SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 1802.01561 · 2018

IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures

Volodymyr Mnih, Karen Simonyan, Vlad Firoiu, Shane Legg, Remi Munos, Koray Kavukcuoglu, Lasse Espeholt, Tim Harley, Hubert Soyer, Yotam Doron, Tom Ward, Iain Dunning

arXiv · PDF · Open in the Atlas

Code that ran

16 of 34 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (6 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 1 violated it and 8 have no contract verdict; 6 failures of our instrument; 18 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/1802.01561.json · badge: /api/ran/1802.01561.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
villinvic/Georges community
Papers with Code archive
7 ran · 7 no recorded run
facebookresearch/torchbeast community
Papers with Code archive
3 ran · 0 no recorded run
google-research/valan community
Papers with Code archive
1 ran · 2 no recorded run
google-deepmind/scalable_agent community
Papers with Code archive
1 ran · 2 no recorded run
seolhokim/DistributedRL-Pytorch-Ray community
Papers with Code archive
1 ran · 1 no recorded run
Sheepsody/Batched-Impala-PyTorch community
Papers with Code archive
1 ran · 1 no recorded run
facebookresearch/gala community
Papers with Code archive
1 ran · 0 no recorded run
deepmind/rlax community
Papers with Code archive
1 ran · 0 no recorded run
urw7rs/spiralpp community
Papers with Code archive
0 ran · 2 no recorded run
seolhokim/SimpleDistributedRL community
Papers with Code archive
0 ran · 2 no recorded run
haje01/impala community
Papers with Code archive
0 ran · 1 no recorded run
FunctionStatusWhere it lives
Categorical Ran
ran on a synthesized input
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("89ed626ce6f952a2")
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 EagerTensor, ndim 1, 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 tensorflow as tf
import numpy as np

# Construct the Categorical distribution
INIT_ARGS = {
    'dim': 5  # dimension of the categorical distribution
}

# For the log_likelihood method
METHOD = 'log_likelihood'

# Create sample data for log_likelihood
# x should be one-hot encoded vectors
# param should be a dict with 'prob' key containing probability distributions
batch_size = 2
dim = 5

# Create one-hot encoded actions (batch_size x dim)
x = tf.constant([
    [1.0, 0.0, 0.0, 0.0, 0.0],
    [0.0, 0.0, 1.0, 0.0, 0.0]
], dtype=tf.float32)

# Create probability distributions (batch_size x dim)
probs = tf.constant([
    [0.1, 0.2, 0.3, 0.25, 0.15],
    [0.05, 0.15, 0.4, 0.2, 0.2]
], dtype=tf.float32)

param = {'prob': probs}

ARGS = {
    'x': x,
    'param': param
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 1, 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)
CategoricalActor Ran
ran on a synthesized input · constructed an object; no result was computed
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("e05eecb1a5750a6c")
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, CategoricalActor: 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 tensorflow as tf
import numpy as np

ARGS = {
    'action_dim': 4,
    'epsilon': 0.1,
    'layer_dims': [8, 6],
    'default_activation': 'elu',
    'name': 'CategoricalActor'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type CategoricalActor — 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 CategoricalActor 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
ConfigLoader Ran
ran on a synthesized input
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("4d9828f13202e523")
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 ConfigParser. 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 configparser
import tempfile
import os

class ConfigLoader:
    """
    Global Config Loader helper class
    """
    parser = configparser.ConfigParser()

    @staticmethod
    def load(config_file):
        ConfigLoader.parser.read(config_file)
        return ConfigLoader.parser


# Create a temporary config file for testing
with tempfile.NamedTemporaryFile(mode='w', suffix='.ini', delete=False) as f:
    f.write('[section1]\n')
    f.write('key1 = value1\n')
    f.write('key2 = value2\n')
    f.write('[section2]\n')
    f.write('key3 = value3\n')
    config_file_path = f.name

try:
    INIT_ARGS = {}
    
    ARGS = {
        'config_file': config_file_path
    }
    
    METHOD = 'load'
    
    # Construct and call
    instance = ConfigLoader(**INIT_ARGS)
    result = instance.load(**ARGS)
    
finally:
    os.unlink(config_file_path)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 ConfigParser — a ConfigParser output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ConfigParser 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
Distribution Ran
ran on a synthesized input · constructed an object; no result was computed
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("6490dcf1e1570c31")
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, Distribution: 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)

# The Distribution class is abstract and takes only 'dim' as a constructor parameter
# We need to construct a concrete subclass to test it

class ConcreteDistribution(object):
    def __init__(self, dim):
        self._dim = dim
        self._tiny = 1e-08

    @property
    def dim(self):
        return self._dim

    def kl(self, old_dist, new_dist):
        """Compute the KL divergence of two distributions"""
        # Simple placeholder implementation
        return torch.tensor(0.0)

    def likelihood_ratio(self, x, old_dist, new_dist):
        # Simple placeholder implementation
        return torch.ones_like(x)

    def entropy(self, dist):
        # Simple placeholder implementation
        return torch.tensor(0.0)

    def log_likelihood_sym(self, x, dist):
        # Simple placeholder implementation
        return torch.zeros_like(x)

    def log_likelihood(self, xs, dist):
        # Simple placeholder implementation
        return torch.zeros(xs.shape[0])


# Create instances for testing
dim_size = 4

dist_instance = ConcreteDistribution(dim=dim_size)

ARGS = {
    "dim": dim_size
}
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 Distribution — 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 Distribution 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
GALA_A2C Ran
ran on a synthesized input · constructed an object; no result was computed
facebookresearch/gala/gala/gala_a2c.py
pointer only (licence: NOASSERTION) · get_code("aad00ea22af3c855")
How it ran

The body of this sample is withheld: GitHub could not classify the licence of this paper's repository, facebookresearch/gala (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

Learner Ran
ran on a synthesized input · constructed an object; no result was computed
seolhokim/DistributedRL-Pytorch-Ray/agents/runners/learners/impala_learner.py
code served (permissive licence) · get_code("3fdd0c84f214c2bb")
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, Learner: 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 torch.optim as optim

torch.manual_seed(0)

# Define a minimal algorithm class that satisfies the usage in Learner
class MinimalAlgorithm(nn.Module):
    def __init__(self, writer, device, state_dim, action_dim, agent_args):
        super().__init__()
        self.writer = writer
        self.device = device
        self.state_dim = state_dim
        self.action_dim = action_dim
        self.agent_args = agent_args
        # Add a simple parameter so the module has parameters for the optimizer
        self.linear = nn.Linear(state_dim, action_dim)
    
    def get_weights(self):
        return [p.clone() for p in self.parameters()]

# Define a minimal writer class
class MinimalWriter:
    def log(self, *args, **kwargs):
        pass

ARGS = {
    'algorithm': MinimalAlgorithm,
    'writer': MinimalWriter(),
    'device': torch.device('cpu'),
    'state_dim': 4,
    'action_dim': 3,
    'agent_args': {
        'lr': 0.001,
        'gamma': 0.99,
        'batch_size': 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 Learner — 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 Learner 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
V Ran
ran on a synthesized input · constructed an object; no result was computed
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("7d75b7a3afdfddc6")
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, V: 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 tensorflow as tf
import numpy as np

# Set seed for reproducibility
np.random.seed(0)
tf.random.set_seed(0)

# Define layer_dims - a list of integers representing hidden layer dimensions
layer_dims = [8, 6, 4]

# Define default_activation - a string specifying the activation function
default_activation = 'elu'

# Define name - a string for the model name
name = 'vf'

# Create the ARGS dictionary with all parameters
ARGS = {
    'layer_dims': layer_dims,
    'default_activation': default_activation,
    'name': name
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type V — 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 V 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
VTrace Ran
ran on a synthesized input · constructed an object; no result was computed
Sheepsody/Batched-Impala-PyTorch/src/IMPALA.py
pointer only (licence: NONE) · get_code("a7ffabde1fe6a3c9")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, Sheepsody/Batched-Impala-PyTorch when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

action_log_probs Ran
ran; our contract draft was wrong, not the code
facebookresearch/torchbeast/torchbeast/core/vtrace.py
code served (permissive licence) · get_code("815b225c5ab9f314")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn.functional as F

torch.manual_seed(0)

# policy_logits: should be a tensor that can be flattened and has a last dimension
# The function flattens it to 2D (batch, features), then applies log_softmax on dim=-1
# actions: should be a tensor of class indices that matches the flattened shape
# After flattening policy_logits to (N, C), actions should be (N,) with values in [0, C)

# Create policy_logits with shape (2, 3, 4)
# After flatten(0, -2) this becomes (6, 4) where 4 is the number of classes
policy_logits = torch.randn(2, 3, 4)

# Create actions with shape (2, 3) containing class indices in range [0, 4)
actions = torch.randint(0, 4, (2, 3))

ARGS = {
    "policy_logits": policy_logits,
    "actions": actions,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 3) — shape (2, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
from_importance_weights Ran
ran; our contract draft was wrong, not the code
facebookresearch/torchbeast/torchbeast/core/vtrace.py
code served (permissive licence) · get_code("123172de8a1bc477")
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 VTraceReturns. 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 collections

torch.manual_seed(0)

# Create small tensors for V-trace computation
# T = sequence length (time steps), B = batch size
T = 5  # time steps
B = 2  # batch size

# log_rhos: log importance weights, shape (T, B)
log_rhos = torch.randn(T, B)

# discounts: discount factors, shape (T, B)
discounts = torch.full((T, B), 0.99)

# rewards: rewards, shape (T, B)
rewards = torch.randn(T, B)

# values: value estimates, shape (T, B)
values = torch.randn(T, B)

# bootstrap_value: bootstrap value for final state, shape (B,)
bootstrap_value = torch.randn(B)

# clip_rho_threshold and clip_pg_rho_threshold are floats
clip_rho_threshold = 1.0
clip_pg_rho_threshold = 1.0

ARGS = {
    'log_rhos': log_rhos,
    'discounts': discounts,
    'rewards': rewards,
    'values': values,
    'bootstrap_value': bootstrap_value,
    'clip_rho_threshold': clip_rho_threshold,
    'clip_pg_rho_threshold': clip_pg_rho_threshold,
}
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 VTraceReturns — a VTraceReturns output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a VTraceReturns output the recorded ndim 3 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
from_logits Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
facebookresearch/torchbeast/torchbeast/core/vtrace.py
code served (permissive licence) · get_code("715b439b8761ea76")
How it ran

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

import torch
import torch.nn.functional as F

torch.manual_seed(0)

# Set up tensor dimensions
batch_size = 2
num_actions = 4
sequence_length = 3

# Create behavior policy logits: shape (sequence_length, batch_size, num_actions)
behavior_policy_logits = torch.randn(sequence_length, batch_size, num_actions)

# Create target policy logits: shape (sequence_length, batch_size, num_actions)
target_policy_logits = torch.randn(sequence_length, batch_size, num_actions)

# Create actions: shape (sequence_length, batch_size) with valid action indices
actions = torch.randint(0, num_actions, (sequence_length, batch_size))

# Create discounts: shape (sequence_length, batch_size)
discounts = torch.full((sequence_length, batch_size), 0.99)

# Create rewards: shape (sequence_length, batch_size)
rewards = torch.randn(sequence_length, batch_size)

# Create values: shape (sequence_length, batch_size)
values = torch.randn(sequence_length, batch_size)

# Create bootstrap_value: shape (batch_size,)
bootstrap_value = torch.randn(batch_size)

ARGS = {
    'behavior_policy_logits': behavior_policy_logits,
    'target_policy_logits': target_policy_logits,
    'actions': actions,
    'discounts': discounts,
    'rewards': rewards,
    'values': values,
    'bootstrap_value': bootstrap_value,
    'clip_rho_threshold': 1.0,
    'clip_pg_rho_threshold': 1.0,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type VTraceFromLogitsReturns — a VTraceFromLogitsReturns output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a VTraceFromLogitsReturns output the recorded ndim 3 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
log_probs_from_logits_and_actions Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
google-research/valan/framework/vtrace.py
code served (permissive licence) · get_code("a872fa719c5242ff")
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 EagerTensor, ndim 2, dtype family float, shape (3, 2) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tensorflow.compat.v2 as tf
import numpy as np

tf.enable_v2_behavior()

# Create small example tensors
# T=3 (time steps), B=2 (batch size), NUM_ACTIONS=4
T = 3
B = 2
NUM_ACTIONS = 4

# policy_logits: shape [T, B, NUM_ACTIONS] with random logits
policy_logits = tf.constant(
    np.random.randn(T, B, NUM_ACTIONS).astype(np.float32)
)

# actions: shape [T, B] with action indices in range [0, NUM_ACTIONS)
actions = tf.constant(
    np.random.randint(0, NUM_ACTIONS, size=(T, B)).astype(np.int32)
)

ARGS = {
    "policy_logits": policy_logits,
    "actions": actions,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 2, dtype family float, shape (3, 2) — shape (3, 2) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
log_probs_from_logits_and_actions Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
google-deepmind/scalable_agent/vtrace.py
code served (permissive licence) · get_code("a75de4dc4c61d7dc")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned EagerTensor, ndim 2, dtype family float, shape (3, 2) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tensorflow as tf
import numpy as np

# Set seed for reproducibility
tf.random.set_seed(0)
np.random.seed(0)

# Define dimensions
T = 3  # time dimension
B = 2  # batch size
NUM_ACTIONS = 4  # number of actions

# Create policy_logits: shape [T, B, NUM_ACTIONS]
policy_logits = tf.random.normal(shape=[T, B, NUM_ACTIONS], dtype=tf.float32)

# Create actions: shape [T, B] with valid action indices (0 to NUM_ACTIONS-1)
actions = tf.constant(
    np.random.randint(0, NUM_ACTIONS, size=[T, B]),
    dtype=tf.int32
)

ARGS = {
    "policy_logits": policy_logits,
    "actions": actions,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EagerTensor, ndim 2, dtype family float, shape (3, 2) — shape (3, 2) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
pairwise_cross_corr Ran
ran, honoured the contract we drafted
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("ba92bb44edd2ada6")
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 ndarray, ndim 2, dtype family int, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tensorflow as tf
import numpy as np
from sklearn.preprocessing import StandardScaler

# Create small test matrices
# La: shape (batch_size, n_features) where n_features = 4
# Lb: shape (batch_size, n_features) where n_features = 4
np.random.seed(0)

batch_size = 2
n_features = 4

La = np.random.randn(batch_size, n_features).astype(np.float32)
Lb = np.random.randn(batch_size, n_features).astype(np.float32)

ARGS = {
    "La": La,
    "Lb": Lb
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family int, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
safe_crossover Ran
ran, violated the contract we drafted
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("d9e1fb4c37c2e728")
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 ndarray, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# safe_crossover takes two array-like arguments (ax and bx)
# and performs a linear interpolation with a random parameter t
# The function uses numpy operations, so we need numpy arrays

np.random.seed(0)

# Create two small numpy arrays for crossover
ax = np.array([1.0, 2.0, 3.0, 4.0])
bx = np.array([5.0, 6.0, 7.0, 8.0])

ARGS = {
    "ax": ax,
    "bx": bx
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
vtrace Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
deepmind/rlax/rlax/_src/vtrace.py
code served (permissive licence) · get_code("e565e1e08272948c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned ArrayImpl, ndim 1, dtype family float, shape (2,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import jax
import jax.numpy as jnp
import chex

# Set seed for reproducibility
jax.config.update('jax_platform_name', 'cpu')

# Create small batch of data
batch_size = 2
seq_len = 4

# All arrays must be rank 1 (as per chex assertions in the function)
# and have matching shapes
v_tm1 = jnp.array([0.5, 0.3], dtype=jnp.float32)
v_t = jnp.array([0.6, 0.4], dtype=jnp.float32)
r_t = jnp.array([1.0, 0.5], dtype=jnp.float32)
discount_t = jnp.array([0.99, 0.99], dtype=jnp.float32)
rho_tm1 = jnp.array([1.2, 0.8], dtype=jnp.float32)

# lambda_ can be scalar or rank 1
lambda_ = 0.9

# clip_rho_threshold is a float
clip_rho_threshold = 1.0

# stop_target_gradients is a bool
stop_target_gradients = True

# sample_mask is optional, rank 1 bool array
sample_mask = jnp.array([True, True], dtype=jnp.bool_)

ARGS = {
    'v_tm1': v_tm1,
    'v_t': v_t,
    'r_t': r_t,
    'discount_t': discount_t,
    'rho_tm1': rho_tm1,
    'lambda_': lambda_,
    'clip_rho_threshold': clip_rho_threshold,
    'stop_target_gradients': stop_target_gradients,
    'sample_mask': sample_mask,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
chex, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (2,) — shape (2,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
AC Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("f34df61f30b311d5")
How it ran

no invocation recorded.

ActionStateProbs Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("103308756a178abe")
How it ran

no invocation recorded.

Default Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("ceddd134b3dd01f7")
How it ran

no invocation recorded.

Impala Not yet run
unverified (harvested, no recorded run)
Sheepsody/Batched-Impala-PyTorch/src/IMPALA.py
pointer only (licence: NONE) · get_code("cd20ae279ac1aa9f")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, Sheepsody/Batched-Impala-PyTorch when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

ImpalaLearner Not yet run
unverified (harvested, no recorded run)
seolhokim/DistributedRL-Pytorch-Ray/agents/runners/learners/impala_learner.py
code served (permissive licence) · get_code("b3a1158468bd53aa")
How it ran

no invocation recorded.

PlayerState Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("cbb6758024239ec3")
How it ran

no invocation recorded.

action_log_probs Not yet run
unverified (harvested, no recorded run)
urw7rs/spiralpp/torchbeast/core/vtrace.py
code served (permissive licence) · get_code("bc2ac55256ed974b")
How it ran

no invocation recorded.

calc_loss_and_backprop Not yet run
unverified (harvested, no recorded run)
haje01/impala/learner.py
code served (permissive licence) · get_code("cb89e31851cc2931")
How it ran

no invocation recorded.

corr_matrix Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("9f5db323eec0eeba")
How it ran

no invocation recorded.

from_importance_weights Not yet run
unverified (harvested, no recorded run)
google-research/valan/framework/vtrace.py
code served (permissive licence) · get_code("86f7ee994e64dc88")
How it ran

no invocation recorded.

from_importance_weights Not yet run
unverified (harvested, no recorded run)
google-deepmind/scalable_agent/vtrace.py
code served (permissive licence) · get_code("e11ee068514828e6")
How it ran

no invocation recorded.

from_logits Not yet run
unverified (harvested, no recorded run)
google-research/valan/framework/vtrace.py
code served (permissive licence) · get_code("c0ca8087e51d7e1d")
How it ran

no invocation recorded.

from_logits Not yet run
unverified (harvested, no recorded run)
urw7rs/spiralpp/torchbeast/core/vtrace.py
code served (permissive licence) · get_code("e26d17c678348c08")
How it ran

no invocation recorded.

from_logits Not yet run
unverified (harvested, no recorded run)
google-deepmind/scalable_agent/vtrace.py
code served (permissive licence) · get_code("2b5d90526b4ad312")
How it ran

no invocation recorded.

get_vtrace_and_advantages Not yet run
unverified (harvested, no recorded run)
seolhokim/SimpleDistributedRL/distributed_reinforcement_learning/algorithms/impala.py
code served (permissive licence) · get_code("b1527fc2ce13e346")
How it ran

no invocation recorded.

layer_crossover Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("a22618d583e622fe")
How it ran

no invocation recorded.

nn_crossover Not yet run
unverified (harvested, no recorded run)
villinvic/Georges/training/RL.py
code served (permissive licence) · get_code("e1569fbaca09c124")
How it ran

no invocation recorded.

train Not yet run
unverified (harvested, no recorded run)
seolhokim/SimpleDistributedRL/distributed_reinforcement_learning/algorithms/impala.py
code served (permissive licence) · get_code("01b62761de35ed90")
How it ran

no invocation recorded.

Also present in other papers' code

Of 34 samples lifted from this paper's linked repositories, 34 carry a function identity and 0 do not. Of the 32 distinct functions with an identity, 5 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 5 later) and 27 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
action_log_probs
brick of this paper's assembly: helper
1 other paper
1 later
2303.10158 (2023-03)
Data-centric Artificial Intelligence: A Survey
from_importance_weights
brick of this paper's assembly: helper
1 other paper
1 later
1910.03552 (2019-10)
TorchBeast: A PyTorch Platform for Distributed RL
from_logits
brick of this paper's assembly: entry
1 other paper
1 later
1910.03552 (2019-10)
TorchBeast: A PyTorch Platform for Distributed RL
from_logits
brick of this paper's assembly: entry
1 other paper
1 later
2303.10158 (2023-03)
Data-centric Artificial Intelligence: A Survey
log_probs_from_logits_and_actions
brick of this paper's assembly: helper
1 other paper
1 later
1910.03552 (2019-10)
TorchBeast: A PyTorch Platform for Distributed RL

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

In this work we aim to solve a large collection of tasks using a single reinforcement learning agent with a single set of parameters. A key challenge is to handle the increased amount of data and extended training time. We have developed a new distributed agent IMPALA (Importance Weighted Actor-Learner Architecture) that not only uses resources more efficiently in singlemachine training but also scales to thousands of machines without sacrificing data efficiency or resource utilisation. We achieve stable learning at high throughput by combining decoupled acting and learning with a novel off-policy correction method called V-trace. We demonstrate the effectiveness of IMPALA for multi-task reinforcement learning on DMLab-30 (a set of 30 tasks from the DeepMind Lab environment (Beattie et al., 2016)) and Atari-57 (all available Atari games in Arcade Learning Environment (Bellemare et al., 2013a)). Our results show that IMPALA is able to achieve better performance than previous agents with less data, and crucially exhibits positive transfer between tasks as a result of its multi-task approach. The source code is publicly available at github.com/deepmind/scalable agent.

For agents

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

get_harvested_code_for_paper("1802.01561")
get_code_for_paper("1802.01561")
have("1802.01561")

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

curl https://syntology.ai/api/ran/1802.01561.json

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

[![Syntology run record](https://syntology.ai/api/ran/1802.01561.svg)](https://syntology.ai/paper/1802.01561)

Connect an agent — have() is free.