SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2402.17414 · CVPR · 2024

Neural Video Compression with Feature Modulation

Bin Li, Jiahao Li, Yan Lu

arXiv · PDF · Open in the Atlas

Code that ran

36 of 43 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (17 of them constructed an object rather than computing a result; 4 honoured the contract we drafted, 0 violated it and 22 have no contract verdict; 10 failures of our instrument; 7 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/2402.17414.json · badge: /api/ran/2402.17414.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
microsoft/dcvc community
Papers with Code archive
36 ran · 7 no recorded run
FunctionStatusWhere it lives
AEHelper Ran
ran on a synthesized input
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("f82bf9fe9efb4ab4")
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 numpy as np

class AEHelper:

    def __init__(self):
        super().__init__()
        self.entropy_coder = None
        self.cdf_group_index = None
        self._offset = None
        self._quantized_cdf = None
        self._cdf_length = None

    def set_cdf_info(self, quantized_cdf, cdf_length, offset):
        self._quantized_cdf = quantized_cdf.cpu().numpy()
        self._cdf_length = cdf_length.reshape(-1).int().cpu().numpy()
        self._offset = offset.reshape(-1).int().cpu().numpy()

    def get_cdf_info(self):
        return (self._quantized_cdf, self._cdf_length, self._offset)


INIT_ARGS = {}

ARGS = {
    'quantized_cdf': torch.randint(0, 256, (4, 3), dtype=torch.uint8),
    'cdf_length': torch.tensor([5, 6, 7, 8], dtype=torch.int32),
    'offset': torch.tensor([0, 1, 2, 3], dtype=torch.int32),
}

METHOD = 'set_cdf_info'

if __name__ == '__main__':
    obj = AEHelper(**INIT_ARGS)
    result = getattr(obj, METHOD)(**ARGS)
    cdf_info = obj.get_cdf_info()
    print("CDF info retrieved successfully")
    print(f"Quantized CDF shape: {cdf_info[0].shape}")
    print(f"CDF length shape: {cdf_info[1].shape}")
    print(f"Offset shape: {cdf_info[2].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 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)
BitEstimator Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("379a90fedf952408")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, BitEstimator: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn
import numpy as np

torch.manual_seed(0)

ARGS = {
    "qp_num": 3,
    "channel": 4,
}
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 BitEstimator — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BitEstimator 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
Bitparm Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("c4a3ec7db0af2724")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, Bitparm: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

# Create a Bitparm module instance with small dimensions
bitparm_module = nn.Sequential()  # Stand-in, though we're testing Bitparm itself

# Define arguments for Bitparm.__init__
qp_num = 3  # Number of QP values (distinct from other dims)
channel = 4  # Number of channels
final = False  # Test non-final case

# Create test inputs for forward pass
x = torch.randn(2, 4, 8, 8)  # batch=2, channels=4, height=8, width=8
index = torch.tensor([0, 1], dtype=torch.long)  # Index tensor for selecting QP parameters

ARGS = {
    "qp_num": qp_num,
    "channel": channel,
    "final": final,
}
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 Bitparm — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Bitparm 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
ContextualDecoder Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("fca5b37ad0d391c2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ContextualDecoder: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build the ContextualDecoder module
decoder = nn.Module()

# Create minimal dummy tensors for the forward pass
# x shape: (batch, g_ch_16x, height, width) = (2, 128, 4, 4)
# After up1: (2, 96, 8, 8)
# After up2: (2, 96, 16, 16)
# context3: (2, 96, 16, 16) - concatenated with feature
# After res1: (2, 192, 16, 16)
# After up3: (2, 64, 32, 32)
# quant_step: scalar or (1,)
# context2: (2, 128, 32, 32) - concatenated with feature * quant_step
# After res2: (2, 128, 32, 32)
# After up4: (2, 32, 64, 64)

x = torch.randn(2, 128, 4, 4)
context2 = torch.randn(2, 128, 32, 32)
context3 = torch.randn(2, 96, 16, 16)
quant_step = torch.tensor(0.5)

ARGS = {
    "inplace": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ContextualDecoder — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ContextualDecoder 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
ContextualEncoder Ran
ran on a synthesized input
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("310fe7d061e997cf")
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
from torch import nn

# Copy the classes from the problem
class ConvFFN3(nn.Module):

    def __init__(self, in_ch, inplace=False):
        super().__init__()
        expansion_factor = 2
        internal_ch = in_ch * expansion_factor
        self.conv = nn.Conv2d(in_ch, internal_ch * 2, 1)
        self.conv_out = nn.Conv2d(internal_ch, in_ch, 1)
        self.relu1 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        self.relu2 = nn.LeakyReLU(negative_slope=0.01, inplace=inplace)

    def forward(self, x):
        identity = x
        x1, x2 = self.conv(x).chunk(2, 1)
        out = self.relu1(x1) + self.relu2(x2)
        return identity + self.conv_out(out)

class DepthConv(nn.Module):

    def __init__(self, in_ch, out_ch, slope=0.01, inplace=False):
        super().__init__()
        self.conv1 = nn.Sequential(nn.Conv2d(in_ch, in_ch, 1), nn.LeakyReLU(negative_slope=slope, inplace=inplace))
        self.depth_conv = nn.Conv2d(in_ch, in_ch, 3, padding=1, groups=in_ch)
        self.conv2 = nn.Conv2d(in_ch, out_ch, 1)
        self.adaptor = None
        if in_ch != out_ch:
            self.adaptor = nn.Conv2d(in_ch, out_ch, 1)

    def forward(self, x):
        identity = x
        if self.adaptor is not None:
            identity = self.adaptor(identity)
        out = self.conv1(x)
        out = self.depth_conv(out)
        out = self.conv2(out)
        return out + identity

class DepthConvBlock4(nn.Module):

    def __init__(self, in_ch, out_ch, slope_depth_conv=0.01, inplace=False):
        super().__init__()
        self.block = nn.Sequential(DepthConv(in_ch, out_ch, slope=slope_depth_conv, inplace=inplace), ConvFFN3(out_ch, inplace=inplace))

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

g_ch_16x = 128
g_ch_1x = 48
g_ch_2x = 64
g_ch_4x = 96
g_ch_8x = 96

class ContextualEncoder(nn.Module):

    def __init__(self, inplace=False):
        super().__init__()
        self.conv1 = nn.Conv2d(g_ch_1x + 3, g_ch_2x, 3, stride=2, padding=1)
        self.res1 = DepthConvBlock4(g_ch_2x * 2, g_ch_2x * 2, inplace=inplace)
        self.conv2 = nn.Conv2d(g_ch_2x * 2, g_ch_4x, 3, stride=2, padding=1)
        self.res2 = DepthConvBlock4(g_ch_4x * 2, g_ch_4x * 2, inplace=inplace)
        self.conv3 = nn.Conv2d(g_ch_4x * 2, g_ch_8x, 3, stride=2, padding=1)
        self.conv4 = nn.Conv2d(g_ch_8x, g_ch_16x, 3, stride=2, padding=1)

    def forward(self, x, context1, context2, context3, quant_step):
        feature = self.conv1(torch.cat([x, context1], dim=1))
        feature = self.res1(torch.cat([feature, context2], dim=1))
        feature = feature * quant_step
        feature = self.conv2(feature)
        feature = self.res2(torch.cat([feature, context3], dim=1))
        feature = self.conv3(feature)
        feature = self.conv4(feature)
        return feature

INIT_ARGS = {
    'inplace': False,
}

ARGS = {
    'x': torch.randn(2, 3, 32, 32),
    'context1': torch.randn(2, g_ch_1x, 32, 32),
    'context2': torch.randn(2, g_ch_2x, 16, 16),
    'context3': torch.randn(2, g_ch_4x, 8, 8),
    'quant_step': torch.tensor(1.0),
}
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 observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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)
ConvFFN Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("419be1df805c2172")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ConvFFN: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'in_ch': 16,
    'slope': 0.1,
    'inplace': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ConvFFN — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ConvFFN 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
ConvFFN3 Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("03692913271d3292")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ConvFFN3: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class ConvFFN3(nn.Module):

    def __init__(self, in_ch, inplace=False):
        super().__init__()
        expansion_factor = 2
        internal_ch = in_ch * expansion_factor
        self.conv = nn.Conv2d(in_ch, internal_ch * 2, 1)
        self.conv_out = nn.Conv2d(internal_ch, in_ch, 1)
        self.relu1 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        self.relu2 = nn.LeakyReLU(negative_slope=0.01, inplace=inplace)

    def forward(self, x):
        identity = x
        x1, x2 = self.conv(x).chunk(2, 1)
        out = self.relu1(x1) + self.relu2(x2)
        return identity + self.conv_out(out)

ARGS = {
    "in_ch": 4,
    "inplace": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ConvFFN3 — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ConvFFN3 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
DepthConv Ran
ran on a synthesized input
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("a6721c08436a1879")
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
from torch import nn

class DepthConv(nn.Module):

    def __init__(self, in_ch, out_ch, slope=0.01, inplace=False):
        super().__init__()
        self.conv1 = nn.Sequential(nn.Conv2d(in_ch, in_ch, 1), nn.LeakyReLU(negative_slope=slope, inplace=inplace))
        self.depth_conv = nn.Conv2d(in_ch, in_ch, 3, padding=1, groups=in_ch)
        self.conv2 = nn.Conv2d(in_ch, out_ch, 1)
        self.adaptor = None
        if in_ch != out_ch:
            self.adaptor = nn.Conv2d(in_ch, out_ch, 1)

    def forward(self, x):
        identity = x
        if self.adaptor is not None:
            identity = self.adaptor(identity)
        out = self.conv1(x)
        out = self.depth_conv(out)
        out = self.conv2(out)
        return out + identity


INIT_ARGS = {
    'in_ch': 4,
    'out_ch': 6,
    'slope': 0.01,
    'inplace': False
}

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

if __name__ == '__main__':
    model = DepthConv(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: (2, 6, 3, 5)")
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 observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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)
DepthConvBlock Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("b14677bd2a05b95f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, DepthConvBlock: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "in_ch": 4,
    "out_ch": 8,
    "slope_depth_conv": 0.01,
    "slope_ffn": 0.1,
    "inplace": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DepthConvBlock — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a DepthConvBlock 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
DepthConvBlock4 Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("5fcc0a5945f66527")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, DepthConvBlock4: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a small test input tensor
# Shape: (batch_size=2, channels=4, height=8, width=8)
test_input = torch.randn(2, 4, 8, 8)

# Define ARGS dict with all parameters for DepthConvBlock4.__init__
ARGS = {
    'in_ch': 4,
    'out_ch': 3,
    'slope_depth_conv': 0.01,
    'inplace': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DepthConvBlock4 — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a DepthConvBlock4 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
FeatureExtractor Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("4d085f78c5a59bdc")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, FeatureExtractor: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Define the constants used in FeatureExtractor
g_ch_1x = 48
g_ch_2x = 64
g_ch_4x = 96

# Define ResBlock class
class ResBlock(nn.Module):
    def __init__(self, channel, slope=0.01, end_with_relu=False, bottleneck=False, inplace=False):
        super().__init__()
        in_channel = channel // 2 if bottleneck else channel
        self.first_layer = nn.LeakyReLU(negative_slope=slope, inplace=False)
        self.conv1 = nn.Conv2d(channel, in_channel, 3, padding=1)
        self.relu = nn.LeakyReLU(negative_slope=slope, inplace=inplace)
        self.conv2 = nn.Conv2d(in_channel, channel, 3, padding=1)
        self.last_layer = self.relu if end_with_relu else nn.Identity()

    def forward(self, x):
        identity = x
        out = self.first_layer(x)
        out = self.conv1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.last_layer(out)
        return identity + out

# Define FeatureExtractor class
class FeatureExtractor(nn.Module):
    def __init__(self, inplace=False):
        super().__init__()
        self.conv1 = nn.Conv2d(g_ch_1x, g_ch_1x, 3, stride=1, padding=1)
        self.res_block1 = ResBlock(g_ch_1x, inplace=inplace)
        self.conv2 = nn.Conv2d(g_ch_1x, g_ch_2x, 3, stride=2, padding=1)
        self.res_block2 = ResBlock(g_ch_2x, inplace=inplace)
        self.conv3 = nn.Conv2d(g_ch_2x, g_ch_4x, 3, stride=2, padding=1)
        self.res_block3 = ResBlock(g_ch_4x, inplace=inplace)

    def forward(self, feature):
        layer1 = self.conv1(feature)
        layer1 = self.res_block1(layer1)
        layer2 = self.conv2(layer1)
        layer2 = self.res_block2(layer2)
        layer3 = self.conv3(layer2)
        layer3 = self.res_block3(layer3)
        return (layer1, layer2, layer3)

# Create arguments for FeatureExtractor
ARGS = {
    'inplace': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FeatureExtractor — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a FeatureExtractor 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
GaussianEncoder Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("659877e766994421")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, GaussianEncoder: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import math
import numpy as np

torch.manual_seed(0)

ARGS = {
    'distribution': 'laplace'
}
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 GaussianEncoder — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a GaussianEncoder 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
MEBasic Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("fe224c85d7424b84")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, MEBasic: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# MEBasic.__init__ takes complexity_level as a parameter
# MEBasic.forward expects input x to be a 4D tensor (batch, channels, height, width)
# The forward method accesses x[:, -2:, :, :] in bypass mode
# and applies convolutions that expect 8 input channels

# Create a small input tensor: batch_size=2, channels=8, height=4, width=3
input_tensor = torch.randn(2, 8, 4, 3)

ARGS = {
    "complexity_level": 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 MEBasic — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a MEBasic 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
MultiScaleContextFusion Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("320925690ef29c06")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, MultiScaleContextFusion: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

# Define the constants and classes needed
g_ch_1x = 48
g_ch_2x = 64
g_ch_4x = 96

def subpel_conv3x3(in_ch, out_ch, r=1):
    """3x3 sub-pixel convolution for up-sampling."""
    return nn.Sequential(nn.Conv2d(in_ch, out_ch * r ** 2, kernel_size=3, padding=1), nn.PixelShuffle(r))

class ResBlock(nn.Module):
    def __init__(self, channel, slope=0.01, end_with_relu=False, bottleneck=False, inplace=False):
        super().__init__()
        in_channel = channel // 2 if bottleneck else channel
        self.first_layer = nn.LeakyReLU(negative_slope=slope, inplace=False)
        self.conv1 = nn.Conv2d(channel, in_channel, 3, padding=1)
        self.relu = nn.LeakyReLU(negative_slope=slope, inplace=inplace)
        self.conv2 = nn.Conv2d(in_channel, channel, 3, padding=1)
        self.last_layer = self.relu if end_with_relu else nn.Identity()

    def forward(self, x):
        identity = x
        out = self.first_layer(x)
        out = self.conv1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.last_layer(out)
        return identity + out

# Create test inputs
# The forward method expects context1, context2, context3 with specific channel dimensions
batch_size = 2
height = 4
width = 3

context1 = torch.randn(batch_size, g_ch_1x, height * 4, width * 4)
context2 = torch.randn(batch_size, g_ch_2x, height * 2, width * 2)
context3 = torch.randn(batch_size, g_ch_4x, height, width)

ARGS = {
    'inplace': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MultiScaleContextFusion — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a MultiScaleContextFusion 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
MvDec Ran
ran on a synthesized input
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("ddc18f98237531b4")
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

# Define the helper functions and classes needed
def conv3x3(in_ch, out_ch, stride=1):
    """3x3 convolution with padding."""
    return nn.Conv2d(in_ch, out_ch, kernel_size=3, stride=stride, padding=1)

def subpel_conv1x1(in_ch, out_ch, r=1):
    """1x1 sub-pixel convolution for up-sampling."""
    return nn.Sequential(nn.Conv2d(in_ch, out_ch * r ** 2, kernel_size=1, padding=0), nn.PixelShuffle(r))

class ConvFFN3(nn.Module):
    def __init__(self, in_ch, inplace=False):
        super().__init__()
        expansion_factor = 2
        internal_ch = in_ch * expansion_factor
        self.conv = nn.Conv2d(in_ch, internal_ch * 2, 1)
        self.conv_out = nn.Conv2d(internal_ch, in_ch, 1)
        self.relu1 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        self.relu2 = nn.LeakyReLU(negative_slope=0.01, inplace=inplace)

    def forward(self, x):
        identity = x
        x1, x2 = self.conv(x).chunk(2, 1)
        out = self.relu1(x1) + self.relu2(x2)
        return identity + self.conv_out(out)

class DepthConv(nn.Module):
    def __init__(self, in_ch, out_ch, slope=0.01, inplace=False):
        super().__init__()
        self.conv1 = nn.Sequential(nn.Conv2d(in_ch, in_ch, 1), nn.LeakyReLU(negative_slope=slope, inplace=inplace))
        self.depth_conv = nn.Conv2d(in_ch, in_ch, 3, padding=1, groups=in_ch)
        self.conv2 = nn.Conv2d(in_ch, out_ch, 1)
        self.adaptor = None
        if in_ch != out_ch:
            self.adaptor = nn.Conv2d(in_ch, out_ch, 1)

    def forward(self, x):
        identity = x
        if self.adaptor is not None:
            identity = self.adaptor(identity)
        out = self.conv1(x)
        out = self.depth_conv(out)
        out = self.conv2(out)
        return out + identity

class DepthConvBlock4(nn.Module):
    def __init__(self, in_ch, out_ch, slope_depth_conv=0.01, inplace=False):
        super().__init__()
        self.block = nn.Sequential(DepthConv(in_ch, out_ch, slope=slope_depth_conv, inplace=inplace), ConvFFN3(out_ch, inplace=inplace))

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

class ResidualBlockUpsample(nn.Module):
    def __init__(self, in_ch, out_ch, upsample=2, inplace=False):
        super().__init__()
        self.subpel_conv = subpel_conv1x1(in_ch, out_ch, upsample)
        self.leaky_relu = nn.LeakyReLU(inplace=inplace)
        self.conv = conv3x3(out_ch, out_ch)
        self.leaky_relu2 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        self.upsample = subpel_conv1x1(in_ch, out_ch, upsample)

    def forward(self, x):
        identity = x
        out = self.subpel_conv(x)
        out = self.leaky_relu(out)
        out = self.conv(out)
        out = self.leaky_relu2(out)
        identity = self.upsample(x)
        out = out + identity
        return out

class MvDec(nn.Module):
    def __init__(self, output_channel, channel, inplace=False):
        super().__init__()
        self.dec_1 = nn.Sequential(DepthConvBlock4(channel, channel, inplace=inplace), ResidualBlockUpsample(channel, channel, 2, inplace=inplace), DepthConvBlock4(channel, channel, inplace=inplace), ResidualBlockUpsample(channel, channel, 2, inplace=inplace), DepthConvBlock4(channel, channel, inplace=inplace))
        self.dec_2 = ResidualBlockUpsample(channel, channel, 2, inplace=inplace)
        self.dec_3 = nn.Sequential(DepthConvBlock4(channel, channel, inplace=inplace), subpel_conv1x1(channel, output_channel, 2))

    def forward(self, x, quant_step):
        feature = self.dec_1(x)
        out = self.dec_2(feature)
        out = out * quant_step
        mv = self.dec_3(out)
        return (mv, feature)

INIT_ARGS = {
    'output_channel': 3,
    'channel': 4,
    'inplace': False,
}

ARGS = {
    'x': torch.randn(2, 4, 8, 8),
    'quant_step': torch.tensor(0.5),
}

if __name__ == '__main__':
    model = MvDec(**INIT_ARGS)
    output = model(**ARGS)
    print("Output shapes:", output[0].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)
MvEnc Ran
ran on a synthesized input
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("0272de7b35cc7549")
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
from torch import nn

# Helper functions and classes from the provided code
def conv1x1(in_ch, out_ch, stride=1):
    """1x1 convolution."""
    return nn.Conv2d(in_ch, out_ch, kernel_size=1, stride=stride)

def conv3x3(in_ch, out_ch, stride=1):
    """3x3 convolution with padding."""
    return nn.Conv2d(in_ch, out_ch, kernel_size=3, stride=stride, padding=1)

class ConvFFN3(nn.Module):
    def __init__(self, in_ch, inplace=False):
        super().__init__()
        expansion_factor = 2
        internal_ch = in_ch * expansion_factor
        self.conv = nn.Conv2d(in_ch, internal_ch * 2, 1)
        self.conv_out = nn.Conv2d(internal_ch, in_ch, 1)
        self.relu1 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        self.relu2 = nn.LeakyReLU(negative_slope=0.01, inplace=inplace)

    def forward(self, x):
        identity = x
        x1, x2 = self.conv(x).chunk(2, 1)
        out = self.relu1(x1) + self.relu2(x2)
        return identity + self.conv_out(out)

class DepthConv(nn.Module):
    def __init__(self, in_ch, out_ch, slope=0.01, inplace=False):
        super().__init__()
        self.conv1 = nn.Sequential(nn.Conv2d(in_ch, in_ch, 1), nn.LeakyReLU(negative_slope=slope, inplace=inplace))
        self.depth_conv = nn.Conv2d(in_ch, in_ch, 3, padding=1, groups=in_ch)
        self.conv2 = nn.Conv2d(in_ch, out_ch, 1)
        self.adaptor = None
        if in_ch != out_ch:
            self.adaptor = nn.Conv2d(in_ch, out_ch, 1)

    def forward(self, x):
        identity = x
        if self.adaptor is not None:
            identity = self.adaptor(identity)
        out = self.conv1(x)
        out = self.depth_conv(out)
        out = self.conv2(out)
        return out + identity

class DepthConvBlock4(nn.Module):
    def __init__(self, in_ch, out_ch, slope_depth_conv=0.01, inplace=False):
        super().__init__()
        self.block = nn.Sequential(DepthConv(in_ch, out_ch, slope=slope_depth_conv, inplace=inplace), ConvFFN3(out_ch, inplace=inplace))

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

class ResidualBlockWithStride(nn.Module):
    def __init__(self, in_ch, out_ch, stride=2, inplace=False):
        super().__init__()
        self.conv1 = conv3x3(in_ch, out_ch, stride=stride)
        self.leaky_relu = nn.LeakyReLU(inplace=inplace)
        self.conv2 = conv3x3(out_ch, out_ch)
        self.leaky_relu2 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        if stride != 1:
            self.downsample = conv1x1(in_ch, out_ch, stride=stride)
        else:
            self.downsample = None

    def forward(self, x):
        identity = x
        out = self.conv1(x)
        out = self.leaky_relu(out)
        out = self.conv2(out)
        out = self.leaky_relu2(out)
        if self.downsample is not None:
            identity = self.downsample(x)
        out = out + identity
        return out

class MvEnc(nn.Module):
    def __init__(self, input_channel, channel, inplace=False):
        super().__init__()
        self.enc_1 = nn.Sequential(ResidualBlockWithStride(input_channel, channel, stride=2, inplace=inplace), DepthConvBlock4(channel, channel, inplace=inplace))
        self.enc_2 = ResidualBlockWithStride(channel, channel, stride=2, inplace=inplace)
        self.adaptor_0 = DepthConvBlock4(channel, channel, inplace=inplace)
        self.adaptor_1 = DepthConvBlock4(channel * 2, channel, inplace=inplace)
        self.enc_3 = nn.Sequential(ResidualBlockWithStride(channel, channel, stride=2, inplace=inplace), DepthConvBlock4(channel, channel, inplace=inplace), nn.Conv2d(channel, channel, 3, stride=2, padding=1))

    def forward(self, x, context, quant_step):
        out = self.enc_1(x)
        out = out * quant_step
        out = self.enc_2(out)
        if context is None:
            out = self.adaptor_0(out)
        else:
            out = self.adaptor_1(torch.cat((out, context), dim=1))
        return self.enc_3(out)

INIT_ARGS = {
    'input_channel': 3,
    'channel': 4,
    'inplace': False,
}

ARGS = {
    'x': torch.randn(2, 3, 64, 64),
    'context': None,
    'quant_step': torch.tensor(0.5),
}
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 observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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)
OffsetDiversity Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("d2f0192d643ab535")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, OffsetDiversity: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn
import torch.nn.functional as F

torch.manual_seed(0)

# Create minimal tensors for testing
B, C_in, H, W = 2, 48, 8, 8
C_aux = 48 + 3 + 2  # g_ch_1x + 3 + 2
offset_num = 2
group_num = 16

# x: input feature map (B, C_in, H, W)
x = torch.randn(B, C_in, H, W)

# aux_feature: auxiliary feature (B, aux_feature_num, H, W)
aux_feature = torch.randn(B, C_aux, H, W)

# flow: optical flow (B, 2, H, W)
flow = torch.randn(B, 2, H, W)

ARGS = {
    "in_channel": 48,
    "aux_feature_num": 48 + 3 + 2,
    "offset_num": 2,
    "group_num": 16,
    "max_residue_magnitude": 40,
    "inplace": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type OffsetDiversity — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a OffsetDiversity 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
ReconGeneration Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("3d55656b67d7cbf4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ReconGeneration: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

# The ReconGeneration class expects:
# - ctx: context tensor with ctx_channel (default g_ch_1x=48) channels
# - res: residual tensor with res_channel (default 32) channels
# Both tensors should have spatial dimensions that work through the UNet2 architecture
# (which uses max_pool and upsampling by factor of 2)

# Create sample tensors: batch=2, spatial dims=8x8 (divisible by 2 for pooling)
batch_size = 2
ctx_channel = 48
res_channel = 32
height = 8
width = 8

ctx_tensor = torch.randn(batch_size, ctx_channel, height, width)
res_tensor = torch.randn(batch_size, res_channel, height, width)

ARGS = {
    "ctx_channel": ctx_channel,
    "res_channel": res_channel,
    "inplace": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ReconGeneration — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ReconGeneration 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
ResBlock Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("75b276630f3fe107")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ResBlock: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'channel': 16,
    'slope': 0.01,
    'end_with_relu': False,
    'bottleneck': False,
    'inplace': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ResBlock — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ResBlock 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
ResidualBlockUpsample Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("c0ac2daecdabac0b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ResidualBlockUpsample: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Define the helper functions needed
def conv3x3(in_ch, out_ch, stride=1):
    """3x3 convolution with padding."""
    return nn.Conv2d(in_ch, out_ch, kernel_size=3, stride=stride, padding=1)

def subpel_conv1x1(in_ch, out_ch, r=1):
    """1x1 sub-pixel convolution for up-sampling."""
    return nn.Sequential(nn.Conv2d(in_ch, out_ch * r ** 2, kernel_size=1, padding=0), nn.PixelShuffle(r))

class ResidualBlockUpsample(nn.Module):
    """Residual block with sub-pixel upsampling on the last convolution.

    Args:
        in_ch (int): number of input channels
        out_ch (int): number of output channels
        upsample (int): upsampling factor (default: 2)
    """

    def __init__(self, in_ch, out_ch, upsample=2, inplace=False):
        super().__init__()
        self.subpel_conv = subpel_conv1x1(in_ch, out_ch, upsample)
        self.leaky_relu = nn.LeakyReLU(inplace=inplace)
        self.conv = conv3x3(out_ch, out_ch)
        self.leaky_relu2 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        self.upsample = subpel_conv1x1(in_ch, out_ch, upsample)

    def forward(self, x):
        identity = x
        out = self.subpel_conv(x)
        out = self.leaky_relu(out)
        out = self.conv(out)
        out = self.leaky_relu2(out)
        identity = self.upsample(x)
        out = out + identity
        return out

# Create arguments for ResidualBlockUpsample
ARGS = {
    "in_ch": 3,
    "out_ch": 4,
    "upsample": 2,
    "inplace": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ResidualBlockUpsample — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ResidualBlockUpsample 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
ResidualBlockWithStride Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("0b4d24dfa9389a82")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ResidualBlockWithStride: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

def conv1x1(in_ch, out_ch, stride=1):
    """1x1 convolution."""
    return nn.Conv2d(in_ch, out_ch, kernel_size=1, stride=stride)

def conv3x3(in_ch, out_ch, stride=1):
    """3x3 convolution with padding."""
    return nn.Conv2d(in_ch, out_ch, kernel_size=3, stride=stride, padding=1)

class ResidualBlockWithStride(nn.Module):
    """Residual block with a stride on the first convolution.

    Args:
        in_ch (int): number of input channels
        out_ch (int): number of output channels
        stride (int): stride value (default: 2)
    """

    def __init__(self, in_ch, out_ch, stride=2, inplace=False):
        super().__init__()
        self.conv1 = conv3x3(in_ch, out_ch, stride=stride)
        self.leaky_relu = nn.LeakyReLU(inplace=inplace)
        self.conv2 = conv3x3(out_ch, out_ch)
        self.leaky_relu2 = nn.LeakyReLU(negative_slope=0.1, inplace=inplace)
        if stride != 1:
            self.downsample = conv1x1(in_ch, out_ch, stride=stride)
        else:
            self.downsample = None

    def forward(self, x):
        identity = x
        out = self.conv1(x)
        out = self.leaky_relu(out)
        out = self.conv2(out)
        out = self.leaky_relu2(out)
        if self.downsample is not None:
            identity = self.downsample(x)
        out = out + identity
        return out

ARGS = {
    'in_ch': 3,
    'out_ch': 4,
    'stride': 2,
    'inplace': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ResidualBlockWithStride — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ResidualBlockWithStride 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
UNet2 Ran
ran on a synthesized input · constructed an object; no result was computed
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("a12a5eee4649c19e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, UNet2: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# UNet2.__init__ takes: in_ch=64, out_ch=64, inplace=False
# UNet2.forward takes: x (a tensor)
# The forward method expects x to be a 4D tensor (batch, channels, height, width)
# in_ch=64 means input has 64 channels
# out_ch=64 means output has 64 channels

# Create a small input tensor: batch_size=2, in_ch=64, height=4, width=3
x = torch.randn(2, 64, 4, 3)

ARGS = {
    'in_ch': 64,
    'out_ch': 64,
    'inplace': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type UNet2 — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a UNet2 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
bilineardownsacling Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("4ee1913b1f8ce86e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 4, 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.functional as F

torch.manual_seed(0)

# Create a small input tensor with shape (batch, channels, height, width)
# Using distinct sizes: batch=2, channels=4, height=8, width=6
inputfeature = torch.randn(2, 4, 8, 6)

ARGS = {
    'inputfeature': inputfeature
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 4, 3) — shape (2, 4, 4, 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
bilinearupsacling Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("2c8c63d56261cb76")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 6, 10) 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

torch.manual_seed(0)

# Create a small input tensor with shape (batch, channels, height, width)
# Using distinct sizes: batch=2, channels=4, height=3, width=5
inputfeature = torch.randn(2, 4, 3, 5)

ARGS = {
    "inputfeature": inputfeature
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 6, 10) — shape (2, 4, 6, 10) 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
block_mc_func Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("da82b58aa9f6478c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 8, 6) 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 testing
batch_size = 2
channels = 4
height = 8
width = 6

# im: image tensor (B, C, H, W)
im = torch.randn(batch_size, channels, height, width, dtype=torch.float32)

# flow: optical flow tensor (B, 2, H, W)
# 2 channels for horizontal and vertical flow
flow = torch.randn(batch_size, 2, height, width, dtype=torch.float32)

ARGS = {
    "im": im,
    "flow": flow,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 8, 6) — shape (2, 4, 8, 6) 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
conv1x1 Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("ff496db6b866af12")
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 Conv2d. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

ARGS = {
    "in_ch": 3,
    "out_ch": 64,
    "stride": 1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Conv2d — a Conv2d output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Conv2d output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
conv3x3 Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("194462a957fb0991")
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 Conv2d. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "in_ch": 3,
    "out_ch": 64,
    "stride": 1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Conv2d — a Conv2d output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Conv2d output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
flow_warp Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("ca37ddb87705e620")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 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

torch.manual_seed(0)

# Create small tensors for testing
# im: image tensor with shape (B, C, H, W)
# flow: optical flow tensor with shape (B, 2, H, W)

im = torch.randn(2, 4, 3, 5, dtype=torch.float32)
flow = torch.randn(2, 2, 3, 5, dtype=torch.float32)

ARGS = {
    'im': im,
    'flow': flow
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 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: none by the harness; any seed is in the setup block
get_downsampled_shape Ran
ran, honoured the contract we drafted
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("0c429d93ea07cde7")
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, ndim 1, dtype family int; 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)

# Analyzing the function:
# - height: int, used in arithmetic operations
# - width: int, used in arithmetic operations
# - p: int, used as divisor/multiplier in arithmetic operations
# The function performs downsampling calculations based on height, width, and a patch size p

ARGS = {
    "height": 64,
    "width": 48,
    "p": 8
}
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, ndim 1, dtype family int — 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
get_padding_size Ran
ran, honoured the contract we drafted
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("058564027e22df3c")
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, ndim 1, dtype family int; 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 numpy as np

torch.manual_seed(0)

ARGS = {
    "height": 100,
    "width": 150,
    "p": 64
}
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, ndim 1, dtype family int — 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
subpel_conv1x1 Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("05d9e671c5f84bdb")
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 Sequential. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

ARGS = {
    'in_ch': 3,
    'out_ch': 64,
    'r': 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 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
subpel_conv3x3 Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("c441c9f39e918e8a")
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 Sequential. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

ARGS = {
    "in_ch": 3,
    "out_ch": 16,
    "r": 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 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
torch_warp Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("4ed6bbdfa756e1be")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 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

torch.manual_seed(0)

# Create small tensors for the function
# feature: used in grid_sample as input, needs shape (B, C, H, W)
# flow: used to get size and access device, needs shape (B, 2, H, W) for optical flow

batch_size = 2
channels = 4
height = 3
width = 5

# feature tensor: (batch, channels, height, width)
feature = torch.randn(batch_size, channels, height, width, dtype=torch.float32)

# flow tensor: (batch, 2, height, width) - 2 channels for x and y flow
flow = torch.randn(batch_size, 2, height, width, dtype=torch.float32)

ARGS = {
    'feature': feature,
    'flow': flow
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 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: none by the harness; any seed is in the setup block
write_ip Ran
ran, honoured the contract we drafted
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("c8819fe99eeef17c")
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 int, ndim 0, dtype family int, 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 enum
import struct
import io

class NalType(enum.IntEnum):
    NAL_SPS = 0
    NAL_I = 1
    NAL_P = 2
    NAL_Ps = 3

def write_bytes(fd, values, fmt='>{:d}s'):
    if len(values) == 0:
        return 0
    fd.write(struct.pack(fmt.format(len(values)), values))
    return len(values)

def write_uchars(fd, values, fmt='>{:d}B'):
    fd.write(struct.pack(fmt.format(len(values)), *values))
    return len(values)

def write_uint_adaptive(f, a):
    if a <= 32767:
        a0 = a & 255
        a1 = a >> 8
        write_uchars(f, (a1, a0))
        return 2
    assert a < 1 << 30
    a0 = a & 255
    a1 = a >> 8 & 255
    a2 = a >> 16 & 255
    a3 = a >> 24 & 255
    a3 = a3 | 1 << 7
    write_uchars(f, (a3, a2, a1, a0))
    return 4

def write_ip(f, is_i_frame, sps_id, bit_stream):
    written = 0
    flag = (int(NalType.NAL_I if is_i_frame else NalType.NAL_P) << 4) + sps_id
    written += write_uchars(f, (flag,))
    written += write_uint_adaptive(f, len(bit_stream))
    written += write_bytes(f, bit_stream)
    return written

# Setup arguments for write_ip
f = io.BytesIO()
is_i_frame = True
sps_id = 2
bit_stream = b'\x00\x01\x02\x03\x04\x05'

ARGS = {
    'f': f,
    'is_i_frame': is_i_frame,
    'sps_id': sps_id,
    'bit_stream': bit_stream
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, 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
write_uchars Ran
ran; our contract draft was wrong, not the code
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("cde57c4548df7988")
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 int, ndim 0, dtype family int, 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 io
import struct

# Create a file-like object to write to
fd = io.BytesIO()

# Create valid unsigned char values (0-255 range)
values = [42, 100, 200, 15, 7]

# Format string for packing (default format)
fmt = '>{:d}B'

ARGS = {
    'fd': fd,
    'values': values,
    'fmt': fmt
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, 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
write_uint_adaptive Ran
ran, honoured the contract we drafted
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("1ec4dcdc3950ddfe")
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 int, ndim 0, dtype family int, 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 struct
import io

def write_uchars(fd, values, fmt='>{:d}B'):
    fd.write(struct.pack(fmt.format(len(values)), *values))
    return len(values)

def write_uint_adaptive(f, a):
    if a <= 32767:
        a0 = a & 255
        a1 = a >> 8
        write_uchars(f, (a1, a0))
        return 2
    assert a < 1 << 30
    a0 = a & 255
    a1 = a >> 8 & 255
    a2 = a >> 16 & 255
    a3 = a >> 24 & 255
    a3 = a3 | 1 << 7
    write_uchars(f, (a3, a2, a1, a0))
    return 4

# Create a file-like object for writing
f = io.BytesIO()

# Test with a small value (fits in 2 bytes)
a = 1000

ARGS = {
    'f': f,
    'a': a
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, 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
CompressionModel Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("b04500057b53dc0d")
How it ran

no invocation recorded.

DMC Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("d0541ab1ef1f01d7")
How it ran

no invocation recorded.

EntropyCoder Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("8b1f88062c054aad")
How it ran

no invocation recorded.

ME_Spynet Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("8a52237f88c790c0")
How it ran

no invocation recorded.

NalType Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("b50de84040ba0329")
How it ran

no invocation recorded.

add_grid_cache Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("a3cc1c3e7b0b88aa")
How it ran

no invocation recorded.

write_bytes Not yet run
unverified (harvested, no recorded run)
microsoft/dcvc/DCVC-family/DCVC-FM/src/models/video_model.py
code served (permissive licence) · get_code("395e86482196ce43")
How it ran

no invocation recorded.

Also present in other papers' code

Of 43 samples lifted from this paper's linked repositories, 43 carry a function identity and 0 do not. Of the 43 distinct functions with an identity, 9 are also present in other papers whose code we hold (5 with at least one paper dated earlier, 0 the same month, 6 later) and 34 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
conv3x3
brick of this paper's assembly: helper
4 other papers
3 dated earlier, 1 later · 1 only as an assembly brick
Deng_Deep_Homography_for_Efficient_Stereo_Image_Compression_CVPR_2021_paper (2021 (year only))
Deep Homography for Efficient Stereo Image Compression
2108.03690 (2021-08)
Enhanced Invertible Encoding for Learned Image Compression
2302.05071 (2023-02)
EVC: Towards Real-Time Neural Image Compression with Mask Decay
and 1 more not named here
conv1x1
brick of this paper's assembly: helper
3 other papers
2 dated earlier, 1 later
Deng_Deep_Homography_for_Efficient_Stereo_Image_Compression_CVPR_2021_paper (2021 (year only))
Deep Homography for Efficient Stereo Image Compression
2108.03690 (2021-08)
Enhanced Invertible Encoding for Learned Image Compression
2510.12479 (2025-10)
MH-LVC: Multi-Hypothesis Temporal Prediction for Learned Conditional Residual Video Coding
subpel_conv3x3
brick of this paper's assembly: helper
3 other papers
3 dated earlier
Deng_Deep_Homography_for_Efficient_Stereo_Image_Compression_CVPR_2021_paper (2021 (year only))
Deep Homography for Efficient Stereo Image Compression
2108.03690 (2021-08)
Enhanced Invertible Encoding for Learned Image Compression
2109.15047 (2021-09)
Deep Contextual Video Compression
get_downsampled_shape
brick of this paper's assembly: helper
2 other papers
2 dated earlier · 2 only as an assembly brick
2109.15047 (2021-09)
Deep Contextual Video Compression
2302.05071 (2023-02)
EVC: Towards Real-Time Neural Image Compression with Mask Decay
AEHelper
brick of this paper's assembly: helper
1 other paper
1 later · 1 only as an assembly brick
2502.20762 (2025-02)
Towards Practical Real-Time Neural Video Compression
Bitparm
brick of this paper's assembly: helper
1 other paper
1 later
2502.20762 (2025-02)
Towards Practical Real-Time Neural Video Compression
bilinearupsacling
brick of this paper's assembly: helper
1 other paper
1 later
2510.12479 (2025-10)
MH-LVC: Multi-Hypothesis Temporal Prediction for Learned Conditional Residual Video Coding
get_padding_size
brick of this paper's assembly: helper
1 other paper
1 later
2508.02072 (2025-08)
HyTIP: Hybrid Temporal Information Propagation for Masked Conditional Residual Video Coding
subpel_conv1x1
brick of this paper's assembly: helper
1 other paper
1 dated earlier · 1 only as an assembly brick
2302.05071 (2023-02)
EVC: Towards Real-Time Neural Image Compression with Mask Decay

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

The emerging conditional coding-based neural video codec (NVC) shows superiority over commonly-used residual coding-based codec and the latest NVC already claims to outperform the best traditional codec. However, there still exist critical problems blocking the practicality of NVC. In this paper, we propose a powerful conditional codingbased NVC that solves two critical problems via feature modulation. The first is how to support a wide quality range in a single model. Previous NVC with this capability only supports about 3.8 dB PSNR range on average. To tackle this limitation, we modulate the latent feature of the current frame via the learnable quantization scaler. During the training, we specially design the uniform quantization parameter sampling mechanism to improve the harmonization of encoding and quantization. This results in a better learning of the quantization scaler and helps our NVC support about 11.4 dB PSNR range. The second is how to make NVC still work under a long prediction chain. We expose that the previous SOTA NVC has an obvious quality degradation problem when using a large intra-period setting. To this end, we propose modulating the temporal feature with a periodically refreshing mechanism to boost the quality. Notably, under single intra-frame setting, our codec can achieve 29.7% bitrate saving over previous SOTA NVC with 16% MACs reduction. Our codec serves as a notable landmark in the journey of NVC evolution. The codes are at https://github.com/microsoft/DCVC.

For agents

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

get_harvested_code_for_paper("2402.17414")
get_code_for_paper("2402.17414")
have("2402.17414")

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

curl https://syntology.ai/api/ran/2402.17414.json

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

[![Syntology run record](https://syntology.ai/api/ran/2402.17414.svg)](https://syntology.ai/paper/2402.17414)

Connect an agent — have() is free.