SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2103.15619 · CVPR · 2021

SetVAE: Learning Hierarchical Composition for Generative Modeling of Set-Structured Data

Jinwoo Kim, Jaehoon Yoo, Juho Lee Seunghoon, Hong Kaist

arXiv · PDF · Open in the Atlas

Code that ran

17 of 27 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (7 of them constructed an object rather than computing a result; 2 honoured the contract we drafted, 0 violated it and 7 have no contract verdict; 8 failures of our instrument; 10 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/2103.15619.json · badge: /api/ran/2103.15619.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
jw9730/setvae marked official in Papers with Code
Papers with Code archive
17 ran · 10 no recorded run
FunctionStatusWhere it lives
ChamferCriterion Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("6787a72363072514")
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, ChamferCriterion: 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

torch.manual_seed(0)

# Create a minimal args object with all required attributes
class Args:
    def __init__(self):
        self.device = 'cpu'
        self.kl_warmup_epochs = 2
        self.beta = 0.1
        self.z_dim = 8
        self.z_scales = [1, 2]

args = Args()

# Create test tensors
batch_size = 2
max_points = 5
feature_dim = 3

# output_set and target_set: [B, N, D] - points in 3D space
output_set = torch.randn(batch_size, max_points, feature_dim)
target_set = torch.randn(batch_size, max_points, feature_dim)

# Masks: [B, N] - False means valid point, True means padded/invalid
output_mask = torch.tensor([
    [False, False, False, True, True],
    [False, False, True, True, True]
])
target_mask = torch.tensor([
    [False, False, False, True, True],
    [False, False, True, True, True]
])

# Create output dict with required keys
output = {
    'set': output_set,
    'set_mask': output_mask,
    'kls': [torch.randn(batch_size, args.z_dim) for _ in args.z_scales]
}

# Package everything into ARGS dict
ARGS = {
    'args': args
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
metrics, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ChamferCriterion — 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 ChamferCriterion 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
CombinedCriterion Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("2dbcaf1aa112361e")
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, CombinedCriterion: 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

torch.manual_seed(0)

# Create a minimal args class with all required attributes
class Args:
    def __init__(self):
        self.device = 'cpu'  # CPU only
        self.kl_warmup_epochs = 2
        self.beta = 0.5
        self.z_dim = 4
        self.z_scales = [1, 2]

# Create sample tensors
batch_size = 2
num_points = 5
feature_dim = 3

# Output from model
output_set = torch.randn(batch_size, num_points, feature_dim)
output_mask = torch.tensor([
    [False, False, True, True, True],
    [False, False, False, True, True]
])

# Target set
target_set = torch.randn(batch_size, num_points, feature_dim)
target_mask = torch.tensor([
    [False, False, True, True, True],
    [False, False, False, True, True]
])

# KL divergence values (one per scale in z_scales)
kls = [
    torch.randn(batch_size, 4),  # First scale
    torch.randn(batch_size, 4),  # Second scale
]

# Output dictionary
output = {
    'set': output_set,
    'set_mask': output_mask,
    'kls': kls
}

# Arguments for CombinedCriterion
ARGS = {
    'args': Args()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
metrics, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type CombinedCriterion — 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 CombinedCriterion 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
EMDCriterion Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("bed8a7b03b512335")
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, EMDCriterion: 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

torch.manual_seed(0)

# Mock the unmask function and match_cost that are imported
def unmask(tensor, mask):
    """Extract unmasked elements from tensor based on mask"""
    # mask is True where data is masked (invalid), False where valid
    # Return only the valid elements
    batch_size = tensor.shape[0]
    results = []
    for i in range(batch_size):
        valid_indices = ~mask[i]
        results.append(tensor[i][valid_indices])
    # Return concatenated for batch processing
    return torch.cat(results, dim=0)

def match_cost(output_set, target_set):
    """Mock match_cost function that returns loss values"""
    # Just return a dummy tensor of losses
    return torch.tensor([0.5, 0.3])

# Create mock args object
class Args:
    def __init__(self):
        self.device = 'cpu'  # Using CPU only
        self.beta = 0.1
        self.kl_warmup_epochs = 10
        self.z_dim = 2
        self.z_scales = [1, 2]

args = Args()

# Create batch of output data
batch_size = 2
num_points = 4
output_set = torch.randn(batch_size, num_points, 3)
output_mask = torch.tensor([[False, False, True, True], [False, False, False, True]], dtype=torch.bool)

# Create batch of target data (matching number of valid points)
# First batch has 2 valid points, second batch has 3 valid points
target_set = torch.randn(batch_size, num_points, 3)
target_mask = torch.tensor([[False, False, True, True], [False, False, False, True]], dtype=torch.bool)

# Create KL losses
kls = [torch.randn(batch_size, args.z_dim), torch.randn(batch_size, args.z_dim)]

# Prepare output dict
output = {
    'set': output_set,
    'set_mask': output_mask,
    'kls': kls
}

# Create ARGS dict for EMDCriterion
ARGS = {
    'args': args
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
metrics, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EMDCriterion — 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 EMDCriterion 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
InitialSet Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("ec91f5dff2cfdd95")
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, InitialSet: 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 math

torch.manual_seed(0)

class MockArgs:
    def __init__(self, train_gmm=True):
        self.train_gmm = train_gmm

ARGS = {
    'args': MockArgs(train_gmm=True),
    'dim_seed': 4,
    'n_mixtures': 3,
    'fixed_gmm': False,
    'train_gmm': True,
    'dim_out': 5,
    'max_outputs': 2
}
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 InitialSet — 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 InitialSet 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
MLP Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("56ed680b39cc62e1")
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, MLP: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'dim_in': 4,
    'dim_hidden': 8,
    'dim_out': 3,
    'n_hidden': 2,
    'activation': 'relu',
    'residual': False,
    'use_bn': False,
    'use_masked_bn': False,
    'dropout_p': 0.1
}
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 MLP — 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 MLP 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
MaskedBatchNorm1d Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("6ccac28dde2e0594")
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, MaskedBatchNorm1d: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create valid arguments for MaskedBatchNorm1d
# The module expects:
# - num_features: int (number of channels C)
# - eps: float (default 1e-5)
# - momentum: float (default 0.1)
# - affine: bool (default True)
# - track_running_stats: bool (default True)

ARGS = {
    'num_features': 4,
    'eps': 1e-5,
    'momentum': 0.1,
    'affine': True,
    'track_running_stats': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MaskedBatchNorm1d — 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 MaskedBatchNorm1d 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
ResidualAttention Ran
ran on a synthesized input · constructed an object; no result was computed
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("d36dbf566a4f76e4")
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, ResidualAttention: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Construct minimal valid arguments for ResidualAttention
ARGS = {
    'dim_q': 4,
    'dim_k': 5,
    'dim_v': 8,
    'num_heads': 2,
    'ln': True,
    'dropout_p': 0.1,
    'slot_att': False,
}

# Test the module can be instantiated and called
if __name__ == "__main__":
    model = nn.Module.__new__(type('ResidualAttention', (nn.Module,), {}))
    
    # Create instance properly
    from torch.nn import Module, Linear, LayerNorm, Dropout
    import math
    import torch.nn.functional as F
    
    class TestResidualAttention(Module):
        def __init__(self, dim_q, dim_k, dim_v, num_heads, ln=False, dropout_p=0.0, slot_att=False):
            super().__init__()
            self.dim_v = dim_v
            self.num_heads = num_heads
            self.dim_split = dim_v // num_heads if dim_v >= num_heads else 1
            self.dropout_p = dropout_p
            self.slot_att = slot_att
            self.fc_q = Linear(dim_q, dim_v, bias=False)
            self.fc_k = Linear(dim_k, dim_v, bias=False)
            self.fc_v = Linear(dim_k, dim_v, bias=False)
            self.fc_o = Linear(dim_v, dim_q, bias=False)
            self.ffn1 = Linear(dim_q, dim_q)
            self.ffn2 = Linear(dim_q, dim_q)
            if ln:
                self.ln_x = LayerNorm(dim_q)
                self.ln_y = LayerNorm(dim_k)
                self.ln_o1 = LayerNorm(dim_q)
                self.ln_o2 = LayerNorm(dim_q)
            if dropout_p > 0:
                self.dropout = Dropout(p=dropout_p)
    
    model = TestResidualAttention(**ARGS)
    
    # Create minimal test inputs
    batch_size = 2
    n_x = 3
    n_y = 4
    
    x = torch.randn(batch_size, n_x, ARGS['dim_q'])
    y = torch.randn(batch_size, n_y, ARGS['dim_k'])
    x_mask = torch.zeros(batch_size, n_x, dtype=torch.bool)
    y_mask = torch.zeros(batch_size, n_y, dtype=torch.bool)
    
    # Verify it can be called
    output = model(x, y, x_mask, y_mask, get_alpha=False, given_alpha=None)
    print(f"Output shape: {output.shape}")
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 ResidualAttention — 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 ResidualAttention 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
_chamfer_loss_v1 Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("c4401e3045759b3c")
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 0, dtype family float, 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
import torch.nn as nn

torch.manual_seed(0)

# Analyze the function:
# - output_set: tensor, flattened to (batch*n, feature_dim), indexed by ~output_mask
# - output_mask: boolean tensor, flattened, used to filter output_set
# - target_set: tensor, same structure as output_set
# - target_mask: boolean tensor, same structure as output_mask
# - sizes: computed from (~output_mask).long().sum(dim=1), used to split tensors
# The function expects batch_size=2 based on dim=1 operations
# output_set and target_set should have shape (batch_size, max_points, feature_dim)
# output_mask and target_mask should have shape (batch_size, max_points)

batch_size = 2
max_points = 5
feature_dim = 4

# Create output_set: (batch_size, max_points, feature_dim)
output_set = torch.randn(batch_size, max_points, feature_dim)

# Create output_mask: (batch_size, max_points)
# False means "keep this point", True means "mask/ignore"
# For batch 0: keep 3 points, for batch 1: keep 2 points
output_mask = torch.tensor([
    [False, False, False, True, True],
    [False, False, True, True, True]
])

# Create target_set: (batch_size, max_points, feature_dim)
target_set = torch.randn(batch_size, max_points, feature_dim)

# Create target_mask: (batch_size, max_points)
# Should match output_mask structure for valid split
target_mask = torch.tensor([
    [False, False, False, True, True],
    [False, False, True, True, True]
])

ARGS = {
    "output_set": output_set,
    "output_mask": output_mask,
    "target_set": target_set,
    "target_mask": target_mask
}
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 0, dtype family float, 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
chamfer_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("42c489cac0ddf783")
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 0, dtype family float, 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
import torch.nn as nn

torch.manual_seed(0)

# Analyze the function:
# - output_set and target_set are tensors that get flattened(0,1), indexed, and split
# - output_mask and target_mask are BoolTensors used for masking
# - The function flattens these to (batch*seq, features) and (batch*seq,)
# - It splits by sizes computed from ~output_mask
# - accelerate parameter is a bool

# The key insight from the code:
# - output_set.flatten(0, 1) and target_set.flatten(0, 1) suggest shape (batch, seq, features)
# - output_mask and target_mask are (batch, seq) boolean masks
# - (~output_mask).long().sum(dim=1) computes valid count per batch item
# - The masking removes padded positions (where mask is True)

# Build small tensors:
batch_size = 2
seq_len = 5
feature_dim = 3

# Create output_set: (batch, seq, features)
output_set = torch.randn(batch_size, seq_len, feature_dim)

# Create output_mask: (batch, seq) - True where invalid/padded
# Make first item have 3 valid points, second have 3 valid points
output_mask = torch.tensor([
    [False, False, False, True, True],   # 3 valid points
    [False, False, False, True, True]    # 3 valid points
], dtype=torch.bool)

# Create target_set: (batch, seq, features)
target_set = torch.randn(batch_size, seq_len, feature_dim)

# Create target_mask: (batch, seq) - must have same valid count per batch item
target_mask = torch.tensor([
    [False, False, False, True, True],   # 3 valid points
    [False, False, False, True, True]    # 3 valid points
], dtype=torch.bool)

# accelerate parameter
accelerate = False

ARGS = {
    'output_set': output_set,
    'output_mask': output_mask,
    'target_set': target_set,
    'target_mask': target_mask,
    'accelerate': accelerate
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
metrics, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, 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
cube_to_sphere Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("46b4d8dba0db14c3")
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 generator. 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.

from math import cos, gamma, pi, sin, sqrt
from typing import Callable, Iterator, List, Optional, Sequence, Tuple

# Create a small set of points in [0, 1]^dim
# cube_to_sphere maps from [0,1]^dim to a dim+1 dimensional sphere
# Let's use 3D cube points (which will map to 4D sphere points)

cube_points = [
    [0.1, 0.2, 0.3],
    [0.5, 0.6, 0.7],
]

ARGS = {
    "cube": cube_points,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type generator — a generator output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a generator 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
get_mask Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("4e612cecc9647edc")
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, 7) 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)

# sizes: Tensor([B,]) where B is batch size
# max_size: int representing the maximum sequence length
sizes = torch.tensor([3, 5])  # batch size 2, with different sizes
max_size = 7

ARGS = {
    "sizes": sizes,
    "max_size": max_size,
}
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, 7) — shape (2, 7) 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_module Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("f5ca3e9834167675")
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 SimpleModel. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

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

# Test case 1: module with 'module' attribute (DataParallel wrapped)
model_with_module = nn.Module()
model_with_module.module = SimpleModel()

# Test case 2: module without 'module' attribute
model_without_module = SimpleModel()

# For this function, we need to test both cases
# Let's use the model with 'module' attribute as the main test
ARGS = {
    'gpu': 0,
    'module': model_with_module
}
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 SimpleModel — a SimpleModel output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a SimpleModel 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
get_pairwise_distance Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("325df0adea9a09ec")
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, 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)

# Create a small tensor with shape (N=2, d=4)
# N is the number of points, d is the dimensionality
x = torch.randn(2, 4)

ARGS = {
    "x": x,
    "p": 2,
}
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, 2) — shape (2, 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
int_sin_m Ran
ran, honoured the contract we drafted
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("0c55dfd2474acb2e")
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 float, ndim 0, dtype family float, 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.

from math import cos, gamma, pi, sin, sqrt

# The function int_sin_m takes two parameters:
# - x: a float representing the upper limit of integration
# - m: an int representing the power of sine

# Let's construct valid arguments by examining the function:
# The function computes the integral of sin^m(t) dt from 0 to x
# It uses recursion and handles base cases for m=0 and m=1
# For m >= 2, it uses the recursive formula

# Valid test case: compute integral of sin^3(t) from 0 to pi/2
x_value = pi / 2  # A reasonable float value (approximately 1.57)
m_value = 3       # A positive integer for the power of sine

ARGS = {
    "x": x_value,
    "m": m_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, 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
inverse_increasing Ran
ran, honoured the contract we drafted
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("6269a6d083ef6af9")
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 float, ndim 0, dtype family float, 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
import numpy as np

torch.manual_seed(0)

# The function requires:
# - func: Callable[[float], float] - a function that takes a float and returns a float
# - target: float - the target value to find the inverse for
# - lower: float - lower bound of search interval
# - upper: float - upper bound of search interval
# - atol: float - absolute tolerance (has default)

# Create a simple monotonically increasing function for testing
# f(x) = 2*x + 1 is monotonically increasing
def monotonic_func(x: float) -> float:
    return 2.0 * x + 1.0

# If we want to find the inverse of target=5 with this function:
# 2*x + 1 = 5 => x = 2
target_value = 5.0
lower_bound = 0.0
upper_bound = 10.0
tolerance = 1e-10

ARGS = {
    "func": monotonic_func,
    "target": target_value,
    "lower": lower_bound,
    "upper": upper_bound,
    "atol": tolerance
}
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 float, ndim 0, dtype family float, 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
masked_fill Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("1d25c9b2a1df8aa1")
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 3, dtype family float, shape (2, 3, 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)

# Construct arguments for masked_fill function
# tensor_bnc: a 3D tensor with shape (batch, n, channels)
tensor_bnc = torch.randn(2, 3, 4)

# mask_bn: a 2D boolean mask with shape (batch, n)
mask_bn = torch.tensor([[True, False, True], [False, True, False]])

# value: float value to fill masked positions with
value = -1.0

ARGS = {
    "tensor_bnc": tensor_bnc,
    "mask_bn": mask_bn,
    "value": 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 Tensor, ndim 3, dtype family float, shape (2, 3, 4) — shape (2, 3, 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
sample_mask Ran
ran; our contract draft was wrong, not the code
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("f3ac4b1a18df2e8f")
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, 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)

# sample_sizes: Tensor([B,]) - a 1D tensor of batch size B containing sample sizes
# max_size: int - the maximum size for the mask
sample_sizes = torch.tensor([2, 3], dtype=torch.long)
max_size = 4

ARGS = {
    "sample_sizes": sample_sizes,
    "max_size": max_size,
}
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, 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
AttentiveBlock Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("e76e125691f86272")
How it ran

no invocation recorded.

DecoderBlock Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("87658bdfa597bd52")
How it ran

no invocation recorded.

DeterministicNetwork Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("508dac2dcd3acee5")
How it ran

no invocation recorded.

ElementwiseMLP Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("3b42633e7587c43c")
How it ran

no invocation recorded.

EncoderBlock Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("6d5bb50995b393f2")
How it ran

no invocation recorded.

ISAB Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("380e877aad19e438")
How it ran

no invocation recorded.

InducedNetwork Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("0c00bb082460dcd5")
How it ran

no invocation recorded.

SetVAE Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("195dea2fed61451a")
How it ran

no invocation recorded.

emd_loss Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("06bb79fd238d6e69")
How it ran

no invocation recorded.

sphere_lattice Not yet run
unverified (harvested, no recorded run)
jw9730/setvae/models/networks.py
code served (permissive licence) · get_code("d226bf1b07b03f2e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 27 samples lifted from this paper's linked repositories, 27 carry a function identity and 0 do not. Of the 27 distinct functions with an identity, 2 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 2 later) and 25 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
int_sin_m
brick of this paper's assembly: helper
1 other paper
1 later
2206.12314 (2022-06)
Learning sparse features can lead to overfitting in neural networks
inverse_increasing
brick of this paper's assembly: helper
1 other paper
1 later
2206.12314 (2022-06)
Learning sparse features can lead to overfitting in neural networks

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

Generative modeling of set-structured data, such as point clouds, requires reasoning over local and global structures at various scales. However, adopting multi-scale frameworks for ordinary sequential data to a set-structured data is nontrivial as it should be invariant to the permutation of its elements. In this paper, we propose SetVAE, a hierarchical variational autoencoder for sets. Motivated by recent progress in set encoding, we build SetVAE upon attentive modules that first partition the set and project the partition back to the original cardinality. Exploiting this module, our hierarchical VAE learns latent variables at multiple scales, capturing coarse-to-fine dependency of the set elements while achieving permutation invariance. We evaluate our model on point cloud generation task and achieve competitive performance to the prior arts with substantially smaller model capacity. We qualitatively demonstrate that our model generalizes to unseen set sizes and learns interesting subset relations without supervision. Our implementation is available at https://github.com/ jw9730/setvae.

For agents

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

get_harvested_code_for_paper("2103.15619")
get_code_for_paper("2103.15619")
have("2103.15619")

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

curl https://syntology.ai/api/ran/2103.15619.json

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

[![Syntology run record](https://syntology.ai/api/ran/2103.15619.svg)](https://syntology.ai/paper/2103.15619)

Connect an agent — have() is free.