Spaces:
Runtime error
Runtime error
Spectrum visalizer.
Browse files- app.py +74 -0
- requirements.txt +4 -0
app.py
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import pipeline
|
3 |
+
import numpy as np
|
4 |
+
import pandas as pd
|
5 |
+
import re
|
6 |
+
import torch
|
7 |
+
|
8 |
+
number_re = re.compile(r"\.[0-9]*\.")
|
9 |
+
|
10 |
+
STATE_DICT = {}
|
11 |
+
DATA = pd.DataFrame()
|
12 |
+
|
13 |
+
|
14 |
+
def scatter_plot_fn(group_name):
|
15 |
+
global DATA
|
16 |
+
df = DATA[DATA.group_name == group_name]
|
17 |
+
return gr.LinePlot.update(
|
18 |
+
value=df,
|
19 |
+
x="rank",
|
20 |
+
y="val",
|
21 |
+
color="layer",
|
22 |
+
tooltip=["val", "rank", "layer"],
|
23 |
+
caption="",
|
24 |
+
)
|
25 |
+
|
26 |
+
|
27 |
+
def find_choices(state_dict):
|
28 |
+
global DATA
|
29 |
+
layered_tensors = [
|
30 |
+
k for k, v in state_dict.items() if number_re.findall(k) and len(v.shape) == 2
|
31 |
+
]
|
32 |
+
choices = set()
|
33 |
+
data = []
|
34 |
+
for name in layered_tensors:
|
35 |
+
group_name = number_re.sub(".{N}.", name)
|
36 |
+
choices.add(group_name)
|
37 |
+
layer = int(number_re.search(name).group()[1:-1])
|
38 |
+
|
39 |
+
svdvals = torch.linalg.svdvals(state_dict[name])
|
40 |
+
svdvals /= svdvals.sum()
|
41 |
+
for rank, val in enumerate(svdvals.tolist()[:20]):
|
42 |
+
data.append((name, layer, group_name, rank, val))
|
43 |
+
data = np.array(data)
|
44 |
+
DATA = pd.DataFrame(data, columns=["name", "layer", "group_name", "rank", "val"])
|
45 |
+
DATA["val"] = DATA["val"].astype("float")
|
46 |
+
DATA["layer"] = DATA["layer"].astype("category")
|
47 |
+
DATA["rank"] = DATA["rank"].astype("int32")
|
48 |
+
return choices
|
49 |
+
|
50 |
+
|
51 |
+
def weights_fn(model_id):
|
52 |
+
global STATE_DICT
|
53 |
+
try:
|
54 |
+
pipe = pipeline(model=model_id)
|
55 |
+
STATE_DICT = pipe.model.state_dict()
|
56 |
+
except Exception as e:
|
57 |
+
print(e)
|
58 |
+
STATE_DICT = {}
|
59 |
+
choices = find_choices(STATE_DICT)
|
60 |
+
return gr.Dropdown.update(choices=choices)
|
61 |
+
|
62 |
+
|
63 |
+
with gr.Blocks() as scatter_plot:
|
64 |
+
with gr.Row():
|
65 |
+
with gr.Column():
|
66 |
+
model_id = gr.Textbox(value="gpt")
|
67 |
+
weights = gr.Dropdown(choices=["qkv", "c_fc"])
|
68 |
+
with gr.Column():
|
69 |
+
plot = gr.LinePlot(show_label=False).style(container=True)
|
70 |
+
model_id.change(weights_fn, inputs=model_id, outputs=weights)
|
71 |
+
weights.change(fn=scatter_plot_fn, inputs=weights, outputs=plot)
|
72 |
+
|
73 |
+
if __name__ == "__main__":
|
74 |
+
scatter_plot.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
transformers
|
2 |
+
torch
|
3 |
+
numpy
|
4 |
+
pandas
|