-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
149 lines (119 loc) · 5.83 KB
/
model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
import math
import numpy as np
from tqdm import tqdm
import torch
from torch import nn
from torch import optim, nn, utils
from torch.utils.data import Dataset, DataLoader
import lightning.pytorch as pl
# import pytorch_lightning as plpl
# from lightning.pytorch.callbacks import ModelCheckpoint
from transformers import XLNetTokenizer, XLNetModel, AutoTokenizer, AlbertModel, AutoModel, ElectraModel, RobertaModel, AlbertTokenizer
class SoftMaxLit(pl.LightningModule):
"""
Reference
https://machinelearningmastery.com/introduction-to-softmax-classifier-in-pytorch/
"""
def __init__(self, n_inputs, n_outputs):
super().__init__()
self.linear = torch.nn.Linear(n_inputs, n_outputs)
self.softmax = nn.Softmax(dim=1)
self.criterion = nn.CrossEntropyLoss()
def forward(self, x):
return self.softmax(self.linear(x))
def training_step(self, batch, batch_idx):
# training_step defines the train loop.
# it is independent of forward
x, y = batch
y_hat = self(x)
loss = self.criterion(y_hat, y)
self.log('train_loss', loss)
return loss
def configure_optimizers(self):
optimizer = optim.AdamW(self.parameters(), lr = 0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0.01, amsgrad=False)
return optimizer
def validation_step(self, batch, batch_idx):
x, y = batch
y_hat = self(x)
loss = self.criterion(y_hat, y)
self.log("val_loss", loss)
def test_step(self, batch, batch_idx):
x, y = batch
y = torch.argmax(y, dim=1)
y_hat = torch.argmax(self(x), dim=1)
accuracy = torch.sum(y == y_hat).item() / (len(y) * 1.0)
self.log('test_acc', accuracy)
class Data(Dataset):
"The data for multi-class classification"
def __init__(self, df, *, x=None, load_batch_size=None, tokenizer=None, pretrained=None):
self.y, self.len = self._get_y_and_len_from_df(df)
if x is not None:
self.x = x
else:
self.x = self._get_x_from_df(df, load_batch_size, tokenizer, pretrained)
def _get_x_from_df(self, df, load_batch_size, tokenizer, pretrained):
docs = df['text'].tolist()
inputs = tokenizer(docs, return_tensors="pt", padding=True)
cls_arr = []
for i, (x, y) in zip(tqdm(range(math.ceil(len(df) / load_batch_size))), self._get_x_y_from_df_with_batch(df, load_batch_size)):
cls = pretrained(**{k: inputs[k][x:y] for k in list(inputs.keys())}).last_hidden_state[:, 0, :].detach()
cls_arr.append(cls)
return torch.concat(cls_arr).type(torch.float32)
def _get_y_and_len_from_df(self, df):
dim_0 = df['text'].shape[0]
matrix = np.zeros((dim_0,2))
for i, y in enumerate(df['label'].tolist()):
matrix[i][y] = 1
return torch.from_numpy(matrix).type(torch.float32), dim_0
def _get_x_y_from_df_with_batch(self, df, step_size):
l = list(range(0, len(df), step_size))
for ind, _ in enumerate(l):
if l[ind] + step_size >= len(df):
yield (l[ind], len(df))
else:
yield (l[ind], l[ind + 1])
def __getitem__(self, idx):
"accessing one element in the dataset by index"
return self.x[idx], self.y[idx]
def __len__(self):
"size of the entire dataset"
return self.len
@staticmethod
def concat(df, datasets):
"concatenate dataset embeddings from x provided they are applied on the same df"
x = torch.cat([dataset.x for dataset in datasets], 1)
return Data(df, x=x)
# MODELS
class TransformerModel():
# # XLNet: https://huggingface.co/docs/transformers/model_doc/xlnet # size = 768
# # ALBERT: https://huggingface.co/docs/transformers/model_doc/albert # size = 768
# # ELECTRA: 256
# # Roberta: 768
MODELS = {
'albert': {'name': 'albert-base-v2', 'dim': 768,'tokenizer': AlbertTokenizer, 'pretrained': AlbertModel},
'electra': {'name': 'google/electra-small-discriminator', 'dim': 256,'tokenizer': AutoTokenizer, 'pretrained': ElectraModel},
'roberta': {'name': 'roberta-base', 'dim': 768,'tokenizer': AutoTokenizer, 'pretrained': RobertaModel},
'xlnet': {'name': 'xlnet-base-cased', 'dim': 768, 'tokenizer': XLNetTokenizer, 'pretrained': XLNetModel},
}
def __init__(self, model_tag):
if model_tag not in list(self.MODELS.keys()):
raise ValueError(f'Invalid model: {model_tag}. Valide models are: {self.MODELS.join(" ")}')
self.model_tag = model_tag
self.dim = self.MODELS[model_tag]['dim']
self.tokenizer = self.MODELS[model_tag]['tokenizer'].from_pretrained(self.MODELS[model_tag]['name'])
self.pretrained = self.MODELS[model_tag]['pretrained'].from_pretrained(self.MODELS[model_tag]['name'])
def dataset(self, df, dev, save=False, delete=False):
# cur_df = df[:100] if dev else df
dataset = Data(df, load_batch_size = 30, tokenizer=self.tokenizer, pretrained=self.pretrained) # 10 > 30 > 40 yes # 4 is the best
if save:
torch.save(dataset.x, f"pretrained--dev={dev}--model={self.model_tag}.pt")
if delete:
del dataset.x
torch.cuda.empty_cache()
return dataset
def get_dataloaders(dataset, batch_size):
train_dataset, val_dataset, test_dataset = utils.data.random_split(dataset,(0.8, 0.1, 0.1))
train_dataloader = DataLoader(dataset = train_dataset, batch_size = batch_size, shuffle=True)
val_dataloader = DataLoader(dataset = val_dataset, batch_size = batch_size, shuffle=True)
test_dataloader = DataLoader(dataset = test_dataset, batch_size = batch_size, shuffle=True)
return {'train': train_dataloader, 'val': val_dataloader, 'test': test_dataloader}