SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 1908.09791 · 2019

Once-for-All: Train One Network and Specialize it for Efficient Deployment

arXiv · PDF · Open in the Atlas

Code that ran

24 of 34 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 3 honoured the contract we drafted, 1 violated it and 8 have no contract verdict; 12 failures of our instrument; 10 no recorded run), measured 2026-11-21; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/1908.09791.json · badge: /api/ran/1908.09791.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
mit-han-lab/once-for-all canonical (classifier label)
classifier basis: census
3 ran · 0 no recorded run
analogdevicesinc/ai8x-training community (classifier label)
classifier basis not recorded on the link
7 ran · 2 no recorded run
MaximIntegratedAI/ai8x-synthesis community (classifier label)
classifier basis not recorded on the link
6 ran · 5 no recorded run
jordan-hs/rss-interference-cvprw2022 community (classifier label)
classifier basis not recorded on the link
3 ran · 0 no recorded run
pprp/ofa-cifar community (classifier label)
classifier basis not recorded on the link
2 ran · 1 no recorded run
twice154/ofa-for-super-resolution extension (classifier label)
classifier basis: classifier_2025
2 ran · 0 no recorded run
seulkiyeom/once-for-all community (classifier label)
classifier basis not recorded on the link
1 ran · 2 no recorded run
FunctionStatusWhere it lives
bn_fuser Ran
ran; our contract draft was wrong, not the code
analogdevicesinc/ai8x-training/batchnormfuser.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("37154c0fb52a4320")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

build_activation Ran
ran; our contract draft was wrong, not the code
mit-han-lab/once-for-all/ofa/utils/pytorch_modules.py
code served (permissive licence) · get_code("da7205ae31eaf8ba")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ReLU. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

torch.manual_seed(0)

class Hswish(nn.Module):

    def __init__(self, inplace=True):
        super(Hswish, self).__init__()
        self.inplace = inplace

    def forward(self, x):
        return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0

    def __repr__(self):
        return 'Hswish()'

class Hsigmoid(nn.Module):

    def __init__(self, inplace=True):
        super(Hsigmoid, self).__init__()
        self.inplace = inplace

    def forward(self, x):
        return F.relu6(x + 3.0, inplace=self.inplace) / 6.0

    def __repr__(self):
        return 'Hsigmoid()'

def build_activation(act_func, inplace=True):
    if act_func == 'relu':
        return nn.ReLU(inplace=inplace)
    elif act_func == 'relu6':
        return nn.ReLU6(inplace=inplace)
    elif act_func == 'tanh':
        return nn.Tanh()
    elif act_func == 'sigmoid':
        return nn.Sigmoid()
    elif act_func == 'h_swish':
        return Hswish(inplace=inplace)
    elif act_func == 'h_sigmoid':
        return Hsigmoid(inplace=inplace)
    elif act_func is None or act_func == 'none':
        return None
    else:
        raise ValueError('do not support: %s' % act_func)

ARGS = {
    'act_func': 'relu',
    'inplace': True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ReLU — a ReLU output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ReLU output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_activation Ran
ran; our contract draft was wrong, not the code
jordan-hs/rss-interference-cvprw2022/ofa/utils/pytorch_modules.py
code served (permissive licence) · get_code("b7d63a3b731686ad")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ReLU. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

torch.manual_seed(0)

class Hswish(nn.Module):

    def __init__(self, inplace=True):
        super(Hswish, self).__init__()
        self.inplace = inplace

    def forward(self, x):
        x = x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0
        return x

    def __repr__(self):
        return 'Hswish()'

class Hsigmoid(nn.Module):

    def __init__(self, inplace=True):
        super(Hsigmoid, self).__init__()
        self.inplace = inplace

    def forward(self, x):
        return F.relu6(x + 3.0, inplace=self.inplace) / 6.0

    def __repr__(self):
        return 'Hsigmoid()'

def build_activation(act_func, inplace=True):
    if act_func == 'relu':
        return nn.ReLU(inplace=inplace)
    elif act_func == 'relu6':
        return nn.ReLU6(inplace=inplace)
    elif act_func == 'tanh':
        return nn.Tanh()
    elif act_func == 'sigmoid':
        return nn.Sigmoid()
    elif act_func == 'h_swish':
        return Hswish(inplace=inplace)
    elif act_func == 'h_sigmoid':
        return Hsigmoid(inplace=inplace)
    elif act_func is None or act_func == 'none':
        return None
    else:
        raise ValueError('do not support: %s' % act_func)

ARGS = {
    'act_func': 'relu',
    'inplace': True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ReLU — a ReLU output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ReLU output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
conv1d Ran
ran on a synthesized input
MaximIntegratedAI/ai8x-synthesis/izer/compute.py
code served (permissive licence) · get_code("d2876e3d44cca1e5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 2, dtype family float, shape (2, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import torch

torch.manual_seed(0)

# Simple 1D convolution test case
# input_size = (in_channels, length)
# output_size = (out_channels, output_length)
# weight shape: (out_channels, in_channels // groups, kernel_size)
# bias shape: (out_channels,)

in_channels = 3
out_channels = 2
length = 8
kernel_size = 3
stride = 1
pad = 1
dilation = 1
groups = 1

# Calculate output length: (length + 2*pad - dilation*(kernel_size-1) - 1) // stride + 1
output_length = (length + 2*pad - dilation*(kernel_size-1) - 1) // stride + 1

# Setup data
data = np.random.randn(in_channels, length).astype(np.float64)

# Setup weight
weight = np.random.randn(out_channels, in_channels // groups, kernel_size).astype(np.float64)

# Setup bias
bias = np.random.randn(out_channels).astype(np.float64)

# Input and output sizes as tuples
input_size = (in_channels, length)
output_size = (out_channels, output_length)

ARGS = {
    'data': data,
    'weight': weight,
    'bias': bias,
    'input_size': input_size,
    'output_size': output_size,
    'kernel_size': kernel_size,
    'stride': stride,
    'pad': pad,
    'dilation': dilation,
    'fractional_stride': 1,
    'output_pad': 0,
    'groups': 1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (2, 8) — shape (2, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
conv2d Ran
ran on a synthesized input
MaximIntegratedAI/ai8x-synthesis/izer/compute.py
code served (permissive licence) · get_code("d2090bf5a862262f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 3, dtype family float, shape (3, 3, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Set seed for reproducibility
np.random.seed(0)

# Define a simple 2D convolution setup with small tensors
# Input: (C, H, W) = (2, 5, 5) - 2 input channels, 5x5 spatial
# Output: (C, H, W) = (3, 3, 3) - 3 output channels, 3x3 spatial
# Kernel: 3x3

in_channels = 2
out_channels = 3
kernel_h = 3
kernel_w = 3
input_h = 5
input_w = 5
stride_h = 1
stride_w = 1
pad_h = 0
pad_w = 0
dilation_h = 1
dilation_w = 1
fractional_stride_h = 1
fractional_stride_w = 1
output_pad_h = 0
output_pad_w = 0
groups = 1

# Calculate output size
output_h = (input_h + 2 * pad_h - dilation_h * (kernel_h - 1) - 1) // stride_h + 1
output_w = (input_w + 2 * pad_w - dilation_w * (kernel_w - 1) - 1) // stride_w + 1

# Create input data: shape (in_channels, input_h, input_w)
data = np.random.randn(in_channels, input_h, input_w).astype(np.float32)

# Create weight: shape (out_channels, in_channels, kernel_h, kernel_w)
weight = np.random.randn(out_channels, in_channels, kernel_h, kernel_w).astype(np.float32)

# Create bias: shape (out_channels,)
bias = np.random.randn(out_channels).astype(np.float32)

# Define sizes as tuples
input_size = (in_channels, input_h, input_w)
output_size = (out_channels, output_h, output_w)
kernel_size = (kernel_h, kernel_w)
stride = (stride_h, stride_w)
pad = (pad_h, pad_w)
dilation = (dilation_h, dilation_w)
fractional_stride = (fractional_stride_h, fractional_stride_w)
output_pad = (output_pad_h, output_pad_w)

ARGS = {
    'data': data,
    'weight': weight,
    'bias': bias,
    'input_size': input_size,
    'output_size': output_size,
    'kernel_size': kernel_size,
    'stride': stride,
    'pad': pad,
    'dilation': dilation,
    'fractional_stride': fractional_stride,
    'output_pad': output_pad,
    'groups': groups,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 3, dtype family float, shape (3, 3, 3) — shape (3, 3, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_to_x4_q7_weights Ran
ran on a synthesized input
MaximIntegratedAI/ai8x-synthesis/attic/armx4weights.py
code served (permissive licence) · get_code("45ea9f999fe65e71")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (96,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a small test weight matrix with dimensions divisible by 4
# Using shape (8, 12) so we have complete 4x4 blocks
weights = np.arange(96, dtype=np.float32).reshape(8, 12)

ARGS = {
    "weights": weights
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (96,) — shape (96,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
device Ran
ran on a synthesized input
analogdevicesinc/ai8x-training/devices.py
code served (permissive licence) · get_code("438c11e6440fc7fb")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import argparse

def device(astring):
    """
    Take die type, or part number, and return the die type.
    """
    s = astring.lower()
    if s.startswith('max'):
        s = s[3:]
    elif s.startswith('ai'):
        s = s[2:]
    try:
        num = int(s)
    except ValueError as exc:
        raise argparse.ArgumentTypeError(astring, 'is not a supported device type') from exc
    if num in [84, 85, 87]:
        dev = num
    elif num == 78000:
        dev = 85
    elif num == 78002:
        dev = 87
    else:
        raise argparse.ArgumentTypeError(astring, 'is not a supported device type')
    return dev


ARGS = {
    'astring': 'max85'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_bn_param Ran
ran; our contract draft was wrong, not the code
jordan-hs/rss-interference-cvprw2022/ofa/utils/my_modules.py
code served (permissive licence) · get_code("f060e6240188a921")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build a simple network with MyConv2d and BatchNorm2d
class SimpleNet(nn.Module):
    def __init__(self):
        super(SimpleNet, self).__init__()
        # We need to define MyConv2d first since it's used in the function
        self.conv = nn.Conv2d(3, 16, kernel_size=3, padding=1)
        self.bn = nn.BatchNorm2d(16)
    
    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        return x

# Create an instance of SimpleNet
net = SimpleNet()

# The function get_bn_param takes only one parameter: net
ARGS = {
    'net': net
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_bn_param Ran
ran; our contract draft was wrong, not the code
mit-han-lab/once-for-all/ofa/utils/my_modules.py
code served (permissive licence) · get_code("324ab791d89fa48c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build a small network with MyConv2d and BatchNorm2d
class SimpleNet(nn.Module):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.conv = nn.Conv2d(3, 4, kernel_size=3, padding=1)
        self.bn = nn.BatchNorm2d(4)
    
    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        return x

# Test with BatchNorm2d network
net_with_bn = SimpleNet()

# Test with GroupNorm network
class SimpleNetWithGN(nn.Module):
    def __init__(self):
        super(SimpleNetWithGN, self).__init__()
        self.conv = nn.Conv2d(3, 4, kernel_size=3, padding=1)
        self.gn = nn.GroupNorm(num_groups=2, num_channels=4)
    
    def forward(self, x):
        x = self.conv(x)
        x = self.gn(x)
        return x

net_with_gn = SimpleNetWithGN()

# Test with MyConv2d and weight standardization
from torch.nn.functional import conv2d

class MyConv2d(nn.Conv2d):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
        super(MyConv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias)
        self.WS_EPS = None

    def weight_standardization(self, weight):
        if self.WS_EPS is not None:
            weight_mean = weight.mean(dim=1, keepdim=True).mean(dim=2, keepdim=True).mean(dim=3, keepdim=True)
            weight = weight - weight_mean
            std = weight.view(weight.size(0), -1).std(dim=1).view(-1, 1, 1, 1) + self.WS_EPS
            weight = weight / std.expand_as(weight)
        return weight

    def forward(self, x):
        if self.WS_EPS is None:
            return super(MyConv2d, self).forward(x)
        else:
            return conv2d(x, self.weight_standardization(self.weight), self.bias, self.stride, self.padding, self.dilation, self.groups)

class SimpleNetWithWS(nn.Module):
    def __init__(self):
        super(SimpleNetWithWS, self).__init__()
        self.conv = MyConv2d(3, 4, kernel_size=3, padding=1)
        self.conv.WS_EPS = 1e-5
        self.bn = nn.BatchNorm2d(4)
    
    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        return x

net_with_ws = SimpleNetWithWS()

ARGS = {
    'net': net_with_bn
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_data_loaders Ran
ran; our contract draft was wrong, not the code
analogdevicesinc/ai8x-training/run_nas_network_search.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("0b1206ef7396e7eb")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_evo_search_params Ran
ran; our contract draft was wrong, not the code
analogdevicesinc/ai8x-training/run_nas_network_search.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("27a78d9679cfa28a")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_same_padding Ran
ran, honoured the contract we drafted
seulkiyeom/once-for-all/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from facebookresearch/AttentiveNAS, is pointer only (licence: NOASSERTION) · get_code("71cbea6f98bde1f9")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from facebookresearch/AttentiveNAS, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

make_divisible Ran
ran, honoured the contract we drafted
jordan-hs/rss-interference-cvprw2022/ofa/utils/pytorch_modules.py
code served (permissive licence) · get_code("addfbda3077115a3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

ARGS = {
    "v": 33,
    "divisor": 8,
    "min_val": None,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_divisible Ran
ran, honoured the contract we drafted
mit-han-lab/once-for-all/ofa/utils/pytorch_modules.py
code served (permissive licence) · get_code("f53d5967171577a0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# Setup for make_divisible function
# The function takes numeric parameters and performs integer arithmetic

ARGS = {
    "v": 100,
    "divisor": 8,
    "min_val": 16
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_args Ran
ran; our contract draft was wrong, not the code
analogdevicesinc/ai8x-training/run_nas_network_search.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("7f34c5377a84222f")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

parse_json Ran
ran on a synthesized input
MaximIntegratedAI/ai8x-synthesis/izer/assets.py
code served (permissive licence) · get_code("a6fe7c130778ca63")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create a temporary JSON file with test data
temp_dir = tempfile.gettempdir()
test_json_file = os.path.join(temp_dir, "test_parse_json.json")

test_data = {
    "name": "test_value",
    "count": 42,
    "nested": {
        "key": "value"
    }
}

with open(test_json_file, 'w', encoding='UTF-8') as f:
    json.dump(test_data, f)

ARGS = {
    "filename": test_json_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
partnum Ran
ran on a synthesized input
analogdevicesinc/ai8x-training/devices.py
code served (permissive licence) · get_code("6649c27f00740fa4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

ARGS = {
    'num': 84
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
passthrough_faker Ran
ran on a synthesized input
MaximIntegratedAI/ai8x-synthesis/izer/add_fake_passthrough.py
code served (permissive licence) · get_code("4ec1082125016453")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Conv2d. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

ARGS = {
    "n_channels": 3,
    "low_memory_footprint": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Conv2d — a Conv2d output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Conv2d output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
rgb2gray Ran
ran, violated the contract we drafted
twice154/ofa-for-super-resolution/ofa/elastic_nn/networks/ofa_mbs4.py
code served (permissive licence) · get_code("e45c8e7895f4cb79")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned ndarray, ndim 2, dtype family int, shape (3, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a small RGB image (height=3, width=4, channels=3)
# Using distinct dimensions: 3 rows, 4 columns, 3 color channels
img = np.array([
    [[255, 0, 0], [0, 255, 0], [0, 0, 255], [128, 128, 128]],
    [[64, 32, 16], [200, 150, 100], [50, 100, 150], [255, 255, 255]],
    [[10, 20, 30], [40, 50, 60], [70, 80, 90], [0, 0, 0]]
], dtype=np.uint8)

ARGS = {
    'img': img
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family int, shape (3, 4) — shape (3, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
rgb2y Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
twice154/ofa-for-super-resolution/ofa/elastic_nn/networks/ofa_mbs4.py
code served (permissive licence) · get_code("c98aa6358ebfe3a4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned ndarray, ndim 2, dtype family int, shape (3, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a small RGB image with uint8 dtype
# Shape: (height=3, width=4, channels=3) for RGB
img = np.array([
    [[255, 0, 0], [0, 255, 0], [0, 0, 255], [128, 128, 128]],
    [[64, 32, 16], [200, 150, 100], [50, 100, 150], [0, 0, 0]],
    [[255, 255, 255], [100, 100, 100], [30, 60, 90], [120, 240, 180]]
], dtype=np.uint8)

ARGS = {
    'img': img
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family int, shape (3, 4) — shape (3, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
shear_x Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
pprp/ofa-cifar/ofa/imagenet_classification/data_providers/cifar10.py
code served (permissive licence) · get_code("3c5071b8a3d0e038")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Image. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import random
from scipy import ndimage
from PIL import Image, ImageEnhance, ImageOps

# Create a small test image as a PIL Image
# Using a 4x3x3 image (height=4, width=3, channels=3 for RGB)
img_array = np.random.randint(0, 256, (4, 3, 3), dtype=np.uint8)
test_img = Image.fromarray(img_array, mode='RGB')

# magnitude parameter: an integer between 0 and 10 (used as index into magnitudes array)
magnitude_value = 5

ARGS = {
    'img': test_img,
    'magnitude': magnitude_value
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
transform_matrix_offset_center Ran
ran; our contract draft was wrong, not the code
pprp/ofa-cifar/ofa/imagenet_classification/data_providers/cifar10.py
code served (permissive licence) · get_code("3e773c6c5bc796d4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned ndarray, ndim 2, dtype family float, shape (3, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a 3x3 transformation matrix (e.g., a rotation matrix)
matrix = np.array([[0.866, -0.5, 0],
                   [0.5, 0.866, 0],
                   [0, 0, 1]])

# Image dimensions
x = 224
y = 160

ARGS = {
    'matrix': matrix,
    'x': x,
    'y': y
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (3, 3) — shape (3, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
unfold_batch Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
analogdevicesinc/ai8x-training/ai8x.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("4510fd0fdc46dfb1")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

write_mapping Ran
ran on a synthesized input
MaximIntegratedAI/ai8x-synthesis/izer/assets.py
code served (permissive licence) · get_code("e5200ffa1b2aced9")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path

# Define MakefileMapping class (from the function signature)
from collections.abc import MutableMapping

class MakefileMapping(MutableMapping):
    """
    This class is a modified dictionary that maps template strings (keys) to
    values while enforcing template pattern matching.  It also handles special
    cases where a value is not a 1:1 match to what should be written to
    the template Makefile.  For example, prepending source files with
    'SRCS +=', etc. is done "on the fly" through this mapping object.

    The key:value rules are as follows:
        * The key is lowercase, and any '.' is replaced is '_'. (relevant for JSON parsing)

        * The value is a tuple with 2 items:
            * index 0: The template string
            * index 1: The value to use when replacing the template string

    Ex:  The setter `MakefileMapping["mykey"]="myvalue"` results in a getter
    for `MakefileMapping["mykey"]` that would return `("##__MYKEY__##", "myvalue")`
    """

    def __init__(self, *args, **kwargs):
        self.d = {}
        self.update(dict(*args, **kwargs))

    def _form_key(self, key):
        return str(key).lower().replace('.', '_')

    def _form_tstring(self, key):
        return f'##__{self._form_key(key).upper()}__##'

    def __setitem__(self, key, value) -> None:
        _value = str(value)
        _key = self._form_key(key)
        if _key == 'srcs':
            if isinstance(value, (list, tuple)):
                tmp = tuple(map(lambda s: f'SRCS += {s}', value))
                _value = '\n'.join(tmp)
            elif isinstance(value, str):
                _value = f'SRCS += {value}\n'
        elif _key == 'vpaths':
            if isinstance(value, (list, tuple)):
                tmp = tuple(map(lambda s: f'VPATH += {s}', value))
                _value = '\n'.join(tmp)
            elif isinstance(value, str):
                _value = f'VPATH += {value}\n'
        elif _key == 'ipaths':
            if isinstance(value, (list, tuple)):
                tmp = tuple(map(lambda s: f'IPATH += {s}', value))
                _value = '\n'.join(tmp)
            elif isinstance(value, str):
                _value = f'IPATH += {value}\n'
        elif _key == 'defines':
            if isinstance(value, (list, tuple)):
                tmp = tuple(map(lambda s: f'PROJ_CFLAGS += -D{s}', value))
                _value = '\n'.join(tmp)
            elif isinstance(value, str):
                _value = f'PROJ_CFLAGS += -D{value}\n'
        elif _key == 'depth':
            if isinstance(value, int):
                _value = '../' * value
        self.d[_key] = (self._form_tstring(key), _value)

    def __getitem__(self, key):
        return self.d[self._form_key(key)]

    def __delitem__(self, key) -> None:
        del self.d[self._form_key(key)]

    def __iter__(self):
        return iter(self.d)

    def __len__(self):
        return len(self.d)

# Create temporary directory and files for testing
tmpdir = tempfile.TemporaryDirectory()
temp_path = Path(tmpdir.name)

# Create a template file with placeholders
template_file = temp_path / "Makefile.template"
template_content = """# Build configuration
##__SRCS__##
##__VPATHS__##
##__IPATHS__##
##__DEFINES__##
##__DEPTH__##
"""
template_file.write_text(template_content)

# Create output directory
out_dir = temp_path / "output"
out_dir.mkdir(exist_ok=True)

# Create a MakefileMapping with sample data
mapping = MakefileMapping()
mapping['srcs'] = ['main.c', 'utils.c']
mapping['vpaths'] = ['src/', 'lib/']
mapping['ipaths'] = ['include/', 'lib/include/']
mapping['defines'] = ['DEBUG', 'VERSION=1']
mapping['depth'] = 3

ARGS = {
    'template_file': template_file,
    'out_path': out_dir,
    'mapping': mapping,
    'overwrite': False,
    'backup': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
accuracy Not yet run
unverified (harvested, no recorded run)
seulkiyeom/once-for-all/utils.py
code served (permissive licence) · get_code("f867572dc454a862")
How it ran

no invocation recorded.

apply_policy Not yet run
unverified (harvested, no recorded run)
pprp/ofa-cifar/ofa/imagenet_classification/data_providers/cifar10.py
code served (permissive licence) · get_code("629948982eaffe6f")
How it ran

no invocation recorded.

convtranspose2d Not yet run
unverified (harvested, no recorded run)
MaximIntegratedAI/ai8x-synthesis/izer/compute.py
code served (permissive licence) · get_code("a18531b8217a23ef")
How it ran

no invocation recorded.

download_url Not yet run
unverified (harvested, no recorded run)
seulkiyeom/once-for-all/utils.py
code served (permissive licence) · get_code("3396d2fc2ff97e58")
How it ran

no invocation recorded.

idx Not yet run
unverified (harvested, no recorded run)
MaximIntegratedAI/ai8x-synthesis/izer/datamem.py
code served (permissive licence) · get_code("024ea8d0d27a2887")
How it ran

no invocation recorded.

load Not yet run
unverified (harvested, no recorded run)
MaximIntegratedAI/ai8x-synthesis/izer/checkpoint.py
code served (permissive licence) · get_code("ddb4421b09db03c5")
How it ran

no invocation recorded.

quantize_clamp Not yet run
unverified (harvested, no recorded run)
analogdevicesinc/ai8x-training/ai8x.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("df5e31805cad8a70")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

quantize_clamp_pool Not yet run
unverified (harvested, no recorded run)
analogdevicesinc/ai8x-training/ai8x.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Nokia-Bell-Labs/data-channel-extension, is pointer only (licence: NOASSERTION) · get_code("a5107b71c0946967")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from Nokia-Bell-Labs/data-channel-extension, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

unpack Not yet run
unverified (harvested, no recorded run)
MaximIntegratedAI/ai8x-synthesis/izer/datamem.py
code served (permissive licence) · get_code("1db0589ff0da8775")
How it ran

no invocation recorded.

used Not yet run
unverified (harvested, no recorded run)
MaximIntegratedAI/ai8x-synthesis/izer/datamem.py
code served (permissive licence) · get_code("e63ea88246d50a08")
How it ran

no invocation recorded.

Also present in other papers' code

Of 34 samples lifted from this paper's linked repositories, 34 carry a function identity and 0 do not. Of the 31 distinct functions with an identity, 15 are also present in other papers whose code we hold (4 with at least one paper dated earlier, 0 the same month, 15 later) and 16 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
accuracy 23 other papers
11 dated earlier, 12 later
1603.07285 (2016-03)
A guide to convolution arithmetic for deep learning
1608.07017 (2016-08)
Ambient Sound Provides Supervision for Visual Learning
1608.08710 (2016-08)
Pruning Filters for Efficient ConvNets
and 20 more not named here
apply_policy 6 other papers
2 dated earlier, 4 later
1805.09501 (2018-05)
AutoAugment: Learning Augmentation Policies from Data
1903.05285 (2019-03)
All You Need is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification
1909.13719 (2019-09)
RandAugment: Practical automated data augmentation with a reduced search space
and 3 more not named here
transform_matrix_offset_center 6 other papers
2 dated earlier, 4 later
1805.09501 (2018-05)
AutoAugment: Learning Augmentation Policies from Data
1903.05285 (2019-03)
All You Need is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification
1909.13719 (2019-09)
RandAugment: Practical automated data augmentation with a reduced search space
and 3 more not named here
shear_x 5 other papers
2 dated earlier, 3 later
1805.09501 (2018-05)
AutoAugment: Learning Augmentation Policies from Data
1903.05285 (2019-03)
All You Need is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification
2311.16450 (2023-11)
Typhoon Intensity Prediction with Vision Transformer
and 2 more not named here
make_divisible 2 other papers
2 later
2106.08630 (2021-06)
HELP: Hardware-Adaptive Efficient Latency Prediction for NAS via Meta-Learning
2206.13329 (2022-06)
Prior-Guided One-shot Neural Architecture Search
bn_fuser 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
build_activation 1 other paper
1 later
2106.08630 (2021-06)
HELP: Hardware-Adaptive Efficient Latency Prediction for NAS via Meta-Learning
get_bn_param 1 other paper
1 later
2106.08630 (2021-06)
HELP: Hardware-Adaptive Efficient Latency Prediction for NAS via Meta-Learning
get_data_loaders 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
get_evo_search_params 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
get_same_padding 1 other paper
1 later
2102.07954 (2021-02)
AlphaNet: Improved Training of Supernets with Alpha-Divergence
parse_args 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
quantize_clamp 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
quantize_clamp_pool 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators
unfold_batch 1 other paper
1 later
2412.06566 (2024-12)
DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-21.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

We address the challenging problem of efficient inference across many devices and resource constraints, especially on edge devices. Conventional approaches either manually design or use neural architecture search (NAS) to find a specialized neural network and train it from scratch for each case, which is computationally prohibitive (causing $CO_2$ emission as much as 5 cars' lifetime) thus unscalable. In this work, we propose to train a once-for-all (OFA) network that supports diverse architectural settings by decoupling training and search, to reduce the cost. We can quickly get a specialized sub-network by selecting from the OFA network without additional training. To efficiently train OFA networks, we also propose a novel progressive shrinking algorithm, a generalized pruning method that reduces the model size across many more dimensions than pruning (depth, width, kernel size, and resolution). It can obtain a surprisingly large number of sub-networks ($> 10^{19}$) that can fit different hardware platforms and latency constraints while maintaining the same level of accuracy as training independently. On diverse edge devices, OFA consistently outperforms state-of-the-art (SOTA) NAS methods (up to 4.0% ImageNet top1 accuracy improvement over MobileNetV3, or same accuracy but 1.5x faster than MobileNetV3, 2.6x faster than EfficientNet w.r.t measured latency) while reducing many orders of magnitude GPU hours and $CO_2$ emission. In particular, OFA achieves a new SOTA 80.0% ImageNet top-1 accuracy under the mobile setting ($<$600M MACs). OFA is the winning solution for the 3rd Low Power Computer Vision Challenge (LPCVC), DSP classification track and the 4th LPCVC, both classification track and detection track. Code and 50 pre-trained models (for many devices & many latency constraints) are released at https://github.com/mit-han-lab/once-for-all.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("1908.09791")
get_code_for_paper("1908.09791")
have("1908.09791")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/1908.09791.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/1908.09791.svg)](https://syntology.ai/paper/1908.09791)

Connect an agent — have() is free.