Spaces:
Sleeping
Sleeping
Mila
commited on
Commit
•
ec3e101
1
Parent(s):
949bc1b
Working version. Needs updates for game
Browse files- .gitattributes +2 -0
- app.py +291 -291
- requirements.txt +5 -5
- train.py +276 -276
.gitattributes
CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
*.csv filter=lfs diff=lfs merge=lfs -text
|
37 |
+
*.safetenstors filter=lfs diff=lfs merge=lfs -text
|
app.py
CHANGED
@@ -1,292 +1,292 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import math
|
3 |
-
import spacy
|
4 |
-
from datasets import load_dataset
|
5 |
-
from sentence_transformers import SentenceTransformer
|
6 |
-
from sentence_transformers import InputExample
|
7 |
-
from sentence_transformers import losses
|
8 |
-
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
|
9 |
-
from transformers import TrainingArguments, Trainer
|
10 |
-
import torch
|
11 |
-
import torch.nn.functional as F
|
12 |
-
from torch.utils.data import DataLoader
|
13 |
-
import numpy as np
|
14 |
-
import evaluate
|
15 |
-
import nltk
|
16 |
-
from nltk.corpus import stopwords
|
17 |
-
import subprocess
|
18 |
-
import sys
|
19 |
-
|
20 |
-
# !pip install https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl
|
21 |
-
# subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl'])
|
22 |
-
# tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
23 |
-
nltk.download('stopwords')
|
24 |
-
nlp = spacy.load("en_core_web_sm")
|
25 |
-
stops = stopwords.words("english")
|
26 |
-
|
27 |
-
# answer = "Pizza"
|
28 |
-
guesses = []
|
29 |
-
answer = "Pizza"
|
30 |
-
|
31 |
-
|
32 |
-
#Mean Pooling - Take attention mask into account for correct averaging
|
33 |
-
def mean_pooling(model_output, attention_mask):
|
34 |
-
token_embeddings = model_output['token_embeddings'] #First element of model_output contains all token embeddings
|
35 |
-
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
36 |
-
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
|
37 |
-
|
38 |
-
|
39 |
-
def normalize(comment, lowercase, remove_stopwords):
|
40 |
-
if lowercase:
|
41 |
-
comment = comment.lower()
|
42 |
-
comment = nlp(comment)
|
43 |
-
lemmatized = list()
|
44 |
-
for word in comment:
|
45 |
-
lemma = word.lemma_.strip()
|
46 |
-
if lemma:
|
47 |
-
if not remove_stopwords or (remove_stopwords and lemma not in stops):
|
48 |
-
lemmatized.append(lemma)
|
49 |
-
return " ".join(lemmatized)
|
50 |
-
|
51 |
-
|
52 |
-
# def tokenize_function(examples):
|
53 |
-
# return tokenizer(examples["text"])
|
54 |
-
|
55 |
-
|
56 |
-
def compute_metrics(eval_pred):
|
57 |
-
logits, labels = eval_pred
|
58 |
-
predictions = np.argmax(logits, axis=-1)
|
59 |
-
metric = evaluate.load("accuracy")
|
60 |
-
return metric.compute(predictions=predictions, references=labels)
|
61 |
-
|
62 |
-
|
63 |
-
def training():
|
64 |
-
dataset_id = "ag_news"
|
65 |
-
dataset = load_dataset(dataset_id)
|
66 |
-
# dataset = dataset["train"]
|
67 |
-
# tokenized_datasets = dataset.map(tokenize_function, batched=True)
|
68 |
-
|
69 |
-
print(f"- The {dataset_id} dataset has {dataset['train'].num_rows} examples.")
|
70 |
-
print(f"- Each example is a {type(dataset['train'][0])} with a {type(dataset['train'][0]['text'])} as value.")
|
71 |
-
print(f"- Examples look like this: {dataset['train'][0]}")
|
72 |
-
|
73 |
-
# small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
|
74 |
-
# small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
|
75 |
-
|
76 |
-
# dataset = dataset["train"].map(tokenize_function, batched=True)
|
77 |
-
# dataset.set_format(type="torch", columns=["input_ids", "token_type_ids", "attention_mask", "label"])
|
78 |
-
# dataset.format['type']
|
79 |
-
|
80 |
-
# print(dataset)
|
81 |
-
|
82 |
-
train_examples = []
|
83 |
-
train_data = dataset["train"]
|
84 |
-
# For agility we only 1/2 of our available data
|
85 |
-
n_examples = dataset["train"].num_rows // 2
|
86 |
-
# n_remaining = dataset["train"].num_rows - n_examples
|
87 |
-
# dataset_clean = {}
|
88 |
-
# # dataset_0 = []
|
89 |
-
# # dataset_1 = []
|
90 |
-
# # dataset_2 = []
|
91 |
-
# # dataset_3 = []
|
92 |
-
# for i in range(n_examples):
|
93 |
-
# dataset_clean[i] = {}
|
94 |
-
# dataset_clean[i]["text"] = normalize(train_data[i]["text"], lowercase=True, remove_stopwords=True)
|
95 |
-
# dataset_clean[i]["label"] = train_data[i]["label"]
|
96 |
-
# if train_data[i]["label"] == 0:
|
97 |
-
# dataset_0.append(dataset_clean[i])
|
98 |
-
# elif train_data[i]["label"] == 1:
|
99 |
-
# dataset_1.append(dataset_clean[i])
|
100 |
-
# elif train_data[i]["label"] == 2:
|
101 |
-
# dataset_2.append(dataset_clean[i])
|
102 |
-
# elif train_data[i]["label"] == 3:
|
103 |
-
# dataset_3.append(dataset_clean[i])
|
104 |
-
# n_0 = len(dataset_0) // 2
|
105 |
-
# n_1 = len(dataset_1) // 2
|
106 |
-
# n_2 = len(dataset_2) // 2
|
107 |
-
# n_3 = len(dataset_3) // 2
|
108 |
-
# print("Label lengths:", len(dataset_0), len(dataset_1), len(dataset_2), len(dataset_3))
|
109 |
-
|
110 |
-
for i in range(n_examples):
|
111 |
-
example = train_data[i]
|
112 |
-
# example_opposite = dataset_clean[-(i)]
|
113 |
-
# print(example["text"])
|
114 |
-
train_examples.append(InputExample(texts=[example['text']], label=example['label']))
|
115 |
-
|
116 |
-
# for i in range(n_0):
|
117 |
-
# example = dataset_0[i]
|
118 |
-
# # example_opposite = dataset_0[-(i)]
|
119 |
-
# # print(example["text"])
|
120 |
-
# train_examples.append(InputExample(texts=[example['text']], label=0))
|
121 |
-
|
122 |
-
# for i in range(n_1):
|
123 |
-
# example = dataset_1[i]
|
124 |
-
# # example_opposite = dataset_1[-(i)]
|
125 |
-
# # print(example["text"])
|
126 |
-
# train_examples.append(InputExample(texts=[example['text']], label=1))
|
127 |
-
|
128 |
-
# for i in range(n_2):
|
129 |
-
# example = dataset_2[i]
|
130 |
-
# # example_opposite = dataset_2[-(i)]
|
131 |
-
# # print(example["text"])
|
132 |
-
# train_examples.append(InputExample(texts=[example['text']], label=2))
|
133 |
-
|
134 |
-
# for i in range(n_3):
|
135 |
-
# example = dataset_3[i]
|
136 |
-
# # example_opposite = dataset_3[-(i)]
|
137 |
-
# # print(example["text"])
|
138 |
-
# train_examples.append(InputExample(texts=[example['text']], label=3))
|
139 |
-
|
140 |
-
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=25)
|
141 |
-
|
142 |
-
print("END DATALOADER")
|
143 |
-
|
144 |
-
# print(train_examples)
|
145 |
-
|
146 |
-
embeddings = finetune(train_dataloader)
|
147 |
-
|
148 |
-
return (dataset['train'].num_rows, type(dataset['train'][0]), type(dataset['train'][0]['text']), dataset['train'][0], embeddings)
|
149 |
-
|
150 |
-
|
151 |
-
def finetune(train_dataloader):
|
152 |
-
# model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
|
153 |
-
model_id = "sentence-transformers/all-MiniLM-L6-v2"
|
154 |
-
model = SentenceTransformer(model_id)
|
155 |
-
|
156 |
-
# training_args = TrainingArguments(output_dir="test_trainer")
|
157 |
-
|
158 |
-
# USE THIS LINK
|
159 |
-
# https://huggingface.co/blog/how-to-train-sentence-transformers
|
160 |
-
|
161 |
-
train_loss = losses.BatchHardSoftMarginTripletLoss(model=model)
|
162 |
-
|
163 |
-
print("BEGIN FIT")
|
164 |
-
|
165 |
-
model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=10)
|
166 |
-
|
167 |
-
model.save("ag_news_model")
|
168 |
-
|
169 |
-
model.save_to_hub("smhavens/all-MiniLM-agNews")
|
170 |
-
# accuracy = compute_metrics(eval, metric)
|
171 |
-
|
172 |
-
# training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")
|
173 |
-
|
174 |
-
# trainer = Trainer(
|
175 |
-
# model=model,
|
176 |
-
# args=training_args,
|
177 |
-
# train_dataset=train,
|
178 |
-
# eval_dataset=eval,
|
179 |
-
# compute_metrics=compute_metrics,
|
180 |
-
# )
|
181 |
-
|
182 |
-
# trainer.train()
|
183 |
-
|
184 |
-
def embeddings():
|
185 |
-
model = SentenceTransformer("ag_news_model")
|
186 |
-
device = torch.device('cuda:0')
|
187 |
-
model = model.to(device)
|
188 |
-
sentences = ["This is an example sentence", "Each sentence is converted"]
|
189 |
-
|
190 |
-
# model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
191 |
-
embeddings = model.encode(sentences)
|
192 |
-
# print(embeddings)
|
193 |
-
|
194 |
-
# Sentences we want sentence embeddings for
|
195 |
-
sentences = ['This is an example sentence', 'Each sentence is converted']
|
196 |
-
|
197 |
-
# Load model from HuggingFace Hub
|
198 |
-
tokenizer = AutoTokenizer.from_pretrained('ag_news_model')
|
199 |
-
# model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
200 |
-
|
201 |
-
# Tokenize sentences
|
202 |
-
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
|
203 |
-
|
204 |
-
# print(model.device)
|
205 |
-
# print(encoded_input["input_ids"].device)
|
206 |
-
# print(encoded_input["attention_mask"].device)
|
207 |
-
# print(encoded_input["token_type_ids"].device)
|
208 |
-
encoded_input["input_ids"] = encoded_input["input_ids"].to(device)
|
209 |
-
encoded_input["attention_mask"] = encoded_input["attention_mask"].to(device)
|
210 |
-
encoded_input['token_type_ids'] = encoded_input['token_type_ids'].to(device)
|
211 |
-
# print(encoded_input)
|
212 |
-
|
213 |
-
# print(encoded_input["input_ids"].device)
|
214 |
-
# print(encoded_input["attention_mask"].device)
|
215 |
-
# print(encoded_input["token_type_ids"].device)
|
216 |
-
|
217 |
-
encoded_input['input'] = {'input_ids':encoded_input['input_ids'], 'attention_mask':encoded_input['attention_mask']}
|
218 |
-
|
219 |
-
# + encoded_input['token_type_ids'] + encoded_input['attention_mask']
|
220 |
-
del encoded_input['input_ids']
|
221 |
-
del encoded_input['token_type_ids']
|
222 |
-
del encoded_input['attention_mask']
|
223 |
-
|
224 |
-
# print(encoded_input)
|
225 |
-
|
226 |
-
# encoded_input.to(device)
|
227 |
-
# Compute token embeddings
|
228 |
-
with torch.no_grad():
|
229 |
-
model_output = model(**encoded_input)
|
230 |
-
|
231 |
-
print(model_output)
|
232 |
-
# Perform pooling
|
233 |
-
sentence_embeddings = mean_pooling(model_output, encoded_input['input']["attention_mask"])
|
234 |
-
|
235 |
-
# Normalize embeddings
|
236 |
-
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
|
237 |
-
|
238 |
-
print("Sentence embeddings:")
|
239 |
-
print(sentence_embeddings)
|
240 |
-
return sentence_embeddings
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
def greet(name):
|
245 |
-
return "Hello " + name + "!!"
|
246 |
-
|
247 |
-
def check_answer(guess:str):
|
248 |
-
global guesses
|
249 |
-
global answer
|
250 |
-
guesses.append(guess)
|
251 |
-
output = ""
|
252 |
-
for guess in guesses:
|
253 |
-
output += ("- " + guess + "\n")
|
254 |
-
output = output[:-1]
|
255 |
-
|
256 |
-
if guess.lower() == answer.lower():
|
257 |
-
return "Correct!", output
|
258 |
-
else:
|
259 |
-
return "Try again!", output
|
260 |
-
|
261 |
-
def main():
|
262 |
-
word1 = "Black"
|
263 |
-
word2 = "White"
|
264 |
-
word3 = "Sun"
|
265 |
-
global answer
|
266 |
-
answer = "Moon"
|
267 |
-
global guesses
|
268 |
-
|
269 |
-
# num_rows, data_type, value, example, embeddings = training()
|
270 |
-
sent_embeddings = embeddings()
|
271 |
-
|
272 |
-
prompt = f"{word1} is to {word2} as {word3} is to ____"
|
273 |
-
with gr.Blocks() as iface:
|
274 |
-
gr.Markdown(prompt)
|
275 |
-
with gr.Tab("Guess"):
|
276 |
-
text_input = gr.Textbox()
|
277 |
-
text_output = gr.Textbox()
|
278 |
-
text_button = gr.Button("Submit")
|
279 |
-
with gr.Accordion("Open for previous guesses"):
|
280 |
-
text_guesses = gr.Textbox()
|
281 |
-
with gr.Tab("Testing"):
|
282 |
-
gr.Markdown(f"""The Embeddings are {sent_embeddings}.""")
|
283 |
-
text_button.click(check_answer, inputs=[text_input], outputs=[text_output, text_guesses])
|
284 |
-
# iface = gr.Interface(fn=greet, inputs="text", outputs="text")
|
285 |
-
iface.launch()
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
if __name__ == "__main__":
|
292 |
main()
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import math
|
3 |
+
import spacy
|
4 |
+
from datasets import load_dataset
|
5 |
+
from sentence_transformers import SentenceTransformer
|
6 |
+
from sentence_transformers import InputExample
|
7 |
+
from sentence_transformers import losses
|
8 |
+
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
|
9 |
+
from transformers import TrainingArguments, Trainer
|
10 |
+
import torch
|
11 |
+
import torch.nn.functional as F
|
12 |
+
from torch.utils.data import DataLoader
|
13 |
+
import numpy as np
|
14 |
+
import evaluate
|
15 |
+
import nltk
|
16 |
+
from nltk.corpus import stopwords
|
17 |
+
import subprocess
|
18 |
+
import sys
|
19 |
+
|
20 |
+
# !pip install https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl
|
21 |
+
# subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl'])
|
22 |
+
# tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
23 |
+
nltk.download('stopwords')
|
24 |
+
nlp = spacy.load("en_core_web_sm")
|
25 |
+
stops = stopwords.words("english")
|
26 |
+
|
27 |
+
# answer = "Pizza"
|
28 |
+
guesses = []
|
29 |
+
answer = "Pizza"
|
30 |
+
|
31 |
+
|
32 |
+
#Mean Pooling - Take attention mask into account for correct averaging
|
33 |
+
def mean_pooling(model_output, attention_mask):
|
34 |
+
token_embeddings = model_output['token_embeddings'] #First element of model_output contains all token embeddings
|
35 |
+
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
36 |
+
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
|
37 |
+
|
38 |
+
|
39 |
+
def normalize(comment, lowercase, remove_stopwords):
|
40 |
+
if lowercase:
|
41 |
+
comment = comment.lower()
|
42 |
+
comment = nlp(comment)
|
43 |
+
lemmatized = list()
|
44 |
+
for word in comment:
|
45 |
+
lemma = word.lemma_.strip()
|
46 |
+
if lemma:
|
47 |
+
if not remove_stopwords or (remove_stopwords and lemma not in stops):
|
48 |
+
lemmatized.append(lemma)
|
49 |
+
return " ".join(lemmatized)
|
50 |
+
|
51 |
+
|
52 |
+
# def tokenize_function(examples):
|
53 |
+
# return tokenizer(examples["text"])
|
54 |
+
|
55 |
+
|
56 |
+
def compute_metrics(eval_pred):
|
57 |
+
logits, labels = eval_pred
|
58 |
+
predictions = np.argmax(logits, axis=-1)
|
59 |
+
metric = evaluate.load("accuracy")
|
60 |
+
return metric.compute(predictions=predictions, references=labels)
|
61 |
+
|
62 |
+
|
63 |
+
def training():
|
64 |
+
dataset_id = "ag_news"
|
65 |
+
dataset = load_dataset(dataset_id)
|
66 |
+
# dataset = dataset["train"]
|
67 |
+
# tokenized_datasets = dataset.map(tokenize_function, batched=True)
|
68 |
+
|
69 |
+
print(f"- The {dataset_id} dataset has {dataset['train'].num_rows} examples.")
|
70 |
+
print(f"- Each example is a {type(dataset['train'][0])} with a {type(dataset['train'][0]['text'])} as value.")
|
71 |
+
print(f"- Examples look like this: {dataset['train'][0]}")
|
72 |
+
|
73 |
+
# small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
|
74 |
+
# small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
|
75 |
+
|
76 |
+
# dataset = dataset["train"].map(tokenize_function, batched=True)
|
77 |
+
# dataset.set_format(type="torch", columns=["input_ids", "token_type_ids", "attention_mask", "label"])
|
78 |
+
# dataset.format['type']
|
79 |
+
|
80 |
+
# print(dataset)
|
81 |
+
|
82 |
+
train_examples = []
|
83 |
+
train_data = dataset["train"]
|
84 |
+
# For agility we only 1/2 of our available data
|
85 |
+
n_examples = dataset["train"].num_rows // 2
|
86 |
+
# n_remaining = dataset["train"].num_rows - n_examples
|
87 |
+
# dataset_clean = {}
|
88 |
+
# # dataset_0 = []
|
89 |
+
# # dataset_1 = []
|
90 |
+
# # dataset_2 = []
|
91 |
+
# # dataset_3 = []
|
92 |
+
# for i in range(n_examples):
|
93 |
+
# dataset_clean[i] = {}
|
94 |
+
# dataset_clean[i]["text"] = normalize(train_data[i]["text"], lowercase=True, remove_stopwords=True)
|
95 |
+
# dataset_clean[i]["label"] = train_data[i]["label"]
|
96 |
+
# if train_data[i]["label"] == 0:
|
97 |
+
# dataset_0.append(dataset_clean[i])
|
98 |
+
# elif train_data[i]["label"] == 1:
|
99 |
+
# dataset_1.append(dataset_clean[i])
|
100 |
+
# elif train_data[i]["label"] == 2:
|
101 |
+
# dataset_2.append(dataset_clean[i])
|
102 |
+
# elif train_data[i]["label"] == 3:
|
103 |
+
# dataset_3.append(dataset_clean[i])
|
104 |
+
# n_0 = len(dataset_0) // 2
|
105 |
+
# n_1 = len(dataset_1) // 2
|
106 |
+
# n_2 = len(dataset_2) // 2
|
107 |
+
# n_3 = len(dataset_3) // 2
|
108 |
+
# print("Label lengths:", len(dataset_0), len(dataset_1), len(dataset_2), len(dataset_3))
|
109 |
+
|
110 |
+
for i in range(n_examples):
|
111 |
+
example = train_data[i]
|
112 |
+
# example_opposite = dataset_clean[-(i)]
|
113 |
+
# print(example["text"])
|
114 |
+
train_examples.append(InputExample(texts=[example['text']], label=example['label']))
|
115 |
+
|
116 |
+
# for i in range(n_0):
|
117 |
+
# example = dataset_0[i]
|
118 |
+
# # example_opposite = dataset_0[-(i)]
|
119 |
+
# # print(example["text"])
|
120 |
+
# train_examples.append(InputExample(texts=[example['text']], label=0))
|
121 |
+
|
122 |
+
# for i in range(n_1):
|
123 |
+
# example = dataset_1[i]
|
124 |
+
# # example_opposite = dataset_1[-(i)]
|
125 |
+
# # print(example["text"])
|
126 |
+
# train_examples.append(InputExample(texts=[example['text']], label=1))
|
127 |
+
|
128 |
+
# for i in range(n_2):
|
129 |
+
# example = dataset_2[i]
|
130 |
+
# # example_opposite = dataset_2[-(i)]
|
131 |
+
# # print(example["text"])
|
132 |
+
# train_examples.append(InputExample(texts=[example['text']], label=2))
|
133 |
+
|
134 |
+
# for i in range(n_3):
|
135 |
+
# example = dataset_3[i]
|
136 |
+
# # example_opposite = dataset_3[-(i)]
|
137 |
+
# # print(example["text"])
|
138 |
+
# train_examples.append(InputExample(texts=[example['text']], label=3))
|
139 |
+
|
140 |
+
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=25)
|
141 |
+
|
142 |
+
print("END DATALOADER")
|
143 |
+
|
144 |
+
# print(train_examples)
|
145 |
+
|
146 |
+
embeddings = finetune(train_dataloader)
|
147 |
+
|
148 |
+
return (dataset['train'].num_rows, type(dataset['train'][0]), type(dataset['train'][0]['text']), dataset['train'][0], embeddings)
|
149 |
+
|
150 |
+
|
151 |
+
def finetune(train_dataloader):
|
152 |
+
# model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
|
153 |
+
model_id = "sentence-transformers/all-MiniLM-L6-v2"
|
154 |
+
model = SentenceTransformer(model_id)
|
155 |
+
|
156 |
+
# training_args = TrainingArguments(output_dir="test_trainer")
|
157 |
+
|
158 |
+
# USE THIS LINK
|
159 |
+
# https://huggingface.co/blog/how-to-train-sentence-transformers
|
160 |
+
|
161 |
+
train_loss = losses.BatchHardSoftMarginTripletLoss(model=model)
|
162 |
+
|
163 |
+
print("BEGIN FIT")
|
164 |
+
|
165 |
+
model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=10)
|
166 |
+
|
167 |
+
model.save("ag_news_model")
|
168 |
+
|
169 |
+
model.save_to_hub("smhavens/all-MiniLM-agNews")
|
170 |
+
# accuracy = compute_metrics(eval, metric)
|
171 |
+
|
172 |
+
# training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")
|
173 |
+
|
174 |
+
# trainer = Trainer(
|
175 |
+
# model=model,
|
176 |
+
# args=training_args,
|
177 |
+
# train_dataset=train,
|
178 |
+
# eval_dataset=eval,
|
179 |
+
# compute_metrics=compute_metrics,
|
180 |
+
# )
|
181 |
+
|
182 |
+
# trainer.train()
|
183 |
+
|
184 |
+
def embeddings():
|
185 |
+
model = SentenceTransformer("ag_news_model")
|
186 |
+
device = torch.device('cuda:0')
|
187 |
+
model = model.to(device)
|
188 |
+
sentences = ["This is an example sentence", "Each sentence is converted"]
|
189 |
+
|
190 |
+
# model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
191 |
+
embeddings = model.encode(sentences)
|
192 |
+
# print(embeddings)
|
193 |
+
|
194 |
+
# Sentences we want sentence embeddings for
|
195 |
+
sentences = ['This is an example sentence', 'Each sentence is converted']
|
196 |
+
|
197 |
+
# Load model from HuggingFace Hub
|
198 |
+
tokenizer = AutoTokenizer.from_pretrained('ag_news_model')
|
199 |
+
# model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
200 |
+
|
201 |
+
# Tokenize sentences
|
202 |
+
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
|
203 |
+
|
204 |
+
# print(model.device)
|
205 |
+
# print(encoded_input["input_ids"].device)
|
206 |
+
# print(encoded_input["attention_mask"].device)
|
207 |
+
# print(encoded_input["token_type_ids"].device)
|
208 |
+
encoded_input["input_ids"] = encoded_input["input_ids"].to(device)
|
209 |
+
encoded_input["attention_mask"] = encoded_input["attention_mask"].to(device)
|
210 |
+
encoded_input['token_type_ids'] = encoded_input['token_type_ids'].to(device)
|
211 |
+
# print(encoded_input)
|
212 |
+
|
213 |
+
# print(encoded_input["input_ids"].device)
|
214 |
+
# print(encoded_input["attention_mask"].device)
|
215 |
+
# print(encoded_input["token_type_ids"].device)
|
216 |
+
|
217 |
+
encoded_input['input'] = {'input_ids':encoded_input['input_ids'], 'attention_mask':encoded_input['attention_mask']}
|
218 |
+
|
219 |
+
# + encoded_input['token_type_ids'] + encoded_input['attention_mask']
|
220 |
+
del encoded_input['input_ids']
|
221 |
+
del encoded_input['token_type_ids']
|
222 |
+
del encoded_input['attention_mask']
|
223 |
+
|
224 |
+
# print(encoded_input)
|
225 |
+
|
226 |
+
# encoded_input.to(device)
|
227 |
+
# Compute token embeddings
|
228 |
+
with torch.no_grad():
|
229 |
+
model_output = model(**encoded_input)
|
230 |
+
|
231 |
+
print(model_output)
|
232 |
+
# Perform pooling
|
233 |
+
sentence_embeddings = mean_pooling(model_output, encoded_input['input']["attention_mask"])
|
234 |
+
|
235 |
+
# Normalize embeddings
|
236 |
+
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
|
237 |
+
|
238 |
+
print("Sentence embeddings:")
|
239 |
+
print(sentence_embeddings)
|
240 |
+
return sentence_embeddings
|
241 |
+
|
242 |
+
|
243 |
+
|
244 |
+
def greet(name):
|
245 |
+
return "Hello " + name + "!!"
|
246 |
+
|
247 |
+
def check_answer(guess:str):
|
248 |
+
global guesses
|
249 |
+
global answer
|
250 |
+
guesses.append(guess)
|
251 |
+
output = ""
|
252 |
+
for guess in guesses:
|
253 |
+
output += ("- " + guess + "\n")
|
254 |
+
output = output[:-1]
|
255 |
+
|
256 |
+
if guess.lower() == answer.lower():
|
257 |
+
return "Correct!", output
|
258 |
+
else:
|
259 |
+
return "Try again!", output
|
260 |
+
|
261 |
+
def main():
|
262 |
+
word1 = "Black"
|
263 |
+
word2 = "White"
|
264 |
+
word3 = "Sun"
|
265 |
+
global answer
|
266 |
+
answer = "Moon"
|
267 |
+
global guesses
|
268 |
+
|
269 |
+
# num_rows, data_type, value, example, embeddings = training()
|
270 |
+
sent_embeddings = embeddings()
|
271 |
+
|
272 |
+
prompt = f"{word1} is to {word2} as {word3} is to ____"
|
273 |
+
with gr.Blocks() as iface:
|
274 |
+
gr.Markdown(prompt)
|
275 |
+
with gr.Tab("Guess"):
|
276 |
+
text_input = gr.Textbox()
|
277 |
+
text_output = gr.Textbox()
|
278 |
+
text_button = gr.Button("Submit")
|
279 |
+
with gr.Accordion("Open for previous guesses"):
|
280 |
+
text_guesses = gr.Textbox()
|
281 |
+
with gr.Tab("Testing"):
|
282 |
+
gr.Markdown(f"""The Embeddings are {sent_embeddings}.""")
|
283 |
+
text_button.click(check_answer, inputs=[text_input], outputs=[text_output, text_guesses])
|
284 |
+
# iface = gr.Interface(fn=greet, inputs="text", outputs="text")
|
285 |
+
iface.launch()
|
286 |
+
|
287 |
+
|
288 |
+
|
289 |
+
|
290 |
+
|
291 |
+
if __name__ == "__main__":
|
292 |
main()
|
requirements.txt
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
-
spacy
|
2 |
-
sentence_transformers
|
3 |
-
transformers
|
4 |
-
torch
|
5 |
-
evaluate
|
6 |
https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.2.0/en_core_web_sm-2.2.0.tar.gz
|
|
|
1 |
+
spacy
|
2 |
+
sentence_transformers
|
3 |
+
transformers
|
4 |
+
torch
|
5 |
+
evaluate
|
6 |
https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.2.0/en_core_web_sm-2.2.0.tar.gz
|
train.py
CHANGED
@@ -1,277 +1,277 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import math
|
3 |
-
import spacy
|
4 |
-
from datasets import load_dataset
|
5 |
-
from sentence_transformers import SentenceTransformer
|
6 |
-
from sentence_transformers import InputExample
|
7 |
-
from sentence_transformers import losses
|
8 |
-
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
|
9 |
-
from transformers import TrainingArguments, Trainer
|
10 |
-
import torch
|
11 |
-
import torch.nn.functional as F
|
12 |
-
from torch.utils.data import DataLoader
|
13 |
-
import numpy as np
|
14 |
-
import evaluate
|
15 |
-
import nltk
|
16 |
-
from nltk.corpus import stopwords
|
17 |
-
import subprocess
|
18 |
-
import sys
|
19 |
-
from transformers import DataCollatorWithPadding
|
20 |
-
|
21 |
-
|
22 |
-
# !pip install https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl
|
23 |
-
# subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl'])
|
24 |
-
# tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
25 |
-
# data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
|
26 |
-
# nltk.download('stopwords')
|
27 |
-
# nlp = spacy.load("en_core_web_sm")
|
28 |
-
# stops = stopwords.words("english")
|
29 |
-
|
30 |
-
# answer = "Pizza"
|
31 |
-
guesses = []
|
32 |
-
answer = "Pizza"
|
33 |
-
|
34 |
-
|
35 |
-
#Mean Pooling - Take attention mask into account for correct averaging
|
36 |
-
def mean_pooling(model_output, attention_mask):
|
37 |
-
token_embeddings = model_output[0] #First element of model_output contains all token embeddings
|
38 |
-
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
39 |
-
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
|
40 |
-
|
41 |
-
|
42 |
-
# def normalize(comment, lowercase, remove_stopwords):
|
43 |
-
# if lowercase:
|
44 |
-
# comment = comment.lower()
|
45 |
-
# comment = nlp(comment)
|
46 |
-
# lemmatized = list()
|
47 |
-
# for word in comment:
|
48 |
-
# lemma = word.lemma_.strip()
|
49 |
-
# if lemma:
|
50 |
-
# if not remove_stopwords or (remove_stopwords and lemma not in stops):
|
51 |
-
# lemmatized.append(lemma)
|
52 |
-
# return " ".join(lemmatized)
|
53 |
-
|
54 |
-
|
55 |
-
# def tokenize_function(examples):
|
56 |
-
# return tokenizer(examples["text"], truncation=True)
|
57 |
-
|
58 |
-
|
59 |
-
def compute_metrics(eval_pred):
|
60 |
-
logits, labels = eval_pred
|
61 |
-
predictions = np.argmax(logits, axis=-1)
|
62 |
-
metric = evaluate.load("accuracy")
|
63 |
-
return metric.compute(predictions=predictions, references=labels)
|
64 |
-
|
65 |
-
|
66 |
-
def training():
|
67 |
-
dataset_id = "ag_news"
|
68 |
-
|
69 |
-
print("GETTING DATASET")
|
70 |
-
dataset = load_dataset(dataset_id)
|
71 |
-
# dataset = dataset["train"]
|
72 |
-
# tokenized_datasets = dataset.map(tokenize_function, batched=True)
|
73 |
-
|
74 |
-
print(f"- The {dataset_id} dataset has {dataset['train'].num_rows} examples.")
|
75 |
-
print(f"- Each example is a {type(dataset['train'][0])} with a {type(dataset['train'][0]['text'])} as value.")
|
76 |
-
print(f"- Examples look like this: {dataset['train'][0]}")
|
77 |
-
|
78 |
-
# small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
|
79 |
-
# small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
|
80 |
-
|
81 |
-
# dataset = dataset["train"].map(tokenize_function, batched=True)
|
82 |
-
# dataset.set_format(type="torch", columns=["input_ids", "token_type_ids", "attention_mask", "label"])
|
83 |
-
# dataset.format['type']
|
84 |
-
|
85 |
-
# tokenized_news = dataset.map(tokenize_function, batched=True)
|
86 |
-
|
87 |
-
# model = AutoModelForSequenceClassification.from_pretrained("sentence-transformers/all-MiniLM-L6-v2", num_labels=2)
|
88 |
-
|
89 |
-
# print(dataset)
|
90 |
-
|
91 |
-
train_examples = []
|
92 |
-
train_data = dataset["train"]
|
93 |
-
# For agility we only 1/2 of our available data
|
94 |
-
n_examples = dataset["train"].num_rows // 2
|
95 |
-
# n_remaining = dataset["train"].num_rows - n_examples
|
96 |
-
# dataset_clean = {}
|
97 |
-
# # dataset_0 = []
|
98 |
-
# # dataset_1 = []
|
99 |
-
# # dataset_2 = []
|
100 |
-
# # dataset_3 = []
|
101 |
-
# for i in range(n_examples):
|
102 |
-
# dataset_clean[i] = {}
|
103 |
-
# dataset_clean[i]["text"] = normalize(train_data[i]["text"], lowercase=True, remove_stopwords=True)
|
104 |
-
# dataset_clean[i]["label"] = train_data[i]["label"]
|
105 |
-
# if train_data[i]["label"] == 0:
|
106 |
-
# dataset_0.append(dataset_clean[i])
|
107 |
-
# elif train_data[i]["label"] == 1:
|
108 |
-
# dataset_1.append(dataset_clean[i])
|
109 |
-
# elif train_data[i]["label"] == 2:
|
110 |
-
# dataset_2.append(dataset_clean[i])
|
111 |
-
# elif train_data[i]["label"] == 3:
|
112 |
-
# dataset_3.append(dataset_clean[i])
|
113 |
-
# n_0 = len(dataset_0) // 2
|
114 |
-
# n_1 = len(dataset_1) // 2
|
115 |
-
# n_2 = len(dataset_2) // 2
|
116 |
-
# n_3 = len(dataset_3) // 2
|
117 |
-
# print("Label lengths:", len(dataset_0), len(dataset_1), len(dataset_2), len(dataset_3))
|
118 |
-
|
119 |
-
for i in range(n_examples):
|
120 |
-
example = train_data[i]
|
121 |
-
# example_opposite = dataset_clean[-(i)]
|
122 |
-
# print(example["text"])
|
123 |
-
train_examples.append(InputExample(texts=[example['text']], label=example['label']))
|
124 |
-
|
125 |
-
# for i in range(n_0):
|
126 |
-
# example = dataset_0[i]
|
127 |
-
# # example_opposite = dataset_0[-(i)]
|
128 |
-
# # print(example["text"])
|
129 |
-
# train_examples.append(InputExample(texts=[example['text']], label=0))
|
130 |
-
|
131 |
-
# for i in range(n_1):
|
132 |
-
# example = dataset_1[i]
|
133 |
-
# # example_opposite = dataset_1[-(i)]
|
134 |
-
# # print(example["text"])
|
135 |
-
# train_examples.append(InputExample(texts=[example['text']], label=1))
|
136 |
-
|
137 |
-
# for i in range(n_2):
|
138 |
-
# example = dataset_2[i]
|
139 |
-
# # example_opposite = dataset_2[-(i)]
|
140 |
-
# # print(example["text"])
|
141 |
-
# train_examples.append(InputExample(texts=[example['text']], label=2))
|
142 |
-
|
143 |
-
# for i in range(n_3):
|
144 |
-
# example = dataset_3[i]
|
145 |
-
# # example_opposite = dataset_3[-(i)]
|
146 |
-
# # print(example["text"])
|
147 |
-
# train_examples.append(InputExample(texts=[example['text']], label=3))
|
148 |
-
|
149 |
-
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=25)
|
150 |
-
|
151 |
-
print("END DATALOADER")
|
152 |
-
|
153 |
-
# print(train_examples)
|
154 |
-
|
155 |
-
embeddings = finetune(train_dataloader)
|
156 |
-
|
157 |
-
return (dataset['train'].num_rows, type(dataset['train'][0]), type(dataset['train'][0]['text']), dataset['train'][0], embeddings)
|
158 |
-
|
159 |
-
|
160 |
-
def finetune(train_dataloader):
|
161 |
-
# model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
|
162 |
-
model_id = "sentence-transformers/all-MiniLM-L6-v2"
|
163 |
-
model = SentenceTransformer(model_id)
|
164 |
-
device = torch.device('cuda:0')
|
165 |
-
model = model.to(device)
|
166 |
-
|
167 |
-
# training_args = TrainingArguments(output_dir="test_trainer")
|
168 |
-
|
169 |
-
# USE THIS LINK
|
170 |
-
# https://huggingface.co/blog/how-to-train-sentence-transformers
|
171 |
-
|
172 |
-
train_loss = losses.BatchHardSoftMarginTripletLoss(model=model)
|
173 |
-
|
174 |
-
print("BEGIN FIT")
|
175 |
-
|
176 |
-
model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=10)
|
177 |
-
|
178 |
-
model.save("ag_news_model")
|
179 |
-
|
180 |
-
model.save_to_hub("smhavens/all-MiniLM-agNews")
|
181 |
-
# accuracy = compute_metrics(eval, metric)
|
182 |
-
|
183 |
-
# training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")
|
184 |
-
|
185 |
-
# trainer = Trainer(
|
186 |
-
# model=model,
|
187 |
-
# args=training_args,
|
188 |
-
# train_dataset=train,
|
189 |
-
# eval_dataset=eval,
|
190 |
-
# compute_metrics=compute_metrics,
|
191 |
-
# )
|
192 |
-
|
193 |
-
# trainer.train()
|
194 |
-
|
195 |
-
# sentences = ["This is an example sentence", "Each sentence is converted"]
|
196 |
-
|
197 |
-
# # model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
198 |
-
# embeddings = model.encode(sentences)
|
199 |
-
# print(embeddings)
|
200 |
-
|
201 |
-
# # Sentences we want sentence embeddings for
|
202 |
-
# sentences = ['This is an example sentence', 'Each sentence is converted']
|
203 |
-
|
204 |
-
# # Load model from HuggingFace Hub
|
205 |
-
# # tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
206 |
-
# # model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
207 |
-
|
208 |
-
# # Tokenize sentences
|
209 |
-
# encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
|
210 |
-
|
211 |
-
# # Compute token embeddings
|
212 |
-
# with torch.no_grad():
|
213 |
-
# model_output = model(**encoded_input)
|
214 |
-
|
215 |
-
# # Perform pooling
|
216 |
-
# sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
|
217 |
-
|
218 |
-
# # Normalize embeddings
|
219 |
-
# sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
|
220 |
-
|
221 |
-
# print("Sentence embeddings:")
|
222 |
-
# print(sentence_embeddings)
|
223 |
-
return 0
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
def greet(name):
|
228 |
-
return "Hello " + name + "!!"
|
229 |
-
|
230 |
-
def check_answer(guess:str):
|
231 |
-
global guesses
|
232 |
-
global answer
|
233 |
-
guesses.append(guess)
|
234 |
-
output = ""
|
235 |
-
for guess in guesses:
|
236 |
-
output += ("- " + guess + "\n")
|
237 |
-
output = output[:-1]
|
238 |
-
|
239 |
-
if guess.lower() == answer.lower():
|
240 |
-
return "Correct!", output
|
241 |
-
else:
|
242 |
-
return "Try again!", output
|
243 |
-
|
244 |
-
def main():
|
245 |
-
print("BEGIN")
|
246 |
-
word1 = "Black"
|
247 |
-
word2 = "White"
|
248 |
-
word3 = "Sun"
|
249 |
-
global answer
|
250 |
-
answer = "Moon"
|
251 |
-
global guesses
|
252 |
-
|
253 |
-
num_rows, data_type, value, example, embeddings = training()
|
254 |
-
|
255 |
-
# prompt = f"{word1} is to {word2} as {word3} is to ____"
|
256 |
-
# with gr.Blocks() as iface:
|
257 |
-
# gr.Markdown(prompt)
|
258 |
-
# with gr.Tab("Guess"):
|
259 |
-
# text_input = gr.Textbox()
|
260 |
-
# text_output = gr.Textbox()
|
261 |
-
# text_button = gr.Button("Submit")
|
262 |
-
# with gr.Accordion("Open for previous guesses"):
|
263 |
-
# text_guesses = gr.Textbox()
|
264 |
-
# with gr.Tab("Testing"):
|
265 |
-
# gr.Markdown(f"""Number of rows in dataset is {num_rows}, with each having type {data_type} and value {value}.
|
266 |
-
# An example is {example}.
|
267 |
-
# The Embeddings are {embeddings}.""")
|
268 |
-
# text_button.click(check_answer, inputs=[text_input], outputs=[text_output, text_guesses])
|
269 |
-
# # iface = gr.Interface(fn=greet, inputs="text", outputs="text")
|
270 |
-
# iface.launch()
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
if __name__ == "__main__":
|
277 |
main()
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import math
|
3 |
+
import spacy
|
4 |
+
from datasets import load_dataset
|
5 |
+
from sentence_transformers import SentenceTransformer
|
6 |
+
from sentence_transformers import InputExample
|
7 |
+
from sentence_transformers import losses
|
8 |
+
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
|
9 |
+
from transformers import TrainingArguments, Trainer
|
10 |
+
import torch
|
11 |
+
import torch.nn.functional as F
|
12 |
+
from torch.utils.data import DataLoader
|
13 |
+
import numpy as np
|
14 |
+
import evaluate
|
15 |
+
import nltk
|
16 |
+
from nltk.corpus import stopwords
|
17 |
+
import subprocess
|
18 |
+
import sys
|
19 |
+
from transformers import DataCollatorWithPadding
|
20 |
+
|
21 |
+
|
22 |
+
# !pip install https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl
|
23 |
+
# subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'https://huggingface.co/spacy/en_core_web_sm/resolve/main/en_core_web_sm-any-py3-none-any.whl'])
|
24 |
+
# tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
25 |
+
# data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
|
26 |
+
# nltk.download('stopwords')
|
27 |
+
# nlp = spacy.load("en_core_web_sm")
|
28 |
+
# stops = stopwords.words("english")
|
29 |
+
|
30 |
+
# answer = "Pizza"
|
31 |
+
guesses = []
|
32 |
+
answer = "Pizza"
|
33 |
+
|
34 |
+
|
35 |
+
#Mean Pooling - Take attention mask into account for correct averaging
|
36 |
+
def mean_pooling(model_output, attention_mask):
|
37 |
+
token_embeddings = model_output[0] #First element of model_output contains all token embeddings
|
38 |
+
input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
39 |
+
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
|
40 |
+
|
41 |
+
|
42 |
+
# def normalize(comment, lowercase, remove_stopwords):
|
43 |
+
# if lowercase:
|
44 |
+
# comment = comment.lower()
|
45 |
+
# comment = nlp(comment)
|
46 |
+
# lemmatized = list()
|
47 |
+
# for word in comment:
|
48 |
+
# lemma = word.lemma_.strip()
|
49 |
+
# if lemma:
|
50 |
+
# if not remove_stopwords or (remove_stopwords and lemma not in stops):
|
51 |
+
# lemmatized.append(lemma)
|
52 |
+
# return " ".join(lemmatized)
|
53 |
+
|
54 |
+
|
55 |
+
# def tokenize_function(examples):
|
56 |
+
# return tokenizer(examples["text"], truncation=True)
|
57 |
+
|
58 |
+
|
59 |
+
def compute_metrics(eval_pred):
|
60 |
+
logits, labels = eval_pred
|
61 |
+
predictions = np.argmax(logits, axis=-1)
|
62 |
+
metric = evaluate.load("accuracy")
|
63 |
+
return metric.compute(predictions=predictions, references=labels)
|
64 |
+
|
65 |
+
|
66 |
+
def training():
|
67 |
+
dataset_id = "ag_news"
|
68 |
+
|
69 |
+
print("GETTING DATASET")
|
70 |
+
dataset = load_dataset(dataset_id)
|
71 |
+
# dataset = dataset["train"]
|
72 |
+
# tokenized_datasets = dataset.map(tokenize_function, batched=True)
|
73 |
+
|
74 |
+
print(f"- The {dataset_id} dataset has {dataset['train'].num_rows} examples.")
|
75 |
+
print(f"- Each example is a {type(dataset['train'][0])} with a {type(dataset['train'][0]['text'])} as value.")
|
76 |
+
print(f"- Examples look like this: {dataset['train'][0]}")
|
77 |
+
|
78 |
+
# small_train_dataset = tokenized_datasets["train"].shuffle(seed=42).select(range(1000))
|
79 |
+
# small_eval_dataset = tokenized_datasets["test"].shuffle(seed=42).select(range(1000))
|
80 |
+
|
81 |
+
# dataset = dataset["train"].map(tokenize_function, batched=True)
|
82 |
+
# dataset.set_format(type="torch", columns=["input_ids", "token_type_ids", "attention_mask", "label"])
|
83 |
+
# dataset.format['type']
|
84 |
+
|
85 |
+
# tokenized_news = dataset.map(tokenize_function, batched=True)
|
86 |
+
|
87 |
+
# model = AutoModelForSequenceClassification.from_pretrained("sentence-transformers/all-MiniLM-L6-v2", num_labels=2)
|
88 |
+
|
89 |
+
# print(dataset)
|
90 |
+
|
91 |
+
train_examples = []
|
92 |
+
train_data = dataset["train"]
|
93 |
+
# For agility we only 1/2 of our available data
|
94 |
+
n_examples = dataset["train"].num_rows // 2
|
95 |
+
# n_remaining = dataset["train"].num_rows - n_examples
|
96 |
+
# dataset_clean = {}
|
97 |
+
# # dataset_0 = []
|
98 |
+
# # dataset_1 = []
|
99 |
+
# # dataset_2 = []
|
100 |
+
# # dataset_3 = []
|
101 |
+
# for i in range(n_examples):
|
102 |
+
# dataset_clean[i] = {}
|
103 |
+
# dataset_clean[i]["text"] = normalize(train_data[i]["text"], lowercase=True, remove_stopwords=True)
|
104 |
+
# dataset_clean[i]["label"] = train_data[i]["label"]
|
105 |
+
# if train_data[i]["label"] == 0:
|
106 |
+
# dataset_0.append(dataset_clean[i])
|
107 |
+
# elif train_data[i]["label"] == 1:
|
108 |
+
# dataset_1.append(dataset_clean[i])
|
109 |
+
# elif train_data[i]["label"] == 2:
|
110 |
+
# dataset_2.append(dataset_clean[i])
|
111 |
+
# elif train_data[i]["label"] == 3:
|
112 |
+
# dataset_3.append(dataset_clean[i])
|
113 |
+
# n_0 = len(dataset_0) // 2
|
114 |
+
# n_1 = len(dataset_1) // 2
|
115 |
+
# n_2 = len(dataset_2) // 2
|
116 |
+
# n_3 = len(dataset_3) // 2
|
117 |
+
# print("Label lengths:", len(dataset_0), len(dataset_1), len(dataset_2), len(dataset_3))
|
118 |
+
|
119 |
+
for i in range(n_examples):
|
120 |
+
example = train_data[i]
|
121 |
+
# example_opposite = dataset_clean[-(i)]
|
122 |
+
# print(example["text"])
|
123 |
+
train_examples.append(InputExample(texts=[example['text']], label=example['label']))
|
124 |
+
|
125 |
+
# for i in range(n_0):
|
126 |
+
# example = dataset_0[i]
|
127 |
+
# # example_opposite = dataset_0[-(i)]
|
128 |
+
# # print(example["text"])
|
129 |
+
# train_examples.append(InputExample(texts=[example['text']], label=0))
|
130 |
+
|
131 |
+
# for i in range(n_1):
|
132 |
+
# example = dataset_1[i]
|
133 |
+
# # example_opposite = dataset_1[-(i)]
|
134 |
+
# # print(example["text"])
|
135 |
+
# train_examples.append(InputExample(texts=[example['text']], label=1))
|
136 |
+
|
137 |
+
# for i in range(n_2):
|
138 |
+
# example = dataset_2[i]
|
139 |
+
# # example_opposite = dataset_2[-(i)]
|
140 |
+
# # print(example["text"])
|
141 |
+
# train_examples.append(InputExample(texts=[example['text']], label=2))
|
142 |
+
|
143 |
+
# for i in range(n_3):
|
144 |
+
# example = dataset_3[i]
|
145 |
+
# # example_opposite = dataset_3[-(i)]
|
146 |
+
# # print(example["text"])
|
147 |
+
# train_examples.append(InputExample(texts=[example['text']], label=3))
|
148 |
+
|
149 |
+
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=25)
|
150 |
+
|
151 |
+
print("END DATALOADER")
|
152 |
+
|
153 |
+
# print(train_examples)
|
154 |
+
|
155 |
+
embeddings = finetune(train_dataloader)
|
156 |
+
|
157 |
+
return (dataset['train'].num_rows, type(dataset['train'][0]), type(dataset['train'][0]['text']), dataset['train'][0], embeddings)
|
158 |
+
|
159 |
+
|
160 |
+
def finetune(train_dataloader):
|
161 |
+
# model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=5)
|
162 |
+
model_id = "sentence-transformers/all-MiniLM-L6-v2"
|
163 |
+
model = SentenceTransformer(model_id)
|
164 |
+
device = torch.device('cuda:0')
|
165 |
+
model = model.to(device)
|
166 |
+
|
167 |
+
# training_args = TrainingArguments(output_dir="test_trainer")
|
168 |
+
|
169 |
+
# USE THIS LINK
|
170 |
+
# https://huggingface.co/blog/how-to-train-sentence-transformers
|
171 |
+
|
172 |
+
train_loss = losses.BatchHardSoftMarginTripletLoss(model=model)
|
173 |
+
|
174 |
+
print("BEGIN FIT")
|
175 |
+
|
176 |
+
model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=10)
|
177 |
+
|
178 |
+
model.save("ag_news_model")
|
179 |
+
|
180 |
+
model.save_to_hub("smhavens/all-MiniLM-agNews")
|
181 |
+
# accuracy = compute_metrics(eval, metric)
|
182 |
+
|
183 |
+
# training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")
|
184 |
+
|
185 |
+
# trainer = Trainer(
|
186 |
+
# model=model,
|
187 |
+
# args=training_args,
|
188 |
+
# train_dataset=train,
|
189 |
+
# eval_dataset=eval,
|
190 |
+
# compute_metrics=compute_metrics,
|
191 |
+
# )
|
192 |
+
|
193 |
+
# trainer.train()
|
194 |
+
|
195 |
+
# sentences = ["This is an example sentence", "Each sentence is converted"]
|
196 |
+
|
197 |
+
# # model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
198 |
+
# embeddings = model.encode(sentences)
|
199 |
+
# print(embeddings)
|
200 |
+
|
201 |
+
# # Sentences we want sentence embeddings for
|
202 |
+
# sentences = ['This is an example sentence', 'Each sentence is converted']
|
203 |
+
|
204 |
+
# # Load model from HuggingFace Hub
|
205 |
+
# # tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
206 |
+
# # model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
|
207 |
+
|
208 |
+
# # Tokenize sentences
|
209 |
+
# encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
|
210 |
+
|
211 |
+
# # Compute token embeddings
|
212 |
+
# with torch.no_grad():
|
213 |
+
# model_output = model(**encoded_input)
|
214 |
+
|
215 |
+
# # Perform pooling
|
216 |
+
# sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
|
217 |
+
|
218 |
+
# # Normalize embeddings
|
219 |
+
# sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
|
220 |
+
|
221 |
+
# print("Sentence embeddings:")
|
222 |
+
# print(sentence_embeddings)
|
223 |
+
return 0
|
224 |
+
|
225 |
+
|
226 |
+
|
227 |
+
def greet(name):
|
228 |
+
return "Hello " + name + "!!"
|
229 |
+
|
230 |
+
def check_answer(guess:str):
|
231 |
+
global guesses
|
232 |
+
global answer
|
233 |
+
guesses.append(guess)
|
234 |
+
output = ""
|
235 |
+
for guess in guesses:
|
236 |
+
output += ("- " + guess + "\n")
|
237 |
+
output = output[:-1]
|
238 |
+
|
239 |
+
if guess.lower() == answer.lower():
|
240 |
+
return "Correct!", output
|
241 |
+
else:
|
242 |
+
return "Try again!", output
|
243 |
+
|
244 |
+
def main():
|
245 |
+
print("BEGIN")
|
246 |
+
word1 = "Black"
|
247 |
+
word2 = "White"
|
248 |
+
word3 = "Sun"
|
249 |
+
global answer
|
250 |
+
answer = "Moon"
|
251 |
+
global guesses
|
252 |
+
|
253 |
+
num_rows, data_type, value, example, embeddings = training()
|
254 |
+
|
255 |
+
# prompt = f"{word1} is to {word2} as {word3} is to ____"
|
256 |
+
# with gr.Blocks() as iface:
|
257 |
+
# gr.Markdown(prompt)
|
258 |
+
# with gr.Tab("Guess"):
|
259 |
+
# text_input = gr.Textbox()
|
260 |
+
# text_output = gr.Textbox()
|
261 |
+
# text_button = gr.Button("Submit")
|
262 |
+
# with gr.Accordion("Open for previous guesses"):
|
263 |
+
# text_guesses = gr.Textbox()
|
264 |
+
# with gr.Tab("Testing"):
|
265 |
+
# gr.Markdown(f"""Number of rows in dataset is {num_rows}, with each having type {data_type} and value {value}.
|
266 |
+
# An example is {example}.
|
267 |
+
# The Embeddings are {embeddings}.""")
|
268 |
+
# text_button.click(check_answer, inputs=[text_input], outputs=[text_output, text_guesses])
|
269 |
+
# # iface = gr.Interface(fn=greet, inputs="text", outputs="text")
|
270 |
+
# iface.launch()
|
271 |
+
|
272 |
+
|
273 |
+
|
274 |
+
|
275 |
+
|
276 |
+
if __name__ == "__main__":
|
277 |
main()
|