SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2107.07933 · ICCV · 2021

Panoptic Segmentation of Satellite Image Time Series with Convolutional Temporal Attention Networks

Vivien Sainte, Fare Garnot

arXiv · PDF · Open in the Atlas

Code that ran

10 of 11 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (10 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 0 failures of our instrument; 1 no recorded run), measured 2026-11-22; 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/2107.07933.json · badge: /api/ran/2107.07933.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
VSainteuf/utae-paps marked official in Papers with Code
Papers with Code archive
10 ran · 1 no recorded run
FunctionStatusWhere it lives
ConvBlock Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("45da329fac60ca4d")
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, ConvBlock: 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 = {
    'nkernels': [3, 16, 32],
    'pad_value': None,
    'norm': 'batch',
    'last_relu': True,
    'padding_mode': 'reflect',
}
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 ConvBlock — 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 ConvBlock 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
ConvLayer Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("e7e11ee365ea4bce")
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, ConvLayer: 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 = {
    'nkernels': [3, 16, 32],
    'norm': 'batch',
    'k': 3,
    's': 1,
    'p': 1,
    'n_groups': 4,
    'last_relu': True,
    'padding_mode': 'reflect'
}
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 ConvLayer — 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 ConvLayer 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
DownConvBlock Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("4a4104e4258a152b")
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, DownConvBlock: 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 = {
    'd_in': 3,
    'd_out': 16,
    'k': 3,
    's': 1,
    'p': 1,
    'pad_value': None,
    'norm': 'batch',
    'padding_mode': 'reflect'
}
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 DownConvBlock — 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 DownConvBlock 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
LTAE2d Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("0c52a14cf35ab27e")
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, LTAE2d: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import numpy as np

torch.manual_seed(0)

class ScaledDotProductAttention(nn.Module):
    def __init__(self, temperature, attn_dropout=0.1):
        super().__init__()
        self.temperature = temperature
        self.dropout = nn.Dropout(attn_dropout)
        self.softmax = nn.Softmax(dim=2)

    def forward(self, q, k, v, pad_mask=None, return_comp=False):
        attn = torch.matmul(q.unsqueeze(1), k.transpose(1, 2))
        attn = attn / self.temperature
        if pad_mask is not None:
            attn = attn.masked_fill(pad_mask.unsqueeze(1), -1000.0)
        if return_comp:
            comp = attn
        attn = self.softmax(attn)
        attn = self.dropout(attn)
        output = torch.matmul(attn, v)
        if return_comp:
            return (output, attn, comp)
        else:
            return (output, attn)

class MultiHeadAttention(nn.Module):
    def __init__(self, n_head, d_k, d_in):
        super().__init__()
        self.n_head = n_head
        self.d_k = d_k
        self.d_in = d_in
        self.Q = nn.Parameter(torch.zeros((n_head, d_k))).requires_grad_(True)
        nn.init.normal_(self.Q, mean=0, std=np.sqrt(2.0 / d_k))
        self.fc1_k = nn.Linear(d_in, n_head * d_k)
        nn.init.normal_(self.fc1_k.weight, mean=0, std=np.sqrt(2.0 / d_k))
        self.attention = ScaledDotProductAttention(temperature=np.power(d_k, 0.5))

    def forward(self, v, pad_mask=None, return_comp=False):
        d_k, d_in, n_head = (self.d_k, self.d_in, self.n_head)
        sz_b, seq_len, _ = v.size()
        q = torch.stack([self.Q for _ in range(sz_b)], dim=1).view(-1, d_k)
        k = self.fc1_k(v).view(sz_b, seq_len, n_head, d_k)
        k = k.permute(2, 0, 1, 3).contiguous().view(-1, seq_len, d_k)
        if pad_mask is not None:
            pad_mask = pad_mask.repeat((n_head, 1))
        v = torch.stack(v.split(v.shape[-1] // n_head, dim=-1)).view(n_head * sz_b, seq_len, -1)
        if return_comp:
            output, attn, comp = self.attention(q, k, v, pad_mask=pad_mask, return_comp=return_comp)
        else:
            output, attn = self.attention(q, k, v, pad_mask=pad_mask, return_comp=return_comp)
        attn = attn.view(n_head, sz_b, 1, seq_len)
        attn = attn.squeeze(dim=2)
        output = output.view(n_head, sz_b, 1, d_in // n_head)
        output = output.squeeze(dim=2)
        if return_comp:
            return (output, attn, comp)
        else:
            return (output, attn)

class PositionalEncoder(nn.Module):
    def __init__(self, d, T=1000, repeat=None, offset=0):
        super(PositionalEncoder, self).__init__()
        self.d = d
        self.T = T
        self.repeat = repeat
        self.denom = torch.pow(T, 2 * (torch.arange(offset, offset + d).float() // 2) / d)
        self.updated_location = False

    def forward(self, batch_positions):
        if not self.updated_location:
            self.denom = self.denom.to(batch_positions.device)
            self.updated_location = True
        sinusoid_table = batch_positions[:, :, None] / self.denom[None, None, :]
        sinusoid_table[:, :, 0::2] = torch.sin(sinusoid_table[:, :, 0::2])
        sinusoid_table[:, :, 1::2] = torch.cos(sinusoid_table[:, :, 1::2])
        if self.repeat is not None:
            sinusoid_table = torch.cat([sinusoid_table for _ in range(self.repeat)], dim=-1)
        return sinusoid_table

ARGS = {
    "in_channels": 128,
    "n_head": 4,
    "d_k": 8,
    "mlp": [64, 32],
    "dropout": 0.2,
    "d_model": 64,
    "T": 1000,
    "return_att": False,
    "positional_encoding": True
}
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 LTAE2d — 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 LTAE2d 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
MultiHeadAttention Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("05217b7e7b5d3120")
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, MultiHeadAttention: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import numpy as np

torch.manual_seed(0)

# Parameters for MultiHeadAttention
n_head = 2
d_k = 4
d_in = 8

# Create a small input tensor for testing
# Shape: (batch_size, seq_len, d_in)
# batch_size = 2, seq_len = 3, d_in = 8
v = torch.randn(2, 3, 8)

# Create a pad_mask (optional parameter)
# Shape: (batch_size, seq_len)
pad_mask = torch.zeros(2, 3, dtype=torch.bool)
pad_mask[0, 2] = True  # Mask the last position in first batch

ARGS = {
    'n_head': n_head,
    'd_k': d_k,
    'd_in': d_in,
}
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 MultiHeadAttention — 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 MultiHeadAttention 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
PositionalEncoder Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("de6db552db8d11c4")
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, PositionalEncoder: 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 PositionalEncoder(nn.Module):

    def __init__(self, d, T=1000, repeat=None, offset=0):
        super(PositionalEncoder, self).__init__()
        self.d = d
        self.T = T
        self.repeat = repeat
        self.denom = torch.pow(T, 2 * (torch.arange(offset, offset + d).float() // 2) / d)
        self.updated_location = False

    def forward(self, batch_positions):
        if not self.updated_location:
            self.denom = self.denom.to(batch_positions.device)
            self.updated_location = True
        sinusoid_table = batch_positions[:, :, None] / self.denom[None, None, :]
        sinusoid_table[:, :, 0::2] = torch.sin(sinusoid_table[:, :, 0::2])
        sinusoid_table[:, :, 1::2] = torch.cos(sinusoid_table[:, :, 1::2])
        if self.repeat is not None:
            sinusoid_table = torch.cat([sinusoid_table for _ in range(self.repeat)], dim=-1)
        return sinusoid_table


# Create batch_positions: shape (batch_size=2, sequence_length=3)
batch_positions = torch.tensor([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]])

ARGS = {
    'd': 4,
    'T': 1000,
    'repeat': None,
    'offset': 0
}

# Test instantiation
encoder = PositionalEncoder(**ARGS)
output = encoder(batch_positions)
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 PositionalEncoder — 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 PositionalEncoder 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
ScaledDotProductAttention Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("2b213e91591f2da8")
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, ScaledDotProductAttention: 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 ScaledDotProductAttention(nn.Module):
    """Scaled Dot-Product Attention
    Modified from github.com/jadore801120/attention-is-all-you-need-pytorch
    """

    def __init__(self, temperature, attn_dropout=0.1):
        super().__init__()
        self.temperature = temperature
        self.dropout = nn.Dropout(attn_dropout)
        self.softmax = nn.Softmax(dim=2)

    def forward(self, q, k, v, pad_mask=None, return_comp=False):
        attn = torch.matmul(q.unsqueeze(1), k.transpose(1, 2))
        attn = attn / self.temperature
        if pad_mask is not None:
            attn = attn.masked_fill(pad_mask.unsqueeze(1), -1000.0)
        if return_comp:
            comp = attn
        attn = self.softmax(attn)
        attn = self.dropout(attn)
        output = torch.matmul(attn, v)
        if return_comp:
            return (output, attn, comp)
        else:
            return (output, attn)


# Construct tiny tensors for testing
batch_size = 2
seq_len_q = 3
seq_len_k = 4
d_model = 5

q = torch.randn(batch_size, seq_len_q, d_model)
k = torch.randn(batch_size, seq_len_k, d_model)
v = torch.randn(batch_size, seq_len_k, d_model)
pad_mask = torch.zeros(batch_size, seq_len_k, dtype=torch.bool)
temperature = 2.0

ARGS = {
    "temperature": temperature,
    "attn_dropout": 0.1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ScaledDotProductAttention — 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 ScaledDotProductAttention 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
Temporal_Aggregator Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("95b9e0afd8ec4312")
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, Temporal_Aggregator: 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 Temporal_Aggregator(nn.Module):

    def __init__(self, mode='mean'):
        super(Temporal_Aggregator, self).__init__()
        self.mode = mode

    def forward(self, x, pad_mask=None, attn_mask=None):
        if pad_mask is not None and pad_mask.any():
            if self.mode == 'att_group':
                n_heads, b, t, h, w = attn_mask.shape
                attn = attn_mask.view(n_heads * b, t, h, w)
                if x.shape[-2] > w:
                    attn = nn.Upsample(size=x.shape[-2:], mode='bilinear', align_corners=False)(attn)
                else:
                    attn = nn.AvgPool2d(kernel_size=w // x.shape[-2])(attn)
                attn = attn.view(n_heads, b, t, *x.shape[-2:])
                attn = attn * (~pad_mask).float()[None, :, :, None, None]
                out = torch.stack(x.chunk(n_heads, dim=2))
                out = attn[:, :, :, None, :, :] * out
                out = out.sum(dim=2)
                out = torch.cat([group for group in out], dim=1)
                return out
            elif self.mode == 'att_mean':
                attn = attn_mask.mean(dim=0)
                attn = nn.Upsample(size=x.shape[-2:], mode='bilinear', align_corners=False)(attn)
                attn = attn * (~pad_mask).float()[:, :, None, None]
                out = (x * attn[:, :, None, :, :]).sum(dim=1)
                return out
            elif self.mode == 'mean':
                out = x * (~pad_mask).float()[:, :, None, None, None]
                out = out.sum(dim=1) / (~pad_mask).sum(dim=1)[:, None, None, None]
                return out
        elif self.mode == 'att_group':
            n_heads, b, t, h, w = attn_mask.shape
            attn = attn_mask.view(n_heads * b, t, h, w)
            if x.shape[-2] > w:
                attn = nn.Upsample(size=x.shape[-2:], mode='bilinear', align_corners=False)(attn)
            else:
                attn = nn.AvgPool2d(kernel_size=w // x.shape[-2])(attn)
            attn = attn.view(n_heads, b, t, *x.shape[-2:])
            out = torch.stack(x.chunk(n_heads, dim=2))
            out = attn[:, :, :, None, :, :] * out
            out = out.sum(dim=2)
            out = torch.cat([group for group in out], dim=1)
            return out
        elif self.mode == 'att_mean':
            attn = attn_mask.mean(dim=0)
            attn = nn.Upsample(size=x.shape[-2:], mode='bilinear', align_corners=False)(attn)
            out = (x * attn[:, :, None, :, :]).sum(dim=1)
            return out
        elif self.mode == 'mean':
            return x.mean(dim=1)

# Setup for 'mean' mode without pad_mask or attn_mask
aggregator_mean = Temporal_Aggregator(mode='mean')
# x shape: (batch, time, channels, height, width) = (2, 3, 4, 5, 6)
x_mean = torch.randn(2, 3, 4, 5, 6)

# Setup for 'att_mean' mode with attn_mask and pad_mask
aggregator_att_mean = Temporal_Aggregator(mode='att_mean')
# x shape: (batch, time, channels, height, width)
x_att_mean = torch.randn(2, 3, 4, 8, 8)
# attn_mask shape: (n_heads, batch, time, height, width)
attn_mask_att_mean = torch.randn(2, 2, 3, 8, 8)
# pad_mask shape: (batch, time) - boolean mask where True means padded
pad_mask_att_mean = torch.tensor([[False, False, True], [False, True, True]])

# Setup for 'att_group' mode
aggregator_att_group = Temporal_Aggregator(mode='att_group')
n_heads = 2
b = 2
t = 3
c_per_head = 4
h = 8
w = 8
# x should have shape compatible with att_group: (batch, time, n_heads*channels, height, width)
x_att_group = torch.randn(b, t, n_heads * c_per_head, h, w)
# attn_mask shape: (n_heads, batch, time, height, width)
attn_mask_att_group = torch.randn(n_heads, b, t, h, w)
pad_mask_att_group = torch.tensor([[False, False, True], [False, True, True]])

ARGS = {
    'mode': 'mean'
}
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 Temporal_Aggregator — 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 Temporal_Aggregator 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
TemporallySharedBlock Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("f56238c8ef11a6d5")
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, TemporallySharedBlock: 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 concrete subclass of TemporallySharedBlock that implements forward()
class ConcreteBlock(nn.Module):
    def __init__(self, pad_value=None):
        super(ConcreteBlock, self).__init__()
        self.pad_value = pad_value
        self.out_shape = None
        # Simple conv layer for the forward pass
        self.conv = nn.Conv2d(in_channels=4, out_channels=3, kernel_size=3, padding=1)
    
    def forward(self, x):
        # x shape: (b*t, c, h, w) or (b, c, h, w)
        return self.conv(x)
    
    def smart_forward(self, input):
        if len(input.shape) == 4:
            return self.forward(input)
        else:
            b, t, c, h, w = input.shape
            if self.pad_value is not None:
                dummy = torch.zeros(input.shape, device=input.device).float()
                self.out_shape = self.forward(dummy.view(b * t, c, h, w)).shape
            out = input.view(b * t, c, h, w)
            if self.pad_value is not None:
                pad_mask = (out == self.pad_value).all(dim=-1).all(dim=-1).all(dim=-1)
                if pad_mask.any():
                    temp = torch.ones(self.out_shape, device=input.device, requires_grad=False) * self.pad_value
                    temp[~pad_mask] = self.forward(out[~pad_mask])
                    out = temp
                else:
                    out = self.forward(out)
            else:
                out = self.forward(out)
            _, c, h, w = out.shape
            out = out.view(b, t, c, h, w)
            return out

# Create the module instance
module = ConcreteBlock(pad_value=None)

ARGS = {
    "pad_value": None
}
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 TemporallySharedBlock — 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 TemporallySharedBlock 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
UpConvBlock Ran
ran on a synthesized input · constructed an object; no result was computed
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("a64d9cde74098274")
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, UpConvBlock: 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 = {
    'd_in': 64,
    'd_out': 32,
    'k': 3,
    's': 2,
    'p': 1,
    'norm': 'batch',
    'd_skip': 32,
    'padding_mode': 'reflect'
}
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 UpConvBlock — 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 UpConvBlock 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
UTAE Not yet run
unverified (harvested, no recorded run)
VSainteuf/utae-paps/src/backbones/utae.py
code served (permissive licence) · get_code("8abb5042ed8b7f79")
How it ran

no invocation recorded.

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 11 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

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

Abstract

Unprecedented access to multi-temporal satellite imagery has opened new perspectives for a variety of Earth observation tasks. Among them, pixel-precise panoptic segmentation of agricultural parcels has major economic and environmental implications. While researchers have explored this problem for single images, we argue that the complex temporal patterns of crop phenology are better addressed with temporal sequences of images. In this paper, we present the first end-to-end, single-stage method for panoptic segmentation of Satellite Image Time Series (SITS). This module can be combined with our novel image sequence encoding network which relies on temporal selfattention to extract rich and adaptive multi-scale spatiotemporal features. We also introduce PASTIS, the first openaccess SITS dataset with panoptic annotations. We demonstrate the superiority of our encoder for semantic segmentation against multiple competing architectures, and set up the first state-of-the-art of panoptic segmentation of SITS. Our implementation and PASTIS are publicly available.

For agents

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

get_harvested_code_for_paper("2107.07933")
get_code_for_paper("2107.07933")
have("2107.07933")

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

curl https://syntology.ai/api/ran/2107.07933.json

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

[![Syntology run record](https://syntology.ai/api/ran/2107.07933.svg)](https://syntology.ai/paper/2107.07933)

Connect an agent — have() is free.