SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2201.05125 · ICLR · 2022

GradMax: Growing Neural Networks using Gradient Information

Thomas Unterthiner, Max Vladymyrov, Utku Evci, Bart Van Merriënboer, Fabian Pedregosa

arXiv · PDF · Open in the Atlas

Code that ran

10 of 13 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 0 failures of our instrument; 3 no recorded run), measured 2026-11-22; 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/2201.05125.json · badge: /api/ran/2201.05125.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
google-research/growneuron canonical (classifier label)
classifier basis not recorded on the link
10 ran · 3 no recorded run
FunctionStatusWhere it lives
Conv2D Ran
ran on a synthesized input · constructed an object; no result was computed
google-research/growneuron/growneuron/cifar/wide_resnet.py
code served (permissive licence) · get_code("2016192a8b36e38a")
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, Conv2D: 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 tensorflow as tf

ARGS = {
    'filters': 32,
    'seed': 42,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Conv2D — 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 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
adjust_epochs Ran
ran on a synthesized input
google-research/growneuron/growneuron/updaters.py
code served (permissive licence) · get_code("45c580fe8487e547")
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

torch.manual_seed(0)

ARGS = {
    "train_epochs": 100,
    "width_scale": 0.5,
    "update_frequency": 10,
    "start_iteration": 1000,
    "n_growth_steps": 5,
    "steps_per_epoch": 50
}
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 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
build_input_fn Ran
ran on a synthesized input
google-research/growneuron/growneuron/cifar/data.py
code served (permissive licence) · get_code("4147f063ec9303d0")
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 function. 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 tensorflow.compat.v2 as tf
import numpy as np

# Create a mock TFDS builder
class MockTFDSBuilder:
    def as_dataset(self, split='train', shuffle_files=False, as_supervised=True):
        # Create a simple dataset with image (4x4x3) and label (scalar)
        images = tf.constant(np.random.rand(8, 4, 4, 3).astype(np.float32))
        labels = tf.constant(np.random.randint(0, 10, 8))
        dataset = tf.data.Dataset.from_tensor_slices((images, labels))
        return dataset

# Create a mock Topology
class MockTopology:
    def __init__(self, num_tpus_per_task=8):
        self.num_tpus_per_task = num_tpus_per_task

# Create a mock InputContext
class MockInputContext:
    def __init__(self, num_input_pipelines=1, input_pipeline_id=0):
        self.num_input_pipelines = num_input_pipelines
        self.input_pipeline_id = input_pipeline_id
    
    def get_per_replica_batch_size(self, global_batch_size):
        return global_batch_size // self.num_input_pipelines

ARGS = {
    'builder': MockTFDSBuilder(),
    'global_batch_size': 4,
    'topology': MockTopology(num_tpus_per_task=8),
    'is_training': True,
    'cache_dataset': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
center_crop Ran
ran on a synthesized input
google-research/growneuron/growneuron/imagenet/data_util.py
code served (permissive licence) · get_code("93886821dd9582ad")
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 EagerTensor, ndim 3, dtype family float, shape (4, 5, 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 tensorflow.compat.v2 as tf

# Enable eager execution for TensorFlow 2.x compatibility
tf.enable_v2_behavior()

# Create a small test image tensor
# Shape: (height, width, channels) = (8, 6, 3)
# Using small distinct dimensions: 8x6x3 for height x width x channels
image = tf.constant([
    [[255, 0, 0], [0, 255, 0], [0, 0, 255], [255, 255, 0], [255, 0, 255], [0, 255, 255]],
    [[100, 100, 100], [50, 50, 50], [150, 150, 150], [200, 200, 200], [10, 10, 10], [230, 230, 230]],
    [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12], [13, 14, 15], [16, 17, 18]],
    [[255, 0, 0], [0, 255, 0], [0, 0, 255], [255, 255, 0], [255, 0, 255], [0, 255, 255]],
    [[100, 100, 100], [50, 50, 50], [150, 150, 150], [200, 200, 200], [10, 10, 10], [230, 230, 230]],
    [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12], [13, 14, 15], [16, 17, 18]],
    [[255, 0, 0], [0, 255, 0], [0, 0, 255], [255, 255, 0], [255, 0, 255], [0, 255, 255]],
    [[100, 100, 100], [50, 50, 50], [150, 150, 150], [200, 200, 200], [10, 10, 10], [230, 230, 230]],
], dtype=tf.uint8)

# Target crop dimensions
height = 4
width = 5

# Crop proportion - retain 0.9 of the image along the less-cropped side
crop_proportion = 0.9

ARGS = {
    "image": image,
    "height": height,
    "width": width,
    "crop_proportion": crop_proportion,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 3, dtype family float, shape (4, 5, 3) — shape (4, 5, 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
distorted_bounding_box_crop Ran
ran on a synthesized input
google-research/growneuron/growneuron/imagenet/data_util.py
code served (permissive licence) · get_code("362ad792817cb093")
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 EagerTensor, ndim 3, dtype family float, shape (2, 2, 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 tensorflow.compat.v2 as tf
import numpy as np

tf.enable_v2_behavior()

# Create a small test image: [height=4, width=3, channels=3]
image = tf.constant(np.random.rand(4, 3, 3).astype(np.float32))

# Create bounding boxes in format [1, num_boxes, 4] where coords are [ymin, xmin, ymax, xmax]
# with num_boxes=1, normalized to [0, 1)
bbox = tf.constant([[[0.1, 0.1, 0.9, 0.9]]], dtype=tf.float32)

ARGS = {
    'image': image,
    'bbox': bbox,
    'min_object_covered': 0.1,
    'aspect_ratio_range': (0.75, 1.33),
    'area_range': (0.05, 1.0),
    'max_attempts': 100,
    'scope': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 3, dtype family float, shape (2, 2, 3) — shape (2, 2, 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
extract_image_patches Ran
ran on a synthesized input
google-research/growneuron/growneuron/growers.py
code served (permissive licence) · get_code("7e0d28e183f601ad")
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 EagerTensor, ndim 4, dtype family float, shape (2, 3, 3, 27) 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 tensorflow as tf

# Create a small batched input tensor
# Shape: [batch=2, height=5, width=5, channels=3]
batch_size = 2
height = 5
width = 5
in_channels = 3

x = tf.constant(
    np.random.randn(batch_size, height, width, in_channels).astype(np.float32)
)

# Define kernel size as a tuple of two integers
kernel_size = (3, 3)

# Define stride as a tuple of two integers
stride = (1, 1)

ARGS = {
    "x": x,
    "kernel_size": kernel_size,
    "stride": stride,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 4, dtype family float, shape (2, 3, 3, 27) — shape (2, 3, 3, 27) 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_activation_fn Ran
ran on a synthesized input
google-research/growneuron/growneuron/layers.py
code served (permissive licence) · get_code("f5060e86ebb597e4")
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 function. 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 tensorflow as tf

ARGS = {
    'actv_fn': 'relu'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
pad_zeros_to Ran
ran on a synthesized input
google-research/growneuron/growneuron/updaters.py
code served (permissive licence) · get_code("5a9107a6392b75f3")
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 EagerTensor, ndim 2, dtype family float, shape (4, 5) 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 tensorflow as tf
import numpy as np

# Create a small 2D tensor (4 rows, 3 columns)
tensor_2d = tf.constant([[1.0, 2.0, 3.0],
                         [4.0, 5.0, 6.0],
                         [7.0, 8.0, 9.0],
                         [10.0, 11.0, 12.0]])

# Define new_shape to be larger than the original tensor shape
# Original shape is (4, 3), new shape should be (4, 5) or (6, 3) or (6, 5)
new_shape = [4, 5]

ARGS = {
    "tensor": tensor_2d,
    "new_shape": new_shape
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 2, dtype family float, shape (4, 5) — shape (4, 5) 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
process_conv_aux_gradient Ran
ran on a synthesized input
google-research/growneuron/growneuron/growers.py
code served (permissive licence) · get_code("e5e6a1a5870d1730")
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 EagerTensor, ndim 2, dtype family float, shape (9, 20) 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 tensorflow as tf

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

# Create a small gradient tensor
# grad should be a 4D tensor that will be transposed to (2, height, width, channels)
# After transpose from (2, 0, 1, 3), original shape should be (height, width, batch=2, channels)
# So we create shape: (batch=2, height=4, width=3, channels=5)
grad = tf.random.normal(shape=(2, 4, 3, 5), dtype=tf.float32)

# Define second_kernel_size as a tuple of two integers
second_kernel_size = (2, 2)

ARGS = {
    "grad": grad,
    "second_kernel_size": second_kernel_size,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 2, dtype family float, shape (9, 20) — shape (9, 20) 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
random_apply Ran
ran on a synthesized input
google-research/growneuron/growneuron/imagenet/data_util.py
code served (permissive licence) · get_code("293ca4cd93119d57")
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 EagerTensor, 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 tensorflow.compat.v2 as tf

tf.enable_v2_behavior()

def simple_func(x):
    """Simple function that adds 1 to input."""
    return x + 1

ARGS = {
    "func": simple_func,
    "p": 0.7,
    "x": tf.constant([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]], dtype=tf.float32)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, 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: none by the harness; any seed is in the setup block
DepthwiseConv2D Not yet run
unverified (harvested, no recorded run)
google-research/growneuron/growneuron/imagenet/mb_v1.py
code served (permissive licence) · get_code("6f50d373a40a2b95")
How it ran

no invocation recorded.

grow_new_layer Not yet run
unverified (harvested, no recorded run)
google-research/growneuron/growneuron/layers.py
code served (permissive licence) · get_code("6c8ab1cdbee83212")
How it ran

no invocation recorded.

grow_new_ln_layer Not yet run
unverified (harvested, no recorded run)
google-research/growneuron/growneuron/layers.py
code served (permissive licence) · get_code("d9acc2750884926d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 2 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 1 later) and 11 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
Conv2D 1 other paper
1 later
2211.06291 (2022-11)
Do Bayesian Neural Networks Need To Be Fully Stochastic?
random_apply 1 other paper
1 dated earlier
2109.12909 (2021-09)
Compressive Visual Representations

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 2026-11-22.

Repositories linked to this paper

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

Abstract

The architecture and the parameters of neural networks are often optimized independently, which requires costly retraining of the parameters whenever the architecture is modified. In this work we instead focus on growing the architecture without requiring costly retraining. We present a method that adds new neurons during training without impacting what is already learned, while improving the training dynamics. We achieve the latter by maximizing the gradients of the new weights and efficiently find the optimal initialization by means of the singular value decomposition (SVD). We call this technique Gradient Maximizing Growth (GradMax) and demonstrate its effectiveness in variety of vision tasks and architectures 1 .

For agents

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

get_harvested_code_for_paper("2201.05125")
get_code_for_paper("2201.05125")
have("2201.05125")

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

curl https://syntology.ai/api/ran/2201.05125.json

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

[![Syntology run record](https://syntology.ai/api/ran/2201.05125.svg)](https://syntology.ai/paper/2201.05125)

Connect an agent — have() is free.