Upload fin_readability_sustainability.py
Browse files
fin_readability_sustainability.py
ADDED
|
@@ -0,0 +1,110 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import transformers
|
| 3 |
+
from torch.utils.data import Dataset, DataLoader
|
| 4 |
+
from transformers import RobertaModel, RobertaTokenizer, BertModel, BertTokenizer
|
| 5 |
+
import pandas as pd
|
| 6 |
+
|
| 7 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 8 |
+
|
| 9 |
+
MAX_LEN = 128
|
| 10 |
+
BATCH_SIZE = 20
|
| 11 |
+
text_col_name = 'sentence'
|
| 12 |
+
|
| 13 |
+
def scoring_data_prep(dataset):
|
| 14 |
+
out = []
|
| 15 |
+
target = []
|
| 16 |
+
mask = []
|
| 17 |
+
|
| 18 |
+
for i in range(len(dataset)):
|
| 19 |
+
rec = dataset[i]
|
| 20 |
+
out.append(rec['ids'].reshape(-1,MAX_LEN))
|
| 21 |
+
mask.append(rec['mask'].reshape(-1,MAX_LEN))
|
| 22 |
+
|
| 23 |
+
out_stack = torch.cat(out, dim = 0)
|
| 24 |
+
mask_stack = torch.cat(mask, dim =0 )
|
| 25 |
+
out_stack = out_stack.to(device, dtype = torch.long)
|
| 26 |
+
mask_stack = mask_stack.to(device, dtype = torch.long)
|
| 27 |
+
|
| 28 |
+
return out_stack, mask_stack
|
| 29 |
+
|
| 30 |
+
class Triage(Dataset):
|
| 31 |
+
"""
|
| 32 |
+
This is a subclass of torch packages Dataset class. It processes input to create ids, masks and targets required for model training.
|
| 33 |
+
"""
|
| 34 |
+
|
| 35 |
+
def __init__(self, dataframe, tokenizer, max_len, text_col_name):
|
| 36 |
+
self.len = len(dataframe)
|
| 37 |
+
self.data = dataframe
|
| 38 |
+
self.tokenizer = tokenizer
|
| 39 |
+
self.max_len = max_len
|
| 40 |
+
self.text_col_name = text_col_name
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def __getitem__(self, index):
|
| 44 |
+
title = str(self.data[self.text_col_name][index])
|
| 45 |
+
title = " ".join(title.split())
|
| 46 |
+
inputs = self.tokenizer.encode_plus(
|
| 47 |
+
title,
|
| 48 |
+
None,
|
| 49 |
+
add_special_tokens=True,
|
| 50 |
+
max_length=self.max_len,
|
| 51 |
+
pad_to_max_length=True, #padding='max_length' #For future version use `padding='max_length'`
|
| 52 |
+
return_token_type_ids=True,
|
| 53 |
+
truncation=True,
|
| 54 |
+
)
|
| 55 |
+
ids = inputs["input_ids"]
|
| 56 |
+
mask = inputs["attention_mask"]
|
| 57 |
+
|
| 58 |
+
return {
|
| 59 |
+
"ids": torch.tensor(ids, dtype=torch.long),
|
| 60 |
+
"mask": torch.tensor(mask, dtype=torch.long),
|
| 61 |
+
|
| 62 |
+
}
|
| 63 |
+
|
| 64 |
+
def __len__(self):
|
| 65 |
+
return self.len
|
| 66 |
+
|
| 67 |
+
class BERTClass(torch.nn.Module):
|
| 68 |
+
def __init__(self, num_class, task):
|
| 69 |
+
super(BERTClass, self).__init__()
|
| 70 |
+
self.num_class = num_class
|
| 71 |
+
if task =="sustanability":
|
| 72 |
+
self.l1 = RobertaModel.from_pretrained("roberta-base")
|
| 73 |
+
else:
|
| 74 |
+
self.l1 = BertModel.from_pretrained("ProsusAI/finbert")
|
| 75 |
+
self.pre_classifier = torch.nn.Linear(768, 768)
|
| 76 |
+
self.dropout = torch.nn.Dropout(0.3)
|
| 77 |
+
self.classifier = torch.nn.Linear(768, self.num_class)
|
| 78 |
+
self.history = dict()
|
| 79 |
+
|
| 80 |
+
def forward(self, input_ids, attention_mask):
|
| 81 |
+
output_1 = self.l1(input_ids=input_ids, attention_mask=attention_mask)
|
| 82 |
+
hidden_state = output_1[0]
|
| 83 |
+
pooler = hidden_state[:, 0]
|
| 84 |
+
pooler = self.pre_classifier(pooler)
|
| 85 |
+
pooler = torch.nn.ReLU()(pooler)
|
| 86 |
+
pooler = self.dropout(pooler)
|
| 87 |
+
output = self.classifier(pooler)
|
| 88 |
+
return output
|
| 89 |
+
|
| 90 |
+
def do_predict(model, tokenizer, test_df):
|
| 91 |
+
test_set = Triage(test_df, tokenizer, MAX_LEN, text_col_name)
|
| 92 |
+
test_params = {'batch_size' : BATCH_SIZE, 'shuffle': False, 'num_workers':0}
|
| 93 |
+
test_loader = DataLoader(test_set, **test_params)
|
| 94 |
+
out_stack, mask_stack = scoring_data_prep(dataset = test_set)
|
| 95 |
+
n = 0
|
| 96 |
+
combined_output = []
|
| 97 |
+
model.eval()
|
| 98 |
+
with torch.no_grad():
|
| 99 |
+
while n < test_df.shape[0]:
|
| 100 |
+
output = model(out_stack[n:n+BATCH_SIZE,:],mask_stack[n:n+BATCH_SIZE,:])
|
| 101 |
+
n = n + BATCH_SIZE
|
| 102 |
+
combined_output.append(output)
|
| 103 |
+
combined_output = torch.cat(combined_output, dim = 0)
|
| 104 |
+
preds = torch.argsort(combined_output, axis = 1, descending = True)
|
| 105 |
+
preds = preds.to('cpu')
|
| 106 |
+
actual_predictions = [i[0] for i in preds.tolist()]
|
| 107 |
+
combined_output = combined_output.to('cpu')
|
| 108 |
+
prob_predictions= [i[1] for i in combined_output.tolist()]
|
| 109 |
+
return (actual_predictions, prob_predictions)
|
| 110 |
+
|