Spaces:
Running on Zero
Running on Zero
Ben commited on
Commit ·
7ed4dd8
1
Parent(s): 912195c
Enable ZeroGPU inference
Browse files
app.py
CHANGED
|
@@ -9,15 +9,16 @@ import numpy as np
|
|
| 9 |
import gymnasium as gym
|
| 10 |
import imageio
|
| 11 |
import gradio as gr
|
|
|
|
| 12 |
from huggingface_hub import hf_hub_download
|
| 13 |
|
| 14 |
-
|
| 15 |
def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
|
| 16 |
nn.init.orthogonal_(layer.weight, std)
|
| 17 |
nn.init.constant_(layer.bias, bias_const)
|
| 18 |
return layer
|
| 19 |
|
| 20 |
-
|
| 21 |
def get_actor_network(state_dim=8, action_dim=4):
|
| 22 |
actor = nn.Sequential(
|
| 23 |
layer_init(nn.Linear(state_dim, 64)),
|
|
@@ -28,7 +29,8 @@ def get_actor_network(state_dim=8, action_dim=4):
|
|
| 28 |
)
|
| 29 |
return actor
|
| 30 |
|
| 31 |
-
|
|
|
|
| 32 |
def simulate_agent(stage_selection):
|
| 33 |
weight_mapping = {
|
| 34 |
"Stage 1: Baseline": "1_baseline.pth",
|
|
@@ -39,20 +41,16 @@ def simulate_agent(stage_selection):
|
|
| 39 |
filename = weight_mapping.get(stage_selection)
|
| 40 |
repo_id = "ben-dlwlrma/Representation-Over-Routing"
|
| 41 |
|
| 42 |
-
# Download weights from HF Hub
|
| 43 |
try:
|
| 44 |
weights_path = hf_hub_download(repo_id=repo_id, filename=filename)
|
| 45 |
except Exception as e:
|
| 46 |
raise gr.Error(f"Weight download failed. Error: {str(e)}")
|
| 47 |
|
| 48 |
-
# Initialize env
|
| 49 |
env = gym.make("LunarLander-v3", render_mode="rgb_array")
|
| 50 |
|
| 51 |
-
|
| 52 |
-
device = torch.device("cpu")
|
| 53 |
actor = get_actor_network(state_dim=8, action_dim=4).to(device)
|
| 54 |
|
| 55 |
-
# Load weights
|
| 56 |
try:
|
| 57 |
actor.load_state_dict(torch.load(weights_path, map_location=device, weights_only=True))
|
| 58 |
actor.eval()
|
|
@@ -85,7 +83,6 @@ def simulate_agent(stage_selection):
|
|
| 85 |
|
| 86 |
env.close()
|
| 87 |
|
| 88 |
-
# Export to MP4
|
| 89 |
video_filename = "eval_output.mp4"
|
| 90 |
fps = 30
|
| 91 |
try:
|
|
@@ -95,7 +92,7 @@ def simulate_agent(stage_selection):
|
|
| 95 |
|
| 96 |
return video_filename
|
| 97 |
|
| 98 |
-
|
| 99 |
with gr.Blocks(title="Representation over Routing", theme=gr.themes.Base()) as demo:
|
| 100 |
gr.Markdown("## Representation over Routing")
|
| 101 |
gr.Markdown("Multi-timescale RL evaluation environment. Select an ablation stage to visualize policy behavior.")
|
|
|
|
| 9 |
import gymnasium as gym
|
| 10 |
import imageio
|
| 11 |
import gradio as gr
|
| 12 |
+
import spaces
|
| 13 |
from huggingface_hub import hf_hub_download
|
| 14 |
|
| 15 |
+
|
| 16 |
def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
|
| 17 |
nn.init.orthogonal_(layer.weight, std)
|
| 18 |
nn.init.constant_(layer.bias, bias_const)
|
| 19 |
return layer
|
| 20 |
|
| 21 |
+
|
| 22 |
def get_actor_network(state_dim=8, action_dim=4):
|
| 23 |
actor = nn.Sequential(
|
| 24 |
layer_init(nn.Linear(state_dim, 64)),
|
|
|
|
| 29 |
)
|
| 30 |
return actor
|
| 31 |
|
| 32 |
+
|
| 33 |
+
@spaces.GPU(duration=60)
|
| 34 |
def simulate_agent(stage_selection):
|
| 35 |
weight_mapping = {
|
| 36 |
"Stage 1: Baseline": "1_baseline.pth",
|
|
|
|
| 41 |
filename = weight_mapping.get(stage_selection)
|
| 42 |
repo_id = "ben-dlwlrma/Representation-Over-Routing"
|
| 43 |
|
|
|
|
| 44 |
try:
|
| 45 |
weights_path = hf_hub_download(repo_id=repo_id, filename=filename)
|
| 46 |
except Exception as e:
|
| 47 |
raise gr.Error(f"Weight download failed. Error: {str(e)}")
|
| 48 |
|
|
|
|
| 49 |
env = gym.make("LunarLander-v3", render_mode="rgb_array")
|
| 50 |
|
| 51 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
| 52 |
actor = get_actor_network(state_dim=8, action_dim=4).to(device)
|
| 53 |
|
|
|
|
| 54 |
try:
|
| 55 |
actor.load_state_dict(torch.load(weights_path, map_location=device, weights_only=True))
|
| 56 |
actor.eval()
|
|
|
|
| 83 |
|
| 84 |
env.close()
|
| 85 |
|
|
|
|
| 86 |
video_filename = "eval_output.mp4"
|
| 87 |
fps = 30
|
| 88 |
try:
|
|
|
|
| 92 |
|
| 93 |
return video_filename
|
| 94 |
|
| 95 |
+
|
| 96 |
with gr.Blocks(title="Representation over Routing", theme=gr.themes.Base()) as demo:
|
| 97 |
gr.Markdown("## Representation over Routing")
|
| 98 |
gr.Markdown("Multi-timescale RL evaluation environment. Select an ablation stage to visualize policy behavior.")
|