SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2002.11896 · NeurIPS · 2020

Gradient Boosted Normalizing Flows

Robert Giaquinto, Arindam Banerjee

arXiv · PDF · Open in the Atlas

Code that ran

20 of 30 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 17 have no contract verdict; 3 failures of our instrument; 10 no recorded run), measured 2026-11-21; 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/2002.11896.json · badge: /api/ran/2002.11896.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
robert-giaquinto/gradient-boosted-normalizing-flows canonical (classifier label)
classifier basis: classifier_2025
20 ran · 10 no recorded run
FunctionStatusWhere it lives
ActNorm1d Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("7315468727393ef4")
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
import torch.nn as nn

class _ActNorm(nn.Module):
    """
    Activation Normalization
    Initialize the bias and scale with a given minibatch,
    so that the output per-channel have zero mean and unit variance for that.
    After initialization, `bias` and `logs` will be trained as parameters.
    """

    def __init__(self, num_features, image_input, scale=1.0):
        super().__init__()
        param_size = [1, num_features, 1, 1] if image_input else [1, num_features]
        self.bias = nn.Parameter(torch.zeros(*param_size))
        self.logs = nn.Parameter(torch.zeros(*param_size))
        self.num_features = num_features
        self.scale = scale
        self.inited = False
        self.image_input = image_input

    def initialize_parameters(self, sample):
        if not self.training:
            raise ValueError('In Eval mode, but ActNorm not initiated')
        with torch.no_grad():
            reduce_dim = [0, 2, 3] if self.image_input else 0
            bias = -torch.mean(sample.clone(), dim=reduce_dim, keepdim=True)
            var = torch.mean((sample.clone() + bias) ** 2, dim=reduce_dim, keepdim=True)
            logs = torch.log(self.scale / (torch.sqrt(var) + 1e-06))
            self.bias.data.copy_(bias.data)
            self.logs.data.copy_(logs.data)
            self.inited = True

    def _center(self, sample, reverse=False):
        if reverse:
            return sample - self.bias
        else:
            return sample + self.bias

    def _scale(self, sample, logdet=None, reverse=False):
        if reverse:
            sample = sample * torch.exp(-self.logs)
        else:
            sample = sample * torch.exp(self.logs)
        if logdet is not None:
            '\n            logs is log_std of `mean of channels`\n            so we need to multiply by number of pixels\n            '
            if self.image_input:
                b, c, h, w = sample.shape
                dlogdet = torch.sum(self.logs) * h * w
            else:
                b, c = sample.shape
                dlogdet = torch.sum(self.logs)
            if reverse:
                dlogdet *= -1
            logdet = logdet + dlogdet
        return (sample, logdet)

    def forward(self, sample, logdet=None, reverse=False):
        self._check_input_dim(sample)
        if not self.inited:
            self.initialize_parameters(sample)
        if reverse:
            sample, logdet = self._scale(sample, logdet, reverse)
            sample = self._center(sample, reverse)
        else:
            sample = self._center(sample, reverse)
            sample, logdet = self._scale(sample, logdet, reverse)
        return (sample, logdet)

class ActNorm1d(_ActNorm):

    def __init__(self, num_features, scale=1.0):
        super().__init__(num_features, False, scale)

    def _check_input_dim(self, sample):
        assert len(sample.size()) == 2
        assert sample.size(1) == self.num_features, '[ActNorm]: input should be in shape as `BF`, feature size should be {} rather than {}'.format(self.num_features, sample.size())


INIT_ARGS = {
    'num_features': 4,
    'scale': 1.0
}

ARGS = {
    'sample': torch.randn(2, 4),
    'logdet': None,
    'reverse': False
}
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)
ActNorm2d Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("5dfed0f3161fd4b1")
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
import torch.nn as nn

class _ActNorm(nn.Module):
    """
    Activation Normalization
    Initialize the bias and scale with a given minibatch,
    so that the output per-channel have zero mean and unit variance for that.
    After initialization, `bias` and `logs` will be trained as parameters.
    """

    def __init__(self, num_features, image_input, scale=1.0):
        super().__init__()
        param_size = [1, num_features, 1, 1] if image_input else [1, num_features]
        self.bias = nn.Parameter(torch.zeros(*param_size))
        self.logs = nn.Parameter(torch.zeros(*param_size))
        self.num_features = num_features
        self.scale = scale
        self.inited = False
        self.image_input = image_input

    def initialize_parameters(self, sample):
        if not self.training:
            raise ValueError('In Eval mode, but ActNorm not initiated')
        with torch.no_grad():
            reduce_dim = [0, 2, 3] if self.image_input else 0
            bias = -torch.mean(sample.clone(), dim=reduce_dim, keepdim=True)
            var = torch.mean((sample.clone() + bias) ** 2, dim=reduce_dim, keepdim=True)
            logs = torch.log(self.scale / (torch.sqrt(var) + 1e-06))
            self.bias.data.copy_(bias.data)
            self.logs.data.copy_(logs.data)
            self.inited = True

    def _center(self, sample, reverse=False):
        if reverse:
            return sample - self.bias
        else:
            return sample + self.bias

    def _scale(self, sample, logdet=None, reverse=False):
        if reverse:
            sample = sample * torch.exp(-self.logs)
        else:
            sample = sample * torch.exp(self.logs)
        if logdet is not None:
            '\n            logs is log_std of `mean of channels`\n            so we need to multiply by number of pixels\n            '
            if self.image_input:
                b, c, h, w = sample.shape
                dlogdet = torch.sum(self.logs) * h * w
            else:
                b, c = sample.shape
                dlogdet = torch.sum(self.logs)
            if reverse:
                dlogdet *= -1
            logdet = logdet + dlogdet
        return (sample, logdet)

    def forward(self, sample, logdet=None, reverse=False):
        self._check_input_dim(sample)
        if not self.inited:
            self.initialize_parameters(sample)
        if reverse:
            sample, logdet = self._scale(sample, logdet, reverse)
            sample = self._center(sample, reverse)
        else:
            sample = self._center(sample, reverse)
            sample, logdet = self._scale(sample, logdet, reverse)
        return (sample, logdet)

class ActNorm2d(_ActNorm):

    def __init__(self, num_features, scale=1.0):
        super().__init__(num_features, True, scale)

    def _check_input_dim(self, sample):
        assert len(sample.size()) == 4
        assert sample.size(1) == self.num_features, '[ActNorm]: input should be in shape as `BCHW`, channels should be {} rather than {}'.format(self.num_features, sample.size())


INIT_ARGS = {
    'num_features': 4,
    'scale': 1.0,
}

ARGS = {
    'sample': torch.randn(2, 4, 3, 5),
    'logdet': torch.tensor(0.0),
    'reverse': False,
}
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)
BatchNorm Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("84ebb2409e0a1cc7")
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
import torch.nn as nn

class BatchNorm(nn.Module):
    """
    RealNVP BatchNorm layer
    """

    def __init__(self, input_size, momentum=0.9, eps=1e-05):
        super().__init__()
        self.momentum = momentum
        self.eps = eps
        self.log_gamma = nn.Parameter(torch.zeros(input_size))
        self.beta = nn.Parameter(torch.zeros(input_size))
        self.register_buffer('running_mean', torch.zeros(input_size))
        self.register_buffer('running_var', torch.ones(input_size))
        self.register_buffer('batch_mean', torch.zeros(input_size))
        self.register_buffer('batch_var', torch.zeros(input_size))

    def forward(self, x):
        if self.training:
            self.batch_mean = x.mean(0)
            self.batch_var = x.var(0)
            self.running_mean.mul_(self.momentum).add_(self.batch_mean.data * (1 - self.momentum))
            self.running_var.mul_(self.momentum).add_(self.batch_var.data * (1 - self.momentum))
            mean = self.batch_mean
            var = self.batch_var
        else:
            mean = self.running_mean
            var = self.running_var
        x_hat = (x - mean) / torch.sqrt(var + self.eps)
        y = self.log_gamma.exp() * x_hat + self.beta
        log_abs_det_jacobian = self.log_gamma - 0.5 * torch.log(var + self.eps)
        return (y, torch.sum(log_abs_det_jacobian.expand_as(x), dim=1))

    def inverse(self, y):
        if self.training:
            mean = self.batch_mean
            var = self.batch_var
        else:
            mean = self.running_mean
            var = self.running_var
        x_hat = (y - self.beta) * torch.exp(-self.log_gamma)
        x = x_hat * torch.sqrt(var + self.eps) + mean
        log_abs_det_jacobian = 0.5 * torch.log(var + self.eps) - self.log_gamma
        return (x, torch.sum(log_abs_det_jacobian.expand_as(x), dim=1))


INIT_ARGS = {
    'input_size': 4,
    'momentum': 0.9,
    'eps': 1e-05
}

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

if __name__ == '__main__':
    bn = BatchNorm(**INIT_ARGS)
    bn.train()
    y, log_det = bn(**ARGS)
    print("Forward output shape:", y.shape)
    print("Log det jacobian shape:", log_det.shape)
    
    bn.eval()
    y_eval, log_det_eval = bn(**ARGS)
    print("Eval forward output shape:", y_eval.shape)
    print("Eval log det jacobian shape:", log_det_eval.shape)
    
    x_inv, log_det_inv = bn.inverse(y)
    print("Inverse output shape:", x_inv.shape)
    print("Inverse log det jacobian shape:", log_det_inv.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)
Conv2d Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("2d7d6fd92e4989a5")
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 4, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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

INIT_ARGS = {
    'in_dim': 3,
    'out_dim': 5,
    'kernel_size': (3, 3),
    'stride': (1, 1),
    'padding': 'same',
    'do_actnorm': True,
    'weight_std': 0.05
}

ARGS = {
    'sample': torch.randn(2, 3, 8, 8)
}
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 4, dtype family float — no output shape is stated: observe_signature.py could not observe it (call:RuntimeError)
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)
Conv2dZeros Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("eeed4e16d893cea5")
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 4, dtype family float, shape (2, 4, 8, 8) 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

def compute_same_pad(kernel_size, stride):
    if isinstance(kernel_size, int):
        kernel_size = [kernel_size]
    if isinstance(stride, int):
        stride = [stride]
    assert len(stride) == len(kernel_size), 'Pass kernel size and stride both as int, or both as equal length iterable'
    return [((k - 1) * s + 1) // 2 for k, s in zip(kernel_size, stride)]

class Conv2dZeros(nn.Module):

    def __init__(self, in_dim, out_dim, kernel_size=(3, 3), stride=(1, 1), padding='same', logscale_factor=3):
        super().__init__()
        if padding == 'same':
            padding = compute_same_pad(kernel_size, stride)
        elif padding == 'valid':
            padding = 0
        self.conv = nn.Conv2d(in_dim, out_dim, kernel_size, stride, padding)
        self.conv.weight.data.zero_()
        self.conv.bias.data.zero_()
        self.logscale_factor = logscale_factor
        self.logs = nn.Parameter(torch.zeros(out_dim, 1, 1))

    def forward(self, sample):
        output = self.conv(sample)
        return output * torch.exp(self.logs * self.logscale_factor)


INIT_ARGS = {
    'in_dim': 3,
    'out_dim': 4,
    'kernel_size': (3, 3),
    'stride': (1, 1),
    'padding': 'same',
    'logscale_factor': 3
}

ARGS = {
    'sample': torch.randn(2, 3, 8, 8)
}

if __name__ == '__main__':
    model = Conv2dZeros(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output dtype: {output.dtype}")
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 4, dtype family float, shape (2, 4, 8, 8) — shape (2, 4, 8, 8) 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)
InvertibleConv1x1 Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("5c8a917425f0def4")
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
import torch.nn as nn
import torch.nn.functional as F

class InvertibleConv1x1(nn.Module):

    def __init__(self, num_dim, LU_decomposed):
        super().__init__()
        w_shape = [num_dim, num_dim]
        w_init = torch.qr(torch.randn(*w_shape))[0]
        if not LU_decomposed:
            self.weight = nn.Parameter(torch.Tensor(w_init))
        else:
            p, lower, upper = torch.lu_unpack(*torch.lu(w_init))
            s = torch.diag(upper)
            sign_s = torch.sign(s)
            log_s = torch.log(torch.abs(s))
            upper = torch.triu(upper, 1)
            l_mask = torch.tril(torch.ones(w_shape), -1)
            eye = torch.eye(*w_shape)
            self.register_buffer('p', p)
            self.register_buffer('sign_s', sign_s)
            self.lower = nn.Parameter(lower)
            self.log_s = nn.Parameter(log_s)
            self.upper = nn.Parameter(upper)
            self.l_mask = l_mask
            self.eye = eye
        self.w_shape = w_shape
        self.LU_decomposed = LU_decomposed

    def get_weight(self, sample, reverse):
        b, c, h, w = sample.shape
        if not self.LU_decomposed:
            dlogdet = torch.slogdet(self.weight)[1] * h * w
            if reverse:
                weight = torch.inverse(self.weight)
            else:
                weight = self.weight
        else:
            self.l_mask = self.l_mask.to(sample.device)
            self.eye = self.eye.to(sample.device)
            lower = self.lower * self.l_mask + self.eye
            u = self.upper * self.l_mask.transpose(0, 1).contiguous()
            u += torch.diag(self.sign_s * torch.exp(self.log_s))
            dlogdet = torch.sum(self.log_s) * h * w
            if reverse:
                u_inv = torch.inverse(u)
                l_inv = torch.inverse(lower)
                p_inv = torch.inverse(self.p)
                weight = torch.matmul(u_inv, torch.matmul(l_inv, p_inv))
            else:
                weight = torch.matmul(self.p, torch.matmul(lower, u))
        return (weight.view(self.w_shape[0], self.w_shape[1], 1, 1), dlogdet)

    def forward(self, sample, logdet=None, reverse=False):
        """
        log-det = log|abs(|W|)| * pixels
        """
        weight, dlogdet = self.get_weight(sample, reverse)
        if not reverse:
            z = F.conv2d(sample, weight)
            if logdet is not None:
                logdet = logdet + dlogdet
            return (z, logdet)
        else:
            z = F.conv2d(sample, weight)
            if logdet is not None:
                logdet = logdet - dlogdet
            return (z, logdet)


INIT_ARGS = {
    'num_dim': 4,
    'LU_decomposed': True
}

ARGS = {
    'sample': torch.randn(2, 4, 3, 3),
    'logdet': torch.tensor(0.0),
    'reverse': False
}

if __name__ == '__main__':
    model = InvertibleConv1x1(**INIT_ARGS)
    output, logdet_out = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Logdet output: {logdet_out}")
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)
LinearZeros Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("11e0e1eaee63a699")
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
import torch.nn as nn

class LinearZeros(nn.Module):

    def __init__(self, in_dim, out_dim, logscale_factor=3):
        super().__init__()
        self.linear = nn.Linear(in_dim, out_dim)
        self.linear.weight.data.zero_()
        self.linear.bias.data.zero_()
        self.logscale_factor = logscale_factor
        self.logs = nn.Parameter(torch.zeros(out_dim))

    def forward(self, sample):
        output = self.linear(sample)
        return output * torch.exp(self.logs * self.logscale_factor)


INIT_ARGS = {
    'in_dim': 4,
    'out_dim': 3,
    'logscale_factor': 3
}

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

if __name__ == '__main__':
    model = LinearZeros(**INIT_ARGS)
    result = model(**ARGS)
    print(f"Output shape: {result.shape}")
    print(f"Output: {result}")
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)
Permute1d Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("ec45833a8bb10b3c")
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 as nn

class PermuteNd(nn.Module):

    def __init__(self, num_dim, shuffle):
        super().__init__()
        self.num_dim = num_dim
        self.indices = torch.arange(self.num_dim - 1, -1, -1, dtype=torch.long)
        self.indices_inverse = torch.zeros(self.num_dim, dtype=torch.long)
        for i in range(self.num_dim):
            self.indices_inverse[self.indices[i]] = i
        if shuffle:
            self.reset_indices()

    def reset_indices(self):
        shuffle_idx = torch.randperm(self.indices.shape[0])
        self.indices = self.indices[shuffle_idx]
        for i in range(self.num_dim):
            self.indices_inverse[self.indices[i]] = i

    def forward(self, forward, reverse=False):
        pass

class Permute1d(PermuteNd):

    def __init__(self, num_dim, shuffle):
        super().__init__(num_dim, shuffle)

    def forward(self, sample, reverse=False):
        assert len(sample.size()) == 2
        if not reverse:
            sample = sample[:, self.indices]
            return sample
        else:
            return sample[:, self.indices_inverse]


INIT_ARGS = {
    'num_dim': 4,
    'shuffle': False
}

ARGS = {
    'sample': torch.randn(2, 4),
    'reverse': False
}
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)
Permute2d Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("3884e401a1dea604")
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 4, dtype family float, shape (2, 4, 3, 5) 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

class PermuteNd(nn.Module):

    def __init__(self, num_dim, shuffle):
        super().__init__()
        self.num_dim = num_dim
        self.indices = torch.arange(self.num_dim - 1, -1, -1, dtype=torch.long)
        self.indices_inverse = torch.zeros(self.num_dim, dtype=torch.long)
        for i in range(self.num_dim):
            self.indices_inverse[self.indices[i]] = i
        if shuffle:
            self.reset_indices()

    def reset_indices(self):
        shuffle_idx = torch.randperm(self.indices.shape[0])
        self.indices = self.indices[shuffle_idx]
        for i in range(self.num_dim):
            self.indices_inverse[self.indices[i]] = i

    def forward(self, forward, reverse=False):
        pass

class Permute2d(PermuteNd):

    def __init__(self, num_dim, shuffle):
        super().__init__(num_dim, shuffle)

    def forward(self, sample, reverse=False):
        assert len(sample.size()) == 4
        if not reverse:
            sample = sample[:, self.indices, :, :]
            return sample
        else:
            return sample[:, self.indices_inverse, :, :]


INIT_ARGS = {
    'num_dim': 4,
    'shuffle': True,
}

ARGS = {
    'sample': torch.randn(2, 4, 3, 5),
    'reverse': False,
}

if __name__ == '__main__':
    model = Permute2d(**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
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 3, 5) — shape (2, 4, 3, 5) 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)
PermuteNd Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("d6de36917b79ca67")
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 NoneType. 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

class PermuteNd(nn.Module):

    def __init__(self, num_dim, shuffle):
        super().__init__()
        self.num_dim = num_dim
        self.indices = torch.arange(self.num_dim - 1, -1, -1, dtype=torch.long)
        self.indices_inverse = torch.zeros(self.num_dim, dtype=torch.long)
        for i in range(self.num_dim):
            self.indices_inverse[self.indices[i]] = i
        if shuffle:
            self.reset_indices()

    def reset_indices(self):
        shuffle_idx = torch.randperm(self.indices.shape[0])
        self.indices = self.indices[shuffle_idx]
        for i in range(self.num_dim):
            self.indices_inverse[self.indices[i]] = i

    def forward(self, forward, reverse=False):
        pass


INIT_ARGS = {
    'num_dim': 4,
    'shuffle': False
}

ARGS = {
    'forward': torch.randn(2, 3, 5, 7),
    'reverse': False
}

if __name__ == '__main__':
    model = PermuteNd(**INIT_ARGS)
    output = model(**ARGS)
    print("Model created and called successfully")
    print(f"Model indices: {model.indices}")
    print(f"Model indices_inverse: {model.indices_inverse}")
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 NoneType — a NoneType output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a NoneType 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)
ReLUNet Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("b27b3f12e9fea1e8")
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
import torch.nn as nn

class ReLUNet(nn.Module):
    """
    Simple fully connected neural network with ReLU activations.

    TODO: change this to a Fully Connected Network with a activation passed as an argument
    """

    def __init__(self, in_dim, out_dim, hidden_dim, num_layers=1):
        super().__init__()
        layers = []
        layers += [nn.Linear(in_dim, hidden_dim)]
        layers += []
        for i in range(num_layers):
            layers += [nn.ReLU(), nn.Linear(hidden_dim, hidden_dim)]
        layers += [nn.ReLU(), nn.Linear(hidden_dim, out_dim)]
        self.network = nn.Sequential(*layers)

    def forward(self, x):
        return self.network(x)


INIT_ARGS = {
    'in_dim': 4,
    'out_dim': 3,
    'hidden_dim': 8,
    'num_layers': 2,
}

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

if __name__ == '__main__':
    model = ReLUNet(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: (2, 3)")
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)
RealNVP Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("55233dad2a63c90e")
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
import torch.nn as nn

def split_feature(tensor, type='split'):
    """
    type = ["split", "cross"]
    """
    image_input = len(tensor.size()) > 2
    C = tensor.size(1)
    if image_input:
        if type == 'split':
            return (tensor[:, :C // 2, ...], tensor[:, C // 2:, ...])
        elif type == 'cross':
            return (tensor[:, 0::2, ...], tensor[:, 1::2, ...])
    elif type == 'split':
        return (tensor[:, :C // 2], tensor[:, C // 2:])
    elif type == 'cross':
        return (tensor[:, 0::2], tensor[:, 1::2])

class RealNVP(nn.Module):
    """
    Non-volume preserving flow.
    [Dinh et. al. 2017]

    Contains coupling layers, representating one RealNVP flow step
    """

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

    def forward(self, x, layers):
        (t_net, s_net, batch_norm), flipped = layers
        if batch_norm is not None:
            x, bn_ldj = batch_norm(x)
        else:
            bn_ldj = 0.0
        if flipped:
            z2, z1 = split_feature(x, 'split')
        else:
            z1, z2 = split_feature(x, 'split')
        shift = t_net(z1)
        scale = s_net(z1)
        z2 = shift + z2 * torch.exp(scale)
        z = torch.cat([z1, z2], dim=1)
        log_det = torch.sum(scale, dim=1) + bn_ldj
        return (z, log_det)

    def inverse(self, z, layers):
        (t_net, s_net, batch_norm), flipped = layers
        if flipped:
            x1, x2 = split_feature(z, 'split')
        else:
            x2, x1 = split_feature(z, 'split')
        shift = t_net(x2)
        scale = s_net(x2)
        x1 = (x1 - shift) * torch.exp(-scale)
        x = torch.cat([x1, x2], dim=1)
        log_det = torch.sum(-scale, dim=1)
        if batch_norm is not None:
            x, bn_ldj = batch_norm.inverse(x)
            log_det = log_det + bn_ldj
        return (x, log_det)


class SimpleNet(nn.Module):
    def __init__(self, in_dim, out_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(in_dim, 8),
            nn.ReLU(),
            nn.Linear(8, out_dim)
        )
    
    def forward(self, x):
        return self.net(x)


INIT_ARGS = {}

t_net = SimpleNet(2, 2)
s_net = SimpleNet(2, 2)
batch_norm = None
layers = ((t_net, s_net, batch_norm), False)

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

if __name__ == '__main__':
    model = RealNVP(**INIT_ARGS)
    z, log_det = model(**ARGS)
    print("Forward output shape:", z.shape)
    print("Log determinant shape:", log_det.shape)
    print("Forward pass successful")
    
    x_reconstructed, log_det_inv = model.inverse(z, layers)
    print("Inverse output shape:", x_reconstructed.shape)
    print("Inverse log determinant shape:", log_det_inv.shape)
    print("Inverse pass successful")
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)
ResidualBlock Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("0a7417c78d382ffd")
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, 8) 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

INIT_ARGS = {
    'hidden_dim': 8,
    'use_batch_norm': True,
    'zero_initialization': True,
}

ARGS = {
    'inputs': torch.randn(2, 8),
}

if __name__ == '__main__':
    block = nn.Module()
    block.__class__ = type('ResidualBlock', (nn.Module,), {
        '__init__': lambda self, hidden_dim, use_batch_norm=True, zero_initialization=True: (
            super(type(self), self).__init__(),
            setattr(self, 'activation', nn.ReLU()),
            setattr(self, 'linear_layers', nn.ModuleList([nn.Linear(hidden_dim, hidden_dim) for _ in range(2)])),
            (lambda: (
                nn.init.uniform_(self.linear_layers[-1].weight, -0.001, 0.001),
                nn.init.uniform_(self.linear_layers[-1].bias, -0.001, 0.001)
            ))() if zero_initialization else None
        )[-1],
        'forward': lambda self, inputs: (
            lambda temps: inputs + temps
        )(self.linear_layers[1](self.activation(self.linear_layers[0](self.activation(inputs)))))
    })
    
    residual_block = nn.Module.__new__(nn.Module)
    residual_block.__class__ = type('ResidualBlock', (nn.Module,), {
        '__init__': lambda self, hidden_dim, use_batch_norm=True, zero_initialization=True: (
            nn.Module.__init__(self),
            setattr(self, 'activation', nn.ReLU()),
            setattr(self, 'linear_layers', nn.ModuleList([nn.Linear(hidden_dim, hidden_dim) for _ in range(2)])),
            (nn.init.uniform_(self.linear_layers[-1].weight, -0.001, 0.001),
             nn.init.uniform_(self.linear_layers[-1].bias, -0.001, 0.001)) if zero_initialization else None
        ),
        'forward': lambda self, inputs: (
            lambda temps: inputs + temps
        )(self.linear_layers[1](self.activation(self.linear_layers[0](self.activation(inputs)))))
    })
    
    class ResidualBlock(nn.Module):
        def __init__(self, hidden_dim, use_batch_norm=True, zero_initialization=True):
            super().__init__()
            self.activation = nn.ReLU()
            self.linear_layers = nn.ModuleList([nn.Linear(hidden_dim, hidden_dim) for _ in range(2)])
            if zero_initialization:
                nn.init.uniform_(self.linear_layers[-1].weight, -0.001, 0.001)
                nn.init.uniform_(self.linear_layers[-1].bias, -0.001, 0.001)

        def forward(self, inputs):
            temps = inputs
            temps = self.activation(temps)
            temps = self.linear_layers[0](temps)
            temps = self.activation(temps)
            temps = self.linear_layers[1](temps)
            return inputs + temps

    block = ResidualBlock(**INIT_ARGS)
    output = block(**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, 8) — shape (2, 8) 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)
ResidualNet Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("28dd15935b6ad2c2")
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
import torch.nn as nn

INIT_ARGS = {
    'in_dim': 4,
    'out_dim': 3,
    'hidden_dim': 8,
    'num_layers': 2,
}

ARGS = {
    'inputs': torch.randn(2, 4),
}
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)
Split2d Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("b46eb8d0c2c169c9")
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
import torch.nn as nn

def compute_same_pad(kernel_size, stride):
    if isinstance(kernel_size, int):
        kernel_size = [kernel_size]
    if isinstance(stride, int):
        stride = [stride]
    assert len(stride) == len(kernel_size), 'Pass kernel size and stride both as int, or both as equal length iterable'
    return [((k - 1) * s + 1) // 2 for k, s in zip(kernel_size, stride)]

def log_normal_diag(x, mean, log_var, average=False, reduce=True, dim=None):
    log_norm = -0.5 * (log_var + (x - mean) * (x - mean) * log_var.exp().reciprocal())
    if reduce:
        if average:
            return torch.mean(log_norm, dim)
        else:
            return torch.sum(log_norm, dim)
    else:
        return log_norm

def split_feature(tensor, type='split'):
    """
    type = ["split", "cross"]
    """
    image_input = len(tensor.size()) > 2
    C = tensor.size(1)
    if image_input:
        if type == 'split':
            return (tensor[:, :C // 2, ...], tensor[:, C // 2:, ...])
        elif type == 'cross':
            return (tensor[:, 0::2, ...], tensor[:, 1::2, ...])
    elif type == 'split':
        return (tensor[:, :C // 2], tensor[:, C // 2:])
    elif type == 'cross':
        return (tensor[:, 0::2], tensor[:, 1::2])

class Conv2dZeros(nn.Module):

    def __init__(self, in_dim, out_dim, kernel_size=(3, 3), stride=(1, 1), padding='same', logscale_factor=3):
        super().__init__()
        if padding == 'same':
            padding = compute_same_pad(kernel_size, stride)
        elif padding == 'valid':
            padding = 0
        self.conv = nn.Conv2d(in_dim, out_dim, kernel_size, stride, padding)
        self.conv.weight.data.zero_()
        self.conv.bias.data.zero_()
        self.logscale_factor = logscale_factor
        self.logs = nn.Parameter(torch.zeros(out_dim, 1, 1))

    def forward(self, sample):
        output = self.conv(sample)
        return output * torch.exp(self.logs * self.logscale_factor)

class Split2d(nn.Module):

    def __init__(self, in_dim):
        super().__init__()
        self.conv = Conv2dZeros(in_dim // 2, in_dim)

    def split2d_prior(self, z):
        h = self.conv(z)
        return split_feature(h, 'cross')

    def forward(self, sample, logdet=0.0, reverse=False, temperature=None):
        if reverse:
            z1 = sample
            z_mu, z_var = self.split2d_prior(z1)
            z2 = torch.normal(z_mu, torch.exp(z_var) * temperature)
            z = torch.cat((z1, z2), dim=1)
            return (z, logdet)
        else:
            z1, z2 = split_feature(sample, 'split')
            z_mu, z_var = self.split2d_prior(z1)
            logdet = log_normal_diag(z2, z_mu, z_var, dim=[1, 2, 3]) + logdet
            return (z1, logdet)

INIT_ARGS = {
    'in_dim': 8
}

ARGS = {
    'sample': torch.randn(2, 8, 4, 3),
    'logdet': 0.0,
    'reverse': False,
    'temperature': 1.0
}

if __name__ == '__main__':
    model = Split2d(**INIT_ARGS)
    output = model(**ARGS)
    print("Output:", output)
    print("Output[0] shape:", output[0].shape)
    print("Output[1] 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)
SqueezeLayer Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("825e0d0a10f8281b")
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
import torch.nn as nn

def squeeze2d(input, factor):
    if factor == 1:
        return input
    B, C, H, W = input.size()
    assert H % factor == 0 and W % factor == 0, 'H or W modulo factor is not 0'
    x = input.view(B, C, H // factor, factor, W // factor, factor)
    x = x.permute(0, 1, 3, 5, 2, 4).contiguous()
    x = x.view(B, C * factor * factor, H // factor, W // factor)
    return x

def unsqueeze2d(input, factor):
    if factor == 1:
        return input
    factor2 = factor ** 2
    B, C, H, W = input.size()
    assert C % factor2 == 0, 'C module factor squared is not 0'
    x = input.view(B, C // factor2, factor, factor, H, W)
    x = x.permute(0, 1, 4, 2, 5, 3).contiguous()
    x = x.view(B, C // factor2, H * factor, W * factor)
    return x

class SqueezeLayer(nn.Module):

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

    def forward(self, sample, logdet=None, reverse=False):
        if reverse:
            output = unsqueeze2d(sample, self.factor)
        else:
            output = squeeze2d(sample, self.factor)
        return (output, logdet)


INIT_ARGS = {
    'factor': 2
}

ARGS = {
    'sample': torch.randn(2, 4, 8, 6),
    'logdet': None,
    'reverse': False
}

if __name__ == '__main__':
    layer = SqueezeLayer(**INIT_ARGS)
    output, logdet = layer(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Logdet: {logdet}")
    
    ARGS['reverse'] = True
    ARGS['sample'] = torch.randn(2, 16, 4, 3)
    output, logdet = layer(**ARGS)
    print(f"Output shape (reversed): {output.shape}")
    print(f"Logdet (reversed): {logdet}")
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)
TanhNet Ran
ran on a synthesized input
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("45e491a0b611d3dc")
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
import torch.nn as nn

class TanhNet(nn.Module):

    def __init__(self, in_dim, out_dim, hidden_dim, num_layers=1):
        super().__init__()
        layers = []
        layers += [nn.Linear(in_dim, hidden_dim)]
        for i in range(num_layers):
            layers += [nn.Tanh(), nn.Linear(hidden_dim, hidden_dim)]
        layers += [nn.Tanh(), nn.Linear(hidden_dim, out_dim)]
        self.network = nn.Sequential(*layers)

    def forward(self, x):
        return self.network(x)


INIT_ARGS = {
    'in_dim': 4,
    'out_dim': 3,
    'hidden_dim': 8,
    'num_layers': 2,
}

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

model = TanhNet(**INIT_ARGS)
output = model(**ARGS)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 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)
log_normal_diag Ran
ran; our contract draft was wrong, not the code
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("de88c452ca85eb2d")
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 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

torch.manual_seed(0)

# Create small tensors for the function
batch_size = 2
dim = 4

x = torch.randn(batch_size, dim)
mean = torch.randn(batch_size, dim)
log_var = torch.randn(batch_size, dim)

ARGS = {
    "x": x,
    "mean": mean,
    "log_var": log_var,
    "average": False,
    "reduce": True,
    "dim": None,
}
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 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
log_normal_standard Ran
ran; our contract draft was wrong, not the code
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("6ee1f71f50ee4b26")
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 0, dtype family float, shape () as a later run recorded it. 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, 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 math

torch.manual_seed(0)

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

# Create arguments dict
ARGS = {
    'x': x,
    'average': False,
    'reduce': True,
    'dim': None,
    'device': None
}
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 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
split_feature Ran
ran; our contract draft was wrong, not the code
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("45f3824dcc7faec1")
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, ndim 5, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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)

# Test case 1: image input with 'split' type
# tensor shape: (batch=2, channels=4, height=3, width=3)
tensor_image_split = torch.randn(2, 4, 3, 3)

# Test case 2: image input with 'cross' type
tensor_image_cross = torch.randn(2, 4, 3, 3)

# Test case 3: non-image input with 'split' type
# tensor shape: (batch=2, features=4)
tensor_non_image_split = torch.randn(2, 4)

# Test case 4: non-image input with 'cross' type
tensor_non_image_cross = torch.randn(2, 4)

# Use the first test case as the primary argument
ARGS = {
    'tensor': tensor_image_split,
    'type': 'split'
}
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, ndim 5, dtype family float — a tuple output has no array shape of its own; none is stated
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
BoostedFlow Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("639270ba7c0b6ccd")
How it ran

no invocation recorded.

ConvNet Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("85e265af1f8b4721")
How it ran

no invocation recorded.

FlowNet Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("e898fc9ab950c09b")
How it ran

no invocation recorded.

FlowStep Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("cfe3b31cd1ff04b1")
How it ran

no invocation recorded.

GenerativeFlow Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("3b26d29c4a8c9004")
How it ran

no invocation recorded.

Glow Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("832c9c612efadf21")
How it ran

no invocation recorded.

RealNVPFlow Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("bd9b9fc7030ba799")
How it ran

no invocation recorded.

_ActNorm Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("0e3e13be358d4e88")
How it ran

no invocation recorded.

_Dequantization Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/models/boosted_flow.py
code served (permissive licence) · get_code("1ec304d03cf1726b")
How it ran

no invocation recorded.

epoch_reporting Not yet run
unverified (harvested, no recorded run)
robert-giaquinto/gradient-boosted-normalizing-flows/density_experiment.py
code served (permissive licence) · get_code("ebb3df3128ed7f4a")
How it ran

no invocation recorded.

Also present in other papers' code

Of 30 samples lifted from this paper's linked repositories, 30 carry a function identity and 0 do not. Of the 30 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 29 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
log_normal_diag
brick of this paper's assembly: helper
2 other papers
2 dated earlier
1803.05649 (2018-03)
Sylvester Normalizing Flows for Variational Inference
1908.04742 (2019-08)
Online Continual Learning with Maximally Interfered Retrieval

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 2026-11-21.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

By chaining a sequence of differentiable invertible transformations, normalizing flows (NF) provide an expressive method of posterior approximation, exact density evaluation, and sampling. The trend in normalizing flow literature has been to devise deeper, more complex transformations to achieve greater flexibility. We propose an alternative: Gradient Boosted Normalizing Flows (GBNF) model a density by successively adding new NF components with gradient boosting. Under the boosting framework, each new NF component optimizes a sample weighted likelihood objective, resulting in new components that are fit to the residuals of the previously trained components. The GBNF formulation results in a mixture model structure, whose flexibility increases as more components are added. Moreover, GBNFs offer a wider, as opposed to strictly deeper, approach that improves existing NFs at the cost of additional training-not more complex transformations. We demonstrate the effectiveness of this technique for density estimation and, by coupling GBNF with a variational autoencoder, generative modeling of images. Our results show that GBNFs outperform their non-boosted analog, and, in some cases, produce better results with smaller, simpler flows.

For agents

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

get_harvested_code_for_paper("2002.11896")
get_code_for_paper("2002.11896")
have("2002.11896")

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

curl https://syntology.ai/api/ran/2002.11896.json

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

[![Syntology run record](https://syntology.ai/api/ran/2002.11896.svg)](https://syntology.ai/paper/2002.11896)

Connect an agent — have() is free.