Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -2,19 +2,18 @@
|
|
2 |
|
3 |
import streamlit as st
|
4 |
import logging
|
|
|
5 |
import pandas as pd
|
6 |
import plotly.express as px
|
7 |
-
import torch
|
8 |
-
from typing import Union, List
|
9 |
-
|
10 |
-
from langdetect import detect, LangDetectException
|
11 |
from transformers import (
|
12 |
AutoTokenizer,
|
13 |
AutoModelForSeq2SeqLM,
|
14 |
pipeline,
|
15 |
BitsAndBytesConfig,
|
16 |
)
|
|
|
17 |
import evaluate
|
|
|
18 |
|
19 |
# ββββββββββ Logging ββββββββββ
|
20 |
logging.basicConfig(
|
@@ -38,7 +37,7 @@ class ModelManager:
|
|
38 |
quantize: bool = True,
|
39 |
default_tgt: str = None,
|
40 |
):
|
41 |
-
#
|
42 |
if quantize and not torch.cuda.is_available():
|
43 |
logger.warning("CUDA unavailable; disabling 8-bit quantization")
|
44 |
quantize = False
|
@@ -50,7 +49,7 @@ class ModelManager:
|
|
50 |
]
|
51 |
self.default_tgt = default_tgt # will auto-pick if None
|
52 |
|
53 |
-
self.
|
54 |
self.tokenizer = None
|
55 |
self.model = None
|
56 |
self.pipeline = None
|
@@ -66,14 +65,10 @@ class ModelManager:
|
|
66 |
logger.info(f"Loading tokenizer for {model_name}")
|
67 |
tok = AutoTokenizer.from_pretrained(model_name, use_fast=True)
|
68 |
if not hasattr(tok, "lang_code_to_id"):
|
69 |
-
raise AttributeError(
|
70 |
-
f"Tokenizer for {model_name} missing lang_code_to_id"
|
71 |
-
)
|
72 |
|
73 |
# Load model (with or without 8-bit)
|
74 |
-
logger.info(
|
75 |
-
f"Loading model {model_name} (8-bit={self.quantize})"
|
76 |
-
)
|
77 |
if self.quantize:
|
78 |
bnb_cfg = BitsAndBytesConfig(load_in_8bit=True)
|
79 |
mdl = AutoModelForSeq2SeqLM.from_pretrained(
|
@@ -96,7 +91,7 @@ class ModelManager:
|
|
96 |
)
|
97 |
|
98 |
# Store and break
|
99 |
-
self.
|
100 |
self.tokenizer = tok
|
101 |
self.model = mdl
|
102 |
self.pipeline = pipe
|
@@ -104,22 +99,19 @@ class ModelManager:
|
|
104 |
|
105 |
# Auto-pick Turkish target code if none specified
|
106 |
if not self.default_tgt:
|
107 |
-
tur_codes = [
|
108 |
-
c for c in self.lang_codes if c.lower().startswith("tr")
|
109 |
-
]
|
110 |
if not tur_codes:
|
111 |
raise ValueError(f"No Turkish code found in {model_name}")
|
112 |
self.default_tgt = tur_codes[0]
|
113 |
logger.info(f"Default target language: {self.default_tgt}")
|
114 |
|
115 |
return
|
|
|
116 |
except Exception as e:
|
117 |
logger.warning(f"Failed to load {model_name}: {e}")
|
118 |
last_err = e
|
119 |
|
120 |
-
raise RuntimeError(
|
121 |
-
f"Could not load any model from {self.candidates}: {last_err}"
|
122 |
-
)
|
123 |
|
124 |
def translate(
|
125 |
self,
|
@@ -138,9 +130,7 @@ class ModelManager:
|
|
138 |
sample = text[0] if isinstance(text, list) else text
|
139 |
try:
|
140 |
iso = detect(sample).lower()
|
141 |
-
candidates = [
|
142 |
-
c for c in self.lang_codes if c.lower().startswith(iso)
|
143 |
-
]
|
144 |
if not candidates:
|
145 |
raise LangDetectException(f"No code for ISO '{iso}'")
|
146 |
exact = [c for c in candidates if c.lower() == iso]
|
@@ -148,9 +138,7 @@ class ModelManager:
|
|
148 |
logger.info(f"Auto-detected src_lang={src}")
|
149 |
except Exception as e:
|
150 |
logger.warning(f"langdetect failed ({e}); defaulting to English")
|
151 |
-
eng_codes = [
|
152 |
-
c for c in self.lang_codes if c.lower().startswith("en")
|
153 |
-
]
|
154 |
src = eng_codes[0] if eng_codes else self.lang_codes[0]
|
155 |
else:
|
156 |
src = src_lang
|
@@ -159,13 +147,14 @@ class ModelManager:
|
|
159 |
|
160 |
def get_info(self):
|
161 |
"""Return metadata for the sidebar display."""
|
|
|
162 |
device = "cpu"
|
163 |
if torch.cuda.is_available() and hasattr(self.model, "device"):
|
164 |
-
|
165 |
-
device =
|
166 |
return {
|
167 |
-
"model": self.
|
168 |
-
"quantized":
|
169 |
"device": device,
|
170 |
"default_tgt": self.default_tgt,
|
171 |
}
|
@@ -175,9 +164,16 @@ class ModelManager:
|
|
175 |
class TranslationEvaluator:
|
176 |
def __init__(self):
|
177 |
self.bleu = evaluate.load("bleu")
|
178 |
-
|
179 |
-
|
180 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
181 |
|
182 |
def evaluate(
|
183 |
self,
|
@@ -186,36 +182,71 @@ class TranslationEvaluator:
|
|
186 |
predictions: List[str],
|
187 |
):
|
188 |
results = {}
|
|
|
189 |
# BLEU
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
194 |
# BERTScore (general)
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
# BERTurk (Turkish)
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
204 |
# COMET
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
210 |
return results
|
211 |
|
212 |
|
213 |
# ββββββββββ Streamlit App ββββββββββ
|
214 |
-
|
215 |
@st.cache_resource
|
216 |
def load_resources():
|
217 |
mgr = ModelManager(quantize=True)
|
218 |
-
ev
|
219 |
return mgr, ev
|
220 |
|
221 |
|
@@ -235,17 +266,17 @@ def process_text(
|
|
235 |
metrics: List[str],
|
236 |
):
|
237 |
out = mgr.translate(src)
|
238 |
-
hyp = out[0]["translation_text"]
|
239 |
scores = ev.evaluate([src], [ref or ""], [hyp])
|
240 |
return {
|
241 |
"source": src,
|
242 |
"reference": ref,
|
243 |
"hypothesis": hyp,
|
244 |
-
**{m: scores
|
245 |
}
|
246 |
|
247 |
|
248 |
-
def
|
249 |
left, right = st.columns(2)
|
250 |
with left:
|
251 |
st.markdown("**Source:**")
|
@@ -257,7 +288,7 @@ def _show_single_results(res: dict):
|
|
257 |
st.write(res["reference"])
|
258 |
with right:
|
259 |
st.markdown("### Scores")
|
260 |
-
df = pd.DataFrame([{
|
261 |
st.table(df)
|
262 |
|
263 |
|
@@ -283,13 +314,13 @@ def process_file(
|
|
283 |
for s, r, h in zip(srcs, refs, hyps):
|
284 |
sc = ev.evaluate([s], [r], [h])
|
285 |
entry = {"src": s, "ref_tr": r, "hyp_tr": h}
|
286 |
-
entry.update({m: sc
|
287 |
results.append(entry)
|
288 |
prog.progress(min(i + batch_size, total) / total)
|
289 |
return pd.DataFrame(results)
|
290 |
|
291 |
|
292 |
-
def
|
293 |
for m in metrics:
|
294 |
st.markdown(f"#### {m} Distribution")
|
295 |
fig = px.histogram(df, x=m)
|
@@ -297,9 +328,7 @@ def _show_batch_viz(df: pd.DataFrame, metrics: List[str]):
|
|
297 |
|
298 |
|
299 |
def main():
|
300 |
-
st.set_page_config(
|
301 |
-
page_title="π€ TranslationβTurkish Quality", layout="wide"
|
302 |
-
)
|
303 |
st.title("π€ Translation β TR Quality & COMET")
|
304 |
st.markdown(
|
305 |
"Translate any language into Turkish and evaluate with BLEU, BERTScore, BERTurk & COMET."
|
@@ -326,7 +355,7 @@ def main():
|
|
326 |
if st.button("Evaluate"):
|
327 |
with st.spinner("Translating & evaluatingβ¦"):
|
328 |
res = process_text(src, ref, mgr, ev, metrics)
|
329 |
-
|
330 |
|
331 |
with tab2:
|
332 |
uploaded = st.file_uploader(
|
@@ -337,10 +366,8 @@ def main():
|
|
337 |
df_res = process_file(uploaded, mgr, ev, metrics, batch_size)
|
338 |
st.markdown("### Batch Results")
|
339 |
st.dataframe(df_res, use_container_width=True)
|
340 |
-
|
341 |
-
st.download_button(
|
342 |
-
"Download CSV", df_res.to_csv(index=False), "results.csv"
|
343 |
-
)
|
344 |
|
345 |
|
346 |
if __name__ == "__main__":
|
|
|
2 |
|
3 |
import streamlit as st
|
4 |
import logging
|
5 |
+
import torch
|
6 |
import pandas as pd
|
7 |
import plotly.express as px
|
|
|
|
|
|
|
|
|
8 |
from transformers import (
|
9 |
AutoTokenizer,
|
10 |
AutoModelForSeq2SeqLM,
|
11 |
pipeline,
|
12 |
BitsAndBytesConfig,
|
13 |
)
|
14 |
+
from langdetect import detect, LangDetectException
|
15 |
import evaluate
|
16 |
+
from typing import Union, List
|
17 |
|
18 |
# ββββββββββ Logging ββββββββββ
|
19 |
logging.basicConfig(
|
|
|
37 |
quantize: bool = True,
|
38 |
default_tgt: str = None,
|
39 |
):
|
40 |
+
# Disable 8-bit if CUDA isn't available
|
41 |
if quantize and not torch.cuda.is_available():
|
42 |
logger.warning("CUDA unavailable; disabling 8-bit quantization")
|
43 |
quantize = False
|
|
|
49 |
]
|
50 |
self.default_tgt = default_tgt # will auto-pick if None
|
51 |
|
52 |
+
self.model_name: str = None
|
53 |
self.tokenizer = None
|
54 |
self.model = None
|
55 |
self.pipeline = None
|
|
|
65 |
logger.info(f"Loading tokenizer for {model_name}")
|
66 |
tok = AutoTokenizer.from_pretrained(model_name, use_fast=True)
|
67 |
if not hasattr(tok, "lang_code_to_id"):
|
68 |
+
raise AttributeError(f"Tokenizer for {model_name} missing lang_code_to_id")
|
|
|
|
|
69 |
|
70 |
# Load model (with or without 8-bit)
|
71 |
+
logger.info(f"Loading model {model_name} (8-bit={self.quantize})")
|
|
|
|
|
72 |
if self.quantize:
|
73 |
bnb_cfg = BitsAndBytesConfig(load_in_8bit=True)
|
74 |
mdl = AutoModelForSeq2SeqLM.from_pretrained(
|
|
|
91 |
)
|
92 |
|
93 |
# Store and break
|
94 |
+
self.model_name = model_name
|
95 |
self.tokenizer = tok
|
96 |
self.model = mdl
|
97 |
self.pipeline = pipe
|
|
|
99 |
|
100 |
# Auto-pick Turkish target code if none specified
|
101 |
if not self.default_tgt:
|
102 |
+
tur_codes = [c for c in self.lang_codes if c.lower().startswith("tr")]
|
|
|
|
|
103 |
if not tur_codes:
|
104 |
raise ValueError(f"No Turkish code found in {model_name}")
|
105 |
self.default_tgt = tur_codes[0]
|
106 |
logger.info(f"Default target language: {self.default_tgt}")
|
107 |
|
108 |
return
|
109 |
+
|
110 |
except Exception as e:
|
111 |
logger.warning(f"Failed to load {model_name}: {e}")
|
112 |
last_err = e
|
113 |
|
114 |
+
raise RuntimeError(f"Could not load any model from {self.candidates}: {last_err}")
|
|
|
|
|
115 |
|
116 |
def translate(
|
117 |
self,
|
|
|
130 |
sample = text[0] if isinstance(text, list) else text
|
131 |
try:
|
132 |
iso = detect(sample).lower()
|
133 |
+
candidates = [c for c in self.lang_codes if c.lower().startswith(iso)]
|
|
|
|
|
134 |
if not candidates:
|
135 |
raise LangDetectException(f"No code for ISO '{iso}'")
|
136 |
exact = [c for c in candidates if c.lower() == iso]
|
|
|
138 |
logger.info(f"Auto-detected src_lang={src}")
|
139 |
except Exception as e:
|
140 |
logger.warning(f"langdetect failed ({e}); defaulting to English")
|
141 |
+
eng_codes = [c for c in self.lang_codes if c.lower().startswith("en")]
|
|
|
|
|
142 |
src = eng_codes[0] if eng_codes else self.lang_codes[0]
|
143 |
else:
|
144 |
src = src_lang
|
|
|
147 |
|
148 |
def get_info(self):
|
149 |
"""Return metadata for the sidebar display."""
|
150 |
+
quantized = getattr(self.model, "is_loaded_in_8bit", False)
|
151 |
device = "cpu"
|
152 |
if torch.cuda.is_available() and hasattr(self.model, "device"):
|
153 |
+
dev = self.model.device
|
154 |
+
device = str(dev) if isinstance(dev, torch.device) else f"cuda:{getattr(dev, 'index', '')}"
|
155 |
return {
|
156 |
+
"model": self.model_name,
|
157 |
+
"quantized": quantized,
|
158 |
"device": device,
|
159 |
"default_tgt": self.default_tgt,
|
160 |
}
|
|
|
164 |
class TranslationEvaluator:
|
165 |
def __init__(self):
|
166 |
self.bleu = evaluate.load("bleu")
|
167 |
+
try:
|
168 |
+
self.bertscore = evaluate.load("bertscore")
|
169 |
+
except Exception as e:
|
170 |
+
logger.error("BERTScore load error: %s", e)
|
171 |
+
self.bertscore = None
|
172 |
+
try:
|
173 |
+
self.comet = evaluate.load("comet", model_id="unbabel/comet-mqm-qe-da")
|
174 |
+
except Exception as e:
|
175 |
+
logger.error("COMET load error: %s", e)
|
176 |
+
self.comet = None
|
177 |
|
178 |
def evaluate(
|
179 |
self,
|
|
|
182 |
predictions: List[str],
|
183 |
):
|
184 |
results = {}
|
185 |
+
|
186 |
# BLEU
|
187 |
+
try:
|
188 |
+
bleu_res = self.bleu.compute(
|
189 |
+
predictions=predictions,
|
190 |
+
references=[[r] for r in references],
|
191 |
+
)
|
192 |
+
results["BLEU"] = float(bleu_res.get("bleu", 0.0))
|
193 |
+
except Exception as e:
|
194 |
+
logger.error("BLEU compute error: %s", e)
|
195 |
+
results["BLEU"] = 0.0
|
196 |
+
|
197 |
# BERTScore (general)
|
198 |
+
if self.bertscore:
|
199 |
+
try:
|
200 |
+
bs = self.bertscore.compute(
|
201 |
+
predictions=predictions, references=references, lang="xx"
|
202 |
+
)
|
203 |
+
f1 = bs.get("f1", [])
|
204 |
+
results["BERTScore"] = float(sum(f1)) / max(len(f1), 1)
|
205 |
+
except Exception as e:
|
206 |
+
logger.error("BERTScore compute error: %s", e)
|
207 |
+
results["BERTScore"] = 0.0
|
208 |
+
else:
|
209 |
+
results["BERTScore"] = 0.0
|
210 |
+
|
211 |
# BERTurk (Turkish)
|
212 |
+
if self.bertscore:
|
213 |
+
try:
|
214 |
+
bs_tr = self.bertscore.compute(
|
215 |
+
predictions=predictions, references=references, lang="tr"
|
216 |
+
)
|
217 |
+
f1_tr = bs_tr.get("f1", [])
|
218 |
+
results["BERTurk"] = float(sum(f1_tr)) / max(len(f1_tr), 1)
|
219 |
+
except Exception as e:
|
220 |
+
logger.error("BERTurk compute error: %s", e)
|
221 |
+
results["BERTurk"] = 0.0
|
222 |
+
else:
|
223 |
+
results["BERTurk"] = 0.0
|
224 |
+
|
225 |
# COMET
|
226 |
+
if self.comet:
|
227 |
+
try:
|
228 |
+
cm = self.comet.compute(
|
229 |
+
srcs=sources, hyps=predictions, refs=references
|
230 |
+
)
|
231 |
+
sc = cm.get("scores", None)
|
232 |
+
if isinstance(sc, list):
|
233 |
+
results["COMET"] = float(sc[0]) if sc else 0.0
|
234 |
+
else:
|
235 |
+
results["COMET"] = float(sc or 0.0)
|
236 |
+
except Exception as e:
|
237 |
+
logger.error("COMET compute error: %s", e)
|
238 |
+
results["COMET"] = 0.0
|
239 |
+
else:
|
240 |
+
results["COMET"] = 0.0
|
241 |
+
|
242 |
return results
|
243 |
|
244 |
|
245 |
# ββββββββββ Streamlit App ββββββββββ
|
|
|
246 |
@st.cache_resource
|
247 |
def load_resources():
|
248 |
mgr = ModelManager(quantize=True)
|
249 |
+
ev = TranslationEvaluator()
|
250 |
return mgr, ev
|
251 |
|
252 |
|
|
|
266 |
metrics: List[str],
|
267 |
):
|
268 |
out = mgr.translate(src)
|
269 |
+
hyp = out[0]["translation_text"] if isinstance(out, list) else out["translation_text"]
|
270 |
scores = ev.evaluate([src], [ref or ""], [hyp])
|
271 |
return {
|
272 |
"source": src,
|
273 |
"reference": ref,
|
274 |
"hypothesis": hyp,
|
275 |
+
**{m: scores.get(m, 0.0) for m in metrics},
|
276 |
}
|
277 |
|
278 |
|
279 |
+
def show_single_results(res: dict, metrics: List[str]):
|
280 |
left, right = st.columns(2)
|
281 |
with left:
|
282 |
st.markdown("**Source:**")
|
|
|
288 |
st.write(res["reference"])
|
289 |
with right:
|
290 |
st.markdown("### Scores")
|
291 |
+
df = pd.DataFrame([{m: res[m] for m in metrics}])
|
292 |
st.table(df)
|
293 |
|
294 |
|
|
|
314 |
for s, r, h in zip(srcs, refs, hyps):
|
315 |
sc = ev.evaluate([s], [r], [h])
|
316 |
entry = {"src": s, "ref_tr": r, "hyp_tr": h}
|
317 |
+
entry.update({m: sc.get(m, 0.0) for m in metrics})
|
318 |
results.append(entry)
|
319 |
prog.progress(min(i + batch_size, total) / total)
|
320 |
return pd.DataFrame(results)
|
321 |
|
322 |
|
323 |
+
def show_batch_viz(df: pd.DataFrame, metrics: List[str]):
|
324 |
for m in metrics:
|
325 |
st.markdown(f"#### {m} Distribution")
|
326 |
fig = px.histogram(df, x=m)
|
|
|
328 |
|
329 |
|
330 |
def main():
|
331 |
+
st.set_page_config(page_title="π€ TranslationβTurkish Quality", layout="wide")
|
|
|
|
|
332 |
st.title("π€ Translation β TR Quality & COMET")
|
333 |
st.markdown(
|
334 |
"Translate any language into Turkish and evaluate with BLEU, BERTScore, BERTurk & COMET."
|
|
|
355 |
if st.button("Evaluate"):
|
356 |
with st.spinner("Translating & evaluatingβ¦"):
|
357 |
res = process_text(src, ref, mgr, ev, metrics)
|
358 |
+
show_single_results(res, metrics)
|
359 |
|
360 |
with tab2:
|
361 |
uploaded = st.file_uploader(
|
|
|
366 |
df_res = process_file(uploaded, mgr, ev, metrics, batch_size)
|
367 |
st.markdown("### Batch Results")
|
368 |
st.dataframe(df_res, use_container_width=True)
|
369 |
+
show_batch_viz(df_res, metrics)
|
370 |
+
st.download_button("Download CSV", df_res.to_csv(index=False), "results.csv")
|
|
|
|
|
371 |
|
372 |
|
373 |
if __name__ == "__main__":
|