Update README.md
Browse files
README.md
CHANGED
@@ -2,21 +2,127 @@
|
|
2 |
license: mit
|
3 |
---
|
4 |
|
5 |
-
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
https://arxiv.org/abs/2104.04258
|
12 |
-
Tim Pearce, Jun Zhu
|
13 |
|
14 |
-
|
15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
https://arxiv.org/abs/2301.10677
|
17 |
Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, Sam Devlin
|
18 |
|
19 |
-
|
20 |
https://arxiv.org/pdf/2405.12399
|
21 |
Eloi Alonso∗, Adam Jelley∗, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce‡, François Fleuret‡
|
22 |
Tweet here: https://twitter.com/EloiAlonso1/status/1844803606064611771
|
|
|
2 |
license: mit
|
3 |
---
|
4 |
|
5 |
+
This repo hosts the dataset presented in:
|
6 |
+
__Counter-Strike Deathmatch with Large-Scale Behavioural Cloning__
|
7 |
+
[Tim Pearce](https://teapearce.github.io/), [Jun Zhu](https://ml.cs.tsinghua.edu.cn/~jun/index.shtml)
|
8 |
+
IEEE Conference on Games (CoG) 2022 [⭐️ Best Paper Award!]
|
9 |
+
ArXiv paper: https://arxiv.org/abs/2104.04258 (Contains some extra experiments not in CoG version)
|
10 |
+
CoG paper: https://ieee-cog.org/2022/assets/papers/paper_45.pdf
|
11 |
+
Four minute introduction video: https://youtu.be/rnz3lmfSHv0
|
12 |
+
Gameplay examples: https://youtu.be/KTY7UhjIMm4
|
13 |
+
Code: https://github.com/TeaPearce/Counter-Strike_Behavioural_Cloning
|
14 |
|
15 |
|
16 |
+
The dataset comprises several different subsets of data as described below.
|
17 |
+
You probably only care about the first one (if you want the largest dataset), or the second or third one (if you care about clean expert data).
|
|
|
|
|
18 |
|
19 |
+
- ```hdf5_dm_july2021_*_to_*.tar```
|
20 |
+
- each .tar file contains 200 .hdf5 files
|
21 |
+
- total files when unzipped: 5500
|
22 |
+
- approx size: 700 GB
|
23 |
+
- map: dust2
|
24 |
+
- gamemode: deathmatch
|
25 |
+
- source: scraped from online servers
|
26 |
+
|
27 |
+
- ```dataset_dm_expert_dust2/hdf5_dm_july2021_expert_*.hdf5```
|
28 |
+
- total files when unzipped: 190
|
29 |
+
- approx size: 24 GB
|
30 |
+
- map: dust2
|
31 |
+
- gamemode: deathmatch
|
32 |
+
- source: manually created, clean actions
|
33 |
+
|
34 |
+
- ```dataset_aim_expert/hdf5_aim_july2021_expert_*.hdf5```
|
35 |
+
- total files when unzipped: 45
|
36 |
+
- approx size: 6 GB
|
37 |
+
- map: aim map
|
38 |
+
- gamemode: aim mode
|
39 |
+
- source: manually created, clean actions
|
40 |
+
|
41 |
+
- ```dataset_dm_expert_othermaps/hdf5_dm_nuke_expert_*.hdf5```
|
42 |
+
- total files when unzipped: 10
|
43 |
+
- approx size: 1 GB
|
44 |
+
- map: nuke
|
45 |
+
- gamemode: deathmatch
|
46 |
+
- source: manually created, clean actions
|
47 |
+
|
48 |
+
- ```dataset_dm_expert_othermaps/hdf5_dm_mirage_expert_*.hdf5```
|
49 |
+
- total files when unzipped: 10
|
50 |
+
- approx size: 1 GB
|
51 |
+
- map: mirage
|
52 |
+
- gamemode: deathmatch
|
53 |
+
- source: manually created, clean actions
|
54 |
+
|
55 |
+
- ```dataset_dm_expert_othermaps/hdf5_dm_inferno_expert_*.hdf5```
|
56 |
+
- total files when unzipped: 10
|
57 |
+
- approx size: 1 GB
|
58 |
+
- map: mirage
|
59 |
+
- gamemode: deathmatch
|
60 |
+
- source: manually created, clean actions
|
61 |
+
|
62 |
+
- ```dataset_metadata/currvarsv2_dm_july2021_*_to_*.npy, currvarsv2_dm_july2021_expert_*_to_*.npy, currvarsv2_dm_mirage_expert_1_to_100.npy, currvarsv2_dm_inferno_expert_1_to_100.npy, currvarsv2_dm_nuke_expert_1_to_100.npy, currvarsv2_aim_july2021_expert_1_to_100.npy```
|
63 |
+
- total files when unzipped: 55 + 2 + 1 + 1 + 1 + 1 = 61
|
64 |
+
- approx size: 6 GB
|
65 |
+
- map: as per filename
|
66 |
+
- gamemode: as per filename
|
67 |
+
- source: as per filename
|
68 |
+
|
69 |
+
- ```location_trackings_backup/```
|
70 |
+
- total files when unzipped: 305
|
71 |
+
- approx size: 0.5 GB
|
72 |
+
- map: dust2
|
73 |
+
- gamemode: deathmatch
|
74 |
+
- source: contains metadata used to compute map coverage analysis
|
75 |
+
- **currvarsv2_agentj22** is the agent trained over the full online dataset
|
76 |
+
- **currvarsv2_agentj22_dmexpert20** is previous model finetuned on the clean expert dust2 dataset
|
77 |
+
- **currvarsv2_bot_capture** is medium difficulty built-in bot
|
78 |
+
|
79 |
+
|
80 |
+
|
81 |
+
### Structure of .hdf5 files (image and action labels -- you probably care about this one):
|
82 |
+
|
83 |
+
Each file contains an ordered sequence of 1000 frames (~1 minute) of play.
|
84 |
+
This contains screenshots, as well as processed action labels.
|
85 |
+
We chose .hdf5 format for fast dataloading, since a subset of frames can be accessed without opening the full file.
|
86 |
+
The lookup keys are as follows (where i is frame number 0-999)
|
87 |
+
- **frame_i_x**: is the image
|
88 |
+
- **frame_i_xaux**: contains actions applied in previous timesteps, as well as health, ammo, and team. see dm_pretrain_preprocess.py for details, note this was not used in our final version of the agent
|
89 |
+
- **frame_i_y**: contains target actions in flattened vector form; [keys_pressed_onehot, Lclicks_onehot, Rclicks_onehot, mouse_x_onehot, mouse_y_onehot]
|
90 |
+
- **frame_i_helperarr**: in format [kill_flag, death_flag], each a binary variable, e.g. [1,0] means the player scored a kill and did not die in that timestep
|
91 |
+
|
92 |
+
### Structure of .npy files (scraped metadata -- you probably don't care about this):
|
93 |
+
|
94 |
+
Each .npy file contains metadata corresponding to 100 .hdf5 files (as indicated by file name)
|
95 |
+
They are dictionaries with keys of format: file_numi_frame_j for file number i, and frame number j in 0-999
|
96 |
+
The values are of format **[curr_vars, infer_a, frame_i_helperarr]** where,
|
97 |
+
- **curr_vars**: contains a dictionary of the metadata originally scraped -- see dm_record_data.py for details
|
98 |
+
- **infer_a**: are inferred actions, [keys_pressed,mouse_x,mouse_y,press_mouse_l,press_mouse_r], with mouse_x and y being continuous values and keys_pressed is in string format
|
99 |
+
- **frame_i_helperarr**: is a repeat of the .hdf5 file
|
100 |
+
|
101 |
+
|
102 |
+
## Trained Models
|
103 |
+
|
104 |
+
Four trained models are provided. There are 'non-stateful' (use during training) and 'stateful' (use at test time) versions of each.
|
105 |
+
Models can be downloaded under ```trained_models.zip```.
|
106 |
+
|
107 |
+
- ```ak47_sub_55k_drop_d4```
|
108 |
+
: Pretrained on AK47 sequences only.
|
109 |
+
- ```ak47_sub_55k_drop_d4_dmexpert_28```
|
110 |
+
: Finetuned on expert deathmatch data.
|
111 |
+
- ```ak47_sub_55k_drop_d4_aimexpertv2_60```
|
112 |
+
: Finetuned on aim mode expert data.
|
113 |
+
- ```July_remoterun7_g9_4k_n32_recipe_ton96__e14```
|
114 |
+
: Pretrained on full dataset.
|
115 |
+
|
116 |
+
|
117 |
+
|
118 |
+
|
119 |
+
## Other works using the dataset:
|
120 |
+
|
121 |
+
- __Imitating Human Behaviour with Diffusion Models, ICLR 2023__
|
122 |
https://arxiv.org/abs/2301.10677
|
123 |
Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, Sam Devlin
|
124 |
|
125 |
+
- __Diffusion for World Modeling: Visual Details Matter in Atari, NeurIPS 2024__
|
126 |
https://arxiv.org/pdf/2405.12399
|
127 |
Eloi Alonso∗, Adam Jelley∗, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce‡, François Fleuret‡
|
128 |
Tweet here: https://twitter.com/EloiAlonso1/status/1844803606064611771
|