Edward Beeching
commited on
Commit
•
d7c6a2a
1
Parent(s):
7978894
Updated README
Browse files
README.md
CHANGED
@@ -13,6 +13,7 @@ We share models trained for one seed (123), whereas the paper contained weights
|
|
13 |
### Usage
|
14 |
|
15 |
```
|
|
|
16 |
conda env create -f conda_env.yml
|
17 |
```
|
18 |
|
@@ -20,7 +21,25 @@ Then, you can use the model like this:
|
|
20 |
|
21 |
```python
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
```
|
|
|
13 |
### Usage
|
14 |
|
15 |
```
|
16 |
+
git clone https://huggingface.co/edbeeching/decision_transformer_atari
|
17 |
conda env create -f conda_env.yml
|
18 |
```
|
19 |
|
|
|
21 |
|
22 |
```python
|
23 |
|
24 |
+
from decision_transform_atari import GPTConfig, GPT
|
25 |
+
|
26 |
+
vocab_size = 4
|
27 |
+
block_size = 90
|
28 |
+
model_type = "reward_conditioned"
|
29 |
+
timesteps = 2654
|
30 |
+
|
31 |
+
mconf = GPTConfig(
|
32 |
+
vocab_size,
|
33 |
+
block_size,
|
34 |
+
n_layer=6,
|
35 |
+
n_head=8,
|
36 |
+
n_embd=128,
|
37 |
+
model_type=model_type,
|
38 |
+
max_timestep=timesteps,
|
39 |
+
)
|
40 |
+
model = GPT(mconf)
|
41 |
+
|
42 |
+
checkpoint_path = "checkpoints/Breakout_123.pth" # or Pong, Qbert, Seaquest
|
43 |
+
checkpoint = torch.load(checkpoint_path)
|
44 |
+
model.load_state_dict(checkpoint)
|
45 |
```
|