Spaces:
Sleeping
Sleeping
Update tasks/text.py
Browse files- tasks/text.py +184 -22
tasks/text.py
CHANGED
@@ -30,7 +30,13 @@ else:
|
|
30 |
device = torch.device("cpu")
|
31 |
|
32 |
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
class ConspiracyClassification(
|
36 |
nn.Module,
|
@@ -65,26 +71,90 @@ class ConspiracyClassification(
|
|
65 |
|
66 |
return outputs
|
67 |
|
68 |
-
class
|
69 |
nn.Module,
|
70 |
PyTorchModelHubMixin,
|
71 |
# optionally, you can add metadata which gets pushed to the model card
|
72 |
-
):
|
73 |
def __init__(self, num_classes):
|
74 |
super().__init__()
|
75 |
-
self.n_classes = num_classes
|
76 |
self.bert = BertForPreTraining.from_pretrained('digitalepidemiologylab/covid-twitter-bert-v2')
|
77 |
self.bert.cls.seq_relationship = nn.Linear(1024, num_classes)
|
78 |
-
|
79 |
-
self.sigmoid = nn.Sigmoid()
|
80 |
|
81 |
-
def forward(self, input_ids,
|
82 |
outputs = self.bert(input_ids = input_ids, token_type_ids = token_type_ids, attention_mask = input_mask)
|
83 |
-
|
84 |
logits = outputs[1]
|
85 |
|
86 |
-
return logits
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
|
89 |
@router.post(ROUTE, tags=["Text Task"],
|
90 |
description=DESCRIPTION)
|
@@ -120,28 +190,20 @@ async def evaluate_text(request: TextEvaluationRequest):
|
|
120 |
# Split dataset
|
121 |
train_test = dataset["train"]
|
122 |
test_dataset = dataset["test"]
|
123 |
-
|
124 |
-
# Start tracking emissions
|
125 |
-
tracker.start()
|
126 |
-
tracker.start_task("inference")
|
127 |
|
128 |
-
#--------------------------------------------------------------------------------------------
|
129 |
-
# YOUR MODEL INFERENCE CODE HERE
|
130 |
-
# Update the code below to replace the random baseline by your model inference within the inference pass where the energy consumption and emissions are tracked.
|
131 |
-
#--------------------------------------------------------------------------------------------
|
132 |
if MODEL =="mlp":
|
133 |
model = ConspiracyClassification.from_pretrained("ypesk/frugal-ai-mlp-baseline")
|
134 |
model = model.to(device)
|
135 |
emb_model = SentenceTransformer("paraphrase-MiniLM-L3-v2")
|
136 |
batch_size = 6
|
137 |
-
|
138 |
test_texts = torch.Tensor(emb_model.encode([t['quote'] for t in test_dataset]))
|
139 |
test_data = TensorDataset(test_texts)
|
140 |
test_sampler = SequentialSampler(test_data)
|
141 |
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
142 |
|
143 |
elif MODEL == "ct":
|
144 |
-
model =
|
145 |
model = model.to(device)
|
146 |
tokenizer = AutoTokenizer.from_pretrained('digitalepidemiologylab/covid-twitter-bert')
|
147 |
|
@@ -161,18 +223,118 @@ async def evaluate_text(request: TextEvaluationRequest):
|
|
161 |
|
162 |
test_sampler = SequentialSampler(test_data)
|
163 |
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
|
|
|
|
|
|
|
|
|
|
|
165 |
model.eval()
|
166 |
-
predictions = []
|
167 |
for batch in tqdm(test_dataloader):
|
168 |
batch = tuple(t.to(device) for t in batch)
|
169 |
with torch.no_grad():
|
170 |
if MODEL =="mlp":
|
171 |
b_texts = batch[0]
|
172 |
logits = model(b_texts)
|
173 |
-
elif MODEL == "
|
|
|
|
|
|
|
174 |
b_input_ids, b_input_mask, b_token_type_ids = batch
|
175 |
-
logits = model(b_input_ids,
|
176 |
|
177 |
logits = logits.detach().cpu().numpy()
|
178 |
predictions.extend(logits.argmax(1))
|
|
|
30 |
device = torch.device("cpu")
|
31 |
|
32 |
|
33 |
+
if torch.cuda.is_available():
|
34 |
+
device = torch.device("cuda")
|
35 |
+
else:
|
36 |
+
device = torch.device("cpu")
|
37 |
+
|
38 |
+
|
39 |
+
MODEL = "modern-large" #mlp, ct, modern-base, modern-large, gte-base, gte-large
|
40 |
|
41 |
class ConspiracyClassification(
|
42 |
nn.Module,
|
|
|
71 |
|
72 |
return outputs
|
73 |
|
74 |
+
class CTBERT(
|
75 |
nn.Module,
|
76 |
PyTorchModelHubMixin,
|
77 |
# optionally, you can add metadata which gets pushed to the model card
|
78 |
+
):
|
79 |
def __init__(self, num_classes):
|
80 |
super().__init__()
|
|
|
81 |
self.bert = BertForPreTraining.from_pretrained('digitalepidemiologylab/covid-twitter-bert-v2')
|
82 |
self.bert.cls.seq_relationship = nn.Linear(1024, num_classes)
|
|
|
|
|
83 |
|
84 |
+
def forward(self, input_ids, input_mask, token_type_ids):
|
85 |
outputs = self.bert(input_ids = input_ids, token_type_ids = token_type_ids, attention_mask = input_mask)
|
|
|
86 |
logits = outputs[1]
|
87 |
|
88 |
+
return logits
|
89 |
+
|
90 |
+
class conspiracyModelBase(
|
91 |
+
nn.Module,
|
92 |
+
PyTorchModelHubMixin,
|
93 |
+
# optionally, you can add metadata which gets pushed to the model card
|
94 |
+
):
|
95 |
+
def __init__(self, num_classes):
|
96 |
+
super().__init__()
|
97 |
+
self.n_classes = num_classes
|
98 |
+
self.bert = ModernBertForSequenceClassification.from_pretrained('answerdotai/ModernBERT-base', num_labels=num_classes)
|
99 |
+
|
100 |
+
def forward(self, input_ids, input_mask):
|
101 |
+
outputs = self.bert(input_ids = input_ids, attention_mask = input_mask)
|
102 |
+
|
103 |
+
return outputs.logits
|
104 |
|
105 |
+
class conspiracyModelLarge(
|
106 |
+
nn.Module,
|
107 |
+
PyTorchModelHubMixin,
|
108 |
+
# optionally, you can add metadata which gets pushed to the model card
|
109 |
+
):
|
110 |
+
def __init__(self, num_classes):
|
111 |
+
super().__init__()
|
112 |
+
self.n_classes = num_classes
|
113 |
+
self.bert = ModernBertForSequenceClassification.from_pretrained('answerdotai/ModernBERT-large', num_labels=num_classes)
|
114 |
+
|
115 |
+
def forward(self, input_ids, input_mask):
|
116 |
+
outputs = self.bert(input_ids = input_ids, attention_mask = input_mask)
|
117 |
+
|
118 |
+
return outputs.logits
|
119 |
+
|
120 |
+
class gteModelLarge(
|
121 |
+
nn.Module,
|
122 |
+
PyTorchModelHubMixin,
|
123 |
+
# optionally, you can add metadata which gets pushed to the model card
|
124 |
+
):
|
125 |
+
def __init__(self, num_classes):
|
126 |
+
super().__init__()
|
127 |
+
self.n_classes = num_classes
|
128 |
+
#self.bert = ModernBertForSequenceClassification.from_pretrained('answerdotai/ModernBERT-large', num_labels=num_classes)
|
129 |
+
self.gte = AutoModel.from_pretrained('Alibaba-NLP/gte-large-en-v1.5', trust_remote_code=True)
|
130 |
+
#self.cls = nn.Linear(768, num_classes)
|
131 |
+
self.cls = nn.Linear(1024, num_classes)
|
132 |
+
|
133 |
+
def forward(self, input_ids, input_mask, input_type_ids):
|
134 |
+
outputs = self.gte(input_ids = input_ids, attention_mask = input_mask, token_type_ids = input_type_ids)
|
135 |
+
embeddings = outputs.last_hidden_state[:, 0]
|
136 |
+
logits = self.cls(embeddings)
|
137 |
+
return logits
|
138 |
+
|
139 |
+
class gteModel(
|
140 |
+
nn.Module,
|
141 |
+
PyTorchModelHubMixin,
|
142 |
+
# optionally, you can add metadata which gets pushed to the model card
|
143 |
+
):
|
144 |
+
def __init__(self, num_classes):
|
145 |
+
super().__init__()
|
146 |
+
self.n_classes = num_classes
|
147 |
+
#self.bert = ModernBertForSequenceClassification.from_pretrained('answerdotai/ModernBERT-large', num_labels=num_classes)
|
148 |
+
self.gte = AutoModel.from_pretrained('Alibaba-NLP/gte-base-en-v1.5', trust_remote_code=True)
|
149 |
+
self.cls = nn.Linear(768, num_classes)
|
150 |
+
#self.cls = nn.Linear(1024, num_classes)
|
151 |
+
|
152 |
+
def forward(self, input_ids, input_mask, input_type_ids):
|
153 |
+
outputs = self.gte(input_ids = input_ids, attention_mask = input_mask, token_type_ids = input_type_ids)
|
154 |
+
embeddings = outputs.last_hidden_state[:, 0]
|
155 |
+
logits = self.cls(embeddings)
|
156 |
+
return logits
|
157 |
+
|
158 |
|
159 |
@router.post(ROUTE, tags=["Text Task"],
|
160 |
description=DESCRIPTION)
|
|
|
190 |
# Split dataset
|
191 |
train_test = dataset["train"]
|
192 |
test_dataset = dataset["test"]
|
|
|
|
|
|
|
|
|
193 |
|
|
|
|
|
|
|
|
|
194 |
if MODEL =="mlp":
|
195 |
model = ConspiracyClassification.from_pretrained("ypesk/frugal-ai-mlp-baseline")
|
196 |
model = model.to(device)
|
197 |
emb_model = SentenceTransformer("paraphrase-MiniLM-L3-v2")
|
198 |
batch_size = 6
|
199 |
+
|
200 |
test_texts = torch.Tensor(emb_model.encode([t['quote'] for t in test_dataset]))
|
201 |
test_data = TensorDataset(test_texts)
|
202 |
test_sampler = SequentialSampler(test_data)
|
203 |
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
204 |
|
205 |
elif MODEL == "ct":
|
206 |
+
model = CTBERT.from_pretrained("ypesk/frugal-ai-ct-bert-baseline")
|
207 |
model = model.to(device)
|
208 |
tokenizer = AutoTokenizer.from_pretrained('digitalepidemiologylab/covid-twitter-bert')
|
209 |
|
|
|
223 |
|
224 |
test_sampler = SequentialSampler(test_data)
|
225 |
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
226 |
+
|
227 |
+
elif MODEL == "modern-base":
|
228 |
+
model = conspiracyModelBase.from_pretrained("ypesk/frugal-ai-modern-base-baseline")
|
229 |
+
model = model.to(device)
|
230 |
+
tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-base")
|
231 |
+
|
232 |
+
test_texts = [t['quote'] for t in test_dataset]
|
233 |
+
|
234 |
+
MAX_LEN = 256 #1024 # < m some tweets will be truncated
|
235 |
+
|
236 |
+
tokenized_test = tokenizer(test_texts, max_length=MAX_LEN, padding='max_length', truncation=True)
|
237 |
+
test_input_ids, test_attention_mask = tokenized_test['input_ids'], tokenized_test['attention_mask']
|
238 |
+
|
239 |
+
test_input_ids = torch.tensor(test_input_ids)
|
240 |
+
test_attention_mask = torch.tensor(test_attention_mask)
|
241 |
+
|
242 |
+
batch_size = 12 #
|
243 |
+
test_data = TensorDataset(test_input_ids, test_attention_mask)
|
244 |
+
|
245 |
+
test_sampler = SequentialSampler(test_data)
|
246 |
+
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
247 |
+
|
248 |
+
elif MODEL == "modern-large":
|
249 |
+
model = conspiracyModelLarge.from_pretrained("ypesk/frugal-ai-modern-large-baseline")
|
250 |
+
model = model.to(device)
|
251 |
+
tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-large")
|
252 |
+
|
253 |
+
test_texts = [t['quote'] for t in test_dataset]
|
254 |
+
|
255 |
+
MAX_LEN = 256 #1024 # < m some tweets will be truncated
|
256 |
+
|
257 |
+
tokenized_test = tokenizer(test_texts, max_length=MAX_LEN, padding='max_length', truncation=True)
|
258 |
+
test_input_ids, test_attention_mask = tokenized_test['input_ids'], tokenized_test['attention_mask']
|
259 |
+
|
260 |
+
test_input_ids = torch.tensor(test_input_ids)
|
261 |
+
test_attention_mask = torch.tensor(test_attention_mask)
|
262 |
+
|
263 |
+
batch_size = 12 #
|
264 |
+
test_data = TensorDataset(test_input_ids, test_attention_mask)
|
265 |
+
|
266 |
+
test_sampler = SequentialSampler(test_data)
|
267 |
+
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
268 |
+
|
269 |
+
elif MODEL == "gte-base":
|
270 |
+
model = gteModel.from_pretrained("ypesk/frugal-ai-gte-base-baseline")
|
271 |
+
model = model.to(device)
|
272 |
+
tokenizer = AutoTokenizer.from_pretrained('Alibaba-NLP/gte-base-en-v1.5')
|
273 |
+
|
274 |
+
test_texts = [t['quote'] for t in test_dataset]
|
275 |
+
|
276 |
+
MAX_LEN = 256 #1024 # < m some tweets will be truncated
|
277 |
+
|
278 |
+
tokenized_test = tokenizer(test_texts, max_length=MAX_LEN, padding='max_length', truncation=True)
|
279 |
+
test_input_ids, test_attention_mask, test_token_type_ids = tokenized_test['input_ids'], tokenized_test['attention_mask'], tokenized_test['token_type_ids']
|
280 |
+
|
281 |
+
test_input_ids = torch.tensor(test_input_ids)
|
282 |
+
test_attention_mask = torch.tensor(test_attention_mask)
|
283 |
+
test_token_type_ids = torch.tensor(test_token_type_ids)
|
284 |
+
|
285 |
+
batch_size = 12 #
|
286 |
+
test_data = TensorDataset(test_input_ids, test_attention_mask, test_token_type_ids)
|
287 |
+
|
288 |
+
test_sampler = SequentialSampler(test_data)
|
289 |
+
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
290 |
+
|
291 |
+
elif MODEL == "gte-large":
|
292 |
+
model = gteModel.from_pretrained("ypesk/frugal-ai-gte-large-baseline")
|
293 |
+
model = model.to(device)
|
294 |
+
tokenizer = AutoTokenizer.from_pretrained('Alibaba-NLP/gte-large-en-v1.5')
|
295 |
+
|
296 |
+
test_texts = [t['quote'] for t in test_dataset]
|
297 |
+
|
298 |
+
MAX_LEN = 256 #1024 # < m some tweets will be truncated
|
299 |
+
|
300 |
+
tokenized_test = tokenizer(test_texts, max_length=MAX_LEN, padding='max_length', truncation=True)
|
301 |
+
test_input_ids, test_attention_mask, test_token_type_ids = tokenized_test['input_ids'], tokenized_test['attention_mask'], tokenized_test['token_type_ids']
|
302 |
+
|
303 |
+
test_input_ids = torch.tensor(test_input_ids)
|
304 |
+
test_attention_mask = torch.tensor(test_attention_mask)
|
305 |
+
test_token_type_ids = torch.tensor(test_token_type_ids)
|
306 |
+
|
307 |
+
batch_size = 12 #
|
308 |
+
test_data = TensorDataset(test_input_ids, test_attention_mask, test_token_type_ids)
|
309 |
+
|
310 |
+
test_sampler = SequentialSampler(test_data)
|
311 |
+
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=batch_size)
|
312 |
+
|
313 |
+
|
314 |
+
|
315 |
+
|
316 |
+
# Start tracking emissions
|
317 |
+
tracker.start()
|
318 |
+
tracker.start_task("inference")
|
319 |
|
320 |
+
#--------------------------------------------------------------------------------------------
|
321 |
+
# YOUR MODEL INFERENCE CODE HERE
|
322 |
+
# Update the code below to replace the random baseline by your model inference within the inference pass where the energy consumption and emissions are tracked.
|
323 |
+
#--------------------------------------------------------------------------------------------
|
324 |
+
|
325 |
model.eval()
|
|
|
326 |
for batch in tqdm(test_dataloader):
|
327 |
batch = tuple(t.to(device) for t in batch)
|
328 |
with torch.no_grad():
|
329 |
if MODEL =="mlp":
|
330 |
b_texts = batch[0]
|
331 |
logits = model(b_texts)
|
332 |
+
elif MODEL == "modern-base" or MODEL=="modern-large":
|
333 |
+
b_input_ids, b_input_mask = batch
|
334 |
+
logits = model(b_input_ids, b_input_mask)
|
335 |
+
elif MODEL == "gte-base" or MODEL=="gte-large" or MODEL=="ct":
|
336 |
b_input_ids, b_input_mask, b_token_type_ids = batch
|
337 |
+
logits = model(b_input_ids, b_input_mask, b_token_type_ids)
|
338 |
|
339 |
logits = logits.detach().cpu().numpy()
|
340 |
predictions.extend(logits.argmax(1))
|