SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.00426 · ICML · 2024

InterpreTabNet: Distilling Predictive Signals from Tabular Data by Salient Feature Interpretation

Rahul Krishnan, Jacob Si, Wendy Cheng, Michael Cooper

arXiv · PDF · Open in the Atlas

Code that ran

10 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (10 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; 5 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/2406.00426.json · badge: /api/ran/2406.00426.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
jacobyhsi/InterpreTabNet marked official in Papers with Code
Papers with Code archive
10 ran · 5 no recorded run
FunctionStatusWhere it lives
EmbeddingGenerator Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("32d61856363edf25")
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, EmbeddingGenerator: 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 numpy as np

torch.manual_seed(0)

# Test case 1: With categorical features
input_dim = 5
cat_dims = [10, 8]  # 2 categorical features with 10 and 8 modalities
cat_idxs = [1, 3]   # at positions 1 and 3
cat_emb_dim = 4     # embedding dimension for each categorical feature

ARGS = {
    "input_dim": input_dim,
    "cat_dims": cat_dims,
    "cat_idxs": cat_idxs,
    "cat_emb_dim": cat_emb_dim,
}
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 EmbeddingGenerator — 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 EmbeddingGenerator 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
Entmax15 Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("534e36abacbedbee")
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, Entmax15: 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
from torch import nn

torch.manual_seed(0)

# Create a small input tensor for testing
# Shape: (batch_size=2, num_classes=4)
input_tensor = torch.randn(2, 4, requires_grad=True)

# The Entmax15 module takes dim as a parameter
dim_value = -1

ARGS = {
    "dim": dim_value
}
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 Entmax15 — 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 Entmax15 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
Entmax15Function Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("8fc8e9f7a0141422")
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, Entmax15Function: 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

torch.manual_seed(0)

# The Entmax15Function.forward() expects:
# - input: a tensor with numerical values
# - dim: an integer specifying which dimension to apply entmax over

# Create a small input tensor (batch_size=2, num_classes=4)
input_tensor = torch.randn(2, 4, requires_grad=True)

# dim parameter specifies the dimension along which to compute entmax
dim_param = -1

ARGS = {
    "input": input_tensor,
    "dim": dim_param,
}
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 Entmax15Function — 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 Entmax15Function 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
FeatTransformer Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("87a7c3a7f5cab073")
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, FeatTransformer: 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
import numpy as np

torch.manual_seed(0)

# Build shared_layers as a ModuleList of Linear layers
shared_layers = nn.ModuleList([
    nn.Linear(4, 8, bias=False),
    nn.Linear(8, 8, bias=False)
])

ARGS = {
    'input_dim': 4,
    'output_dim': 8,
    'shared_layers': shared_layers,
    'n_glu_independent': 2,
    'virtual_batch_size': 128,
    'momentum': 0.02
}
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 FeatTransformer — 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 FeatTransformer 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
GBN Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("7ad035217e9fa718")
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, GBN: 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
from torch.nn import Linear, BatchNorm1d, ReLU, functional
import numpy as np

torch.manual_seed(0)

# GBN takes input_dim, virtual_batch_size, and momentum as parameters
# Create a small test case with batch_size=2 and input_dim=4

input_dim = 4
virtual_batch_size = 128
momentum = 0.01

ARGS = {
    'input_dim': input_dim,
    'virtual_batch_size': virtual_batch_size,
    'momentum': momentum
}
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 GBN — 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 GBN 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
GLU_Block Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("bda5ab5fc12e96df")
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, GLU_Block: 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
import numpy as np

torch.manual_seed(0)

# Build shared_layers as a list of Linear modules
# The GLU_Block uses shared_layers[0], shared_layers[1], etc.
# Each should be a Linear layer with appropriate dimensions
shared_layers = [
    nn.Linear(4, 8, bias=False),  # First layer: input_dim=4 to 2*output_dim=8
    nn.Linear(8, 8, bias=False),  # Subsequent layers: output_dim=4 to 2*output_dim=8
]

ARGS = {
    'input_dim': 4,
    'output_dim': 4,
    'n_glu': 2,
    'first': False,
    'shared_layers': shared_layers,
    'virtual_batch_size': 128,
    'momentum': 0.02,
}
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 GLU_Block — 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 GLU_Block 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
GLU_Layer Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("5861b1716a0a0046")
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, GLU_Layer: 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
from torch.nn import Linear, BatchNorm1d, ReLU, functional
import numpy as np

torch.manual_seed(0)

class GBN(torch.nn.Module):
    """
    Ghost Batch Normalization
    https://arxiv.org/abs/1705.08741
    """

    def __init__(self, input_dim, virtual_batch_size=128, momentum=0.01):
        super(GBN, self).__init__()
        self.input_dim = input_dim
        self.virtual_batch_size = virtual_batch_size
        self.bn = BatchNorm1d(self.input_dim, momentum=momentum)

    def forward(self, x):
        chunks = x.chunk(int(np.ceil(x.shape[0] / self.virtual_batch_size)), 0)
        res = [self.bn(x_) for x_ in chunks]
        return torch.cat(res, dim=0)

def initialize_glu(module, input_dim, output_dim):
    gain_value = np.sqrt((input_dim + output_dim) / np.sqrt(input_dim))
    torch.nn.init.xavier_normal_(module.weight, gain=gain_value)
    return

class GLU_Layer(torch.nn.Module):

    def __init__(self, input_dim, output_dim, fc=None, virtual_batch_size=128, momentum=0.02):
        super(GLU_Layer, self).__init__()
        self.output_dim = output_dim
        if fc:
            self.fc = fc
        else:
            self.fc = Linear(input_dim, 2 * output_dim, bias=False)
        initialize_glu(self.fc, input_dim, 2 * output_dim)
        self.bn = GBN(2 * output_dim, virtual_batch_size=virtual_batch_size, momentum=momentum)

    def forward(self, x):
        x = self.fc(x)
        x = self.bn(x)
        out = torch.mul(x[:, :self.output_dim], torch.sigmoid(x[:, self.output_dim:]))
        return out

input_dim = 4
output_dim = 3
fc = Linear(input_dim, 2 * output_dim, bias=False)
initialize_glu(fc, input_dim, 2 * output_dim)
virtual_batch_size = 128
momentum = 0.02

ARGS = {
    "input_dim": input_dim,
    "output_dim": output_dim,
    "fc": fc,
    "virtual_batch_size": virtual_batch_size,
    "momentum": momentum,
}
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 GLU_Layer — 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 GLU_Layer 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
Sparsemax Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("9ad4e49330a00585")
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, Sparsemax: 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
from torch import nn
from torch.autograd import Function

class SparsemaxFunction(Function):
    """
    An implementation of sparsemax (Martins & Astudillo, 2016). See
    :cite:`DBLP:journals/corr/MartinsA16` for detailed description.
    By Ben Peters and Vlad Niculae
    """

    @staticmethod
    def forward(ctx, input, dim=-1):
        """sparsemax: normalizing sparse transform (a la softmax)

        Parameters
        ----------
        ctx : torch.autograd.function._ContextMethodMixin
        input : torch.Tensor
            any shape
        dim : int
            dimension along which to apply sparsemax

        Returns
        -------
        output : torch.Tensor
            same shape as input

        """
        ctx.dim = dim
        max_val, _ = input.max(dim=dim, keepdim=True)
        input = input - max_val
        tau, supp_size = SparsemaxFunction._threshold_and_support(input, dim=dim)
        output = torch.clamp(input - tau, min=0)
        ctx.save_for_backward(supp_size, output)
        return output

    @staticmethod
    def backward(ctx, grad_output):
        supp_size, output = ctx.saved_tensors
        dim = ctx.dim
        grad_input = grad_output.clone()
        grad_input[output == 0] = 0
        v_hat = grad_input.sum(dim=dim) / supp_size.to(output.dtype).squeeze()
        v_hat = v_hat.unsqueeze(dim)
        grad_input = torch.where(output != 0, grad_input - v_hat, grad_input)
        return (grad_input, None)

    @staticmethod
    def _threshold_and_support(input, dim=-1):
        """Sparsemax building block: compute the threshold

        Parameters
        ----------
        input: torch.Tensor
            any dimension
        dim : int
            dimension along which to apply the sparsemax

        Returns
        -------
        tau : torch.Tensor
            the threshold value
        support_size : torch.Tensor

        """
        input_srt, _ = torch.sort(input, descending=True, dim=dim)
        input_cumsum = input_srt.cumsum(dim) - 1
        rhos = _make_ix_like(input, dim)
        support = rhos * input_srt > input_cumsum
        support_size = support.sum(dim=dim).unsqueeze(dim)
        tau = input_cumsum.gather(dim, support_size - 1)
        tau = tau / support_size.to(input.dtype)
        return (tau, support_size)

def _make_ix_like(input, dim=0):
    d = input.size(dim)
    rho = torch.arange(1, d + 1, device=input.device, dtype=input.dtype)
    view = [1] * input.dim()
    view[0] = -1
    return rho.view(view).transpose(0, dim)

sparsemax = SparsemaxFunction.apply

class Sparsemax(nn.Module):

    def __init__(self, dim=-1):
        self.dim = dim
        super(Sparsemax, self).__init__()

    def forward(self, input):
        return sparsemax(input, self.dim)

torch.manual_seed(0)

ARGS = {
    "dim": -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 Sparsemax — 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 Sparsemax 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
SparsemaxFunction Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("fd64f23efb357294")
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, SparsemaxFunction: 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
from torch.autograd import Function

torch.manual_seed(0)

# The SparsemaxFunction.forward() and backward() methods need:
# - ctx: context object (provided by autograd)
# - input: a tensor (any shape, but we'll use batch_size=2, dim=4)
# - dim: dimension along which to apply sparsemax (default -1)

# Create a small input tensor
input_tensor = torch.randn(2, 4, requires_grad=True)

# The dim parameter specifies which dimension to apply sparsemax along
dim_value = -1

ARGS = {
    'input': input_tensor,
    'dim': dim_value,
}
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 SparsemaxFunction — 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 SparsemaxFunction 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
TabNetEncoder Ran
ran on a synthesized input · constructed an object; no result was computed
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("e2353d85a4193ce6")
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, TabNetEncoder: 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
from torch.nn import Linear, BatchNorm1d, ReLU, functional, Identity
import numpy as np

torch.manual_seed(0)

# Mock sparsemax module since it's not standard
class MockSparsemax(nn.Module):
    def __init__(self, dim=-1):
        super().__init__()
        self.dim = dim
    
    def forward(self, x):
        return torch.softmax(x, dim=self.dim)

class MockEntmax15(nn.Module):
    def __init__(self, dim=-1):
        super().__init__()
        self.dim = dim
    
    def forward(self, x):
        return torch.softmax(x, dim=self.dim)

# Mock sparsemax module
class MockSparseMaxModule:
    Sparsemax = MockSparsemax
    Entmax15 = MockEntmax15

# Mock GBN (Ghost Batch Normalization)
class GBN(nn.Module):
    def __init__(self, output_dim, virtual_batch_size=128, momentum=0.02):
        super().__init__()
        self.bn = BatchNorm1d(output_dim, momentum=momentum)
    
    def forward(self, x):
        return self.bn(x)

# Mock initialize_non_glu function
def initialize_non_glu(layer, input_dim, output_dim):
    pass

# Mock GLU_Block
class GLU_Block(nn.Module):
    def __init__(self, input_dim, output_dim, first=True, shared_layers=None, n_glu=1, virtual_batch_size=128, momentum=0.02):
        super().__init__()
        self.fc = Linear(input_dim, output_dim)
    
    def forward(self, x):
        return self.fc(x)

# Now define the AttentiveTransformer class (as given)
class AttentiveTransformer(torch.nn.Module):

    def __init__(self, input_dim, output_dim, virtual_batch_size=128, momentum=0.02, mask_type='sparsemax'):
        super(AttentiveTransformer, self).__init__()
        self.fc = Linear(input_dim, output_dim, bias=False)
        initialize_non_glu(self.fc, input_dim, output_dim)
        self.bn = GBN(output_dim, virtual_batch_size=virtual_batch_size, momentum=momentum)
        if mask_type == 'sparsemax':
            self.selector = MockSparsemax(dim=-1)
        elif mask_type == 'entmax':
            self.selector = MockEntmax15(dim=-1)
        elif mask_type == 'relu':
            self.selector = ReLU()
        elif mask_type == 'softmax':
            self.selector = None
        else:
            raise NotImplementedError('Please choose either sparsemax or entmax as masktype')

    def forward(self, priors, processed_feat):
        x = self.fc(processed_feat)
        x = self.bn(x)
        x = torch.mul(x, priors)
        x = torch.nn.functional.softmax(x, dim=-1)
        return x

# Create valid arguments for AttentiveTransformer
ARGS = {
    'input_dim': 4,
    'output_dim': 3,
    'virtual_batch_size': 128,
    'momentum': 0.02,
    'mask_type': 'sparsemax'
}
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 TabNetEncoder — 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 TabNetEncoder 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
AttentiveTransformer Not yet run
unverified (harvested, no recorded run)
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("072b418dfe2df0ba")
How it ran

no invocation recorded.

TabNet Not yet run
unverified (harvested, no recorded run)
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("ccadc817f4a9d1a2")
How it ran

no invocation recorded.

TabNetNoEmbeddings Not yet run
unverified (harvested, no recorded run)
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("440701a70b477120")
How it ran

no invocation recorded.

initialize_glu Not yet run
unverified (harvested, no recorded run)
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("7690965dab7e21e6")
How it ran

no invocation recorded.

initialize_non_glu Not yet run
unverified (harvested, no recorded run)
jacobyhsi/InterpreTabNet/pytorch_tabnet/tab_network.py
code served (permissive licence) · get_code("ed1b885a712fb350")
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, 1 are also present in other papers whose code we hold (1 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.

FunctionAlso present inEarliest-datedOthers named
Entmax15Function
brick of this paper's assembly: helper
1 other paper
1 dated earlier
2106.01613 (2021-06)
NODE-GAM: Neural Generalized Additive Model for Interpretable Deep Learning

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

Tabular data are omnipresent in various sectors of industries. Neural networks for tabular data such as TabNet have been proposed to make predictions while leveraging the attention mechanism for interpretability. However, the inferred attention masks are often dense, making it challenging to come up with rationales about the predictive signal. To remedy this, we propose InterpreTab-Net, a variant of the TabNet model that models the attention mechanism as a latent variable sampled from a Gumbel-Softmax distribution. This enables us to regularize the model to learn distinct concepts in the attention masks via a KL Divergence regularizer. It prevents overlapping feature selection by promoting sparsity which maximizes the model's efficacy and improves interpretability to determine the important features when predicting the outcome. To assist in the interpretation of feature interdependencies from our model, we employ a large language model (GPT-4) and use prompt engineering to map from the learned feature mask onto natural language text describing the learned signal. Through comprehensive experiments on real-world datasets, we demonstrate that InterpreTabNet outperforms previous methods for interpreting tabular data while attaining competitive accuracy.

For agents

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

get_harvested_code_for_paper("2406.00426")
get_code_for_paper("2406.00426")
have("2406.00426")

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

curl https://syntology.ai/api/ran/2406.00426.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.00426.svg)](https://syntology.ai/paper/2406.00426)

Connect an agent — have() is free.