SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2307.03217 · NeurIPS · 2023

Quantification of Uncertainty with Adversarial Models

Günter Klambauer, Sepp Hochreiter, Kajetan Schweighofer, Lukas Aichberger, Mykyta Ielanskyi

arXiv · PDF · Open in the Atlas

Code that ran

8 of 14 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; 0 honoured the contract we drafted, 0 violated it and 3 have no contract verdict; 5 failures of our instrument; 6 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/2307.03217.json · badge: /api/ran/2307.03217.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
ml-jku/quam community
Papers with Code archive
8 ran · 6 no recorded run
FunctionStatusWhere it lives
BN2DND Ran
ran on a synthesized input
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("706813ea88c4046c")
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 BatchNorm2d. 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

class NDModule:
    @property
    def n(self):
        raise NotImplementedError
    
    def check_compat(self, m: nn.Module) -> bool:
        raise NotImplementedError
    
    def mirror_nth_member(self, idx: int, other):
        raise NotImplementedError
    
    def get_nth_member(self, idx: int) -> nn.Module:
        raise NotImplementedError
    
    def set_nth_member(self, idx: int, member: nn.Module):
        raise NotImplementedError
    
    def unbatch(self):
        return [self.get_nth_member(i) for i in range(self.n)]


class BN2DND(nn.BatchNorm2d, NDModule):

    def __init__(self, bn2d: nn.BatchNorm2d, n_groups: int):
        device = next(bn2d.parameters()).device
        dtype = next(bn2d.parameters()).dtype
        self.multiplicity = n_groups
        super().__init__(bn2d.num_features * n_groups, eps=bn2d.eps, momentum=bn2d.momentum, affine=bn2d.affine, track_running_stats=bn2d.track_running_stats, device=device, dtype=dtype)
        nuweights = bn2d.weight.data.repeat(n_groups)
        assert nuweights.shape == self.weight.data.shape, 'Something is up with repeating the weights!'
        self.weight.data = nuweights
        nubias = bn2d.bias.data.repeat(n_groups)
        assert nubias.shape == self.bias.data.shape, 'Something is up with repeating the biases!'
        self.bias.data = nubias
        if bn2d.track_running_stats:
            self.running_mean.data = bn2d.running_mean.data.repeat(n_groups)
            self.running_var.data = bn2d.running_var.data.repeat(n_groups)

    @property
    def n(self):
        return self.multiplicity

    @torch.no_grad()
    def check_compat(self, m: nn.Module) -> bool:
        if isinstance(m, nn.BatchNorm2d):
            return True
        else:
            return False

    @torch.no_grad()
    def mirror_nth_member(self, idx: int, other):
        n_in = self.num_features // self.n
        self.weight.data[n_in * idx:n_in * (idx + 1)] = other.weight.data[n_in * idx:n_in * (idx + 1)]
        self.bias.data[n_in * idx:n_in * (idx + 1)] = other.bias.data[n_in * idx:n_in * (idx + 1)]
        if self.track_running_stats:
            self.running_mean.data[n_in * idx:n_in * (idx + 1)] = other.running_mean.data[n_in * idx:n_in * (idx + 1)]
            self.running_var.data[n_in * idx:n_in * (idx + 1)] = other.running_var.data[n_in * idx:n_in * (idx + 1)]

    @torch.no_grad()
    def get_nth_member(self, idx: int) -> nn.Module:
        n_in = self.num_features // self.n
        retval = nn.BatchNorm2d(n_in, eps=self.eps, momentum=self.momentum, affine=self.affine, track_running_stats=self.track_running_stats, device=self.weight.data.device, dtype=self.weight.data.dtype)
        retval.weight.data = self.weight.data[n_in * idx:n_in * (idx + 1)].clone()
        retval.bias.data = self.bias.data[n_in * idx:n_in * (idx + 1)].clone()
        if self.track_running_stats:
            retval.running_mean.data = self.running_mean.data[n_in * idx:n_in * (idx + 1)].clone()
            retval.running_var.data = self.running_var.data[n_in * idx:n_in * (idx + 1)].clone()
        return retval

    @torch.no_grad()
    def set_nth_member(self, idx: int, member: nn.Module):
        n_in = self.num_features // self.n
        self.weight.data[n_in * idx:n_in * (idx + 1)] = member.weight.data
        self.bias.data[n_in * idx:n_in * (idx + 1)] = member.bias.data
        if self.track_running_stats:
            self.running_mean.data[n_in * idx:n_in * (idx + 1)] = member.running_mean.data
            self.running_var.data[n_in * idx:n_in * (idx + 1)] = member.running_var.data


base_bn = nn.BatchNorm2d(4, eps=1e-5, momentum=0.1, affine=True, track_running_stats=True)

INIT_ARGS = {
    'bn2d': base_bn,
    'n_groups': 3
}

ARGS = {
    'idx': 1
}

METHOD = 'get_nth_member'

bn2d_nd = BN2DND(**INIT_ARGS)
result = getattr(bn2d_nd, METHOD)(**ARGS)
print(f"Result type: {type(result)}")
print(f"Result num_features: {result.num_features}")
print(f"BN2DND n property: {bn2d_nd.n}")
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 BatchNorm2d — a BatchNorm2d output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a BatchNorm2d 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Conv2DND Ran
ran on a synthesized input
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("08a64a462fa9a870")
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 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

# Define a base Conv2d to wrap
base_conv = nn.Conv2d(
    in_channels=4,
    out_channels=3,
    kernel_size=3,
    bias=True,
    stride=1,
    padding=1,
    dilation=1
)

INIT_ARGS = {
    'conv': base_conv,
    'n_groups': 2
}

# For get_nth_member, we call it with idx=0
ARGS = {
    'idx': 0
}

METHOD = 'get_nth_member'

# Construct and call
from torch import nn
from typing import Callable, List, Dict, Union
from abc import ABC, abstractmethod, abstractproperty

class NDModule(ABC):

    @property
    @abstractmethod
    def n(self):
        pass

    @abstractmethod
    def check_compat(self, m: nn.Module) -> bool:
        pass

    @abstractmethod
    def mirror_nth_member(self, idx: int, other):
        pass

    @abstractmethod
    def get_nth_member(self, idx: int) -> nn.Module:
        pass

    @abstractmethod
    def set_nth_member(self, idx: int, member: nn.Module):
        pass

    def unbatch(self) -> List[nn.Module]:
        return [self.get_nth_member_as_linear(i) for i in range(self.n)]

class Conv2DND(nn.Conv2d, NDModule):

    def __init__(self, conv: nn.Conv2d, n_groups: int):
        device = next(conv.parameters()).device
        dtype = next(conv.parameters()).dtype
        has_bias = hasattr(conv, 'bias') and conv.bias is not None
        self.multiplicity = n_groups
        super().__init__(conv.in_channels * n_groups, conv.out_channels * n_groups, conv.kernel_size, bias=has_bias, stride=conv.stride, padding=conv.padding, dilation=conv.dilation, device=device, dtype=dtype, groups=n_groups)
        nuweights = conv.weight.data.repeat(n_groups, 1, 1, 1)
        assert nuweights.shape == self.weight.data.shape, 'Something is up with repeating the weights!'
        self.weight.data = nuweights
        if has_bias:
            nubias = conv.bias.data.repeat(n_groups)
            assert nubias.shape == self.bias.data.shape, 'Something is up with repeating the weights the biases!'
            self.bias.data = nubias

    @property
    def n(self):
        return self.multiplicity

    def check_compat(self, m: nn.Module) -> bool:
        if isinstance(m, nn.Conv2d):
            return True
        else:
            return False

    @torch.no_grad()
    def mirror_nth_member(self, idx: int, other):
        n_in = self.in_channels // self.n
        n_out = self.out_channels // self.n
        self.weight.data[n_out * idx:n_out * (idx + 1), :, :, :] = other.weight.data[n_out * idx:n_out * (idx + 1), :, :, :]
        if hasattr(self, 'bias') and self.bias is not None:
            self.bias.data[n_out * idx:n_out * (idx + 1)] = other.bias.data[n_out * idx:n_out * (idx + 1)]

    @torch.no_grad()
    def get_nth_member(self, idx: int) -> nn.Module:
        n_in = self.in_channels // self.n
        n_out = self.out_channels // self.n
        retval = nn.Conv2d(self.in_channels // self.n, self.out_channels // self.n, self.kernel_size, bias=hasattr(self, 'bias'), stride=self.stride, padding=self.padding, dilation=self.dilation, device=self.weight.data.device, dtype=self.weight.data.dtype)
        retval.weight.data = self.weight.data[n_out * idx:n_out * (idx + 1)]
        if hasattr(self, 'bias') and self.bias is not None:
            retval.bias.data = self.bias.data[n_out * idx:n_out * (idx + 1)]
        return retval

    @torch.no_grad()
    def set_nth_member(self, idx: int, member: nn.Conv2d):
        n_in = self.in_channels // self.n
        n_out = self.out_channels // self.n
        self.weight.data[n_out * idx:n_out * (idx + 1)] = member.weight.data
        if hasattr(self, 'bias') and self.bias is not None:
            self.bias.data[n_out * idx:n_out * (idx + 1)] = member.bias.data

instance = Conv2DND(**INIT_ARGS)
result = instance.get_nth_member(**ARGS)
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
LinearND Ran
ran on a synthesized input
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("26ddd38b22c99269")
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 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 torch
from torch import nn

# Create a base Linear layer to pass to LinearND
base_linear = nn.Linear(in_features=4, out_features=3, bias=True)

# INIT_ARGS: arguments for LinearND.__init__
INIT_ARGS = {
    'linear': base_linear,
    'n': 2
}

# ARGS: arguments for forward method
# Input shape should be (batch_size, n, in_features) or (batch_size, n*in_features)
# With n=2 and in_features=4, we create input of shape (2, 2, 4)
ARGS = {
    'input': torch.randn(2, 2, 4)
}

# Construct and call the class
from typing import Callable, List, Dict, Union
from abc import ABC, abstractmethod

class NDModule(ABC):

    @property
    @abstractmethod
    def n(self):
        pass

    @abstractmethod
    def check_compat(self, m: nn.Module) -> bool:
        pass

    @abstractmethod
    def mirror_nth_member(self, idx: int, other):
        pass

    @abstractmethod
    def get_nth_member(self, idx: int) -> nn.Module:
        pass

    @abstractmethod
    def set_nth_member(self, idx: int, member: nn.Module):
        pass

    def unbatch(self) -> List[nn.Module]:
        return [self.get_nth_member(i) for i in range(self.n)]

class LinearND(nn.Module, NDModule):

    def __init__(self, linear: nn.Linear, n: int) -> None:
        super(LinearND, self).__init__()
        device = linear.weight.device
        dtype = linear.weight.dtype
        self.multiplicity = n
        self.in_features = linear.in_features
        self.out_features = linear.out_features
        self.weight = nn.Parameter(linear.weight.data.unsqueeze(0).repeat(n, 1, 1).to(device).to(dtype))
        if linear.bias is not None:
            self.bias = nn.Parameter(linear.bias.data.unsqueeze(0).repeat(n, 1).to(device).to(dtype))
        return

    @property
    def n(self):
        return self.multiplicity

    def check_compat(self, m: nn.Module) -> bool:
        if isinstance(m, nn.Linear):
            return True
        else:
            return False

    @torch.no_grad()
    def get_nth_member(self, idx: int) -> nn.Linear:
        retval = nn.Linear(self.in_features, self.out_features, bias=hasattr(self, 'bias'))
        retval.weight.data = self.weight.data[idx, :, :]
        if hasattr(self, 'bias'):
            retval.bias.data = self.bias.data[idx, :]
        return retval

    @torch.no_grad()
    def set_nth_member(self, idx: int, member: nn.Linear):
        self.weight.data[idx, :, :] = member.weight.data
        if hasattr(self, 'bias'):
            self.bias.data[idx, :] = member.bias.data

    @torch.no_grad()
    def mirror_nth_member(self, idx: int, other):
        self.weight.data[idx, :, :] = other.weight.data[idx, :, :]
        if hasattr(self, 'bias'):
            self.bias.data[idx, :] = other.bias.data[idx, :]

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        if input.ndim == 2:
            input = input.reshape(input.shape[0], self.n, -1).transpose(0, 1)
        return torch.matmul(input, self.weight.transpose(1, 2)).add(self.bias.unsqueeze(1)) if hasattr(self, 'bias') else torch.matmul(input, self.weight.transpose(1, 2))


# Instantiate and call
instance = LinearND(**INIT_ARGS)
output = instance(**ARGS)
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
_forward Ran
ran; our contract draft was wrong, not the code
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("91857222d88bc678")
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
from torch import nn

torch.manual_seed(0)

# Create a simple model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Create a simple loss function
loss_fn = nn.MSELoss()

# Create model and move to CPU (default)
model = SimpleModel()

# Create sample input and target tensors
# x: batch_size=2, feature_dim=4
x = torch.randn(2, 4)

# y: batch_size=2, output_dim=3 (matches model output)
y = torch.randn(2, 3)

ARGS = {
    'x': x,
    'y': y,
    'model': model,
    'loss': loss_fn,
    'device_movements_async': 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 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
_forward_result Ran
ran; our contract draft was wrong, not the code
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("68960454ca149bef")
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 2, dtype family float, shape (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 torch
from torch import nn

torch.manual_seed(0)

# Create a small model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def forward(self, x):
        return self.linear(x)

# Instantiate the model on CPU
model = SimpleModel()

# Create a small input tensor
x = torch.randn(2, 4)

# Define arguments
ARGS = {
    'x': x,
    'model': model,
    'device_movements_async': 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 Tensor, ndim 2, dtype family float, shape (2, 3) — shape (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
batch_output_for_ce_loss_computation Ran
ran; our contract draft was wrong, not the code
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("27a0f5aa23396728")
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) and returned Tensor, ndim 3, dtype family float, shape (4, 3, 2) 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 takes a tensor and permutes its dimensions from (0, 1, 2) to (1, 2, 0)
# For a typical batch output used in CE loss computation:
# - dimension 0 is batch size
# - dimension 1 is number of classes
# - dimension 2 is sequence length (or spatial dimension)
# We'll create a small tensor with distinct sizes for each dimension

output = torch.randn(2, 4, 3)  # batch=2, classes=4, sequence_length=3

ARGS = {
    "output": output
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 3, dtype family float, shape (4, 3, 2) — shape (4, 3, 2) 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
feed_batch_model_image_multiple Ran
ran; our contract draft was wrong, not the code
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("79b796e77278df4c")
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 4, dtype family float, shape (2, 15, 4, 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

torch.manual_seed(0)

# Create a small image tensor
# Shape: (batch_size=2, channels=3, height=4, width=4)
img = torch.randn(2, 3, 4, 4)

# Number of models to repeat for
n_models = 5

ARGS = {
    "img": img,
    "n_models": n_models,
}
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 4, dtype family float, shape (2, 15, 4, 4) — shape (2, 15, 4, 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
feed_batch_model_vector_data Ran
ran; our contract draft was wrong, not the code
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("375d7ff0eb5b12ce")
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) and returned Tensor, ndim 3, dtype family float, shape (3, 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

torch.manual_seed(0)

# Create a small vector data tensor
# Shape: (batch_size=2, vector_dim=4)
vd = torch.randn(2, 4)

# Number of models to expand to
n_models = 3

ARGS = {
    "vd": vd,
    "n_models": n_models,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 3, dtype family float, shape (3, 2, 4) — shape (3, 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
ModelBatchedCELoss Not yet run
unverified (harvested, no recorded run)
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("6f835ba0e01b2840")
How it ran

no invocation recorded.

NDModule Not yet run
unverified (harvested, no recorded run)
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("b45f2865111fc2a0")
How it ran

no invocation recorded.

batch_target_for_ce_loss_computation Not yet run
unverified (harvested, no recorded run)
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("3b8baf6ab55b878c")
How it ran

no invocation recorded.

batch_that_net Not yet run
unverified (harvested, no recorded run)
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("a3c6cc2ebe0a4249")
How it ran

no invocation recorded.

predict_quam Not yet run
unverified (harvested, no recorded run)
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("149d4c9ec5226399")
How it ran

no invocation recorded.

set_submodule Not yet run
unverified (harvested, no recorded run)
ml-jku/quam/source/methods/quam.py
code served (permissive licence) · get_code("eeb07bae5055e8a0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 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 14 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

Quantifying uncertainty is important for actionable predictions in real-world applications. A crucial part of predictive uncertainty quantification is the estimation of epistemic uncertainty, which is defined as an integral of the product between a divergence function and the posterior. Current methods such as Deep Ensembles or MC dropout underperform at estimating the epistemic uncertainty, since they primarily consider the posterior when sampling models. We suggest Quantification of Uncertainty with Adversarial Models (QUAM) to better estimate the epistemic uncertainty. QUAM identifies regions where the whole product under the integral is large, not just the posterior. Consequently, QUAM has lower approximation error of the epistemic uncertainty compared to previous methods. Models for which the product is large correspond to adversarial models (not adversarial examples!). Adversarial models have both a high posterior as well as a high divergence between their predictions and that of a reference model. Our experiments show that QUAM excels in capturing epistemic uncertainty for deep learning models and outperforms previous methods on challenging tasks in the vision domain.

For agents

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

get_harvested_code_for_paper("2307.03217")
get_code_for_paper("2307.03217")
have("2307.03217")

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

curl https://syntology.ai/api/ran/2307.03217.json

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

[![Syntology run record](https://syntology.ai/api/ran/2307.03217.svg)](https://syntology.ai/paper/2307.03217)

Connect an agent — have() is free.