Spaces:
Runtime error
Adding infinite sampling functionality + updating lm code to take multiple prompts
Browse files- Added some basic ability to continue building the level and iteratively sampling
- Added a change in lm.py that allows for multiple prompts to be passed in
After pressing generate level button we get two sample options:
![Screenshot from 2023-02-16 19-31-18.png](https://s3.amazonaws.com/moonup/production/uploads/1676604892732-6244b415b0d6be7d12c6be01.png)
After pressing "Sample Choice 1", it appends the new portion to the existing level
![Screenshot from 2023-02-16 19-31-46.png](https://s3.amazonaws.com/moonup/production/uploads/1676604896368-6244b415b0d6be7d12c6be01.png)
Right now its far from a perfect implementation as I just hard coded the number of samples + buttons, but this lets us mix and match prompts while generating the level, which is cool.
Let me know what you think :)
- app.py +105 -23
- mario_gpt/lm.py +1 -1
@@ -46,25 +46,79 @@ def make_html_file(generated_level):
|
|
46 |
</html>''')
|
47 |
return f"demo-{unique_id}.html"
|
48 |
|
49 |
-
def
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
prompts=prompts,
|
56 |
num_steps=level_size,
|
57 |
temperature=temperature,
|
58 |
-
use_tqdm=True
|
|
|
59 |
)
|
60 |
-
|
61 |
-
|
62 |
-
|
|
|
|
|
63 |
gradio_html = f'''<div>
|
64 |
<iframe width=512 height=512 style="margin: 0 auto" src="static/{filename}"></iframe>
|
65 |
<p style="text-align:center">Press the arrow keys to move. Press <code>a</code> to run, <code>s</code> to jump and <code>d</code> to shoot fireflowers</p>
|
66 |
</div>'''
|
67 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
|
69 |
with gr.Blocks().queue() as demo:
|
70 |
gr.Markdown('''### Playable demo for MarioGPT: Open-Ended Text2Level Generation through Large Language Models
|
@@ -83,27 +137,55 @@ with gr.Blocks().queue() as demo:
|
|
83 |
|
84 |
with gr.Accordion(label="Advanced settings", open=False):
|
85 |
temperature = gr.Number(value=2.0, label="temperature: Increase these for more diverse, but lower quality, generations")
|
86 |
-
level_size = gr.Number(value=
|
87 |
-
|
88 |
-
|
|
|
|
|
|
|
89 |
with gr.Row():
|
90 |
with gr.Box():
|
91 |
-
level_play = gr.HTML()
|
92 |
-
level_image = gr.Image()
|
93 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
gr.Examples(
|
95 |
examples=[
|
96 |
-
["many", "many", "some", "high"],
|
97 |
["no", "some", "many", "high", 2.0],
|
98 |
-
["many", "many", "little", "low", 2.
|
99 |
-
["no", "no", "many", "high", 2.
|
100 |
],
|
101 |
-
inputs=[pipes, enemies, blocks, elevation],
|
102 |
outputs=[level_image, level_play],
|
103 |
-
fn=
|
104 |
cache_examples=True,
|
105 |
)
|
106 |
|
107 |
app.mount("/static", StaticFiles(directory="static", html=True), name="static")
|
108 |
app = gr.mount_gradio_app(app, demo, "/", gradio_api_url="http://localhost:7860/")
|
109 |
-
uvicorn.run(app, host="0.0.0.0", port=7860)
|
|
|
46 |
</html>''')
|
47 |
return f"demo-{unique_id}.html"
|
48 |
|
49 |
+
def trim_level(level):
|
50 |
+
mod = level.shape[-1] % 14
|
51 |
+
if mod > 0:
|
52 |
+
return level[:, :-mod]
|
53 |
+
return level
|
54 |
+
|
55 |
+
def reset_state(seed_state):
|
56 |
+
length = len(seed_state)
|
57 |
+
print(f"Resetting state with {length} levels!")
|
58 |
+
for _ in range(length):
|
59 |
+
seed_state.pop()
|
60 |
+
|
61 |
+
def _generate_level(prompts, seed, level_size, temperature):
|
62 |
+
print(f"Using prompts: {prompts}")
|
63 |
+
generated_levels = mario_lm.sample(
|
64 |
prompts=prompts,
|
65 |
num_steps=level_size,
|
66 |
temperature=temperature,
|
67 |
+
use_tqdm=True,
|
68 |
+
seed = seed
|
69 |
)
|
70 |
+
generated_levels = trim_level(generated_levels)
|
71 |
+
return generated_levels
|
72 |
+
|
73 |
+
def _make_gradio_html(level):
|
74 |
+
filename = make_html_file(level)
|
75 |
gradio_html = f'''<div>
|
76 |
<iframe width=512 height=512 style="margin: 0 auto" src="static/{filename}"></iframe>
|
77 |
<p style="text-align:center">Press the arrow keys to move. Press <code>a</code> to run, <code>s</code> to jump and <code>d</code> to shoot fireflowers</p>
|
78 |
</div>'''
|
79 |
+
return gradio_html
|
80 |
+
|
81 |
+
def initialize_generate(pipes, enemies, blocks, elevation, temperature = 2.4, level_size = 1400):
|
82 |
+
prompts = [f"{pipes} pipes, {enemies} enemies, {blocks} blocks, {elevation} elevation"]
|
83 |
+
generated_levels = _generate_level(prompts, None, level_size, temperature)
|
84 |
+
level = generated_levels.squeeze().detach().cpu()
|
85 |
+
img = convert_level_to_png(level, TILE_DIR, mario_lm.tokenizer)[0]
|
86 |
+
return [img, _make_gradio_html(level)]
|
87 |
+
|
88 |
+
def generate_choices(pipes, enemies, blocks, elevation, temperature = 2.4, level_size = 1400, prompt = "", seed_state = []):
|
89 |
+
NUM_SAMPLES = 2
|
90 |
+
if prompt == "":
|
91 |
+
prompt = f"{pipes} pipes, {enemies} enemies, {blocks} blocks, {elevation} elevation"
|
92 |
+
prompts = [prompt] * NUM_SAMPLES
|
93 |
+
|
94 |
+
seed = None
|
95 |
+
if len(seed_state) > 0:
|
96 |
+
seed = torch.cat(seed_state).squeeze()[-48*14:].view(1, -1).repeat(NUM_SAMPLES, 1) # context length
|
97 |
+
|
98 |
+
generated_levels = _generate_level(prompts, seed, level_size, temperature).detach().cpu().squeeze()
|
99 |
+
level_choices = [generated_level[-level_size:] for generated_level in generated_levels]
|
100 |
+
level_choice_images = [convert_level_to_png(generated_level[-level_size:], TILE_DIR, mario_lm.tokenizer)[0] for generated_level in generated_levels]
|
101 |
+
|
102 |
+
# level choices + separate images
|
103 |
+
return [level_choices, *level_choice_images]
|
104 |
+
|
105 |
+
def update_level_state(choice_id, level_choices, seed_state):
|
106 |
+
num_choice = int(choice_id)
|
107 |
+
level_choice = level_choices[num_choice]
|
108 |
+
|
109 |
+
# append level choice to seed state
|
110 |
+
seed_state.append(level_choice)
|
111 |
+
|
112 |
+
# get new level from concatenation
|
113 |
+
level = torch.cat(seed_state).squeeze()
|
114 |
+
|
115 |
+
# final image and gradio html
|
116 |
+
img = convert_level_to_png(level, TILE_DIR, mario_lm.tokenizer)[0]
|
117 |
+
gradio_html = _make_gradio_html(level)
|
118 |
+
|
119 |
+
# return img, gradio html, seed state, level_choice, choice_image_1, choice_image_2, current_level_size
|
120 |
+
return img, gradio_html, seed_state, None, None, None, level.shape[-1]
|
121 |
+
|
122 |
|
123 |
with gr.Blocks().queue() as demo:
|
124 |
gr.Markdown('''### Playable demo for MarioGPT: Open-Ended Text2Level Generation through Large Language Models
|
|
|
137 |
|
138 |
with gr.Accordion(label="Advanced settings", open=False):
|
139 |
temperature = gr.Number(value=2.0, label="temperature: Increase these for more diverse, but lower quality, generations")
|
140 |
+
level_size = gr.Number(value=1400, precision=0, label="level_size")
|
141 |
+
|
142 |
+
generate_btn = gr.Button("Generate Level")
|
143 |
+
reset_btn = gr.Button("Reset Level")
|
144 |
+
|
145 |
+
|
146 |
with gr.Row():
|
147 |
with gr.Box():
|
148 |
+
level_play = gr.HTML()
|
149 |
+
level_image = gr.Image(label="Current Level")
|
150 |
+
with gr.Box():
|
151 |
+
with gr.Column():
|
152 |
+
level_choice1_image = gr.Image(label="Sample Choice 1")
|
153 |
+
level_choice1_btn = gr.Button("Sample Choice 1")
|
154 |
+
with gr.Column():
|
155 |
+
level_choice2_image = gr.Image(label="Sample Choice 2")
|
156 |
+
level_choice2_btn = gr.Button("Sample Choice 2")
|
157 |
+
current_level_size = gr.Number(0, visible=True, label="Current Level Size")
|
158 |
+
|
159 |
+
|
160 |
+
seed_state = gr.State([])
|
161 |
+
state_choices = gr.State(None)
|
162 |
+
|
163 |
+
image_choice_1_id = gr.Number(0, visible=False)
|
164 |
+
image_choice_2_id = gr.Number(1, visible=False)
|
165 |
+
|
166 |
+
# choice buttons
|
167 |
+
level_choice1_btn.click(fn=update_level_state, inputs=[image_choice_1_id, state_choices, seed_state], outputs=[level_image, level_play, seed_state, state_choices, level_choice1_image, level_choice2_image, current_level_size])
|
168 |
+
level_choice2_btn.click(fn=update_level_state, inputs=[image_choice_2_id, state_choices, seed_state], outputs=[level_image, level_play, seed_state, state_choices, level_choice1_image, level_choice2_image, current_level_size])
|
169 |
+
|
170 |
+
# generate_btn
|
171 |
+
generate_btn.click(fn=generate_choices, inputs=[pipes, enemies, blocks, elevation, temperature, level_size, text_prompt, seed_state], outputs=[state_choices, level_choice1_image, level_choice2_image])
|
172 |
+
|
173 |
+
# reset btn
|
174 |
+
reset_btn.click(fn=reset_state, inputs=[seed_state], outputs=[])
|
175 |
+
|
176 |
gr.Examples(
|
177 |
examples=[
|
178 |
+
["many", "many", "some", "high", 2.0],
|
179 |
["no", "some", "many", "high", 2.0],
|
180 |
+
["many", "many", "little", "low", 2.4],
|
181 |
+
["no", "no", "many", "high", 2.8],
|
182 |
],
|
183 |
+
inputs=[pipes, enemies, blocks, elevation, temperature, level_size],
|
184 |
outputs=[level_image, level_play],
|
185 |
+
fn=initialize_generate,
|
186 |
cache_examples=True,
|
187 |
)
|
188 |
|
189 |
app.mount("/static", StaticFiles(directory="static", html=True), name="static")
|
190 |
app = gr.mount_gradio_app(app, demo, "/", gradio_api_url="http://localhost:7860/")
|
191 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|
@@ -105,7 +105,7 @@ class MarioLM:
|
|
105 |
self.lm.eval()
|
106 |
with torch.no_grad():
|
107 |
if seed is None:
|
108 |
-
seed = self.tokenizer("X", return_tensors="pt").input_ids.view(1, 1)
|
109 |
out = seed.to(self.device)
|
110 |
if encoder_hidden_states is None:
|
111 |
if prompts is not None:
|
|
|
105 |
self.lm.eval()
|
106 |
with torch.no_grad():
|
107 |
if seed is None:
|
108 |
+
seed = self.tokenizer("X", return_tensors="pt").input_ids.view(1, 1).repeat(len(prompts), 1)
|
109 |
out = seed.to(self.device)
|
110 |
if encoder_hidden_states is None:
|
111 |
if prompts is not None:
|