Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,6 +1,5 @@
|
|
1 |
import gc
|
2 |
import os
|
3 |
-
import sys
|
4 |
import warnings
|
5 |
from types import SimpleNamespace
|
6 |
|
@@ -14,6 +13,7 @@ from generation_utils import (
|
|
14 |
decode_output,
|
15 |
save_multiple_predictions,
|
16 |
)
|
|
|
17 |
from torch.utils.data import DataLoader
|
18 |
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
19 |
from utils import seed_everything
|
@@ -111,6 +111,8 @@ with st.sidebar:
|
|
111 |
model_options = ["sagawa/ReactionT5v2-yield"] # default as requested
|
112 |
model_help = "Default model for yield prediction."
|
113 |
input_max_length_default = 400
|
|
|
|
|
114 |
|
115 |
model_name_or_path = st.selectbox(
|
116 |
"Model",
|
@@ -118,15 +120,15 @@ with st.sidebar:
|
|
118 |
index=0,
|
119 |
help=model_help,
|
120 |
)
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
|
131 |
seed = st.number_input(
|
132 |
"Random seed",
|
@@ -187,9 +189,12 @@ def load_tokenizer(model_ref: str):
|
|
187 |
|
188 |
|
189 |
@st.cache_resource(show_spinner=True)
|
190 |
-
def load_model(model_ref: str, device_str: str):
|
191 |
resolved = os.path.abspath(model_ref) if os.path.exists(model_ref) else model_ref
|
192 |
-
|
|
|
|
|
|
|
193 |
model.to(torch.device(device_str))
|
194 |
model.eval()
|
195 |
return model
|
@@ -253,14 +258,22 @@ if run:
|
|
253 |
else:
|
254 |
# Build config object expected by your dataset/utils
|
255 |
CFG = SimpleNamespace(
|
256 |
-
|
257 |
-
|
|
|
|
|
|
|
258 |
model_name_or_path=model_name_or_path,
|
259 |
input_column="input",
|
260 |
-
input_max_length=int(input_max_length)
|
261 |
-
|
262 |
-
|
263 |
-
|
|
|
|
|
|
|
|
|
|
|
264 |
seed=int(seed),
|
265 |
batch_size=int(batch_size),
|
266 |
)
|
@@ -272,7 +285,7 @@ if run:
|
|
272 |
try:
|
273 |
tokenizer = load_tokenizer(CFG.model_name_or_path)
|
274 |
CFG.tokenizer = tokenizer
|
275 |
-
model = load_model(CFG.model_name_or_path, device.type)
|
276 |
status.update(label="Model ready.", state="complete")
|
277 |
except Exception as e:
|
278 |
st.session_state["last_error"] = f"Failed to load model: {e}"
|
@@ -296,51 +309,60 @@ if run:
|
|
296 |
drop_last=False,
|
297 |
)
|
298 |
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
for i, inputs in enumerate(dataloader, start=1):
|
306 |
-
inputs = {k: v.to(device) for k, v in inputs.items()}
|
307 |
-
with torch.no_grad():
|
308 |
-
output = model.generate(
|
309 |
-
**inputs,
|
310 |
-
min_length=CFG.output_min_length,
|
311 |
-
max_length=CFG.output_max_length,
|
312 |
-
num_beams=CFG.num_beams,
|
313 |
-
num_return_sequences=CFG.num_return_sequences,
|
314 |
-
return_dict_in_generate=True,
|
315 |
-
output_scores=True,
|
316 |
-
)
|
317 |
-
sequences, scores = decode_output(output, CFG)
|
318 |
-
all_sequences.extend(sequences)
|
319 |
-
if scores:
|
320 |
-
all_scores.extend(scores)
|
321 |
-
|
322 |
-
del output
|
323 |
-
if device.type == "cuda":
|
324 |
-
torch.cuda.empty_cache()
|
325 |
-
gc.collect()
|
326 |
-
|
327 |
-
progress.progress(i / total, text=f"Generating predictions... {i}/{total}")
|
328 |
-
info_placeholder.caption(f"Processed batch {i} of {total}")
|
329 |
-
|
330 |
-
progress.empty()
|
331 |
-
info_placeholder.empty()
|
332 |
-
|
333 |
-
# Save predictions
|
334 |
-
try:
|
335 |
-
output_df = save_multiple_predictions(
|
336 |
-
input_df, all_sequences, all_scores, CFG
|
337 |
-
)
|
338 |
st.session_state["results_df"] = output_df
|
339 |
st.success("Prediction complete.")
|
340 |
-
|
341 |
-
|
342 |
-
|
343 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
344 |
|
345 |
# ------------------------------
|
346 |
# Results
|
|
|
1 |
import gc
|
2 |
import os
|
|
|
3 |
import warnings
|
4 |
from types import SimpleNamespace
|
5 |
|
|
|
13 |
decode_output,
|
14 |
save_multiple_predictions,
|
15 |
)
|
16 |
+
from models import ReactionT5Yield2
|
17 |
from torch.utils.data import DataLoader
|
18 |
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
|
19 |
from utils import seed_everything
|
|
|
111 |
model_options = ["sagawa/ReactionT5v2-yield"] # default as requested
|
112 |
model_help = "Default model for yield prediction."
|
113 |
input_max_length_default = 400
|
114 |
+
from task_yield.train import preprocess_df
|
115 |
+
from task_yield.prediction import inference_fn
|
116 |
|
117 |
model_name_or_path = st.selectbox(
|
118 |
"Model",
|
|
|
120 |
index=0,
|
121 |
help=model_help,
|
122 |
)
|
123 |
+
if task != "yield prediction":
|
124 |
+
num_beams = st.slider(
|
125 |
+
"Beam size",
|
126 |
+
min_value=1,
|
127 |
+
max_value=10,
|
128 |
+
value=5,
|
129 |
+
step=1,
|
130 |
+
help="Number of beams for beam search.",
|
131 |
+
)
|
132 |
|
133 |
seed = st.number_input(
|
134 |
"Random seed",
|
|
|
189 |
|
190 |
|
191 |
@st.cache_resource(show_spinner=True)
|
192 |
+
def load_model(model_ref: str, device_str: str, task: str):
|
193 |
resolved = os.path.abspath(model_ref) if os.path.exists(model_ref) else model_ref
|
194 |
+
if task != "yield prediction":
|
195 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(resolved)
|
196 |
+
else:
|
197 |
+
model = ReactionT5Yield2.from_pretrained(resolved)
|
198 |
model.to(torch.device(device_str))
|
199 |
model.eval()
|
200 |
return model
|
|
|
258 |
else:
|
259 |
# Build config object expected by your dataset/utils
|
260 |
CFG = SimpleNamespace(
|
261 |
+
task=task,
|
262 |
+
num_beams=int(num_beams) if task != "yield prediction" else None,
|
263 |
+
num_return_sequences=int(num_beams)
|
264 |
+
if task != "yield prediction"
|
265 |
+
else None, # tie to beams by default
|
266 |
model_name_or_path=model_name_or_path,
|
267 |
input_column="input",
|
268 |
+
input_max_length=int(input_max_length)
|
269 |
+
if task != "yield prediction"
|
270 |
+
else None,
|
271 |
+
output_max_length=int(output_max_length)
|
272 |
+
if task != "yield prediction"
|
273 |
+
else None,
|
274 |
+
output_min_length=int(output_min_length)
|
275 |
+
if task != "yield prediction"
|
276 |
+
else None,
|
277 |
seed=int(seed),
|
278 |
batch_size=int(batch_size),
|
279 |
)
|
|
|
285 |
try:
|
286 |
tokenizer = load_tokenizer(CFG.model_name_or_path)
|
287 |
CFG.tokenizer = tokenizer
|
288 |
+
model = load_model(CFG.model_name_or_path, device.type, task)
|
289 |
status.update(label="Model ready.", state="complete")
|
290 |
except Exception as e:
|
291 |
st.session_state["last_error"] = f"Failed to load model: {e}"
|
|
|
309 |
drop_last=False,
|
310 |
)
|
311 |
|
312 |
+
if task == "yield prediction":
|
313 |
+
# Use custom inference function for yield prediction
|
314 |
+
prediction = inference_fn(dataloader, model, CFG)
|
315 |
+
output_df = input_df.copy()
|
316 |
+
output_df["prediction"] = prediction
|
317 |
+
output_df["prediction"] = output_df["prediction"].clip(lower=0.0, upper=100.0)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
318 |
st.session_state["results_df"] = output_df
|
319 |
st.success("Prediction complete.")
|
320 |
+
else:
|
321 |
+
# Generation loop with progress
|
322 |
+
all_sequences, all_scores = [], []
|
323 |
+
total = len(dataloader)
|
324 |
+
progress = st.progress(0, text="Generating predictions...")
|
325 |
+
info_placeholder = st.empty()
|
326 |
+
|
327 |
+
for i, inputs in enumerate(dataloader, start=1):
|
328 |
+
inputs = {k: v.to(device) for k, v in inputs.items()}
|
329 |
+
with torch.no_grad():
|
330 |
+
output = model.generate(
|
331 |
+
**inputs,
|
332 |
+
min_length=CFG.output_min_length,
|
333 |
+
max_length=CFG.output_max_length,
|
334 |
+
num_beams=CFG.num_beams,
|
335 |
+
num_return_sequences=CFG.num_return_sequences,
|
336 |
+
return_dict_in_generate=True,
|
337 |
+
output_scores=True,
|
338 |
+
)
|
339 |
+
sequences, scores = decode_output(output, CFG)
|
340 |
+
all_sequences.extend(sequences)
|
341 |
+
if scores:
|
342 |
+
all_scores.extend(scores)
|
343 |
+
|
344 |
+
del output
|
345 |
+
if device.type == "cuda":
|
346 |
+
torch.cuda.empty_cache()
|
347 |
+
gc.collect()
|
348 |
+
|
349 |
+
progress.progress(i / total, text=f"Generating predictions... {i}/{total}")
|
350 |
+
info_placeholder.caption(f"Processed batch {i} of {total}")
|
351 |
+
|
352 |
+
progress.empty()
|
353 |
+
info_placeholder.empty()
|
354 |
+
|
355 |
+
# Save predictions
|
356 |
+
try:
|
357 |
+
output_df = save_multiple_predictions(
|
358 |
+
input_df, all_sequences, all_scores, CFG
|
359 |
+
)
|
360 |
+
st.session_state["results_df"] = output_df
|
361 |
+
st.success("Prediction complete.")
|
362 |
+
except Exception as e:
|
363 |
+
st.session_state["last_error"] = f"Failed to assemble output: {e}"
|
364 |
+
st.error(st.session_state["last_error"])
|
365 |
+
st.stop()
|
366 |
|
367 |
# ------------------------------
|
368 |
# Results
|