Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
Update: Codestyle
Browse files
app.py
CHANGED
|
@@ -1,42 +1,38 @@
|
|
| 1 |
# Standardbibliotheken
|
| 2 |
-
import os
|
| 3 |
-
import
|
| 4 |
-
import
|
| 5 |
-
import
|
| 6 |
-
import
|
| 7 |
-
import
|
| 8 |
|
| 9 |
# Machine Learning / NLP
|
| 10 |
-
import torch
|
| 11 |
-
import fasttext
|
| 12 |
-
#
|
| 13 |
-
#
|
| 14 |
-
import
|
| 15 |
-
import
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
#
|
| 19 |
-
import
|
| 20 |
-
from
|
| 21 |
-
from
|
| 22 |
-
|
|
|
|
|
|
|
| 23 |
|
| 24 |
# UI / Serving
|
| 25 |
-
import gradio as gr
|
| 26 |
-
|
| 27 |
-
import deepl
|
| 28 |
|
| 29 |
# Projektspezifische Module
|
| 30 |
from lib.bert_regressor import BertMultiHeadRegressor
|
| 31 |
from lib.bert_regressor_utils import (
|
| 32 |
-
#
|
| 33 |
-
predict_flavours,
|
| 34 |
-
#predict_is_review,
|
| 35 |
-
#TARGET_COLUMNS,
|
| 36 |
-
#ICONS
|
| 37 |
)
|
| 38 |
-
from lib.wheel import build_svg_with_values
|
| 39 |
-
from lib.examples import EXAMPLES
|
| 40 |
|
| 41 |
### Stettings ####################################################################
|
| 42 |
|
|
@@ -76,14 +72,14 @@ model_flavours.to(device).eval()
|
|
| 76 |
|
| 77 |
ID = LanguageIdentifier.from_modelstring(model, norm_probs=True)
|
| 78 |
|
| 79 |
-
def
|
| 80 |
t = (text or "").strip()
|
| 81 |
if len(t) < min_chars:
|
| 82 |
return True, 0.0
|
| 83 |
lang, prob = ID.classify(t) # prob ∈ [0,1]
|
| 84 |
return (lang == "en" and prob >= threshold), float(prob)
|
| 85 |
|
| 86 |
-
def
|
| 87 |
deepl_client = deepl.Translator(DEEPL_API_KEY)
|
| 88 |
result = deepl_client.translate_text(text, target_lang=target_lang)
|
| 89 |
return result.text
|
|
@@ -102,11 +98,11 @@ def predict(review: str):
|
|
| 102 |
return "Please enter a review.", {}
|
| 103 |
|
| 104 |
# Check for lang of text
|
| 105 |
-
review_is_eng, review_lang_prob =
|
| 106 |
|
| 107 |
# Abort if text is not english
|
| 108 |
if not review_is_eng:
|
| 109 |
-
review =
|
| 110 |
html_out += f"""<div style='border-radius: 2px; padding: 1px 5px; background-color: rgb(255, 237, 213);'>Your text has been automatically translated</div>
|
| 111 |
<p>{review}</p>
|
| 112 |
"""
|
|
@@ -146,7 +142,7 @@ def predict(review: str):
|
|
| 146 |
def random_text():
|
| 147 |
return random.choice(EXAMPLES)
|
| 148 |
|
| 149 |
-
def
|
| 150 |
if torch.cuda.is_available():
|
| 151 |
return f"<span style='border-radius: 2px; padding: 1px 5px; background-color: rgb(220, 252, 231);'>Runs on GPU: {torch.cuda.get_device_name(0)}</span>"
|
| 152 |
else:
|
|
@@ -167,7 +163,7 @@ with gr.Blocks(css=custom_css) as demo:
|
|
| 167 |
<h3>Automatically turns Whisky Tasting Notes into Flavour Wheels.</h3>
|
| 168 |
<p>This model is a fine-tuned version of <a href='https://huggingface.co/microsoft/deberta-v3-base'>microsoft/deberta-v3-base</a> designed to analyze English whisky tasting notes. It predicts the intensity of eight sensory categories — <strong>grainy</strong>, <strong>grassy</strong>, <strong>fragrant</strong>, <strong>fruity</strong>, <strong>peated</strong>, <strong>woody</strong>, <strong>winey</strong> and <strong>off-notes</strong> — on a continuous scale from 0 (none) to 4 (extreme).</p>
|
| 169 |
""")
|
| 170 |
-
gr.HTML(f"<span style='color: var(--block-title-text-color)'>{
|
| 171 |
|
| 172 |
with gr.Row(): # alles nebeneinander
|
| 173 |
with gr.Column(scale=1): # linke Seite: Input
|
|
|
|
| 1 |
# Standardbibliotheken
|
| 2 |
+
import os # Umgebungsvariablen (z.B. HF_TOKEN)
|
| 3 |
+
import time # Timing / Performance-Messung
|
| 4 |
+
import random # Zufallswerte (z.B. Beispiel-Reviews)
|
| 5 |
+
import html # HTML-Escaping für sichere Ausgabe in Gradio
|
| 6 |
+
import types # Monkeypatching von Instanzen (fastText .predict)
|
| 7 |
+
import numpy as np # Numerische Arrays und Wahrscheinlichkeiten
|
| 8 |
|
| 9 |
# Machine Learning / NLP
|
| 10 |
+
import torch # PyTorch: Modelle, Tensoren, Devices
|
| 11 |
+
import fasttext # Sprach-ID-Modell (lid.176)
|
| 12 |
+
# Folgende sind notwendig, auch wenn sie nicht explizit genutzt werden:
|
| 13 |
+
import sentencepiece # Pflicht für SentencePiece-basierte Tokenizer (z.B. DeBERTa v3)
|
| 14 |
+
import tiktoken # Optionaler Converter (verhindert Fallback-Fehler bei Tokenizer)
|
| 15 |
+
from langid.langid import LanguageIdentifier, model # Alternative Sprach-ID
|
| 16 |
+
|
| 17 |
+
# Hugging Face Ökosystem
|
| 18 |
+
import spaces # HF Spaces-Dekoratoren (@spaces.GPU)
|
| 19 |
+
from transformers import AutoTokenizer # Tokenizer laden (use_fast=False für DeBERTa v3)
|
| 20 |
+
from huggingface_hub import hf_hub_download # Download von Dateien/Weights aus dem HF Hub
|
| 21 |
+
from safetensors.torch import load_file # Sicheres & schnelles Laden von Weights (.safetensors)
|
| 22 |
+
|
| 23 |
+
# Übersetzung
|
| 24 |
+
import deepl # DeepL API für automatische Übersetzung
|
| 25 |
|
| 26 |
# UI / Serving
|
| 27 |
+
import gradio as gr # Web-UI für Demo/Spaces
|
|
|
|
|
|
|
| 28 |
|
| 29 |
# Projektspezifische Module
|
| 30 |
from lib.bert_regressor import BertMultiHeadRegressor
|
| 31 |
from lib.bert_regressor_utils import (
|
| 32 |
+
predict_flavours, # Hauptfunktion: Vorhersage der 8 Aromenachsen
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
)
|
| 34 |
+
from lib.wheel import build_svg_with_values # SVG-Rendering für Flavour Wheel
|
| 35 |
+
from lib.examples import EXAMPLES # Beispiel-Reviews (vordefiniert)
|
| 36 |
|
| 37 |
### Stettings ####################################################################
|
| 38 |
|
|
|
|
| 72 |
|
| 73 |
ID = LanguageIdentifier.from_modelstring(model, norm_probs=True)
|
| 74 |
|
| 75 |
+
def _is_eng(text: str, min_chars: int = 6, threshold: float = 0.1):
|
| 76 |
t = (text or "").strip()
|
| 77 |
if len(t) < min_chars:
|
| 78 |
return True, 0.0
|
| 79 |
lang, prob = ID.classify(t) # prob ∈ [0,1]
|
| 80 |
return (lang == "en" and prob >= threshold), float(prob)
|
| 81 |
|
| 82 |
+
def _translate_en(text: str, target_lang: str = "EN-GB"):
|
| 83 |
deepl_client = deepl.Translator(DEEPL_API_KEY)
|
| 84 |
result = deepl_client.translate_text(text, target_lang=target_lang)
|
| 85 |
return result.text
|
|
|
|
| 98 |
return "Please enter a review.", {}
|
| 99 |
|
| 100 |
# Check for lang of text
|
| 101 |
+
review_is_eng, review_lang_prob = _is_eng(review)
|
| 102 |
|
| 103 |
# Abort if text is not english
|
| 104 |
if not review_is_eng:
|
| 105 |
+
review = _translate_en(review)
|
| 106 |
html_out += f"""<div style='border-radius: 2px; padding: 1px 5px; background-color: rgb(255, 237, 213);'>Your text has been automatically translated</div>
|
| 107 |
<p>{review}</p>
|
| 108 |
"""
|
|
|
|
| 142 |
def random_text():
|
| 143 |
return random.choice(EXAMPLES)
|
| 144 |
|
| 145 |
+
def _get_device_info():
|
| 146 |
if torch.cuda.is_available():
|
| 147 |
return f"<span style='border-radius: 2px; padding: 1px 5px; background-color: rgb(220, 252, 231);'>Runs on GPU: {torch.cuda.get_device_name(0)}</span>"
|
| 148 |
else:
|
|
|
|
| 163 |
<h3>Automatically turns Whisky Tasting Notes into Flavour Wheels.</h3>
|
| 164 |
<p>This model is a fine-tuned version of <a href='https://huggingface.co/microsoft/deberta-v3-base'>microsoft/deberta-v3-base</a> designed to analyze English whisky tasting notes. It predicts the intensity of eight sensory categories — <strong>grainy</strong>, <strong>grassy</strong>, <strong>fragrant</strong>, <strong>fruity</strong>, <strong>peated</strong>, <strong>woody</strong>, <strong>winey</strong> and <strong>off-notes</strong> — on a continuous scale from 0 (none) to 4 (extreme).</p>
|
| 165 |
""")
|
| 166 |
+
gr.HTML(f"<span style='color: var(--block-title-text-color)'>{_get_device_info()}</span>")
|
| 167 |
|
| 168 |
with gr.Row(): # alles nebeneinander
|
| 169 |
with gr.Column(scale=1): # linke Seite: Input
|