如何在我的 GPU 而不是 Spyder 或 Jupyter Notebook 中的 CPU 上运行 BERT/FinBert？

Question

我无法在 GPU 而不是 CPU 上运行我的代码。

我正在尝试运行 FinBert 代码，该代码将文本的每个句子分类为正面、负面或中性。但是，这段代码需要很长时间才能在我的 CPU 上运行。因此，我试图在我的 GPU 上运行它，但到目前为止我还不知道该怎么做。

这是我要在 GPU 上执行的代码：

# first I am importing the necessary packages 
import pandas as pd
from transformers import BertTokenizer, BertForSequenceClassification
from transformers import pipeline 
import pandas as pd
import nltk

# then I load the pre-trained FinBert Model
finbert = BertForSequenceClassification.from_pretrained('yiyanghkust/finbert- 
tone',num_labels=3)
tokenizer = BertTokenizer.from_pretrained('yiyanghkust/finbert-tone')
# build model and pipeline
nlp = pipeline("sentiment-analysis", model=finbert, tokenizer=tokenizer)

然后我运行以下 for 循环：

for i in range(0,len(data)):
print(i)
# access the text i from the data set
temp = data.iloc[i,1]
# tokenize the text to get the sentences of the text 
sentences = nltk.sent_tokenize(temp)
# apply FinBert-Model onto the sentences and save results in the variable "results"
results = nlp(sentences)
# get the filename of call i from clean_data 

# set j to 0 for the next for-loop
j = 0
# reset positive, neutral, and negative after each iteration over the second for-loop back to 0 to not inflate the count 
positive = 0
neutral = 0 
negative = 0 
# Create for loop to check for each sentence within the text whether FinBert classifies this sentence 
# as positive, neutral or negative 
for j in range (0,len(results)):
    label = results[j]["label"]
    if label == "Positive":
        positive = positive + 1
    elif label == "Neutral": 
        neutral = neutral + 1 
    else:
        negative = negative + 1  

# Calculate the sentiment scores
percentage_positive = positive / len(results)
percentage_negative = negative / len(results)
net_score = percentage_positive - percentage_negative 

# save the results in a DataFrame previously created 
results.iloc[i,1] = percentage_positive 
results.iloc[i,2] = percentage_negative 
results.iloc[i,3] = net_score 

我有 NVIDIA GeForce MX130。

如果我运行以下代码，它说有可用的 GPU：

import tensorflow as tf
print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

output: Num GPUs Available:  1

这段代码打印出来：

tf.config.list_logical_devices('GPU')

output: [LogicalDevice(name='/device:GPU:0', device_type='GPU')]

我尝试以多种方式在 GPU 上运行代码，包括这个：

tf.debugging.set_log_device_placement(True)
# Place ops on the GPU
with tf.device('/GPU:0'):
   -> my code here 

并且运行以下代码也显示我的 GPU 可用：

import torch
use_cuda = torch.cuda.is_available()
print(use_cuda)

output: True


if use_cuda:
print('__CUDNN VERSION:', torch.backends.cudnn.version())
print('__Number CUDA Devices:', torch.cuda.device_count())
print('__CUDA Device Name:',torch.cuda.get_device_name(0))
print('__CUDA Device Total Memory 
[GB]:',torch.cuda.get_device_properties(0).total_memory/1e9)

output: 
[GB]:',torch.cuda.get_device_properties(0).total_memory/1e9)
__CUDNN VERSION: 8302
__Number CUDA Devices: 1
__CUDA Device Name: GeForce MX130
__CUDA Device Total Memory [GB]: 4.294967296

但我仍然无法在 GPU 上运行代码。 我错过了什么？

最好的， 凯尔

Answer 1

在PyTorch中，用户必须自己手动将模型、数据等移动到cuda。由于您的模型是 PyTorch，因此您需要将模型和数据都移动到

"cuda"

。别担心，你正朝着正确的方向前进：

finbert = BertForSequenceClassification.from_pretrained('yiyanghkust/finbert- tone',num_labels=3) 
finbert.to("cuda")
...

nlp = pipeline("sentiment-analysis", model=finbert, tokenizer=tokenizer, device=0)

如果它仍然引发

RuntimeError

但具有相同的消息，您也可以添加带有

to.("cuda")

的行。