Spaces:
Running
Running
Update src/streamlit_app.py
Browse files- src/streamlit_app.py +141 -38
src/streamlit_app.py
CHANGED
@@ -1,40 +1,143 @@
|
|
1 |
-
import
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
4 |
import streamlit as st
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
In the meantime, below is an example of what you can do with just a few lines of code:
|
14 |
-
"""
|
15 |
-
|
16 |
-
num_points = st.slider("Number of points in spiral", 1, 10000, 1100)
|
17 |
-
num_turns = st.slider("Number of turns in spiral", 1, 300, 31)
|
18 |
-
|
19 |
-
indices = np.linspace(0, 1, num_points)
|
20 |
-
theta = 2 * np.pi * num_turns * indices
|
21 |
-
radius = indices
|
22 |
-
|
23 |
-
x = radius * np.cos(theta)
|
24 |
-
y = radius * np.sin(theta)
|
25 |
-
|
26 |
-
df = pd.DataFrame({
|
27 |
-
"x": x,
|
28 |
-
"y": y,
|
29 |
-
"idx": indices,
|
30 |
-
"rand": np.random.randn(num_points),
|
31 |
-
})
|
32 |
-
|
33 |
-
st.altair_chart(alt.Chart(df, height=700, width=700)
|
34 |
-
.mark_point(filled=True)
|
35 |
-
.encode(
|
36 |
-
x=alt.X("x", axis=None),
|
37 |
-
y=alt.Y("y", axis=None),
|
38 |
-
color=alt.Color("idx", legend=None, scale=alt.Scale()),
|
39 |
-
size=alt.Size("rand", legend=None, scale=alt.Scale(range=[1, 150])),
|
40 |
-
))
|
|
|
1 |
+
import os
|
2 |
+
# turn off Streamlit’s automatic file-watching
|
3 |
+
os.environ["STREAMLIT_SERVER_ENABLE_FILE_WATCHER"] = "false"
|
4 |
+
|
5 |
+
import sys
|
6 |
+
import types
|
7 |
+
import torch # now safe to import
|
8 |
import streamlit as st
|
9 |
+
import numpy as np
|
10 |
+
|
11 |
+
# Prevent Streamlit from trying to walk torch.classes' non-standard __path__
|
12 |
+
if isinstance(getattr(sys.modules.get("torch"), "classes", None), types.ModuleType):
|
13 |
+
torch.classes.__path__ = []
|
14 |
+
|
15 |
+
# pip install tiktoken transformers
|
16 |
+
import tiktoken
|
17 |
+
from transformers import GPT2TokenizerFast
|
18 |
+
|
19 |
+
st.set_page_config(page_title="Embedding Dimension Visualizer", layout="wide")
|
20 |
+
st.title("🔍 Embedding Dimension Visualizer")
|
21 |
+
|
22 |
+
# ---- THEORY EXPANDER ----
|
23 |
+
with st.expander("📖 Theory: Tokenization, BPE & Positional Encoding"):
|
24 |
+
st.markdown("""
|
25 |
+
**1️⃣ Tokenization**
|
26 |
+
Splits raw text into atomic units (“tokens”).
|
27 |
+
|
28 |
+
**2️⃣ Byte-Pair Encoding (BPE)**
|
29 |
+
Iteratively merges the most frequent pair of symbols to build a subword vocabulary.
|
30 |
+
E.g. "embedding" → ["em", "bed", "ding"]
|
31 |
+
|
32 |
+
**3️⃣ Positional Encoding**
|
33 |
+
We add a deterministic sinusoidal vector to each token embedding so the model knows position.
|
34 |
+
""")
|
35 |
+
st.markdown("For embedding dimension \(d\), position \(pos\) and channel index \(i\):")
|
36 |
+
st.latex(r"""\mathrm{PE}_{(pos,\,2i)} = \sin\!\Bigl(\frac{pos}{10000^{2i/d}}\Bigr)""")
|
37 |
+
st.latex(r"""\mathrm{PE}_{(pos,\,2i+1)} = \cos\!\Bigl(\frac{pos}{10000^{2i/d}}\Bigr)""")
|
38 |
+
st.markdown("""
|
39 |
+
- \(pos\) starts at 0 for the first token
|
40 |
+
- Even channels use \(\sin\), odd channels use \(\cos\)
|
41 |
+
- This injects unique, smoothly varying positional signals into each embedding
|
42 |
+
""")
|
43 |
+
|
44 |
+
|
45 |
+
# ---- Sidebar ----
|
46 |
+
with st.sidebar:
|
47 |
+
st.header("Settings")
|
48 |
+
input_text = st.text_input("Enter text to embed", value="Hello world!")
|
49 |
+
dim = st.number_input(
|
50 |
+
"Embedding dimensions",
|
51 |
+
min_value=2,
|
52 |
+
max_value=1536,
|
53 |
+
value=3,
|
54 |
+
step=1,
|
55 |
+
help="Choose 2, 3, 512, 768, 1536, etc."
|
56 |
+
)
|
57 |
+
tokenizer_choice = st.selectbox(
|
58 |
+
"Choose tokenizer",
|
59 |
+
["tiktoken", "openai", "huggingface"],
|
60 |
+
help="Which tokenization scheme to demo."
|
61 |
+
)
|
62 |
+
generate = st.button("Generate / Reset Embedding")
|
63 |
+
|
64 |
+
if not generate:
|
65 |
+
st.info("Adjust the settings in the sidebar and click **Generate / Reset Embedding** to see the tokens and sliders.")
|
66 |
+
st.stop()
|
67 |
+
|
68 |
+
# ---- Tokenize ----
|
69 |
+
if tokenizer_choice in ("tiktoken", "openai"):
|
70 |
+
model_name = "gpt2" if tokenizer_choice=="tiktoken" else "gpt-3.5-turbo"
|
71 |
+
enc = tiktoken.encoding_for_model(model_name)
|
72 |
+
token_ids = enc.encode(input_text)
|
73 |
+
token_strs = [enc.decode([tid]) for tid in token_ids]
|
74 |
+
else:
|
75 |
+
hf_tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
|
76 |
+
token_ids = hf_tokenizer.encode(input_text)
|
77 |
+
token_strs = hf_tokenizer.convert_ids_to_tokens(token_ids)
|
78 |
+
|
79 |
+
st.subheader("🪶 Tokens and IDs")
|
80 |
+
for i, (tok, tid) in enumerate(zip(token_strs, token_ids), start=1):
|
81 |
+
st.write(f"**{i}.** `{tok}` → ID **{tid}**")
|
82 |
+
|
83 |
+
st.write("---")
|
84 |
+
st.subheader("📊 Embedding + Positional Encoding per Token")
|
85 |
+
st.write(f"Input: `{input_text}` | Tokenizer: **{tokenizer_choice}** | Dims per token: **{dim}**")
|
86 |
+
if dim > 20:
|
87 |
+
st.warning("Showing >20 sliders per block may be unwieldy; consider smaller dims for teaching.")
|
88 |
+
|
89 |
+
# helper for sinusoidal positional encoding
|
90 |
+
def get_positional_encoding(position: int, d_model: int) -> np.ndarray:
|
91 |
+
pe = np.zeros(d_model, dtype=float)
|
92 |
+
for i in range(d_model):
|
93 |
+
angle = position / np.power(10000, (2 * (i // 2)) / d_model)
|
94 |
+
pe[i] = np.sin(angle) if (i % 2 == 0) else np.cos(angle)
|
95 |
+
return pe
|
96 |
+
|
97 |
+
# ---- For each token, three slider‐blocks ----
|
98 |
+
for t_idx, tok in enumerate(token_strs, start=1):
|
99 |
+
emb = np.random.uniform(-1.0, 1.0, size=dim)
|
100 |
+
pe = get_positional_encoding(t_idx - 1, dim)
|
101 |
+
combined = emb + pe
|
102 |
+
|
103 |
+
with st.expander(f"Token {t_idx}: `{tok}`"):
|
104 |
+
st.markdown("**1️⃣ Embedding**")
|
105 |
+
for d in range(dim):
|
106 |
+
st.slider(
|
107 |
+
label=f"Emb Dim {d+1}",
|
108 |
+
min_value=-1.0, max_value=1.0,
|
109 |
+
value=float(emb[d]),
|
110 |
+
key=f"t{t_idx}_emb{d+1}",
|
111 |
+
disabled=True
|
112 |
+
)
|
113 |
+
|
114 |
+
st.markdown("**2️⃣ Positional Encoding (sin / cos)**")
|
115 |
+
for d in range(dim):
|
116 |
+
st.slider(
|
117 |
+
label=f"PE Dim {d+1}",
|
118 |
+
min_value=-1.0, max_value=1.0,
|
119 |
+
value=float(pe[d]),
|
120 |
+
key=f"t{t_idx}_pe{d+1}",
|
121 |
+
disabled=True
|
122 |
+
)
|
123 |
+
|
124 |
+
st.markdown("**3️⃣ Embedding + Positional Encoding**")
|
125 |
+
for d in range(dim):
|
126 |
+
st.slider(
|
127 |
+
label=f"Sum Dim {d+1}",
|
128 |
+
min_value=-2.0, max_value=2.0,
|
129 |
+
value=float(combined[d]),
|
130 |
+
key=f"t{t_idx}_sum{d+1}",
|
131 |
+
disabled=True
|
132 |
+
)
|
133 |
+
|
134 |
+
# ---- NEW FINAL SECTION ----
|
135 |
+
st.write("---")
|
136 |
+
st.subheader("Final Input Embedding Plus Positional Encoding Ready to Send to ATtention Heads")
|
137 |
|
138 |
+
for t_idx, tid in enumerate(token_ids, start=1):
|
139 |
+
with st.expander(f"Token ID {tid}"):
|
140 |
+
for d in range(1, dim+1):
|
141 |
+
# pull the “sum” value out of session state
|
142 |
+
val = st.session_state.get(f"t{t_idx}_sum{d}", None)
|
143 |
+
st.write(f"Dim {d}: {val:.4f}" if val is not None else f"Dim {d}: N/A")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|