returning contiguous output instead). In the example posted above, layers that does not support channels last will stop the memory format propagation. In spite of that, as we have converted the model to channels last format, that means each convolution layer, which has its 4 dimensional weight in channels last memory format, will restore channels last memory format and benefit from faster kernels.

But operators that does not support channels last does introduce overhead by permutation. Optionally, you can investigate and identify operators in your model that does not support channels last, if you want to improve the performance of converted model.

That means you need to verify the list of used operators against supported operators list https://github.com/pytorch/pytorch/wiki/Operators-with-Channels-Last-support, or introduce memory format checks into eager execution mode and run your model.

After running the code below, operators will raise an exception if the output of the operator doesn’t match the memory format of the input.

def contains_cl(args):
    for t in args:
        if isinstance(t, torch.Tensor):
            if t.is_contiguous(memory_format=torch.channels_last) and not t.is_contiguous():
                return True
        elif isinstance(t, list) or isinstance(t, tuple):
            if contains_cl(list(t)):
                return True
    return False


def print_inputs(args, indent=""):
    for t in args:
        if isinstance(t, torch.Tensor):
            print(indent, t.stride(), t.shape, t.device, t.dtype)
        elif isinstance(t, list) or isinstance(t, tuple):
            print(indent, type(t))
            print_inputs(list(t), indent=indent + "    ")
        else:
            print(indent, t)


def check_wrapper(fn):
    name = fn.__name__

    def check_cl(*args, **kwargs):
        was_cl = contains_cl(args)
        try:
            result = fn(*args, **kwargs)
        except Exception as e:
            print("`{}` inputs are:".format(name))
            print_inputs(args)
            print("-------------------")
            raise e
        failed = False
        if was_cl:
            if isinstance(result, torch.Tensor):
                if result.dim() == 4 and not result.is_contiguous(memory_format=torch.channels_last):
                    print(
                        "`{}` got channels_last input, but output is not channels_last:".format(name),
                        result.shape,
                        result.stride(),
                        result.device,
                        result.dtype,
                    )
                    failed = True
        if failed and True:
            print("`{}` inputs are:".format(name))
            print_inputs(args)
            raise Exception("Operator `{}` lost channels_last property".format(name))
        return result

    return check_cl


old_attrs = dict()


def attribute(m):
    old_attrs[m] = dict()
    for i in dir(m):
        e = getattr(m, i)
        exclude_functions = ["is_cuda", "has_names", "numel", "stride", "Tensor", "is_contiguous", "__class__"]
        if i not in exclude_functions and not i.startswith("_") and "__call__" in dir(e):
            try:
                old_attrs[m][i] = e
                setattr(m, i, check_wrapper(e))
            except Exception as e:
                print(i)
                print(e)


attribute(torch.Tensor)
attribute(torch.nn.functional)
attribute(torch)

If you found an operator that doesn’t support channels last tensors and you want to contribute, feel free to use following developers guide https://github.com/pytorch/pytorch/wiki/Writing-memory-format-aware-operators.

Code below is to recover the attributes of torch.

for (m, attrs) in old_attrs.items():
    for (k, v) in attrs.items():
        setattr(m, k, v)

Work to do

There are still many things to do, such as:

• Resolving ambiguity of N1HW and NC11 Tensors;

• Testing of Distributed Training support;

• Improving operators coverage.

If you have feedback and/or suggestions for improvement, please let us know by creating an issue.