Datasets:

ArXiv:
License:
kxzxvbk commited on
Commit
cacac16
1 Parent(s): 09658da

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +53 -46
README.md CHANGED
@@ -3,78 +3,86 @@ license: apache-2.0
3
  ---
4
  # Dataset Card for Pong-v4-expert-MCTS
5
  ## Table of Contents
6
- - [Dataset Description](#dataset-description)
7
- - [Dataset Structure](#dataset-structure)
8
- - [Data Instances](#data-instances)
9
- - [Data Fields](#data-fields)
10
- - [Data Splits](#data-splits)
11
- - [Data Creation](#Data-Creation)
12
- - [Curation Rationale](##Curation-Rationale)
13
- - [Source Data](##Source-Data)
14
- - [Initial Data Collection and Normalization](###Initial-Data-Collection-and-Normalization)
15
- - [Who are the source data producers?](### Who-are-the-source-data-producers?)
16
- - [Annotations](###Annotations)
17
- - [Considerations for Using the Data](#Considerations-for-Using-the-Data)
 
18
  - [Social Impact of Dataset](##Social-Impact-of-Dataset)
19
  - [Known Limitations](##Known-Limitations)
20
- - [Additional Information](#Additional-Information)
21
  - [Licensing Information](##Licensing-Information)
22
  - [Citation Information](##Citation-Information)
23
  - [Contributions](##Contributions)
24
 
25
  ## Supported Tasks and Baseline
26
 
27
- - This dataset supports the training for Procedure Cloning algorithm.
 
28
 
29
- - Baseline
 
 
30
 
31
- | Length for procedure sequence | Return |
32
- | ----------------------------- | ------ |
33
- | 0 | 20 |
34
- | 4 | -21 |
35
 
36
- ## Dataset Description
 
 
37
 
38
- This dataset includes 8 episodes of pong-v4 environment. The expert policy is EfficientZero, which is able to generate MCTS hidden states.
39
 
40
- ## Dataset Structure
41
- ### Data Instances
42
- A data point comprises tuples of sequences of (observations, actions, hidden_states):
43
- ```
44
- {
45
- "obs":datasets.Array3D(),
46
- "actions":int,
47
- "hidden_state":datasets.Array3D(),
48
- }
49
- ```
50
- ## Source Data
51
 
52
  ### Data Fields
53
 
54
- - `obs`: An Array3D containing observations from 8 trajectories of an evaluated agent. The data type is uint8 and each value is in 0 to 255.
55
  - `actions`: An integer containing actions from 8 trajectories of an evaluated agent. This value is from 0 to 5. Details about this environment can be viewed at [Pong - Gym Documentation](https://www.gymlibrary.dev/environments/atari/pong/).
56
  - `hidden_state`: An Array3D containing corresponding hidden states generated by EfficientZero, from 8 trajectories of an evaluated agent. The data type is float32.
57
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
  ### Data Splits
59
  There is only a training set for this dataset, as evaluation is undertaken by interacting with a simulator.
60
 
61
- ## Data Creation
62
-
63
- ### Curation Rationale
64
 
65
- - This dataset includes expert data generated by EfficientZero. Since it contains hidden states for each observation, it is suitable for Imitation Learning methods that learn from a sequence like [Procedure Cloning (PC)](https://arxiv.org/abs/2205.10816).
66
- ### Source Data
67
- #### Initial Data Collection and Normalization
68
  - This dataset is collected by EfficientZero policy.
69
  - The standard for expert data is that each return of 8 episodes is over 20.
70
  - No normalization is previously applied ( i.e. each value of observation is a uint8 scalar in [0, 255] )
71
- #### Who are the source language producers?
72
- - [@kxzxvbk](https://huggingface.co/kxzxvbk)
73
 
74
- #### Annotations
75
- - The format of observation picture is [H, W, C], where the channel dimension is the last dimension of the tensor.
 
76
 
77
- ## Considerations for Using the Data
78
 
79
  ### Social Impact of Dataset
80
 
@@ -83,11 +91,10 @@ There is only a training set for this dataset, as evaluation is undertaken by in
83
  - This dataset can potentially promote the research for sequence based imitation learning algorithms.
84
 
85
  ### Known Limitations
 
86
  - This dataset is only used for academic research.
87
  - For any commercial use or other cooperation, please contact: opendilab@pjlab.org.cn
88
 
89
- ## Additional Information
90
-
91
  ### License
92
  This dataset is under [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
93
 
 
3
  ---
4
  # Dataset Card for Pong-v4-expert-MCTS
5
  ## Table of Contents
6
+
7
+ - [Supported Tasks and Baseline](#support-tasks-and-baseline)
8
+
9
+ - [Data Usage](#data-usage)
10
+
11
+ - [Data Discription](##data-description)
12
+ - [Data Fields](##data-fields)
13
+ - [Data Splits](##data-splits)
14
+ - [Initial Data Collection and Normalization](##Initial-Data-Collection-and-Normalization)
15
+
16
+ - [Additional Information](#Additional-Information)
17
+
18
+ - [Who are the source data producers?](## Who-are-the-source-data-producers?)
19
  - [Social Impact of Dataset](##Social-Impact-of-Dataset)
20
  - [Known Limitations](##Known-Limitations)
21
+
22
  - [Licensing Information](##Licensing-Information)
23
  - [Citation Information](##Citation-Information)
24
  - [Contributions](##Contributions)
25
 
26
  ## Supported Tasks and Baseline
27
 
28
+ - This dataset supports the training for [Procedure Cloning (PC )](https://arxiv.org/abs/2205.10816) algorithm.
29
+ - Baselines when sequence length for decision is 0:
30
 
31
+ | Train loss | Test Acc | Reward |
32
+ | -------------------------------------------------- | -------- | ------ |
33
+ | <img src="./img/sup_loss.png" style="zoom:50%;" /> | 0.90 | 20 |
34
 
35
+ - Baselines when sequence length for decision is 4:
 
 
 
36
 
37
+ | Train action loss | Train hidden state loss | Train acc (auto-regressive mode) | Reward |
38
+ | ----------------------------------------------------- | ------------------------------------------------- | --------------------------------------------------- | ------ |
39
+ | <img src="./img/action_loss.png" style="zoom:50%;" /> | <img src="./img/hs_loss.png" style="zoom:50%;" /> | <img src="./img/train_acc.png" style="zoom:50%;" /> | -21 |
40
 
41
+ ## Data Usage
42
 
43
+ ### Data description
44
+
45
+ This dataset includes 8 episodes of pong-v4 environment. The expert policy is [EfficientZero]([[2111.00210\] Mastering Atari Games with Limited Data (arxiv.org)](https://arxiv.org/abs/2111.00210)), which is able to generate MCTS hidden states. Because of the contained hidden states for each observation, this dataset is suitable for Imitation Learning methods that learn from a sequence like PC.
 
 
 
 
 
 
 
 
46
 
47
  ### Data Fields
48
 
49
+ - `obs`: An Array3D containing observations from 8 trajectories of an evaluated agent. The data type is uint8 and each value is in 0 to 255. The shape of this tensor is [96, 96, 3], that is, the channel dimension in the last dimension.
50
  - `actions`: An integer containing actions from 8 trajectories of an evaluated agent. This value is from 0 to 5. Details about this environment can be viewed at [Pong - Gym Documentation](https://www.gymlibrary.dev/environments/atari/pong/).
51
  - `hidden_state`: An Array3D containing corresponding hidden states generated by EfficientZero, from 8 trajectories of an evaluated agent. The data type is float32.
52
 
53
+ This is an example for loading the data using iterator:
54
+
55
+ ```python
56
+ from safetensors import saveopen
57
+
58
+ def generate_examples(self, filepath):
59
+ data = {}
60
+ with safe_open(filepath, framework="pt", device="cpu") as f:
61
+ for key in f.keys():
62
+ data[key] = f.get_tensor(key)
63
+
64
+ for idx in range(len(data['obs'])):
65
+ yield idx, {
66
+ 'observation': data['obs'][idx],
67
+ 'action': data['actions'][idx],
68
+ 'hidden_state': data['hidden_state'][idx],
69
+ }
70
+ ```
71
+
72
  ### Data Splits
73
  There is only a training set for this dataset, as evaluation is undertaken by interacting with a simulator.
74
 
75
+ ### Initial Data Collection and Normalization
 
 
76
 
 
 
 
77
  - This dataset is collected by EfficientZero policy.
78
  - The standard for expert data is that each return of 8 episodes is over 20.
79
  - No normalization is previously applied ( i.e. each value of observation is a uint8 scalar in [0, 255] )
 
 
80
 
81
+ ## Additional Information
82
+
83
+ ### Who are the source language producers?
84
 
85
+ [@kxzxvbk](https://huggingface.co/kxzxvbk)
86
 
87
  ### Social Impact of Dataset
88
 
 
91
  - This dataset can potentially promote the research for sequence based imitation learning algorithms.
92
 
93
  ### Known Limitations
94
+
95
  - This dataset is only used for academic research.
96
  - For any commercial use or other cooperation, please contact: opendilab@pjlab.org.cn
97
 
 
 
98
  ### License
99
  This dataset is under [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
100