Skip to content

Pass-through Wrappers

Stateless wrappers that transform observations, actions, or rewards without carrying any state between steps.

envrax.wrappers.jit_wrapper.JitWrapper

Bases: Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]

Wrap a JaxEnv so that reset and step are compiled with jax.jit on construction.

Parameters:

Name Type Description Default
env JaxEnv

Environment to wrap.

 required
cache_dir Path | str | None

Directory for the persistent XLA compilation cache. Defaults to ~/.cache/envrax/xla_cache. Pass None to disable.

 required
pre_warm bool

Run a dummy reset + step immediately to trigger XLA compilation. Set to False to defer compilation to the first real call or an explicit compile() call. Default is True.

 required
Source code in envrax/wrappers/jit_wrapper.py
Python
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
class JitWrapper(Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]):
    """
    Wrap a `JaxEnv` so that `reset` and `step` are compiled with
    `jax.jit` on construction.

    Parameters
    ----------
    env : JaxEnv
        Environment to wrap.
    cache_dir : Path | str | None (optional)
        Directory for the persistent XLA compilation cache.
        Defaults to `~/.cache/envrax/xla_cache`. Pass `None` to disable.
    pre_warm : bool (optional)
        Run a dummy `reset` + `step` immediately to trigger XLA compilation.
        Set to `False` to defer compilation to the first real call or an
        explicit `compile()` call. Default is `True`.
    """

    def __init__(
        self,
        env: JaxEnv[ObsSpaceT, ActSpaceT, StateT, ConfigT],
        cache_dir: pathlib.Path | str | None = DEFAULT_CACHE_DIR,
        *,
        pre_warm: bool = True,
    ) -> None:
        super().__init__(env)
        setup_cache(cache_dir)

        self._jit_reset = jax.jit(env.reset)
        self._jit_step = jax.jit(env.step)

        if pre_warm:
            self.compile()

    def compile(self) -> None:
        """
        Trigger XLA compilation by running a dummy `reset` + `step`.

        Safe to call multiple times — subsequent calls are near-instant
        because JAX caches the compiled kernels in memory.
        """
        _key = jax.random.key(0)
        _, _state = self._jit_reset(_key)
        self._jit_step(_state, self._env.action_space.sample(_key))

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        return self._jit_reset(rng)

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        return self._jit_step(state, action)

compile()

Trigger XLA compilation by running a dummy reset + step.

Safe to call multiple times — subsequent calls are near-instant because JAX caches the compiled kernels in memory.

Source code in envrax/wrappers/jit_wrapper.py
Python
46
47
48
49
50
51
52
53
54
55
def compile(self) -> None:
    """
    Trigger XLA compilation by running a dummy `reset` + `step`.

    Safe to call multiple times — subsequent calls are near-instant
    because JAX caches the compiled kernels in memory.
    """
    _key = jax.random.key(0)
    _, _state = self._jit_reset(_key)
    self._jit_step(_state, self._env.action_space.sample(_key))

envrax.wrappers.clip_reward.ClipReward

Bases: Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]

Clip rewards to the sign of the reward: {−1, 0, +1}.

Parameters:

Name Type Description Default
env JaxEnv

Inner environment to wrap.

 required
Source code in envrax/wrappers/clip_reward.py
Python
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
class ClipReward(Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]):
    """
    Clip rewards to the sign of the reward: `{−1, 0, +1}`.

    Parameters
    ----------
    env : JaxEnv
        Inner environment to wrap.
    """

    def __init__(self, env: JaxEnv[ObsSpaceT, ActSpaceT, StateT, ConfigT]) -> None:
        super().__init__(env)

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        return self._env.reset(rng)

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Advance the environment by one step and clip the reward to `{−1, 0, +1}`.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs  : chex.Array
            Observation from the inner step
        new_state : StateT
            Updated environment state
        reward  : chex.Array
            Reward clipped to sign: `{−1.0, 0.0, +1.0}`
        done  : chex.Array
            Terminal flag from the inner step
        info : Dict[str, Any]
            Info dict from the inner step
        """
        obs, new_state, reward, done, info = self._env.step(state, action)
        reward = jnp.sign(reward).astype(jnp.float32)
        return obs, new_state, reward, done, info

step(state, action)

Advance the environment by one step and clip the reward to {−1, 0, +1}.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Observation from the inner step
new_state StateT

Updated environment state
reward Array

Reward clipped to sign: {−1.0, 0.0, +1.0}
done Array

Terminal flag from the inner step
info Dict[str, Any]

Info dict from the inner step
Source code in envrax/wrappers/clip_reward.py
Python
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Advance the environment by one step and clip the reward to `{−1, 0, +1}`.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs  : chex.Array
        Observation from the inner step
    new_state : StateT
        Updated environment state
    reward  : chex.Array
        Reward clipped to sign: `{−1.0, 0.0, +1.0}`
    done  : chex.Array
        Terminal flag from the inner step
    info : Dict[str, Any]
        Info dict from the inner step
    """
    obs, new_state, reward, done, info = self._env.step(state, action)
    reward = jnp.sign(reward).astype(jnp.float32)
    return obs, new_state, reward, done, info

envrax.wrappers.discount.EpisodeDiscount

Bases: Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]

Convert the boolean done signal to a float32 episode discount.

The 4th return value of step() changes from bool to float32: 1.0 while the episode is running, 0.0 on termination.

Parameters:

Name Type Description Default
env JaxEnv

Inner environment to wrap.

 required
Source code in envrax/wrappers/discount.py
Python
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
class EpisodeDiscount(Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]):
    """
    Convert the boolean `done` signal to a float32 episode discount.

    The 4th return value of `step()` changes from `bool` to `float32`:
    `1.0` while the episode is running, `0.0` on termination.

    Parameters
    ----------
    env : JaxEnv
        Inner environment to wrap.
    """

    def __init__(self, env: JaxEnv[ObsSpaceT, ActSpaceT, StateT, ConfigT]) -> None:
        super().__init__(env)

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        return self._env.reset(rng)

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Advance the environment and return a float32 discount instead of done.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs  : chex.Array
            Observation from the inner step
        new_state : StateT
            Updated environment state
        reward  : chex.Array
            Reward from the inner step (unchanged)
        discount  : chex.Array
            `1.0` if the episode continues, `0.0` if it terminated
        info : Dict[str, Any]
            Info dict from the inner step
        """
        obs, new_state, reward, done, info = self._env.step(state, action)
        discount = jnp.where(done, jnp.float32(0.0), jnp.float32(1.0))
        return obs, new_state, reward, discount, info

step(state, action)

Advance the environment and return a float32 discount instead of done.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Observation from the inner step
new_state StateT

Updated environment state
reward Array

Reward from the inner step (unchanged)
discount Array

1.0 if the episode continues, 0.0 if it terminated
info Dict[str, Any]

Info dict from the inner step
Source code in envrax/wrappers/discount.py
Python
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Advance the environment and return a float32 discount instead of done.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs  : chex.Array
        Observation from the inner step
    new_state : StateT
        Updated environment state
    reward  : chex.Array
        Reward from the inner step (unchanged)
    discount  : chex.Array
        `1.0` if the episode continues, `0.0` if it terminated
    info : Dict[str, Any]
        Info dict from the inner step
    """
    obs, new_state, reward, done, info = self._env.step(state, action)
    discount = jnp.where(done, jnp.float32(0.0), jnp.float32(1.0))
    return obs, new_state, reward, discount, info

envrax.wrappers.expand_dims.ExpandDims

Bases: Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]

Add a trailing size-1 dimension to reward and done.

Transforms scalar outputs from step() so that reward and done have shape (..., 1) instead of (...).

Parameters:

Name Type Description Default
env JaxEnv

Inner environment to wrap.

 required
Source code in envrax/wrappers/expand_dims.py
Python
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
class ExpandDims(Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]):
    """
    Add a trailing size-1 dimension to `reward` and `done`.

    Transforms scalar outputs from `step()` so that `reward` and `done`
    have shape `(..., 1)` instead of `(...)`.

    Parameters
    ----------
    env : JaxEnv
        Inner environment to wrap.
    """

    def __init__(self, env: JaxEnv[ObsSpaceT, ActSpaceT, StateT, ConfigT]) -> None:
        super().__init__(env)

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        return self._env.reset(rng)

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Advance the environment and expand `reward` and `done`.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs  : chex.Array
            Observation after the step (unchanged)
        new_state : StateT
            Updated environment state
        reward  : chex.Array
            Reward with a trailing size-1 dimension
        done  : chex.Array
            Terminal flag with a trailing size-1 dimension
        info : Dict[str, Any]
            Auxiliary info dict (unchanged)
        """
        obs, new_state, reward, done, info = self._env.step(state, action)
        return (
            obs,
            new_state,
            jnp.expand_dims(reward, -1),
            jnp.expand_dims(done, -1),
            info,
        )

step(state, action)

Advance the environment and expand reward and done.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Observation after the step (unchanged)
new_state StateT

Updated environment state
reward Array

Reward with a trailing size-1 dimension
done Array

Terminal flag with a trailing size-1 dimension
info Dict[str, Any]

Auxiliary info dict (unchanged)
Source code in envrax/wrappers/expand_dims.py
Python
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Advance the environment and expand `reward` and `done`.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs  : chex.Array
        Observation after the step (unchanged)
    new_state : StateT
        Updated environment state
    reward  : chex.Array
        Reward with a trailing size-1 dimension
    done  : chex.Array
        Terminal flag with a trailing size-1 dimension
    info : Dict[str, Any]
        Auxiliary info dict (unchanged)
    """
    obs, new_state, reward, done, info = self._env.step(state, action)
    return (
        obs,
        new_state,
        jnp.expand_dims(reward, -1),
        jnp.expand_dims(done, -1),
        info,
    )

envrax.wrappers.grayscale.GrayscaleObservation

Bases: Wrapper[Box, ActSpaceT, StateT, ConfigT]

Convert RGB observations to grayscale using the NTSC luminance formula.

Wraps any environment whose reset / step return uint8[H, W, 3] observations and converts them to uint8[H, W].

Parameters:

Name Type Description Default
env JaxEnv

Inner environment to wrap. Must have a Box observation space of shape (H, W, 3) and dtype uint8.

 required
Source code in envrax/wrappers/grayscale.py
Python
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
class GrayscaleObservation(Wrapper[Box, ActSpaceT, StateT, ConfigT]):
    """
    Convert RGB observations to grayscale using the NTSC luminance formula.

    Wraps any environment whose `reset` / `step` return `uint8[H, W, 3]`
    observations and converts them to `uint8[H, W]`.

    Parameters
    ----------
    env : JaxEnv
        Inner environment to wrap. Must have a `Box` observation space
        of shape `(H, W, 3)` and dtype `uint8`.
    """

    def __init__(self, env: JaxEnv[Box, ActSpaceT, StateT, ConfigT]) -> None:
        super().__init__(env)
        require_box(
            env,
            type(self).__name__,
            rank=3,
            last_dim=3,
            dtype=jnp.uint8,
        )

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        """
        Reset the inner environment and convert the observation to grayscale.

        Parameters
        ----------
        rng : chex.PRNGKey
            JAX PRNG key

        Returns
        -------
        obs  : chex.Array
            Grayscale observation
        state : StateT
            Inner environment state
        """
        obs, state = self._env.reset(rng)
        return to_gray(obs), state

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Step the inner environment and convert the observation to grayscale.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs  : chex.Array
            Grayscale observation
        new_state : StateT
            Updated environment state
        reward  : chex.Array
            Reward from the inner step
        done  : chex.Array
            Terminal flag from the inner step
        info : Dict[str, Any]
            Info dict from the inner step
        """
        obs, new_state, reward, done, info = self._env.step(state, action)
        return to_gray(obs), new_state, reward, done, info

    @property
    def observation_space(self) -> Box:
        inner = self._env.observation_space
        h, w = inner.shape[0], inner.shape[1]
        return Box(low=inner.low, high=inner.high, shape=(h, w), dtype=inner.dtype)

reset(rng)

Reset the inner environment and convert the observation to grayscale.

Parameters:

Name Type Description Default
rng PRNGKey

JAX PRNG key

 required

Returns:

Name Type Description
obs Array

Grayscale observation
state StateT

Inner environment state
Source code in envrax/wrappers/grayscale.py
Python
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
    """
    Reset the inner environment and convert the observation to grayscale.

    Parameters
    ----------
    rng : chex.PRNGKey
        JAX PRNG key

    Returns
    -------
    obs  : chex.Array
        Grayscale observation
    state : StateT
        Inner environment state
    """
    obs, state = self._env.reset(rng)
    return to_gray(obs), state

step(state, action)

Step the inner environment and convert the observation to grayscale.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Grayscale observation
new_state StateT

Updated environment state
reward Array

Reward from the inner step
done Array

Terminal flag from the inner step
info Dict[str, Any]

Info dict from the inner step
Source code in envrax/wrappers/grayscale.py
Python
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Step the inner environment and convert the observation to grayscale.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs  : chex.Array
        Grayscale observation
    new_state : StateT
        Updated environment state
    reward  : chex.Array
        Reward from the inner step
    done  : chex.Array
        Terminal flag from the inner step
    info : Dict[str, Any]
        Info dict from the inner step
    """
    obs, new_state, reward, done, info = self._env.step(state, action)
    return to_gray(obs), new_state, reward, done, info

envrax.wrappers.normalize_obs.NormalizeObservation

Bases: Wrapper[Box, ActSpaceT, StateT, ConfigT]

Normalises pixel observations from uint8 [0, 255] to float32 [0, 1].

Divides observations by 255.0 and casts to float32.

Parameters:

Name Type Description Default
env JaxEnv

Environment to wrap. Must have a Box observation space with dtype uint8.

 required
Source code in envrax/wrappers/normalize_obs.py
Python
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
class NormalizeObservation(Wrapper[Box, ActSpaceT, StateT, ConfigT]):
    """
    Normalises pixel observations from `uint8 [0, 255]` to `float32 [0, 1]`.

    Divides observations by 255.0 and casts to float32.

    Parameters
    ----------
    env : JaxEnv
        Environment to wrap. Must have a `Box` observation space with
        dtype `uint8`.
    """

    def __init__(self, env: JaxEnv[Box, ActSpaceT, StateT, ConfigT]) -> None:
        super().__init__(env)
        require_box(env, type(self).__name__, dtype=jnp.uint8)

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        """
        Reset and return a normalised initial observation.

        Parameters
        ----------
        rng : chex.PRNGKey
            JAX PRNG key

        Returns
        -------
        obs  : chex.Array
            Normalised observation in `[0, 1]`
        state : StateT
            Inner environment state
        """
        obs, state = self._env.reset(rng)
        return obs.astype(jnp.float32) / jnp.float32(255.0), state

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Step and return a normalised observation.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs  : chex.Array
            Normalised observation in `[0, 1]`
        new_state : StateT
            Updated environment state
        reward  : chex.Array
            Step reward
        done  : chex.Array
            Terminal flag
        info : Dict[str, Any]
            Environment metadata
        """
        obs, new_state, reward, done, info = self._env.step(state, action)
        return (
            obs.astype(jnp.float32) / jnp.float32(255.0),
            new_state,
            reward,
            done,
            info,
        )

    @property
    def observation_space(self) -> Box:
        inner = self._env.observation_space
        return Box(low=0.0, high=1.0, shape=inner.shape, dtype=jnp.float32)

reset(rng)

Reset and return a normalised initial observation.

Parameters:

Name Type Description Default
rng PRNGKey

JAX PRNG key

 required

Returns:

Name Type Description
obs Array

Normalised observation in [0, 1]
state StateT

Inner environment state
Source code in envrax/wrappers/normalize_obs.py
Python
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
    """
    Reset and return a normalised initial observation.

    Parameters
    ----------
    rng : chex.PRNGKey
        JAX PRNG key

    Returns
    -------
    obs  : chex.Array
        Normalised observation in `[0, 1]`
    state : StateT
        Inner environment state
    """
    obs, state = self._env.reset(rng)
    return obs.astype(jnp.float32) / jnp.float32(255.0), state

step(state, action)

Step and return a normalised observation.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Normalised observation in [0, 1]
new_state StateT

Updated environment state
reward Array

Step reward
done Array

Terminal flag
info Dict[str, Any]

Environment metadata
Source code in envrax/wrappers/normalize_obs.py
Python
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Step and return a normalised observation.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs  : chex.Array
        Normalised observation in `[0, 1]`
    new_state : StateT
        Updated environment state
    reward  : chex.Array
        Step reward
    done  : chex.Array
        Terminal flag
    info : Dict[str, Any]
        Environment metadata
    """
    obs, new_state, reward, done, info = self._env.step(state, action)
    return (
        obs.astype(jnp.float32) / jnp.float32(255.0),
        new_state,
        reward,
        done,
        info,
    )

envrax.wrappers.resize.ResizeObservation

Bases: Wrapper[Box, ActSpaceT, StateT, ConfigT]

Resize observations to (h, w) using bilinear interpolation.

Handles both:

  • Grayscale — uint8[H, W] → uint8[h, w]
  • RGB — uint8[H, W, C] → uint8[h, w, C]

The channel dimension is preserved automatically; no pre-processing step is required. For DQN-style pipelines, apply GrayscaleObservation first so the output is uint8[h, w] before stacking.

Parameters:

Name Type Description Default
env JaxEnv

Inner environment returning uint8[H, W] or uint8[H, W, C] observations.

 required
h int

Output height in pixels. Default is 84.

 required
w int

Output width in pixels. Default is 84.

 required
Source code in envrax/wrappers/resize.py
Python
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
class ResizeObservation(Wrapper[Box, ActSpaceT, StateT, ConfigT]):
    """
    Resize observations to `(h, w)` using bilinear interpolation.

    Handles both:

    - **Grayscale** — `uint8[H, W]` → `uint8[h, w]`
    - **RGB** — `uint8[H, W, C]` → `uint8[h, w, C]`

    The channel dimension is preserved automatically; no pre-processing step
    is required.  For DQN-style pipelines, apply `GrayscaleObservation` first
    so the output is `uint8[h, w]` before stacking.

    Parameters
    ----------
    env : JaxEnv
        Inner environment returning `uint8[H, W]` or `uint8[H, W, C]` observations.
    h : int (optional)
        Output height in pixels. Default is `84`.
    w : int (optional)
        Output width in pixels. Default is `84`.
    """

    def __init__(
        self,
        env: JaxEnv[Box, ActSpaceT, StateT, ConfigT],
        *,
        h: int = 84,
        w: int = 84,
    ) -> None:
        super().__init__(env)
        require_box(env, type(self).__name__, rank=(2, 3), dtype=jnp.uint8)
        self._h = h
        self._w = w

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        """
        Reset the inner environment and resize the observation.

        Parameters
        ----------
        rng : chex.PRNGKey
            JAX PRNG key

        Returns
        -------
        obs  : chex.Array
            Resized observation
        state : StateT
            Inner environment state
        """
        obs, state = self._env.reset(rng)
        return resize(obs, self._h, self._w), state

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Step the inner environment and resize the observation.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs  : chex.Array
            Resized observation
        new_state : StateT
            Updated environment state
        reward  : chex.Array
            Reward from the inner step
        done  : chex.Array
            Terminal flag from the inner step
        info : Dict[str, Any]
            Info dict from the inner step
        """
        obs, new_state, reward, done, info = self._env.step(state, action)
        return resize(obs, self._h, self._w), new_state, reward, done, info

    @property
    def observation_space(self) -> Box:
        inner = self._env.observation_space

        if len(inner.shape) == 3:
            shape = (self._h, self._w, inner.shape[-1])
        else:
            shape = (self._h, self._w)

        return Box(low=inner.low, high=inner.high, shape=shape, dtype=inner.dtype)

reset(rng)

Reset the inner environment and resize the observation.

Parameters:

Name Type Description Default
rng PRNGKey

JAX PRNG key

 required

Returns:

Name Type Description
obs Array

Resized observation
state StateT

Inner environment state
Source code in envrax/wrappers/resize.py
Python
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
    """
    Reset the inner environment and resize the observation.

    Parameters
    ----------
    rng : chex.PRNGKey
        JAX PRNG key

    Returns
    -------
    obs  : chex.Array
        Resized observation
    state : StateT
        Inner environment state
    """
    obs, state = self._env.reset(rng)
    return resize(obs, self._h, self._w), state

step(state, action)

Step the inner environment and resize the observation.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Resized observation
new_state StateT

Updated environment state
reward Array

Reward from the inner step
done Array

Terminal flag from the inner step
info Dict[str, Any]

Info dict from the inner step
Source code in envrax/wrappers/resize.py
Python
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Step the inner environment and resize the observation.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs  : chex.Array
        Resized observation
    new_state : StateT
        Updated environment state
    reward  : chex.Array
        Reward from the inner step
    done  : chex.Array
        Terminal flag from the inner step
    info : Dict[str, Any]
        Info dict from the inner step
    """
    obs, new_state, reward, done, info = self._env.step(state, action)
    return resize(obs, self._h, self._w), new_state, reward, done, info

envrax.wrappers.record_video.RecordVideo

Bases: Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]

Save episode frames to MP4 based on configurable triggers.

Not JIT/vmap-compatible. Intended for evaluation, logging, and training visualisation.

Three optional triggers control when recording is active. They are OR'd together — if any trigger returns True, that episode is recorded. When no triggers are provided, every episode is recorded.

Each completed recording is written to <output_dir>/episode_<NNNN>.mp4 via imageio.

Requires imageio with the ffmpeg plugin (pip install "imageio[ffmpeg]").

Parameters:

Name Type Description Default
env JaxEnv

Inner environment to wrap that has a render() method.

 required
output_dir str | Path

Directory where MP4 files are saved. Created automatically if it does not exist. Default is runs/recordings

 required
fps int

Frames per second for the saved video. Default is 30.

 required
episode_trigger Callable[[int], bool]

Called with the episode count at each reset(). If True, record this episode. Useful for "record every Nth episode". Default is None

 required
step_trigger Callable[[int], bool]

Called with the global step count at each step(). If True, start recording from this step until the episode ends. Default is None

 required
recording_trigger Callable[[], bool]

Zero-arg callable checked at each reset(). If True, record this episode. Useful for meta-learning where the framework controls when to record via an external flag. Default is None

 required

Raises:

Name Type Description
render_missing TypeError

If the unwrapped environment does not implement render().
Source code in envrax/wrappers/record_video.py
Python
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
class RecordVideo(Wrapper[ObsSpaceT, ActSpaceT, StateT, ConfigT]):
    """
    Save episode frames to MP4 based on configurable triggers.

    **Not JIT/vmap-compatible.** Intended for evaluation, logging, and
    training visualisation.

    Three optional triggers control when recording is active. They are
    OR'd together — if any trigger returns `True`, that episode is
    recorded. When no triggers are provided, every episode is recorded.

    Each completed recording is written to
    `<output_dir>/episode_<NNNN>.mp4` via `imageio`.

    Requires `imageio` with the `ffmpeg` plugin
    (`pip install "imageio[ffmpeg]"`).

    Parameters
    ----------
    env : JaxEnv
        Inner environment to wrap that has a `render()` method.
    output_dir : str | Path (optional)
        Directory where MP4 files are saved. Created automatically if
        it does not exist. Default is `runs/recordings`
    fps : int (optional)
        Frames per second for the saved video. Default is `30`.
    episode_trigger : Callable[[int], bool] (optional)
        Called with the episode count at each `reset()`. If `True`,
        record this episode. Useful for "record every Nth episode".
        Default is `None`
    step_trigger : Callable[[int], bool] (optional)
        Called with the global step count at each `step()`. If `True`,
        start recording from this step until the episode ends.
        Default is `None`
    recording_trigger : Callable[[], bool] (optional)
        Zero-arg callable checked at each `reset()`. If `True`, record
        this episode. Useful for meta-learning where the framework
        controls when to record via an external flag.
        Default is `None`

    Raises
    ------
    render_missing : TypeError
        If the unwrapped environment does not implement `render()`.
    """

    def __init__(
        self,
        env: JaxEnv[ObsSpaceT, ActSpaceT, StateT, ConfigT],
        *,
        output_dir: str | Path = "runs/recordings",
        fps: int = 30,
        episode_trigger: Callable[[int], bool] | None = None,
        step_trigger: Callable[[int], bool] | None = None,
        recording_trigger: Callable[[], bool] | None = None,
    ) -> None:
        super().__init__(env)

        # Fail fast if the env doesn't support rendering
        if type(self.unwrapped).render is JaxEnv.render:
            raise TypeError(
                f"RecordVideo requires an environment that implements render(). "
                f"{type(self.unwrapped).__name__} does not."
            )

        self.output_dir = Path(output_dir)
        self.fps = fps
        self.output_dir.mkdir(parents=True, exist_ok=True)

        self._episode_trigger = episode_trigger
        self._step_trigger = step_trigger
        self._recording_trigger = recording_trigger
        self._has_triggers = any(
            t is not None for t in (episode_trigger, step_trigger, recording_trigger)
        )

        self._frames: List[np.ndarray] = []
        self._episode_id: int = 0
        self._global_step: int = 0
        self._recording: bool = False

    @property
    def recording(self) -> bool:
        """Whether the current episode is being recorded."""
        return self._recording

    def _should_record_episode(self) -> bool:
        """Check episode-level triggers (OR logic)."""
        if not self._has_triggers:
            return True  # no triggers → record everything

        if self._episode_trigger is not None and self._episode_trigger(
            self._episode_id
        ):
            return True

        if self._recording_trigger is not None and self._recording_trigger():
            return True

        return False

    def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
        """
        Reset the environment and optionally begin a new recording.

        Parameters
        ----------
        rng : chex.PRNGKey
            JAX PRNG key

        Returns
        -------
        obs : chex.Array
            First observation
        state : StateT
            Initial environment state
        """
        _ensure_not_traced(rng)
        obs, state = self._env.reset(rng)

        self._recording = self._should_record_episode()

        if self._recording:
            self._frames = [np.asarray(self._env.render(state))]
        else:
            self._frames = []

        return obs, state

    def step(
        self,
        state: StateT,
        action: chex.Array,
    ) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
        """
        Advance the environment by one step and record the frame if active.

        If a `step_trigger` is provided and fires, recording starts
        mid-episode and continues until the episode ends.

        Flushes accumulated frames to an MP4 file when `done` is `True`.

        Parameters
        ----------
        state : StateT
            Current environment state
        action : chex.Array
            Action to take in the environment

        Returns
        -------
        obs : chex.Array
            Observation after the step
        new_state : StateT
            Updated environment state
        reward : chex.Array
            Reward for this step
        done : chex.Array
            `True` when the episode has ended
        info : Dict[str, Any]
            Pass-through info dict from the inner environment
        """
        _ensure_not_traced(action)
        obs, new_state, reward, done, info = self._env.step(state, action)
        self._global_step += 1

        # Mid-episode trigger: start recording if step_trigger fires
        if (
            not self._recording
            and self._step_trigger is not None
            and self._step_trigger(self._global_step)
        ):
            self._recording = True

        if self._recording:
            self._frames.append(np.asarray(self._env.render(new_state)))

        if bool(done) and self._recording:
            self._flush()
            self._recording = False

        if bool(done):
            self._episode_id += 1

        return obs, new_state, reward, done, info

    def _flush(self) -> None:
        """Write accumulated frames to an MP4 file."""
        if not self._frames:
            return

        try:
            import imageio
        except ImportError as exc:
            raise ImportError(
                "imageio is required for video recording. "
                'Install it with: pip install "imageio[ffmpeg]"'
            ) from exc

        path = self.output_dir / f"episode_{self._episode_id:04d}.mp4"
        frames: list[Any] = self._frames
        imageio.mimwrite(str(path), frames, fps=self.fps)
        self._frames = []

recording property

Whether the current episode is being recorded.

reset(rng)

Reset the environment and optionally begin a new recording.

Parameters:

Name Type Description Default
rng PRNGKey

JAX PRNG key

 required

Returns:

Name Type Description
obs Array

First observation
state StateT

Initial environment state
Source code in envrax/wrappers/record_video.py
Python
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
def reset(self, rng: chex.PRNGKey) -> Tuple[chex.Array, StateT]:
    """
    Reset the environment and optionally begin a new recording.

    Parameters
    ----------
    rng : chex.PRNGKey
        JAX PRNG key

    Returns
    -------
    obs : chex.Array
        First observation
    state : StateT
        Initial environment state
    """
    _ensure_not_traced(rng)
    obs, state = self._env.reset(rng)

    self._recording = self._should_record_episode()

    if self._recording:
        self._frames = [np.asarray(self._env.render(state))]
    else:
        self._frames = []

    return obs, state

step(state, action)

Advance the environment by one step and record the frame if active.

If a step_trigger is provided and fires, recording starts mid-episode and continues until the episode ends.

Flushes accumulated frames to an MP4 file when done is True.

Parameters:

Name Type Description Default
state StateT

Current environment state

 required
action Array

Action to take in the environment

 required

Returns:

Name Type Description
obs Array

Observation after the step
new_state StateT

Updated environment state
reward Array

Reward for this step
done Array

True when the episode has ended
info Dict[str, Any]

Pass-through info dict from the inner environment
Source code in envrax/wrappers/record_video.py
Python
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
def step(
    self,
    state: StateT,
    action: chex.Array,
) -> Tuple[chex.Array, StateT, chex.Array, chex.Array, Dict[str, Any]]:
    """
    Advance the environment by one step and record the frame if active.

    If a `step_trigger` is provided and fires, recording starts
    mid-episode and continues until the episode ends.

    Flushes accumulated frames to an MP4 file when `done` is `True`.

    Parameters
    ----------
    state : StateT
        Current environment state
    action : chex.Array
        Action to take in the environment

    Returns
    -------
    obs : chex.Array
        Observation after the step
    new_state : StateT
        Updated environment state
    reward : chex.Array
        Reward for this step
    done : chex.Array
        `True` when the episode has ended
    info : Dict[str, Any]
        Pass-through info dict from the inner environment
    """
    _ensure_not_traced(action)
    obs, new_state, reward, done, info = self._env.step(state, action)
    self._global_step += 1

    # Mid-episode trigger: start recording if step_trigger fires
    if (
        not self._recording
        and self._step_trigger is not None
        and self._step_trigger(self._global_step)
    ):
        self._recording = True

    if self._recording:
        self._frames.append(np.asarray(self._env.render(new_state)))

    if bool(done) and self._recording:
        self._flush()
        self._recording = False

    if bool(done):
        self._episode_id += 1

    return obs, new_state, reward, done, info