Spaces:
Running
on
Zero
Running
on
Zero
# app.py β encoder-only demo for bert-beatrix-2048 | |
# launch: python app.py | |
# ----------------------------------------------- | |
import json, re, sys, math | |
from pathlib import Path, PurePosixPath | |
import torch, torch.nn.functional as F | |
import gradio as gr | |
import spaces | |
from huggingface_hub import snapshot_download | |
from bert_handler import create_handler_from_checkpoint | |
# ------------------------------------------------------------------ | |
# 0. Download & patch HF checkpoint -------------------------------- | |
REPO_ID = "AbstractPhil/bert-beatrix-2048" | |
LOCAL_CKPT = "bert-beatrix-2048" | |
snapshot_download( | |
repo_id=REPO_ID, | |
revision="main", | |
local_dir=LOCAL_CKPT, | |
local_dir_use_symlinks=False, | |
) | |
# β strip repo prefix in auto_map (one-time) | |
cfg_path = Path(LOCAL_CKPT) / "config.json" | |
with cfg_path.open() as f: cfg = json.load(f) | |
amap = cfg.get("auto_map", {}) | |
for k,v in amap.items(): | |
if "--" in v: | |
amap[k] = PurePosixPath(v.split("--",1)[1]).as_posix() | |
cfg["auto_map"] = amap | |
with cfg_path.open("w") as f: json.dump(cfg,f,indent=2) | |
# ------------------------------------------------------------------ | |
# 1. Load model & components -------------------------------------- | |
handler, full_model, tokenizer = create_handler_from_checkpoint(LOCAL_CKPT) | |
full_model = full_model.eval().cuda() | |
encoder = full_model.bert.encoder | |
embeddings = full_model.bert.embeddings | |
emb_ln = full_model.bert.emb_ln | |
emb_drop = full_model.bert.emb_drop | |
mlm_head = full_model.cls # prediction head | |
# ------------------------------------------------------------------ | |
# 2. Symbolic roles ------------------------------------------------- | |
SYMBOLIC_ROLES = [ | |
"<subject>", "<subject1>", "<subject2>", "<pose>", "<emotion>", | |
"<surface>", "<lighting>", "<material>", "<accessory>", "<footwear>", | |
"<upper_body_clothing>", "<hair_style>", "<hair_length>", "<headwear>", | |
"<texture>", "<pattern>", "<grid>", "<zone>", "<offset>", | |
"<object_left>", "<object_right>", "<relation>", "<intent>", "<style>", | |
"<fabric>", "<jewelry>", | |
] | |
if any(tokenizer.convert_tokens_to_ids(t)==tokenizer.unk_token_id | |
for t in SYMBOLIC_ROLES): | |
sys.exit("β tokenizer missing special tokens") | |
# Quick helpers | |
MASK = tokenizer.mask_token | |
# ------------------------------------------------------------------ | |
# 3. Encoder-plus-MLM logic --------------------------------------- | |
def cosine(a,b): | |
return torch.nn.functional.cosine_similarity(a,b,dim=-1) | |
def pool_accuracy(ids, logits, pool_mask): | |
""" | |
ids : (S,) gold token ids | |
logits : (S,V) MLM logits | |
pool_mask : bool (S,) which tokens belong to the candidate pool | |
returns accuracy over masked positions only (if none, return 0) | |
""" | |
idx = pool_mask.nonzero(as_tuple=False).flatten() | |
if idx.numel()==0: return 0.0 | |
preds = logits.argmax(-1)[idx] | |
gold = ids[idx] | |
return (preds==gold).float().mean().item() | |
def encode_and_trace(text, selected_roles): | |
if not selected_roles: | |
selected_roles = SYMBOLIC_ROLES | |
# Convert symbolic role tokens to IDs | |
sel_ids = [tokenizer.convert_tokens_to_ids(t) for t in selected_roles] | |
sel_ids_tensor = torch.tensor(sel_ids, device="cuda").unsqueeze(0) # shape: (1, R) | |
# Tokenize user prompt | |
batch = tokenizer(text, return_tensors="pt").to("cuda") | |
input_ids, attention_mask = batch.input_ids, batch.attention_mask | |
S = input_ids.shape[1] | |
# === Shared encoder logic with RoPE === | |
def encode(input_ids, attn_mask): | |
x = embeddings(input_ids) # (B, S, H) | |
if emb_ln: x = emb_ln(x) | |
if emb_drop: x = emb_drop(x) | |
ext = full_model.bert.get_extended_attention_mask(attn_mask, input_ids.shape) | |
return encoder(x, attention_mask=ext)[0] # (B, S, H) | |
# Encode prompt | |
encoded_prompt = encode(input_ids, attention_mask)[0] # (S, H) | |
# Encode symbolic roles through same pipeline | |
symbolic_attn = torch.ones_like(sel_ids_tensor) | |
encoded_roles = encode(sel_ids_tensor, symbolic_attn)[0] # (R, H) | |
# === Symbolic classification via cosine similarity === | |
# Compare each token to each symbolic role β shape: (S, R) | |
token_exp = encoded_prompt.unsqueeze(1).expand(-1, encoded_roles.size(0), -1) # (S, R, H) | |
role_exp = encoded_roles.unsqueeze(0).expand(encoded_prompt.size(0), -1, -1) # (S, R, H) | |
sim = F.cosine_similarity(token_exp, role_exp, dim=-1) # β (S, R) | |
argmax_ids = sim.argmax(dim=-1) # (S,) | |
max_scores = sim.max(dim=-1).values # (S,) | |
predicted_roles = [selected_roles[i] for i in argmax_ids.tolist()] | |
decoded_tokens = tokenizer.convert_ids_to_tokens(input_ids[0]) | |
# === Build readable trace | |
role_trace = [ | |
f"{tok:<15} β {role:<22} score={score:.4f}" | |
for tok, role, score in zip(decoded_tokens, predicted_roles, max_scores.tolist()) | |
] | |
# === Final output | |
res_json = { | |
"Prompt": text, | |
"Predicted symbolic roles": predicted_roles, | |
"Max alignment score": f"{max_scores.max().item():.4f}", | |
"Per-token classification": role_trace | |
} | |
return json.dumps(res_json, indent=2), f"{max_scores.max().item():.4f}", len(selected_roles) | |
# ------------------------------------------------------------------ | |
# 4. Gradio UI ----------------------------------------------------- | |
def build_interface(): | |
with gr.Blocks(title="π§ Symbolic Encoder Inspector") as demo: | |
gr.Markdown("## π§ Symbolic Encoder Inspector") | |
with gr.Row(): | |
with gr.Column(): | |
txt = gr.Textbox(label="Prompt", lines=3) | |
roles= gr.CheckboxGroup( | |
choices=SYMBOLIC_ROLES, label="Roles", | |
value=SYMBOLIC_ROLES # pre-checked | |
) | |
btn = gr.Button("Run") | |
with gr.Column(): | |
out_json = gr.Textbox(label="Result JSON") | |
out_max = gr.Textbox(label="Max cos") | |
out_cnt = gr.Textbox(label="# roles") | |
btn.click(encode_and_trace, [txt,roles], [out_json,out_max,out_cnt]) | |
return demo | |
if __name__=="__main__": | |
build_interface().launch() |