stemjara commited on
Commit
bbd8497
·
verified ·
1 Parent(s): 50b8725

Card: continuous time-of-day test set (hora_continua/)

Browse files
Files changed (1) hide show
  1. README.md +25 -2
README.md CHANGED
@@ -32,12 +32,17 @@ world, and for 16 363 of them the **same frame rendered with the Sundial Lite sh
32
  were captured with a Fabric mod that teleports the player through a plan of poses with the world
33
  frozen (no weather, fixed time of day).
34
 
 
 
 
35
  Capture tool (the mod, its configuration, the shader options and the contribution validator):
36
  **<https://github.com/stevenijm777/nrmc-capture>**
37
 
38
  *Resumen en español: 23 828 fotogramas de Minecraft 1.21.1 sin shaders, con profundidad, cámara,
39
  bloque por píxel y la sombra del sol trazada en el mundo guardado; 16 363 de ellos también con el
40
- shader Sundial Lite, para aprender el paso de vanilla a shader.*
 
 
41
 
42
  ## Paired renders: Sundial Lite
43
 
@@ -91,6 +96,18 @@ pose), so they were moved to **train**; the second-round test sets are `plains_2
91
  `desert_2_test` and `snowy_plains_2_test`. `swamp_2` has only its training set here. Nothing of
92
  the first round changed: same scenes, same shards, same split.
93
 
 
 
 
 
 
 
 
 
 
 
 
 
94
  **Known overlap.** Four cells of 512×512 blocks hold poses of both splits, near the edges of their
95
  areas:
96
 
@@ -122,6 +139,11 @@ WebDataset shards, up to 500 scenes and about 1 GB each:
122
  `manifest.jsonl` has one line per scene (pose, biome, time, split, `shard` and `sundial0_shard`,
123
  null when there is no render, and `cell`, the 512×512 cell of the map).
124
 
 
 
 
 
 
125
  The dataset viewer shows `preview/`: a 640×360 JPEG thumbnail and the manifest line of every
126
  first-round scene. It is only for browsing; the data is in `data/` and `sundial0/`.
127
 
@@ -157,7 +179,8 @@ the frame.
157
  ## Loading
158
 
159
  Download the shards (`hf download stemjara/minecraft-vanilla-frames --repo-type dataset
160
- --include "data/*" --include "sundial0/*"`) and read them with `webdataset` or `tarfile`:
 
161
 
162
  ```python
163
  import io, json, numpy as np, webdataset as wds
 
32
  were captured with a Fabric mod that teleports the player through a plan of poses with the world
33
  frozen (no weather, fixed time of day).
34
 
35
+ A separate test set, `hora_continua/`, adds 4 020 frames of a forest in the second world with the
36
+ sun moving in small steps through the day, each with its Sundial Lite render (see *Continuous time of day* below).
37
+
38
  Capture tool (the mod, its configuration, the shader options and the contribution validator):
39
  **<https://github.com/stevenijm777/nrmc-capture>**
40
 
41
  *Resumen en español: 23 828 fotogramas de Minecraft 1.21.1 sin shaders, con profundidad, cámara,
42
  bloque por píxel y la sombra del sol trazada en el mundo guardado; 16 363 de ellos también con el
43
+ shader Sundial Lite, para aprender el paso de vanilla a shader. Aparte, `hora_continua/` trae 4 020
44
+ fotogramas de prueba de un bosque del segundo mundo con el sol avanzando cada 50 ticks, todos con su
45
+ render de Sundial Lite.*
46
 
47
  ## Paired renders: Sundial Lite
48
 
 
96
  `desert_2_test` and `snowy_plains_2_test`. `swamp_2` has only its training set here. Nothing of
97
  the first round changed: same scenes, same shards, same split.
98
 
99
+ ### Continuous time of day (`hora_continua/`)
100
+
101
+ A test set apart from the splits above, for how a model behaves between the seven captured times:
102
+ 20 positions in a forest of `minesr_test2` (cells `x-9_z-1`, `x-8_z-1`, `x-8_z0`, `x-7_z-1`), each
103
+ captured at 201 times of day, from 1000 to 11000 ticks every 50 ticks: 4 020 scenes, all with
104
+ their Sundial Lite render. It was captured with mod 0.6.1 (no animals) and nothing of it is near
105
+ the training data, which is all in the other world. In vanilla the sky and the light change only
106
+ at some of these times, so many frames of one position are the same image (between 1 and 79
107
+ distinct vanilla frames per position); the Sundial render is different at every time.
108
+
109
+ It has its own folder and manifest, so `data/`, `sundial0/` and `manifest.jsonl` are unchanged.
110
+
111
  **Known overlap.** Four cells of 512×512 blocks hold poses of both splits, near the edges of their
112
  areas:
113
 
 
139
  `manifest.jsonl` has one line per scene (pose, biome, time, split, `shard` and `sundial0_shard`,
140
  null when there is no render, and `cell`, the 512×512 cell of the map).
141
 
142
+ The continuous-time test set has the same layout under `hora_continua/`:
143
+ `hora_continua/data/test/minesr_test2-forest-<nnn>.tar`,
144
+ `hora_continua/sundial0/test/minesr_test2-forest-<nnn>.tar` and `hora_continua/manifest.jsonl`
145
+ (its shard paths are relative to `hora_continua/`).
146
+
147
  The dataset viewer shows `preview/`: a 640×360 JPEG thumbnail and the manifest line of every
148
  first-round scene. It is only for browsing; the data is in `data/` and `sundial0/`.
149
 
 
179
  ## Loading
180
 
181
  Download the shards (`hf download stemjara/minecraft-vanilla-frames --repo-type dataset
182
+ --include "data/*" --include "sundial0/*"`, and `--include "hora_continua/*"` for the
183
+ continuous-time test set) and read them with `webdataset` or `tarfile`:
184
 
185
  ```python
186
  import io, json, numpy as np, webdataset as wds