Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,55 +1,36 @@
|
|
1 |
-
#
|
|
|
|
|
|
|
2 |
|
3 |
-
import spaces
|
4 |
-
from bert_handler import create_handler_from_checkpoint
|
5 |
import torch
|
6 |
import gradio as gr
|
7 |
-
import
|
8 |
-
from
|
9 |
-
from huggingface_hub import snapshot_download
|
10 |
-
|
11 |
-
# Load checkpoint using BERTHandler (loads tokenizer and full model)
|
12 |
-
checkpoint_path = snapshot_download(
|
13 |
-
repo_id="AbstractPhil/bert-beatrix-2048",
|
14 |
-
revision="main",
|
15 |
-
local_dir="bert-beatrix-2048",
|
16 |
-
local_dir_use_symlinks=False
|
17 |
-
)
|
18 |
-
handler, model, tokenizer = create_handler_from_checkpoint(checkpoint_path)
|
19 |
-
model = model.eval().cuda()
|
20 |
-
|
21 |
-
# Extract encoder only (NomicBertModel -> encoder)
|
22 |
-
encoder = model.bert.encoder
|
23 |
-
embeddings = model.bert.embeddings
|
24 |
-
emb_ln = model.bert.emb_ln
|
25 |
-
emb_drop = model.bert.emb_drop
|
26 |
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
|
|
33 |
|
34 |
-
|
35 |
-
x = embeddings(input_ids)
|
36 |
-
x = emb_ln(x)
|
37 |
-
x = emb_drop(x)
|
38 |
-
encoded = encoder(x, attention_mask=attention_mask.bool())
|
39 |
|
40 |
-
|
41 |
-
|
42 |
|
43 |
-
|
44 |
-
|
|
|
|
|
|
|
45 |
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
}
|
51 |
-
|
52 |
-
symbolic_roles = [
|
53 |
"<subject>", "<subject1>", "<subject2>", "<pose>", "<emotion>",
|
54 |
"<surface>", "<lighting>", "<material>", "<accessory>", "<footwear>",
|
55 |
"<upper_body_clothing>", "<hair_style>", "<hair_length>", "<headwear>",
|
@@ -58,26 +39,93 @@ symbolic_roles = [
|
|
58 |
"<fabric>", "<jewelry>"
|
59 |
]
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
def build_interface():
|
62 |
-
with gr.Blocks() as demo:
|
63 |
-
|
|
|
|
|
|
|
|
|
64 |
with gr.Row():
|
65 |
with gr.Column():
|
66 |
-
input_text = gr.Textbox(
|
67 |
-
|
68 |
-
|
69 |
-
|
|
|
|
|
|
|
|
|
70 |
)
|
71 |
run_btn = gr.Button("Encode & Trace")
|
72 |
with gr.Column():
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
|
77 |
-
run_btn.click(
|
|
|
|
|
|
|
|
|
78 |
|
79 |
return demo
|
80 |
|
|
|
81 |
if __name__ == "__main__":
|
82 |
demo = build_interface()
|
83 |
demo.launch()
|
|
|
1 |
+
# app.py – encoder-only demo for bert-beatrix-2048
|
2 |
+
# -----------------------------------------------
|
3 |
+
# launch: python app.py
|
4 |
+
# (gradio UI appears at http://localhost:7860)
|
5 |
|
|
|
|
|
6 |
import torch
|
7 |
import gradio as gr
|
8 |
+
import spaces
|
9 |
+
from bert_handler import create_handler_from_checkpoint
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
+
# ------------------------------------------------------------------
|
12 |
+
# 1. Model / tokenizer -------------------------------------------------
|
13 |
+
# ------------------------------------------------------------------
|
14 |
+
#
|
15 |
+
# • We load one repo *once*, via its canonical name.
|
16 |
+
# • BERTHandler handles the VRAM-safe cleanup & guarantees that the
|
17 |
+
# tokenizer already contains all special tokens saved in the checkpoint.
|
18 |
|
19 |
+
REPO_ID = "AbstractPhil/bert-beatrix-2048"
|
|
|
|
|
|
|
|
|
20 |
|
21 |
+
handler, full_model, tokenizer = create_handler_from_checkpoint(REPO_ID)
|
22 |
+
full_model = full_model.eval().cuda()
|
23 |
|
24 |
+
# Grab the encoder + embedding stack only
|
25 |
+
encoder = full_model.bert.encoder
|
26 |
+
embeddings = full_model.bert.embeddings
|
27 |
+
emb_ln = full_model.bert.emb_ln
|
28 |
+
emb_drop = full_model.bert.emb_drop
|
29 |
|
30 |
+
# ------------------------------------------------------------------
|
31 |
+
# 2. Symbolic token set -------------------------------------------
|
32 |
+
# ------------------------------------------------------------------
|
33 |
+
SYMBOLIC_ROLES = [
|
|
|
|
|
|
|
34 |
"<subject>", "<subject1>", "<subject2>", "<pose>", "<emotion>",
|
35 |
"<surface>", "<lighting>", "<material>", "<accessory>", "<footwear>",
|
36 |
"<upper_body_clothing>", "<hair_style>", "<hair_length>", "<headwear>",
|
|
|
39 |
"<fabric>", "<jewelry>"
|
40 |
]
|
41 |
|
42 |
+
# Quick sanity check – should *never* be unk
|
43 |
+
missing = [tok for tok in SYMBOLIC_ROLES
|
44 |
+
if tokenizer.convert_tokens_to_ids(tok) == tokenizer.unk_token_id]
|
45 |
+
if missing:
|
46 |
+
raise RuntimeError(f"Tokenizer is missing {len(missing)} special tokens: {missing}")
|
47 |
+
|
48 |
+
# ------------------------------------------------------------------
|
49 |
+
# 3. Encoder-only inference util ----------------------------------
|
50 |
+
# ------------------------------------------------------------------
|
51 |
+
@spaces.GPU
|
52 |
+
def encode_and_trace(text: str, selected_roles: list[str]):
|
53 |
+
"""
|
54 |
+
• encodes `text`
|
55 |
+
• pulls out the hidden states for any of the `selected_roles`
|
56 |
+
• returns some quick stats so we can verify everything’s wired up
|
57 |
+
"""
|
58 |
+
with torch.no_grad():
|
59 |
+
batch = tokenizer(text, return_tensors="pt").to("cuda")
|
60 |
+
inp_ids, attn_mask = batch.input_ids, batch.attention_mask
|
61 |
+
|
62 |
+
# --- embedding + LayerNorm/dropout ---
|
63 |
+
x = embeddings(inp_ids)
|
64 |
+
x = emb_drop(emb_ln(x))
|
65 |
+
|
66 |
+
# --- proper *additive* attention mask ---
|
67 |
+
ext_mask = full_model.bert.get_extended_attention_mask(
|
68 |
+
attn_mask, x.shape[:-1]
|
69 |
+
)
|
70 |
+
|
71 |
+
encoded = encoder(x, attention_mask=ext_mask) # (B, S, H)
|
72 |
+
|
73 |
+
# --- pick out the positions that match selected_roles ---
|
74 |
+
sel_ids = {tokenizer.convert_tokens_to_ids(t) for t in selected_roles}
|
75 |
+
ids_list = inp_ids.squeeze(0).tolist() # python ints
|
76 |
+
keep_mask = torch.tensor([tid in sel_ids for tid in ids_list],
|
77 |
+
device=encoded.device)
|
78 |
+
|
79 |
+
tokens_found = [tokenizer.convert_ids_to_tokens([tid])[0]
|
80 |
+
for tid in ids_list if tid in sel_ids]
|
81 |
+
if keep_mask.any():
|
82 |
+
repr_vec = encoded.squeeze(0)[keep_mask].mean(0)
|
83 |
+
norm_val = f"{repr_vec.norm().item():.4f}"
|
84 |
+
else:
|
85 |
+
norm_val = "0.0000"
|
86 |
+
|
87 |
+
return {
|
88 |
+
"Symbolic Tokens": ", ".join(tokens_found) or "(none)",
|
89 |
+
"Embedding Norm": norm_val,
|
90 |
+
"Symbolic Token Count": int(keep_mask.sum().item()),
|
91 |
+
}
|
92 |
+
|
93 |
+
# ------------------------------------------------------------------
|
94 |
+
# 4. Gradio UI -----------------------------------------------------
|
95 |
+
# ------------------------------------------------------------------
|
96 |
def build_interface():
|
97 |
+
with gr.Blocks(title="🧠 Symbolic Encoder Inspector") as demo:
|
98 |
+
|
99 |
+
gr.Markdown("## 🧠 Symbolic Encoder Inspector\n"
|
100 |
+
"Paste some text containing the special `<role>` tokens and "
|
101 |
+
"inspect their encoder representations.")
|
102 |
+
|
103 |
with gr.Row():
|
104 |
with gr.Column():
|
105 |
+
input_text = gr.Textbox(
|
106 |
+
label="Input with Symbolic Tokens",
|
107 |
+
placeholder="Example: A <subject> wearing <upper_body_clothing> …",
|
108 |
+
lines=3,
|
109 |
+
)
|
110 |
+
role_selector = gr.CheckboxGroup(
|
111 |
+
choices=SYMBOLIC_ROLES,
|
112 |
+
label="Trace these symbolic roles"
|
113 |
)
|
114 |
run_btn = gr.Button("Encode & Trace")
|
115 |
with gr.Column():
|
116 |
+
out_tokens = gr.Textbox(label="Symbolic Tokens Found")
|
117 |
+
out_norm = gr.Textbox(label="Mean Norm")
|
118 |
+
out_count = gr.Textbox(label="Token Count")
|
119 |
|
120 |
+
run_btn.click(
|
121 |
+
fn=encode_and_trace,
|
122 |
+
inputs=[input_text, role_selector],
|
123 |
+
outputs=[out_tokens, out_norm, out_count],
|
124 |
+
)
|
125 |
|
126 |
return demo
|
127 |
|
128 |
+
|
129 |
if __name__ == "__main__":
|
130 |
demo = build_interface()
|
131 |
demo.launch()
|