SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2004.07802 · ICLR · 2021

Geometry-Aware Gradient Algorithms for Neural Architecture Search

Ameet Talwalkar, Liam Li, Mikhail Khodak, Maria-Florina Balcan

arXiv · PDF · Open in the Atlas

Code that ran

12 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (9 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 1 failure of our instrument; 3 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/2004.07802.json · badge: /api/ran/2004.07802.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
NivNayman/XNAS community
Papers with Code archive
7 ran · 1 no recorded run
liamcli/gaea_release marked official in Papers with Code
Papers with Code archive
3 ran · 2 no recorded run
automl/nasbench-1shot1 application (classifier label)
classifier basis: classifier_2025
2 ran · 0 no recorded run
FunctionStatusWhere it lives
AdaptiveLR Ran
ran on a synthesized input · constructed an object; no result was computed
liamcli/gaea_release/cnn/architect/architect_edarts.py
pointer only (licence: NONE) · get_code("0ae6d8b3f048531d")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, liamcli/gaea_release when it was checked. 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.

Architect Ran
ran on a synthesized input · constructed an object; no result was computed
automl/nasbench-1shot1/optimizers/darts/architect.py
code served (permissive licence) · get_code("b0d6e9e65ecc665d")
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 an instance of its own class, Architect: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 numpy as np
from torch.autograd import Variable

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.conv = nn.Conv2d(3, 4, kernel_size=3, padding=1)
        self.arch_param = nn.Parameter(torch.randn(2, 4))
    
    def forward(self, x):
        return self.conv(x)
    
    def _loss(self, input, target):
        output = self.forward(input)
        return torch.nn.functional.mse_loss(output, target)
    
    def arch_parameters(self):
        return [self.arch_param]
    
    def new(self):
        return SimpleModel()

class ArgsConfig:
    def __init__(self):
        self.momentum = 0.9
        self.weight_decay = 1e-4
        self.arch_learning_rate = 0.01
        self.arch_weight_decay = 1e-3

model = SimpleModel()
args = ArgsConfig()

# Create dummy data
input_train = torch.randn(2, 3, 8, 8)
target_train = torch.randn(2, 4, 8, 8)
input_valid = torch.randn(2, 3, 8, 8)
target_valid = torch.randn(2, 4, 8, 8)

# Create network optimizer
network_optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=1e-4)

# Do one step to populate optimizer state
loss = model._loss(input_train, target_train)
loss.backward()
network_optimizer.step()
network_optimizer.zero_grad()

ARGS = {
    'model': model,
    'args': args
}
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 Architect — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Architect 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
DilConv Ran
ran on a synthesized input · constructed an object; no result was computed
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("132f546bb7ef86ba")
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 an instance of its own class, DilConv: the run constructed the object and did not call it, so no result was computed. 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)

ARGS = {
    'C_in': 3,
    'C_out': 16,
    'kernel_size': 3,
    'stride': 1,
    'padding': 1,
    'dilation': 1,
    'affine': 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 DilConv — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a DilConv 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
FactorizedReduce Ran
ran on a synthesized input · constructed an object; no result was computed
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("5acfdb2e37fd4c79")
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 an instance of its own class, FactorizedReduce: the run constructed the object and did not call it, so no result was computed. 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)

# FactorizedReduce expects:
# - C_in: input channels (int)
# - C_out: output channels (int, must be even)
# - affine: whether to use affine in BatchNorm2d (bool, default True)

ARGS = {
    'C_in': 8,
    'C_out': 16,
    'affine': 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 FactorizedReduce — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a FactorizedReduce 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
History Ran
ran on a synthesized input · constructed an object; no result was computed
liamcli/gaea_release/cnn/architect/architect_edarts.py
pointer only (licence: NONE) · get_code("2b798f4102f93552")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, liamcli/gaea_release when it was checked. 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.

Identity Ran
ran on a synthesized input
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("c6c6288219c9d532")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 2, dtype family float, shape (2, 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 torch
import torch.nn as nn

class Identity(nn.Module):

    def __init__(self):
        super(Identity, self).__init__()

    def forward(self, x):
        return x


INIT_ARGS = {}

ARGS = {
    'x': torch.randn(2, 4)
}

if __name__ == '__main__':
    model = Identity(**INIT_ARGS)
    output = model(**ARGS)
    print(output)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ReLUConvBN Ran
ran on a synthesized input · constructed an object; no result was computed
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("2b614488c8cb14d3")
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 an instance of its own class, ReLUConvBN: the run constructed the object and did not call it, so no result was computed. 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)

# Parameters for ReLUConvBN
C_in = 3
C_out = 16
kernel_size = 3
stride = 1
padding = 1
affine = True

ARGS = {
    'C_in': C_in,
    'C_out': C_out,
    'kernel_size': kernel_size,
    'stride': stride,
    'padding': padding,
    'affine': affine
}
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 ReLUConvBN — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ReLUConvBN 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
SepConv Ran
ran on a synthesized input · constructed an object; no result was computed
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("290cc06f55a8d0ed")
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 an instance of its own class, SepConv: the run constructed the object and did not call it, so no result was computed. 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)

ARGS = {
    'C_in': 4,
    'C_out': 8,
    'kernel_size': 3,
    'stride': 1,
    'padding': 1,
    'affine': 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 SepConv — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a SepConv 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
SepConvRes Ran
ran on a synthesized input · constructed an object; no result was computed
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("c585344417dc92cc")
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 an instance of its own class, SepConvRes: the run constructed the object and did not call it, so no result was computed. 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)

ARGS = {
    'C_in': 3,
    'C_out': 16,
    'kernel_size': 3,
    'stride': 1,
    'padding': 1,
    'affine': 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 SepConvRes — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a SepConvRes 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
Zero Ran
ran on a synthesized input · constructed an object; no result was computed
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("bc41443a9046202e")
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 an instance of its own class, Zero: the run constructed the object and did not call it, so no result was computed. 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)

# Create a small input tensor for testing
# Shape: (batch=2, channels=4, height=8, width=8)
x = torch.randn(2, 4, 8, 8)

# Test with stride=1
stride_1 = 1

# Test with stride=2
stride_2 = 2

# Define ARGS dict with stride parameter
ARGS = {
    'stride': stride_1
}
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 Zero — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Zero 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
_concat Ran
ran, honoured the contract we drafted
automl/nasbench-1shot1/optimizers/darts/architect.py
code served (permissive licence) · get_code("acb824e0877d8cbd")
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 Tensor, ndim 1, dtype family float, shape (19,) 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

torch.manual_seed(0)

# The function _concat takes a list/iterable of tensors and concatenates them
# after flattening each one. We need to provide a list of tensors.

# Create several small tensors with distinct shapes
tensor1 = torch.randn(2, 3)
tensor2 = torch.randn(4, 2)
tensor3 = torch.randn(5)

xs = [tensor1, tensor2, tensor3]

ARGS = {
    "xs": xs
}
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 Tensor, ndim 1, dtype family float, shape (19,) — shape (19,) 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
normalize Ran
ran; our contract draft was wrong, not the code
liamcli/gaea_release/cnn/architect/architect_edarts.py
pointer only (licence: NONE) · get_code("35516df482cccb4b")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, liamcli/gaea_release when it was checked. 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.

Architect Not yet run
unverified (harvested, no recorded run)
liamcli/gaea_release/cnn/architect/architect_edarts.py
pointer only (licence: NONE) · get_code("e70f49be6527cff9")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, liamcli/gaea_release when it was checked. 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.

ArchitectEDARTS Not yet run
unverified (harvested, no recorded run)
liamcli/gaea_release/cnn/architect/architect_edarts.py
pointer only (licence: NONE) · get_code("411c70cf0de59d37")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, liamcli/gaea_release when it was checked. 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.

Cell Not yet run
unverified (harvested, no recorded run)
NivNayman/XNAS/model.py
code served (permissive licence) · get_code("f4c89acbb418f526")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 6 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 0 the same month, 2 later) and 9 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
Identity
brick of this paper's assembly: model
26 other papers
3 dated earlier, 23 later · 22 only as an assembly brick
1512.03385 (2015-12)
Deep Residual Learning for Image Recognition
1806.09055 (2018-06)
DARTS: Differentiable Architecture Search
2002.05709 (2020-02)
A Simple Framework for Contrastive Learning of Visual Representations
and 23 more not named here
_concat
brick of this paper's assembly: helper
5 other papers
1 dated earlier, 4 later · 5 only as an assembly brick
1806.09055 (2018-06)
DARTS: Differentiable Architecture Search
2107.00860 (2021-07)
Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets
2108.01289 (2021-08)
AdvRush: Searching for Adversarially Robust Neural Architectures
and 2 more not named here
DilConv
brick of this paper's assembly: model
1 other paper
1 dated earlier
1806.09055 (2018-06)
DARTS: Differentiable Architecture Search
FactorizedReduce
brick of this paper's assembly: model
1 other paper
1 dated earlier
1806.09055 (2018-06)
DARTS: Differentiable Architecture Search
ReLUConvBN
brick of this paper's assembly: model
1 other paper
1 dated earlier
1806.09055 (2018-06)
DARTS: Differentiable Architecture Search
Zero
brick of this paper's assembly: model
1 other paper
1 dated earlier
1806.09055 (2018-06)
DARTS: Differentiable Architecture Search

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

Recent state-of-the-art methods for neural architecture search (NAS) exploit gradient-based optimization by relaxing the problem into continuous optimization over architectures and shared-weights, a noisy process that remains poorly understood. We argue for the study of single-level empirical risk minimization to understand NAS with weight-sharing, reducing the design of NAS methods to devising optimizers and regularizers that can quickly obtain high-quality solutions to this problem. Invoking the theory of mirror descent, we present a geometry-aware framework that exploits the underlying structure of this optimization to return sparse architectural parameters, leading to simple yet novel algorithms that enjoy fast convergence guarantees and achieve state-of-the-art accuracy on the latest NAS benchmarks in computer vision. Notably, we exceed the best published results for both CIFAR and ImageNet on both the DARTS search space and NAS-Bench-201; on the latter we achieve near-oracle-optimal performance on CIFAR-10 and CIFAR-100. Together, our theory and experiments demonstrate a principled way to co-design optimizers and continuous relaxations of discrete NAS search spaces.

For agents

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

get_harvested_code_for_paper("2004.07802")
get_code_for_paper("2004.07802")
have("2004.07802")

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

curl https://syntology.ai/api/ran/2004.07802.json

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

[![Syntology run record](https://syntology.ai/api/ran/2004.07802.svg)](https://syntology.ai/paper/2004.07802)

Connect an agent — have() is free.