Spaces:
Running on Zero
Running on Zero
add normalize_text
Browse files- gepard_inference/engine.py +36 -1
- index.html +8 -6
gepard_inference/engine.py
CHANGED
|
@@ -10,7 +10,9 @@ the model is never reloaded between requests.
|
|
| 10 |
|
| 11 |
from __future__ import annotations
|
| 12 |
|
|
|
|
| 13 |
import os
|
|
|
|
| 14 |
from dataclasses import dataclass, field
|
| 15 |
from pathlib import Path
|
| 16 |
from typing import Dict, List, NamedTuple, Optional, Tuple
|
|
@@ -22,6 +24,37 @@ import yaml
|
|
| 22 |
from .runner import GepardRunner
|
| 23 |
from .speakers import SpeakerLibrary
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
class Example(NamedTuple):
|
| 27 |
"""A demo row for the UI examples table.
|
|
@@ -295,7 +328,9 @@ class GepardEngine:
|
|
| 295 |
"""
|
| 296 |
if self.runner is None:
|
| 297 |
raise RuntimeError("GepardEngine.load() must be called before synthesize()")
|
| 298 |
-
|
|
|
|
|
|
|
| 299 |
if not text:
|
| 300 |
raise ValueError("text is empty")
|
| 301 |
|
|
|
|
| 10 |
|
| 11 |
from __future__ import annotations
|
| 12 |
|
| 13 |
+
import html
|
| 14 |
import os
|
| 15 |
+
import re
|
| 16 |
from dataclasses import dataclass, field
|
| 17 |
from pathlib import Path
|
| 18 |
from typing import Dict, List, NamedTuple, Optional, Tuple
|
|
|
|
| 24 |
from .runner import GepardRunner
|
| 25 |
from .speakers import SpeakerLibrary
|
| 26 |
|
| 27 |
+
# Tag-like sequences only: "<" must be followed by a letter or "/". This spares
|
| 28 |
+
# legitimate prose such as "2 < 3" or "<3" while catching "<div>", "</span>",
|
| 29 |
+
# "<br/>". Applied AFTER html.unescape so escaped markup (<script>) is
|
| 30 |
+
# caught too.
|
| 31 |
+
_HTML_TAG_RE = re.compile(r"</?[A-Za-z][^>]*>")
|
| 32 |
+
# Zero-width / BOM characters that are not caught by \s.
|
| 33 |
+
_ZERO_WIDTH_RE = re.compile(r"[\u200B-\u200D\uFEFF]")
|
| 34 |
+
_WHITESPACE_RE = re.compile(r"\s+")
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def normalize_text(text: Optional[str]) -> str:
|
| 38 |
+
"""Clean arbitrary client text down to what the tokenizer expects.
|
| 39 |
+
|
| 40 |
+
``/synthesize`` is a public API (browser JS, ``gradio_client``, curl), so
|
| 41 |
+
the input may carry markup or invisible characters the model never saw in
|
| 42 |
+
training. This is the single authoritative choke point right before the
|
| 43 |
+
text reaches the model.
|
| 44 |
+
|
| 45 |
+
Steps: decode HTML entities -> strip HTML-ish tags -> drop zero-width
|
| 46 |
+
chars -> collapse all whitespace (newlines, tabs, (narrow) nbsp) to single
|
| 47 |
+
spaces -> trim. ``\\s`` matches ``\\u00A0`` in Python's Unicode mode, so nbsp
|
| 48 |
+
is handled by the whitespace collapse.
|
| 49 |
+
"""
|
| 50 |
+
if not text:
|
| 51 |
+
return ""
|
| 52 |
+
text = html.unescape(text) # & ' -> chars
|
| 53 |
+
text = _HTML_TAG_RE.sub(" ", text) # <div>, </span>, <br/> -> space
|
| 54 |
+
text = _ZERO_WIDTH_RE.sub("", text) # zero-width chars -> drop
|
| 55 |
+
text = _WHITESPACE_RE.sub(" ", text) # newlines/tabs/nbsp/runs -> space
|
| 56 |
+
return text.strip()
|
| 57 |
+
|
| 58 |
|
| 59 |
class Example(NamedTuple):
|
| 60 |
"""A demo row for the UI examples table.
|
|
|
|
| 328 |
"""
|
| 329 |
if self.runner is None:
|
| 330 |
raise RuntimeError("GepardEngine.load() must be called before synthesize()")
|
| 331 |
+
# Authoritative cleanup right before the model — protects every caller
|
| 332 |
+
# (browser JS, gradio_client, curl), not just the frontend.
|
| 333 |
+
text = normalize_text(text)
|
| 334 |
if not text:
|
| 335 |
raise ValueError("text is empty")
|
| 336 |
|
index.html
CHANGED
|
@@ -828,12 +828,14 @@ let wordSpans = [];
|
|
| 828 |
let hlRAF = null;
|
| 829 |
let isPlaying = false;
|
| 830 |
|
|
|
|
|
|
|
| 831 |
const sliderSpec = [
|
| 832 |
-
{ key: "temperature", label: "Temperature", fmt: v => Number(v).toFixed(2) },
|
| 833 |
-
{ key: "top_k", label: "Top-K", fmt: v => String(Math.round(v)) },
|
| 834 |
-
{ key: "max_frames", label: "Max Frames", fmt: v => String(Math.round(v)) },
|
| 835 |
-
{ key: "repetition_penalty", label: "Rep. Penalty", fmt: v => Number(v).toFixed(2) },
|
| 836 |
-
{ key: "repetition_window", label: "Rep. Window", fmt: v => String(Math.round(v)) },
|
| 837 |
];
|
| 838 |
|
| 839 |
const pad = (n, w = 2) => String(Math.floor(n)).padStart(w, "0");
|
|
@@ -1018,7 +1020,7 @@ const buildSliders = () => {
|
|
| 1018 |
const min = cfg.limits[`min_${spec.key}`];
|
| 1019 |
const max = cfg.limits[`max_${spec.key}`];
|
| 1020 |
const def = cfg.defaults[spec.key];
|
| 1021 |
-
const step = spec.
|
| 1022 |
|
| 1023 |
// Slider row
|
| 1024 |
const row = document.createElement("div"); row.className = "slider";
|
|
|
|
| 828 |
let hlRAF = null;
|
| 829 |
let isPlaying = false;
|
| 830 |
|
| 831 |
+
// step values mirror the original Blocks UI's validated grid — e.g. max_frames
|
| 832 |
+
// must stay on the 43-frame codec chunk; off-grid values degrade generation.
|
| 833 |
const sliderSpec = [
|
| 834 |
+
{ key: "temperature", label: "Temperature", step: 0.05, fmt: v => Number(v).toFixed(2) },
|
| 835 |
+
{ key: "top_k", label: "Top-K", step: 1, fmt: v => String(Math.round(v)) },
|
| 836 |
+
{ key: "max_frames", label: "Max Frames", step: 43, fmt: v => String(Math.round(v)) },
|
| 837 |
+
{ key: "repetition_penalty", label: "Rep. Penalty", step: 0.01, fmt: v => Number(v).toFixed(2) },
|
| 838 |
+
{ key: "repetition_window", label: "Rep. Window", step: 4, fmt: v => String(Math.round(v)) },
|
| 839 |
];
|
| 840 |
|
| 841 |
const pad = (n, w = 2) => String(Math.floor(n)).padStart(w, "0");
|
|
|
|
| 1020 |
const min = cfg.limits[`min_${spec.key}`];
|
| 1021 |
const max = cfg.limits[`max_${spec.key}`];
|
| 1022 |
const def = cfg.defaults[spec.key];
|
| 1023 |
+
const step = spec.step;
|
| 1024 |
|
| 1025 |
// Slider row
|
| 1026 |
const row = document.createElement("div"); row.className = "slider";
|