SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2402.02887 · CVPR · 2024

Time-, Memory-and Parameter-Efficient Visual Adaptation

Alexey Gritsenko, Cordelia Schmid, Anurag Arnab, Otniel-Bogdan Mercea

arXiv · PDF · Open in the Atlas

Code that ran

2 of 2 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 1 has no contract verdict; 1 failure of our instrument; 0 no recorded run), measured 2026-11-22; 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/2402.02887.json · badge: /api/ran/2402.02887.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/scenic community
Papers with Code archive
2 ran · 0 no recorded run
FunctionStatusWhere it lives
KnowledgeBase Ran
ran on a synthesized input · constructed an object; no result was computed
google-research/scenic/scenic/projects/knowledge_visual_language/models/local_memory.py
code served (permissive licence) · get_code("82b639d2d769e409")
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, KnowledgeBase: 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 jax
import jax.numpy as jnp
import numpy as np
from flax import linen as nn
from flax.core.frozen_dict import FrozenDict


class DummyFlaxModel(nn.Module):
    """Minimal Flax model for encode_knowledge testing."""
    
    def encode_knowledge(self, retr_texts, retr_images, bsz, train, random_drop_image):
        """Encode knowledge and return keys, values, mask, and dummy outputs."""
        batch_size = retr_images.shape[0]
        seq_len = retr_images.shape[1]
        
        # Return: keys_head, compressed_val, compressed_mask, dummy1, dummy2
        keys_head = jnp.ones((batch_size, seq_len, 8))
        compressed_val = jnp.ones((batch_size, seq_len, 16))
        compressed_mask = jnp.ones((batch_size, seq_len))
        dummy1 = jnp.zeros((batch_size,))
        dummy2 = jnp.zeros((batch_size,))
        
        return keys_head, compressed_val, compressed_mask, dummy1, dummy2


class DummyTrainState:
    """Minimal train state object."""
    
    def __init__(self):
        self.params = FrozenDict({})
        self.model_state = FrozenDict({})
        self.global_step = 0
        self.rng = jax.random.PRNGKey(0)


class DummyDataset:
    """Minimal dataset for KB initialization."""
    
    def __init__(self, batch_size=2, example_per_shard=8):
        self.batch_size = batch_size
        self.example_per_shard = example_per_shard
        self.train_iter = self._create_iterator()
        self.meta_data = {
            'batch_size': batch_size,
            'example_per_shard': example_per_shard,
            'image_spec': 'image_spec_placeholder',
            'knowledge_spec': 'knowledge_spec_placeholder',
        }
    
    def _create_iterator(self):
        """Create infinite iterator over dummy batches."""
        while True:
            yield {
                'image': np.random.randn(1, self.batch_size, 3, 4).astype(np.float32),
                'knowledge_tokens': np.random.randn(1, self.batch_size, 12).astype(np.float32),
            }


flax_model = DummyFlaxModel()
train_state = DummyTrainState()

kb_datasets = {
    'dataset_0': DummyDataset(batch_size=2, example_per_shard=8),
    'dataset_1': DummyDataset(batch_size=2, example_per_shard=8),
}

ARGS = {
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, flax, jax, numpy, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type KnowledgeBase — 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 KnowledgeBase 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
static_encode_knowledge Ran
ran; our contract draft was wrong, not the code
google-research/scenic/scenic/projects/knowledge_visual_language/models/local_memory.py
code served (permissive licence) · get_code("4851f6314ae20c77")
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 tuple. 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 numpy as np

torch.manual_seed(0)

# Create a minimal Flax-like model class that implements the required interface
class MockFlaxModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.embed_dim = 4
        self.seq_len = 8
        self.compressed_len = 3
    
    def apply(self, variables, retr_texts, retr_images, bsz, train, random_drop_image, method):
        """Mock apply method that simulates encode_knowledge"""
        # Return the expected outputs: keys_head, compressed_val, compressed_mask, _, _
        keys_head = torch.randn(bsz, self.embed_dim)  # (batch_size, embed_dim)
        compressed_val = torch.randn(bsz, self.compressed_len, self.embed_dim)  # (batch_size, compressed_len, embed_dim)
        compressed_mask = torch.ones(bsz, self.compressed_len)  # (batch_size, compressed_len)
        dummy1 = torch.tensor(0.0)
        dummy2 = torch.tensor(0.0)
        return keys_head, compressed_val, compressed_mask, dummy1, dummy2
    
    def encode_knowledge(self):
        """Placeholder encode_knowledge method"""
        pass

# Create a minimal TrainState class
class MockTrainState:
    def __init__(self):
        self.params = {
            'encoder': torch.randn(4, 4),
            'decoder': torch.randn(4, 4)
        }
        self.model_state = {
            'batch_norm': torch.randn(4)
        }
        self.global_step = torch.tensor(0)
        self.rng = torch.Generator()
        self.optimizer = None

# Create knowledge batch with required keys
knowledge_batch = {
    'knowledge_tokens': torch.randint(0, 100, (2, 8)),  # (batch_size=2, seq_len=8)
    'image': torch.randn(2, 3, 16, 16)  # (batch_size=2, channels=3, height=16, width=16)
}

# Create train state
train_state = MockTrainState()

# Create flax model
flax_model = MockFlaxModel()

# Define ARGS dictionary
ARGS = {
    'knowledge_batch': knowledge_batch,
    'train_state': train_state,
    'flax_model': flax_model
}
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 tuple — a tuple output has no array shape of its own; none is 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

Also present in other papers' code

Of 2 samples lifted from this paper's linked repositories, 2 carry a function identity and 0 do not. Of the 2 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 2 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.

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

As foundation models become more popular, there is a growing need to efficiently finetune them for downstream tasks. Although numerous adaptation methods have been proposed, they are designed to be efficient only in terms of how many parameters are trained. They, however, typically still require backpropagating gradients throughout the model, meaning that their training-time and -memory cost does not reduce as significantly. We propose an adaptation method which does not backpropagate gradients through the backbone. We achieve this by designing a lightweight network in parallel that operates on features from the frozen, pretrained backbone. As a result, our method is efficient not only in terms of parameters, but also in training-time and memory usage. Our approach achieves state-of-the-art accuracy-parameter tradeoffs on the popular VTAB benchmark, and we further show how we outperform prior works with respect to trainingtime and -memory usage too. We further demonstrate the training efficiency and scalability of our method by adapting a vision transformer backbone of 4 billion parameters for the computationally demanding task of video classification, without any intricate model parallelism. Here, we outperform a prior adaptor-based method which could only scale to a 1 billion parameter backbone, or fully-finetuning a smaller backbone, with the same GPU and less training time. To facilitate further research, we release code at https://github.com/google-research/scenic.

For agents

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

get_harvested_code_for_paper("2402.02887")
get_code_for_paper("2402.02887")
have("2402.02887")

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

curl https://syntology.ai/api/ran/2402.02887.json

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

[![Syntology run record](https://syntology.ai/api/ran/2402.02887.svg)](https://syntology.ai/paper/2402.02887)

Connect an agent — have() is free.