SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2102.03150 · ICML · 2021

Equivariant message passing for the prediction of tensorial properties and molecular spectra

Kristof Schütt, Oliver Unke, Michael Gastegger

arXiv · PDF · Open in the Atlas

Code that ran

22 of 33 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; 2 honoured the contract we drafted, 0 violated it and 11 have no contract verdict; 9 failures of our instrument; 11 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/2102.03150.json · badge: /api/ran/2102.03150.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
learningmatter-mit/NeuralForceField community
Papers with Code archive
22 ran · 11 no recorded run
FunctionStatusWhere it lives
CosineEnvelope Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("8ac6bce450fc85e3")
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 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
from torch import nn
import numpy as np

from torch import nn
import numpy as np
import torch

class CosineEnvelope(nn.Module):

    def __init__(self, cutoff):
        super().__init__()
        self.cutoff = cutoff

    def forward(self, d):
        output = 0.5 * (torch.cos(np.pi * d / self.cutoff) + 1)
        exclude = d >= self.cutoff
        output[exclude] = 0
        return output


INIT_ARGS = {
    'cutoff': 5.0
}

ARGS = {
    'd': torch.tensor([[0.5, 1.0, 2.5, 4.0], [0.1, 3.0, 5.5, 6.0]], dtype=torch.float32)
}

if __name__ == '__main__':
    envelope = CosineEnvelope(**INIT_ARGS)
    result = envelope(**ARGS)
    print(result)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Dense Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("dc349a5755ad2883")
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 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
from functools import partial
from torch.nn.init import constant_, xavier_uniform_

DEFAULT_DROPOUT_RATE = 0.0

zeros_initializer = partial(constant_, val=0.0)

class Dense(nn.Linear):
    """Applies a dense layer with activation: :math:`y = activation(Wx + b)`

    Args:
        in_features (int): number of input feature
        out_features (int): number of output features
        bias (bool): If set to False, the layer will not adapt the bias. (default: True)
        activation (callable): activation function (default: None)
        weight_init (callable): function that takes weight tensor and initializes (default: xavier)
        bias_init (callable): function that takes bias tensor and initializes (default: zeros initializer)
    """

    def __init__(self, in_features, out_features, bias=True, activation=None, dropout_rate=DEFAULT_DROPOUT_RATE, weight_init=xavier_uniform_, bias_init=zeros_initializer):
        self.weight_init = weight_init
        self.bias_init = bias_init
        super().__init__(in_features, out_features, bias)
        self.activation = activation
        self.dropout = nn.Dropout(p=dropout_rate)

    def reset_parameters(self):
        """
        Reinitialize model parameters.
        """
        self.weight_init(self.weight)
        if self.bias is not None:
            self.bias_init(self.bias)

    def forward(self, inputs):
        """
        Args:
            inputs (dict of torch.Tensor): SchNetPack format dictionary of input tensors.

        Returns:
            torch.Tensor: Output of the dense layer.
        """
        self.to(inputs.device)
        y = super().forward(inputs)
        if hasattr(self, 'dropout'):
            y = self.dropout(y)
        if self.activation:
            y = self.activation(y)
        return y


INIT_ARGS = {
    'in_features': 4,
    'out_features': 3,
    'bias': True,
    'activation': nn.ReLU(),
    'dropout_rate': 0.1,
    'weight_init': xavier_uniform_,
    'bias_init': zeros_initializer,
}

ARGS = {
    'inputs': torch.randn(2, 4),
}

if __name__ == '__main__':
    layer = Dense(**INIT_ARGS)
    output = layer(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
DistanceEmbed Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("6af9e88a6c54be01")
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 2, dtype family float, shape (6, 12) 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
import numpy as np
from functools import partial
from torch.nn.init import constant_, xavier_uniform_

from torch import nn
import numpy as np
import torch
from functools import partial
from torch.nn.init import constant_, xavier_uniform_

class CosineEnvelope(nn.Module):

    def __init__(self, cutoff):
        super().__init__()
        self.cutoff = cutoff

    def forward(self, d):
        output = 0.5 * (torch.cos(np.pi * d / self.cutoff) + 1)
        exclude = d >= self.cutoff
        output[exclude] = 0
        return output

DEFAULT_DROPOUT_RATE = 0.0

class PainnRadialBasis(nn.Module):

    def __init__(self, n_rbf, cutoff, learnable_k):
        super().__init__()
        self.n = torch.arange(1, n_rbf + 1).float()
        if learnable_k:
            self.n = nn.Parameter(self.n)
        self.cutoff = cutoff

    def forward(self, dist):
        """
        Args:
            d (torch.Tensor): tensor of distances
        """
        shape_d = dist.unsqueeze(-1)
        n = self.n.to(dist.device)
        coef = n * np.pi / self.cutoff
        device = shape_d.device
        denom = torch.where(shape_d == 0, torch.tensor(1.0, device=device), shape_d)
        num = torch.where(shape_d == 0, coef, torch.sin(coef * shape_d))
        output = torch.where(shape_d >= self.cutoff, torch.tensor(0.0, device=device), num / denom)
        return output

zeros_initializer = partial(constant_, val=0.0)

class Dense(nn.Linear):
    """Applies a dense layer with activation: :math:`y = activation(Wx + b)`

    Args:
        in_features (int): number of input feature
        out_features (int): number of output features
        bias (bool): If set to False, the layer will not adapt the bias. (default: True)
        activation (callable): activation function (default: None)
        weight_init (callable): function that takes weight tensor and initializes (default: xavier)
        bias_init (callable): function that takes bias tensor and initializes (default: zeros initializer)
    """

    def __init__(self, in_features, out_features, bias=True, activation=None, dropout_rate=DEFAULT_DROPOUT_RATE, weight_init=xavier_uniform_, bias_init=zeros_initializer):
        self.weight_init = weight_init
        self.bias_init = bias_init
        super().__init__(in_features, out_features, bias)
        self.activation = activation
        self.dropout = nn.Dropout(p=dropout_rate)

    def reset_parameters(self):
        """
        Reinitialize model parameters.
        """
        self.weight_init(self.weight)
        if self.bias is not None:
            self.bias_init(self.bias)

    def forward(self, inputs):
        """
        Args:
            inputs (dict of torch.Tensor): SchNetPack format dictionary of input tensors.

        Returns:
            torch.Tensor: Output of the dense layer.
        """
        self.to(inputs.device)
        y = super().forward(inputs)
        if hasattr(self, 'dropout'):
            y = self.dropout(y)
        if self.activation:
            y = self.activation(y)
        return y

class DistanceEmbed(nn.Module):

    def __init__(self, n_rbf, cutoff, feat_dim, learnable_k, dropout):
        super().__init__()
        rbf = PainnRadialBasis(n_rbf=n_rbf, cutoff=cutoff, learnable_k=learnable_k)
        dense = Dense(in_features=n_rbf, out_features=3 * feat_dim, bias=True, dropout_rate=dropout)
        self.block = nn.Sequential(rbf, dense)
        self.f_cut = CosineEnvelope(cutoff=cutoff)

    def forward(self, dist):
        rbf_feats = self.block(dist)
        envelope = self.f_cut(dist).reshape(-1, 1)
        output = rbf_feats * envelope
        return output


INIT_ARGS = {
    'n_rbf': 8,
    'cutoff': 5.0,
    'feat_dim': 4,
    'learnable_k': False,
    'dropout': 0.1,
}

ARGS = {
    'dist': torch.tensor([0.5, 1.5, 2.5, 3.5, 4.5, 5.5], dtype=torch.float32),
}

if __name__ == '__main__':
    model = DistanceEmbed(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, 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 (6, 12) — shape (6, 12) 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)
EmbeddingBlock Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("952ca748b3f1516e")
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 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

class EmbeddingBlock(nn.Module):

    def __init__(self, feat_dim):
        super().__init__()
        self.atom_embed = nn.Embedding(100, feat_dim, padding_idx=0)
        self.feat_dim = feat_dim

    def forward(self, z_number, **kwargs):
        num_atoms = z_number.shape[0]
        s_i = self.atom_embed(z_number)
        v_i = torch.zeros(num_atoms, self.feat_dim, 3).to(s_i.device)
        return (s_i, v_i)


INIT_ARGS = {
    'feat_dim': 4
}

ARGS = {
    'z_number': torch.tensor([1, 5, 8, 12], dtype=torch.long)
}

if __name__ == '__main__':
    model = EmbeddingBlock(**INIT_ARGS)
    output = model(**ARGS)
    print(f"s_i shape: {output[0].shape}")
    print(f"v_i shape: {output[1].shape}")
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
InvariantDense Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("7f75d01bd99cc6dd")
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 2, dtype family float, shape (2, 12) 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
import numpy as np
import torch.nn.functional as F
from functools import partial
from torch.nn.init import constant_, xavier_uniform_

DEFAULT_DROPOUT_RATE = 0.0

class LearnableSwish(torch.nn.Module):

    def __init__(self, alpha=1.0, beta=1.702):
        super().__init__()
        self.alpha_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([alpha])) - 1))
        self.beta_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([beta])) - 1))

    @property
    def alpha(self):
        return F.softplus(self.alpha_inv)

    @property
    def beta(self):
        return F.softplus(self.beta_inv)

    def forward(self, x):
        device = x.device
        alpha = self.alpha.to(device)
        beta = self.beta.to(device)
        output = alpha / beta * F.silu(beta * x)
        return output

class Swish(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x):
        return F.silu(x)

class shifted_softplus(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, input):
        return F.softplus(input) - np.log(2.0)

zeros_initializer = partial(constant_, val=0.0)

class Dense(nn.Linear):

    def __init__(self, in_features, out_features, bias=True, activation=None, dropout_rate=DEFAULT_DROPOUT_RATE, weight_init=xavier_uniform_, bias_init=zeros_initializer):
        self.weight_init = weight_init
        self.bias_init = bias_init
        super().__init__(in_features, out_features, bias)
        self.activation = activation
        self.dropout = nn.Dropout(p=dropout_rate)

    def reset_parameters(self):
        self.weight_init(self.weight)
        if self.bias is not None:
            self.bias_init(self.bias)

    def forward(self, inputs):
        self.to(inputs.device)
        y = super().forward(inputs)
        if hasattr(self, 'dropout'):
            y = self.dropout(y)
        if self.activation:
            y = self.activation(y)
        return y

layer_types = {'linear': torch.nn.Linear, 'Tanh': torch.nn.Tanh, 'ReLU': torch.nn.ReLU, 'Dense': Dense, 'shifted_softplus': shifted_softplus, 'sigmoid': torch.nn.Sigmoid, 'Dropout': torch.nn.Dropout, 'LeakyReLU': torch.nn.LeakyReLU, 'ELU': torch.nn.ELU, 'swish': Swish, 'learnable_swish': LearnableSwish, 'softplus': torch.nn.Softplus}

def to_module(activation):
    return layer_types[activation]()

class InvariantDense(nn.Module):

    def __init__(self, dim, dropout, activation='swish'):
        super().__init__()
        self.layers = nn.Sequential(Dense(in_features=dim, out_features=dim, bias=True, dropout_rate=dropout, activation=to_module(activation)), Dense(in_features=dim, out_features=3 * dim, bias=True, dropout_rate=dropout))

    def forward(self, s_j):
        output = self.layers(s_j)
        return output


INIT_ARGS = {
    'dim': 4,
    'dropout': 0.1,
    'activation': 'swish'
}

ARGS = {
    's_j': torch.randn(2, 4)
}

if __name__ == '__main__':
    model = InvariantDense(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: (2, 12)")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, 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, 12) — shape (2, 12) 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)
LearnableSwish Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("0d21a7509e419d0b")
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 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
import torch.nn.functional as F
from torch import nn

class LearnableSwish(torch.nn.Module):

    def __init__(self, alpha=1.0, beta=1.702):
        super().__init__()
        self.alpha_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([alpha])) - 1))
        self.beta_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([beta])) - 1))

    @property
    def alpha(self):
        return F.softplus(self.alpha_inv)

    @property
    def beta(self):
        return F.softplus(self.beta_inv)

    def forward(self, x):
        device = x.device
        alpha = self.alpha.to(device)
        beta = self.beta.to(device)
        output = alpha / beta * F.silu(beta * x)
        return output


INIT_ARGS = {
    'alpha': 1.0,
    'beta': 1.702
}

ARGS = {
    'x': torch.randn(2, 4)
}

if __name__ == '__main__':
    module = LearnableSwish(**INIT_ARGS)
    output = module(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
MolFpPool Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("398bfbf43b702ff8")
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 dict. 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
from torch import nn
from torch.nn import LeakyReLU, Linear, ModuleDict, ReLU, Sequential, Softmax
from torch.autograd import grad
from functools import partial
from torch.nn.init import constant_, xavier_uniform_
import torch.nn.functional as F

DEFAULT_DROPOUT_RATE = 0.0

class LearnableSwish(torch.nn.Module):

    def __init__(self, alpha=1.0, beta=1.702):
        super().__init__()
        self.alpha_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([alpha])) - 1))
        self.beta_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([beta])) - 1))

    @property
    def alpha(self):
        return F.softplus(self.alpha_inv)

    @property
    def beta(self):
        return F.softplus(self.beta_inv)

    def forward(self, x):
        device = x.device
        alpha = self.alpha.to(device)
        beta = self.beta.to(device)
        output = alpha / beta * F.silu(beta * x)
        return output

class Swish(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x):
        return F.silu(x)

def compute_grad(inputs, output, allow_unused=False):
    assert inputs.requires_grad
    gradspred, = grad(output, inputs, grad_outputs=output.data.new(output.shape).fill_(1), create_graph=True, retain_graph=True, allow_unused=allow_unused)
    return gradspred

class shifted_softplus(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, input):
        return F.softplus(input) - np.log(2.0)

zeros_initializer = partial(constant_, val=0.0)

class Dense(nn.Linear):

    def __init__(self, in_features, out_features, bias=True, activation=None, dropout_rate=DEFAULT_DROPOUT_RATE, weight_init=xavier_uniform_, bias_init=zeros_initializer):
        self.weight_init = weight_init
        self.bias_init = bias_init
        super().__init__(in_features, out_features, bias)
        self.activation = activation
        self.dropout = nn.Dropout(p=dropout_rate)

    def reset_parameters(self):
        self.weight_init(self.weight)
        if self.bias is not None:
            self.bias_init(self.bias)

    def forward(self, inputs):
        self.to(inputs.device)
        y = super().forward(inputs)
        if hasattr(self, 'dropout'):
            y = self.dropout(y)
        if self.activation:
            y = self.activation(y)
        return y

layer_types = {'linear': torch.nn.Linear, 'Tanh': torch.nn.Tanh, 'ReLU': torch.nn.ReLU, 'Dense': Dense, 'shifted_softplus': shifted_softplus, 'sigmoid': torch.nn.Sigmoid, 'Dropout': torch.nn.Dropout, 'LeakyReLU': torch.nn.LeakyReLU, 'ELU': torch.nn.ELU, 'swish': Swish, 'learnable_swish': LearnableSwish, 'softplus': torch.nn.Softplus}

class MolFpPool(nn.Module):

    def __init__(self, feat_dim, mol_fp_act, num_out_layers, out_dim, **kwargs):
        super().__init__()
        feat_num = [int(feat_dim / num_out_layers ** m) for m in range(num_out_layers)]
        mol_fp_layers = [Dense(in_features=feat_num[i], out_features=feat_num[i + 1], activation=layer_types[mol_fp_act]()) for i in range(num_out_layers - 1)]
        mol_fp_layers.append(Dense(in_features=feat_num[-1], out_features=out_dim, activation=None))
        self.mol_fp_nn = Sequential(*mol_fp_layers)

    def forward(self, batch, xyz, atomwise_output, grad_keys, out_keys):
        N = batch['num_atoms'].detach().cpu().tolist()
        results = {}
        for key in out_keys:
            batched_feats = atomwise_output['features']
            split_feats = torch.split(batched_feats, N)
            all_outputs = []
            learned_feats = []
            for feats in split_feats:
                mol_fp = feats.sum(0)
                output = self.mol_fp_nn(mol_fp)
                all_outputs.append(output)
            results[key] = torch.stack(all_outputs)
        return results

INIT_ARGS = {
    'feat_dim': 16,
    'mol_fp_act': 'ReLU',
    'num_out_layers': 2,
    'out_dim': 5,
}

ARGS = {
    'batch': {'num_atoms': torch.tensor([3, 2])},
    'xyz': torch.randn(5, 3),
    'atomwise_output': {'features': torch.randn(5, 16)},
    'grad_keys': [],
    'out_keys': ['energy'],
}

if __name__ == '__main__':
    model = MolFpPool(**INIT_ARGS)
    result = model(**ARGS)
    print("Result keys:", result.keys())
    print("Result shape:", result['energy'].shape)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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)
PainnRadialBasis Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("b65822c621376024")
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 2, dtype family float, shape (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
from torch import nn
import numpy as np

from torch import nn
import numpy as np
import torch

class PainnRadialBasis(nn.Module):

    def __init__(self, n_rbf, cutoff, learnable_k):
        super().__init__()
        self.n = torch.arange(1, n_rbf + 1).float()
        if learnable_k:
            self.n = nn.Parameter(self.n)
        self.cutoff = cutoff

    def forward(self, dist):
        """
        Args:
            d (torch.Tensor): tensor of distances
        """
        shape_d = dist.unsqueeze(-1)
        n = self.n.to(dist.device)
        coef = n * np.pi / self.cutoff
        device = shape_d.device
        denom = torch.where(shape_d == 0, torch.tensor(1.0, device=device), shape_d)
        num = torch.where(shape_d == 0, coef, torch.sin(coef * shape_d))
        output = torch.where(shape_d >= self.cutoff, torch.tensor(0.0, device=device), num / denom)
        return output


INIT_ARGS = {
    'n_rbf': 4,
    'cutoff': 5.0,
    'learnable_k': False,
}

ARGS = {
    'dist': torch.tensor([0.5, 1.0, 2.0, 3.5]),
}

if __name__ == '__main__':
    model = PainnRadialBasis(**INIT_ARGS)
    output = model(**ARGS)
    print(output)
    print(output.shape)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, 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 (4, 4) — shape (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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ScaleShift Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("802d844e902371ef")
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 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
from torch import nn

class ScaleShift(nn.Module):
    """Scale and shift layer for standardization.
    .. math::
       y = x \\times \\sigma + \\mu
    Args:
        means (dict): dictionary of mean values
        stddevs (dict): dictionary of standard deviations
    """

    def __init__(self, means=None, stddevs=None):
        super().__init__()
        means = means if means is not None else {}
        stddevs = stddevs if stddevs is not None else {}
        self.means = means
        self.stddevs = stddevs

    def forward(self, inp, key):
        """Compute layer output.
        Args:
            inp (torch.Tensor): input data.
            key (str): key to look up mean and stddev.
        Returns:
            torch.Tensor: layer output.
        """
        stddev = self.stddevs.get(key, 1.0)
        mean = self.means.get(key, 0.0)
        out = inp * stddev + mean
        return out


INIT_ARGS = {
    "means": {"feature_a": 2.5, "feature_b": -1.0},
    "stddevs": {"feature_a": 0.5, "feature_b": 2.0}
}

ARGS = {
    "inp": torch.randn(2, 4),
    "key": "feature_a"
}
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Swish Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("7f836403bedf830d")
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 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
import torch.nn.functional as F

class Swish(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x):
        return F.silu(x)


INIT_ARGS = {}

ARGS = {
    'x': torch.randn(2, 4)
}

if __name__ == '__main__':
    model = Swish(**INIT_ARGS)
    output = model(**ARGS)
    print(output)
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
add_embedding Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("5cfab56ad64af6fe")
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 dict. 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)

# The function expects:
# - atomwise_out: a dict with a 'features' key containing a tensor
# - all_results: a dict that will be modified to include 'embedding'

# Create a small features tensor (batch_size=2, feature_dim=4)
features_tensor = torch.randn(2, 4)

# Create atomwise_out dict with 'features' key
atomwise_out = {
    'features': features_tensor
}

# Create all_results dict (can be empty or have other keys)
all_results = {
    'other_key': 'other_value'
}

ARGS = {
    'atomwise_out': atomwise_out,
    'all_results': all_results
}
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 dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
att_readout_probs Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("2eece0db59be49c0")
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 function. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn.functional as F

torch.manual_seed(0)

# The function att_readout_probs takes a single parameter 'name'
# which should be a string that is either 'softmax' or 'square'
# Let's test with 'softmax'

ARGS = {
    'name': 'softmax'
}
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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_grad Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("2ac65d5ac1bf4b38")
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 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
from torch.autograd import grad

torch.manual_seed(0)

# Create a small tensor that requires gradients
inputs = torch.randn(2, 4, requires_grad=True)

# Create a scalar output by reducing a computation graph
# We need output to be a scalar for the grad function to work properly
intermediate = inputs.sum(dim=1)  # shape (2,)
output = intermediate.mean()  # scalar

# allow_unused parameter
allow_unused = False

ARGS = {
    "inputs": inputs,
    "output": output,
    "allow_unused": allow_unused,
}
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 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
gen Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("07fc732b54bba7cd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from itertools import repeat

torch.manual_seed(0)

# Create a small source tensor (batch_size=2, feature_dim=4)
src = torch.randn(2, 4)

# Create an index tensor that will be expanded
# This is a 1D tensor with values indicating target positions
index = torch.tensor([0, 1, 0, 2])

# Optional: dim parameter (default -1, which means last dimension)
dim = -1

# Optional: out tensor (if None, will be created in the function)
out = None

# Optional: dim_size (if None, will be inferred from index.max())
dim_size = None

# Optional: fill_value for initializing the output tensor
fill_value = 0

ARGS = {
    "src": src,
    "index": index,
    "dim": dim,
    "out": out,
    "dim_size": dim_size,
    "fill_value": fill_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_offsets Ran
ran, honoured the contract we drafted
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("18fd70ccb86319a7")
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

torch.manual_seed(0)

# The function expects:
# 1. batch: a dict with 'nxyz' key (required) and optionally a key parameter
# 2. key: a string to look up in the batch dict

# Construct a batch dict with 'nxyz' tensor and optional offset tensors
batch = {
    'nxyz': torch.randn(2, 4),  # Small tensor on CPU
    'offsets_dense': torch.randn(2, 3),  # Dense offset tensor
    'offsets_sparse': torch.sparse_coo_tensor(
        indices=torch.tensor([[0, 1], [0, 1]]),
        values=torch.tensor([1.0, 2.0]),
        size=(2, 3)
    )  # Sparse offset tensor that can be converted to dense
}

key = 'offsets_dense'

ARGS = {
    'batch': batch,
    'key': key
}
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
get_rij Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("2cfb01eec4fba085")
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

torch.manual_seed(0)

# Create a batch dictionary with required 'nxyz' key
batch = {
    'nxyz': torch.randn(2, 4),  # Small tensor on CPU
    'offsets': torch.randn(5, 3)  # Small offsets tensor
}

# Create xyz coordinates: shape (num_atoms, 3)
xyz = torch.randn(6, 3)

# Create neighbor pairs: shape (num_pairs, 2)
# Each row is [atom_i, atom_j] indices
nbrs = torch.tensor([
    [0, 1],
    [1, 2],
    [2, 3],
    [3, 4],
    [4, 5]
], dtype=torch.long)

# Create cutoff as a scalar tensor
cutoff = torch.tensor(5.0)

ARGS = {
    'xyz': xyz,
    'batch': batch,
    'nbrs': nbrs,
    'cutoff': cutoff
}
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
make_directed Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("3e1aa28156810aba")
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 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

torch.manual_seed(0)

# nbr_list is a 2D tensor where each row contains two indices
# The function checks if there are pairs where first > second and pairs where second > first
# to determine if the graph is directed
nbr_list = torch.tensor([
    [0, 1],
    [2, 1]
], dtype=torch.long)

ARGS = {
    "nbr_list": nbr_list
}
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 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
norm Ran
ran, honoured the contract we drafted
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("3df02af32c33da5a")
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 1, dtype family float, shape (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
import numpy as np

torch.manual_seed(0)

# The function 'norm' takes a single parameter 'vec'
# It performs: (vec ** 2 + EPS).sum(-1) ** 0.5
# This means:
# - vec is squared element-wise
# - EPS is added
# - sum over the last dimension
# - square root is taken
# So vec should be a tensor with at least 1 dimension

# Create a small tensor with shape (2, 4) - batch of 2, dimension 4
vec = torch.randn(2, 4)

ARGS = {
    "vec": vec
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 1, dtype family float, shape (2,) — shape (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
preprocess_r Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("6565f4a40973667d")
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 tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a small tensor of interatomic vectors
# Shape: (n_nbrs, 3) where n_nbrs=2 (number of neighbors) and 3 is spatial dimension
r_ij = torch.randn(2, 3, dtype=torch.float32)

ARGS = {
    "r_ij": r_ij
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
scatter_add Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("04e4a832f331db07")
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
import torch.nn as nn

torch.manual_seed(0)

# Create test tensors
src = torch.randn(2, 4)  # Source tensor: batch_size=2, features=4
index = torch.tensor([0, 1, 0, 2])  # Index tensor: 1D, size=4

ARGS = {
    "src": src,
    "index": index,
    "dim": -1,
    "out": None,
    "dim_size": None,
    "fill_value": 0
}
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
shifted_softplus Ran
ran on a synthesized input
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("42c43fca053545ea")
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 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 numpy as np
import torch
import torch.nn.functional as F

class shifted_softplus(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, input):
        return F.softplus(input) - np.log(2.0)


INIT_ARGS = {}

ARGS = {
    'input': torch.randn(2, 4)
}

if __name__ == '__main__':
    instance = shifted_softplus(**INIT_ARGS)
    output = instance(**ARGS)
    print(output)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
to_module Ran
ran; our contract draft was wrong, not the code
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("a929e9679ef4134f")
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 Swish. 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
from torch import nn
from functools import partial
from torch.nn.init import constant_, xavier_uniform_
import torch.nn.functional as F

torch.manual_seed(0)

DEFAULT_DROPOUT_RATE = 0.0

class LearnableSwish(torch.nn.Module):

    def __init__(self, alpha=1.0, beta=1.702):
        super().__init__()
        self.alpha_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([alpha])) - 1))
        self.beta_inv = nn.Parameter(torch.log(torch.exp(torch.Tensor([beta])) - 1))

    @property
    def alpha(self):
        return F.softplus(self.alpha_inv)

    @property
    def beta(self):
        return F.softplus(self.beta_inv)

    def forward(self, x):
        device = x.device
        alpha = self.alpha.to(device)
        beta = self.beta.to(device)
        output = alpha / beta * F.silu(beta * x)
        return output

class Swish(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x):
        return F.silu(x)

class shifted_softplus(torch.nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, input):
        return F.softplus(input) - np.log(2.0)

zeros_initializer = partial(constant_, val=0.0)

class Dense(nn.Linear):
    """Applies a dense layer with activation: :math:`y = activation(Wx + b)`

    Args:
        in_features (int): number of input feature
        out_features (int): number of output features
        bias (bool): If set to False, the layer will not adapt the bias. (default: True)
        activation (callable): activation function (default: None)
        weight_init (callable): function that takes weight tensor and initializes (default: xavier)
        bias_init (callable): function that takes bias tensor and initializes (default: zeros initializer)
    """

    def __init__(self, in_features, out_features, bias=True, activation=None, dropout_rate=DEFAULT_DROPOUT_RATE, weight_init=xavier_uniform_, bias_init=zeros_initializer):
        self.weight_init = weight_init
        self.bias_init = bias_init
        super().__init__(in_features, out_features, bias)
        self.activation = activation
        self.dropout = nn.Dropout(p=dropout_rate)

    def reset_parameters(self):
        """
        Reinitialize model parameters.
        """
        self.weight_init(self.weight)
        if self.bias is not None:
            self.bias_init(self.bias)

    def forward(self, inputs):
        """
        Args:
            inputs (dict of torch.Tensor): SchNetPack format dictionary of input tensors.

        Returns:
            torch.Tensor: Output of the dense layer.
        """
        self.to(inputs.device)
        y = super().forward(inputs)
        if hasattr(self, 'dropout'):
            y = self.dropout(y)
        if self.activation:
            y = self.activation(y)
        return y

layer_types = {'linear': torch.nn.Linear, 'Tanh': torch.nn.Tanh, 'ReLU': torch.nn.ReLU, 'Dense': Dense, 'shifted_softplus': shifted_softplus, 'sigmoid': torch.nn.Sigmoid, 'Dropout': torch.nn.Dropout, 'LeakyReLU': torch.nn.LeakyReLU, 'ELU': torch.nn.ELU, 'swish': Swish, 'learnable_swish': LearnableSwish, 'softplus': torch.nn.Softplus}

def to_module(activation):
    return layer_types[activation]()

ARGS = {
    'activation': 'swish'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Swish — a Swish output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Swish 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
AttentionPool Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("55d8c70e377b4a86")
How it ran

no invocation recorded.

InvariantMessage Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("f09c070242efedbf")
How it ran

no invocation recorded.

MeanPool Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("670901a6573e1934")
How it ran

no invocation recorded.

MessageBase Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("573a7b800820d725")
How it ran

no invocation recorded.

MessageBlock Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("9f1980fc08094e16")
How it ran

no invocation recorded.

Painn Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("450568a63ba4bedb")
How it ran

no invocation recorded.

ReadoutBlock Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("05495644c9588ed2")
How it ran

no invocation recorded.

SumPool Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("4b3ac9cd9d28db18")
How it ran

no invocation recorded.

UpdateBlock Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("1418df20e13d6afe")
How it ran

no invocation recorded.

add_stress Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("d31febe9a5aa1b51")
How it ran

no invocation recorded.

sum_and_grad Not yet run
unverified (harvested, no recorded run)
learningmatter-mit/NeuralForceField/nff/nn/models/painn.py
code served (permissive licence) · get_code("5d9228f747be613d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 33 samples lifted from this paper's linked repositories, 33 carry a function identity and 0 do not. Of the 33 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 33 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

Message passing neural networks have become a method of choice for learning on graphs, in particular the prediction of chemical properties and the acceleration of molecular dynamics studies. While they readily scale to large training data sets, previous approaches have proven to be less data efficient than kernel methods. We identify limitations of invariant representations as a major reason and extend the message passing formulation to rotationally equivariant representations. On this basis, we propose the polarizable atom interaction neural network (PaiNN) and improve on common molecule benchmarks over previous networks, while reducing model size and inference time. We leverage the equivariant atomwise representations obtained by PaiNN for the prediction of tensorial properties. Finally, we apply this to the simulation of molecular spectra, achieving speedups of 4-5 orders of magnitude compared to the electronic structure reference.

For agents

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

get_harvested_code_for_paper("2102.03150")
get_code_for_paper("2102.03150")
have("2102.03150")

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

curl https://syntology.ai/api/ran/2102.03150.json

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

[![Syntology run record](https://syntology.ai/api/ran/2102.03150.svg)](https://syntology.ai/paper/2102.03150)

Connect an agent — have() is free.