Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,69 +1,96 @@
|
|
1 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import gradio as gr
|
3 |
-
from model_utils import
|
4 |
from visualize import (
|
5 |
visualize_attention,
|
6 |
-
|
7 |
-
|
8 |
-
compare_model_sizes
|
9 |
)
|
10 |
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
-
def run_visualizer(model_name, text, layer, head):
|
22 |
-
model_info = load_model_info(model_name)
|
23 |
-
attention_plot = visualize_attention(model_info, text, layer, head)
|
24 |
-
token_heatmap = visualize_token_embeddings(model_info, text)
|
25 |
-
token_plot = plot_tokenization(model_info, text)
|
26 |
-
model_stats = get_model_stats(model_info)
|
27 |
|
28 |
-
|
|
|
|
|
|
|
|
|
29 |
|
30 |
-
def run_comparison_chart():
|
31 |
-
return compare_model_sizes(MODEL_CHOICES.values())
|
32 |
|
|
|
33 |
with gr.Blocks() as demo:
|
34 |
-
gr.Markdown(""
|
35 |
-
|
36 |
-
Explore attention heads, token embeddings, and tokenizer behavior across popular transformer models.
|
37 |
-
""")
|
38 |
|
39 |
with gr.Row():
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
42 |
|
43 |
with gr.Row():
|
44 |
-
|
45 |
-
|
|
|
46 |
|
47 |
-
|
48 |
|
49 |
-
with gr.
|
50 |
-
|
51 |
-
embedding_output = gr.Plot(label="Token Embedding Heatmap")
|
52 |
|
53 |
-
with gr.
|
54 |
-
|
55 |
-
model_output = gr.JSON(label="Model Details")
|
56 |
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
)
|
62 |
|
63 |
-
with gr.Accordion("π Compare Model Sizes", open=False):
|
64 |
-
compare_btn = gr.Button("Generate Comparison Chart")
|
65 |
-
comparison_output = gr.Plot()
|
66 |
-
compare_btn.click(fn=run_comparison_chart, outputs=comparison_output)
|
67 |
|
68 |
-
|
|
|
|
|
69 |
|
|
|
1 |
+
# π§ Summary of Features:
|
2 |
+
# π½ Dropdown to choose from 7 prebuilt models
|
3 |
+
# π Updates attention layer & head slider limits based on the model
|
4 |
+
# π Visualizes attention maps for selected head/layer
|
5 |
+
# π§© Tokenized words preview
|
6 |
+
# π Embeddings reduced to 2D using PCA
|
7 |
+
# π¦ Model size bar chart across models
|
8 |
+
# π Full model config info in JSON viewer
|
9 |
+
|
10 |
import gradio as gr
|
11 |
+
from model_utils import MODEL_OPTIONS, load_model, get_model_info
|
12 |
from visualize import (
|
13 |
visualize_attention,
|
14 |
+
show_tokenization,
|
15 |
+
show_embeddings,
|
16 |
+
compare_model_sizes,
|
17 |
)
|
18 |
|
19 |
+
# Initial load
|
20 |
+
DEFAULT_MODEL_NAME = list(MODEL_OPTIONS.values())[0]
|
21 |
+
tokenizer, model = load_model(DEFAULT_MODEL_NAME)
|
22 |
+
|
23 |
+
# Shared state
|
24 |
+
current_tokenizer = tokenizer
|
25 |
+
current_model = model
|
26 |
+
|
27 |
+
|
28 |
+
def update_model(selected_model_name):
|
29 |
+
global current_tokenizer, current_model
|
30 |
+
model_id = MODEL_OPTIONS[selected_model_name]
|
31 |
+
current_tokenizer, current_model = load_model(model_id)
|
32 |
+
info = get_model_info(current_model)
|
33 |
+
|
34 |
+
# Update layer/head sliders based on model
|
35 |
+
num_layers = info.get("Number of Layers", 1)
|
36 |
+
num_heads = info.get("Number of Attention Heads", 1)
|
37 |
+
|
38 |
+
return (
|
39 |
+
info,
|
40 |
+
gr.update(maximum=num_layers - 1, value=0),
|
41 |
+
gr.update(maximum=num_heads - 1, value=0),
|
42 |
+
)
|
43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
+
def run_all_visualizations(text, layer, head):
|
46 |
+
attention_fig = visualize_attention(current_tokenizer, current_model, text, layer, head)
|
47 |
+
token_fig = show_tokenization(current_tokenizer, text)
|
48 |
+
embedding_fig = show_embeddings(current_tokenizer, current_model, text)
|
49 |
+
return attention_fig, token_fig, embedding_fig
|
50 |
|
|
|
|
|
51 |
|
52 |
+
# UI
|
53 |
with gr.Blocks() as demo:
|
54 |
+
gr.Markdown("## π Transformer Explorer")
|
55 |
+
gr.Markdown("Explore attention, tokenization, and embedding visualizations for various transformer models.")
|
|
|
|
|
56 |
|
57 |
with gr.Row():
|
58 |
+
model_dropdown = gr.Dropdown(
|
59 |
+
label="Choose a model",
|
60 |
+
choices=list(MODEL_OPTIONS.keys()),
|
61 |
+
value=list(MODEL_OPTIONS.keys())[0],
|
62 |
+
)
|
63 |
+
model_info = gr.JSON(label="Model Info")
|
64 |
|
65 |
with gr.Row():
|
66 |
+
text_input = gr.Textbox(label="Enter text", value="The quick brown fox jumps over the lazy dog.")
|
67 |
+
layer_slider = gr.Slider(label="Layer", minimum=0, maximum=11, step=1, value=0)
|
68 |
+
head_slider = gr.Slider(label="Head", minimum=0, maximum=11, step=1, value=0)
|
69 |
|
70 |
+
run_button = gr.Button("Run Visualizations")
|
71 |
|
72 |
+
with gr.Tab("π Attention"):
|
73 |
+
attention_plot = gr.Plot()
|
|
|
74 |
|
75 |
+
with gr.Tab("π§© Tokenization"):
|
76 |
+
token_plot = gr.Plot()
|
|
|
77 |
|
78 |
+
with gr.Tab("π Embeddings"):
|
79 |
+
embedding_plot = gr.Plot()
|
80 |
+
|
81 |
+
with gr.Tab("π¦ Model Size Comparison"):
|
82 |
+
model_compare_plot = gr.Plot(value=compare_model_sizes())
|
83 |
+
|
84 |
+
# Event binding
|
85 |
+
model_dropdown.change(fn=update_model, inputs=[model_dropdown], outputs=[model_info, layer_slider, head_slider])
|
86 |
+
run_button.click(
|
87 |
+
fn=run_all_visualizations,
|
88 |
+
inputs=[text_input, layer_slider, head_slider],
|
89 |
+
outputs=[attention_plot, token_plot, embedding_plot],
|
90 |
)
|
91 |
|
|
|
|
|
|
|
|
|
92 |
|
93 |
+
if __name__ == "__main__":
|
94 |
+
demo.launch()
|
95 |
+
|
96 |
|