Table of Contents

Class Transform

Namespace
Gym.Rendering
Assembly
Gym.Rendering.dll

The part of pygame.transform Gymnasium's envs use: Flip(Surface, bool, bool), which turns a y-up drawing into a frame and mirrors sprites, and Scale(Surface, int, int)/Smoothscale(Surface, int, int), which size the toy_text tiles and Pendulum's torque arrow.

public static class Transform
Inheritance
Transform
Inherited Members

Remarks

Scale(Surface, int, int) and Smoothscale(Surface, int, int) are Gym.NET's own implementations of pygame's behaviour (decision D4: no pygame code is used), derived from and proven against live pygame-ce 2.5.8 pixel for pixel, alpha included.

Reference: src_c/transform.c (pygame-ce 2.5.8).

Methods

Flip(Surface, bool, bool)

pygame.transform.flip(surface, flip_x, flip_y): a new surface, mirrored left-right and/or top-bottom. The source is unchanged. Gymnasium's classic envs draw in a y-up world and flip vertically before returning a frame; Taxi and Pendulum mirror their sprites. Per-pixel alpha is kept.

public static Surface Flip(Surface surface, bool flipX, bool flipY)

Parameters

surface Surface

The surface to mirror.

flipX bool

Mirror the columns (left becomes right).

flipY bool

Mirror the rows (top becomes bottom).

Returns

Surface

The mirrored copy.

Exceptions

ArgumentNullException

Thrown when surface is null.

Scale(Surface, int, int)

pygame.transform.scale(surface, (width, height)): a new surface of the given size, each pixel a copy of one source pixel (nearest neighbour, no blending), per-pixel alpha kept. Gymnasium's toy_text envs size their tiles and sprites with it.

public static Surface Scale(Surface surface, int width, int height)

Parameters

surface Surface

The surface to scale.

width int

The new width; 0 gives an empty surface. pygame truncates a float size (int()), as callers must.

height int

The new height; 0 gives an empty surface.

Returns

Surface

The scaled copy.

Remarks

Destination column x takes source column (x * step + step / 2) >> 16 with step = (sourceWidth << 16) / width in integer arithmetic, and rows likewise: sampling at the middle of each 16.16 fixed-point step, which is what pygame-ce 2.5.8 picks (measured on 2,011 size pairs, shrinking and growing).

Exceptions

ArgumentNullException

Thrown when surface is null.

ValueError

Thrown for a negative size, with pygame's message (Cannot scale to negative size).

Smoothscale(Surface, int, int)

pygame.transform.smoothscale(surface, (width, height)) for a shrinking size: a new surface whose pixels average the source pixels they cover (area averaging), each channel of each pixel separately, per-pixel alpha included. Pendulum shrinks its torque arrow with it every frame.

public static Surface Smoothscale(Surface surface, int width, int height)

Parameters

surface Surface

The surface to scale.

width int

The new width, at most the source's; 0 gives an empty surface. pygame truncates a float size (int()).

height int

The new height, at most the source's; 0 gives an empty surface.

Returns

Surface

The scaled copy.

Remarks

The filter, as measured against pygame-ce 2.5.8 (exact on 800 random rows and 300 random images, shrinking one or both axes): the horizontal pass runs first (every row narrowed), then the vertical one, each pass writing 8-bit channels. Along an axis of n pixels shrunk to d, a destination pixel spans span = 0x4000 * n / d source pixels in 2.14 fixed point; a whole source pixel adds its value, the pixel straddling a boundary adds (v * w) >> 14 to each side for its 14-bit weight w, and a finished pixel is (sum * (2^30 / span)) >> 16. The 14-bit weights and the truncations are what make the result differ from an exact average (a constant 255 row shrinks to 254), and they are kept on purpose.

Exceptions

ArgumentNullException

Thrown when surface is null.

ValueError

Thrown for a negative size, with pygame's message (Cannot scale to negative size).

NotSupportedException

Thrown when either dimension grows: pygame expands with a different (interpolating) filter, which no env in this release needs, so it is deferred rather than approximated.