File size: 5,127 Bytes
225a1cb | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 | # -*- coding: utf-8 -*-
"""code_10-testing dev transfer model.ipynb
Automatically generated by Colaboratory.
Original file is located at
https://colab.research.google.com/drive/1YYajoZ_9v-0JQNl3QkTtdQlUDBVPxKDM
"""
!pip install transformers torch
!pip install accelerate -U
from google.colab import files
import zipfile
import os
uploaded = files.upload()
for fn in uploaded.keys():
print('User uploaded file "{name}" with length {length} bytes'.format(
name=fn, length=len(uploaded[fn])))
if fn.endswith('.zip'):
with zipfile.ZipFile(fn, 'r') as zip_ref:
zip_ref.extractall()
print(f"Files from '{fn}' unzipped in the current directory!")
else:
print(f"'{fn}' is not a zip file.")
os.listdir()
import json
import numpy as np
import torch
from transformers import (BertTokenizer, BertForSequenceClassification,
RobertaTokenizer, RobertaForSequenceClassification,
XLNetTokenizer, XLNetForSequenceClassification,
DebertaTokenizer, DebertaForSequenceClassification,
BartTokenizer, BartForSequenceClassification,
TrainingArguments, Trainer)
from sklearn.model_selection import train_test_split
from torch.utils.data import Dataset
from sklearn.metrics import classification_report
def compute_metrics(p):
pred_labels = np.argmax(p.predictions, axis=1)
true_labels = p.label_ids
metrics = classification_report(true_labels, pred_labels, output_dict=True)
accuracy = metrics['accuracy']
weighted_avg = metrics['weighted avg']
precision = weighted_avg['precision']
f1 = weighted_avg['f1-score']
return {
'accuracy': accuracy,
'f1': f1,
'precision': precision
}
# Load surprise data
with open('surprise.data', 'r') as data_file:
data = [json.loads(line) for line in data_file]
with open('surprise.solution', 'r') as solution_file:
solutions = [json.loads(line) for line in solution_file]
intent_map = {item['indoml_id']: item['intent'] for item in solutions}
indoml_ids = [item['indoml_id'] for item in data]
intents = [intent_map[indoml_id] for indoml_id in indoml_ids]
utt = [item['utt'] for item in data]
label_map = {intent: i for i, intent in enumerate(set(intents))}
train_labels = [label_map[label] for label in intents]
utt_train, utt_test, labels_train, labels_test = train_test_split(utt, train_labels, test_size=0.3, random_state=42)
max_length = 64
class IntentDataset(Dataset):
def __init__(self, encodings, labels):
self.encodings = encodings
self.labels = labels
def __getitem__(self, idx):
item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
item['labels'] = torch.tensor(self.labels[idx])
return item
def __len__(self):
return len(self.labels)
# Common training arguments
training_args = TrainingArguments(
output_dir='./results',
save_strategy="epoch",
num_train_epochs=12,
per_device_train_batch_size=16,
per_device_eval_batch_size=64,
warmup_steps=500,
weight_decay=0.01,
logging_dir='./logs',
logging_steps=50,
learning_rate=2e-5,
evaluation_strategy="epoch",
)
from transformers import BertConfig
# Assuming num_classes is defined earlier in the code. If not, define it:
num_classes = len(label_map)
# Transfer
model_path = '/content/drive/MyDrive/bert_model/model.pth'
tokenizer_path = '/content/drive/MyDrive/bert_tokenizer'
transfer_tokenizer = BertTokenizer(vocab_file='/content/drive/MyDrive/bert_tokenizer/vocab.txt')
config_original = BertConfig.from_pretrained('bert-base-multilingual-cased', num_labels=60)
transfer_model = BertForSequenceClassification(config=config_original)
model_state_dict = torch.load(model_path)
transfer_model.load_state_dict(model_state_dict)
transfer_model.num_labels = num_classes
transfer_model.classifier = torch.nn.Linear(transfer_model.config.hidden_size, num_classes)
# Create encodings for transfer learning
transfer_train_encodings = transfer_tokenizer(utt_train, truncation=True, padding=True, max_length=max_length, return_tensors='pt')
transfer_test_encodings = transfer_tokenizer(utt_test, truncation=True, padding=True, max_length=max_length, return_tensors='pt')
transfer_train_dataset = IntentDataset(transfer_train_encodings, labels_train)
transfer_test_dataset = IntentDataset(transfer_test_encodings, labels_test)
# Optional: Freeze all layers except the last few
for param in transfer_model.parameters():
param.requires_grad = False
# Unfreeze the classifier layers
for param in transfer_model.classifier.parameters():
param.requires_grad = True
# Train the transfer model
transfer_trainer = Trainer(
model=transfer_model,
args=training_args,
train_dataset=transfer_train_dataset,
eval_dataset=transfer_test_dataset,
compute_metrics=compute_metrics
)
transfer_trainer.train()
transfer_results = transfer_trainer.evaluate()
# Display results for the transfer model
display_results("Transfer BERT", transfer_results) |