From 6c9a8020aee2cc21de40e4d1647141f6354dfba4 Mon Sep 17 00:00:00 2001
From: MGniew <m.f.gniewkowski@gmail.com>
Date: Mon, 27 Mar 2023 00:31:27 +0200
Subject: [PATCH] poleval
---
.gitignore | 1 +
data/classification/.gitignore | 1 +
data/explanations/.gitignore | 1 +
data/models/.gitignore | 1 +
dvc.lock | 56 +++++++++++
dvc.yaml | 3 +
text_attacks/models/poleval.py | 173 ++++++++++++++++++++++++++++++++-
7 files changed, 233 insertions(+), 3 deletions(-)
diff --git a/.gitignore b/.gitignore
index 853a525..92e0643 100644
--- a/.gitignore
+++ b/.gitignore
@@ -25,6 +25,7 @@ share/python-wheels/
.installed.cfg
*.egg
MANIFEST
+tmp/
# PyInstaller
# Usually these files are written by a python script from a template
diff --git a/data/classification/.gitignore b/data/classification/.gitignore
index e695872..27e43f9 100644
--- a/data/classification/.gitignore
+++ b/data/classification/.gitignore
@@ -1,3 +1,4 @@
/enron_spam
/wiki_pl
/20_news
+/poleval
diff --git a/data/explanations/.gitignore b/data/explanations/.gitignore
index e695872..27e43f9 100644
--- a/data/explanations/.gitignore
+++ b/data/explanations/.gitignore
@@ -1,3 +1,4 @@
/enron_spam
/wiki_pl
/20_news
+/poleval
diff --git a/data/models/.gitignore b/data/models/.gitignore
index ea22867..f37e1f6 100644
--- a/data/models/.gitignore
+++ b/data/models/.gitignore
@@ -1,3 +1,4 @@
/enron_spam
/20_news
/wiki_pl
+/poleval
diff --git a/dvc.lock b/dvc.lock
index 3db2d0d..bbd4afc 100644
--- a/dvc.lock
+++ b/dvc.lock
@@ -639,3 +639,59 @@ stages:
md5: d9595138a6aca9cad4c39b1c5fdbfce1.dir
size: 38396600
nfiles: 2
+ get_model@poleval:
+ cmd: PYTHONPATH=. python experiments/scripts/get_model.py --dataset_name poleval
+ --output_dir data/models/poleval
+ deps:
+ - path: data/preprocessed/poleval
+ md5: b0ea9f0ad1dba6d3b474c0a3cedf866e.dir
+ size: 2812175
+ nfiles: 3
+ - path: experiments/scripts/get_model.py
+ md5: 5050f51b4019bba97af47971f6c7cab4
+ size: 747
+ outs:
+ - path: data/models/poleval/
+ md5: 8f806cb1b2eb0dd097811d42e4bf9c2d.dir
+ size: 501609312
+ nfiles: 7
+ explain@poleval:
+ cmd: PYTHONPATH=. python experiments/scripts/explain.py --dataset_name poleval
+ --output_dir data/explanations/poleval
+ deps:
+ - path: data/models/poleval
+ md5: 8f806cb1b2eb0dd097811d42e4bf9c2d.dir
+ size: 501609312
+ nfiles: 7
+ - path: data/preprocessed/poleval
+ md5: b0ea9f0ad1dba6d3b474c0a3cedf866e.dir
+ size: 2812175
+ nfiles: 3
+ - path: experiments/scripts/explain.py
+ md5: 4d82a557627f59c884f52fd7994ed80a
+ size: 4617
+ outs:
+ - path: data/explanations/poleval/
+ md5: 246ec6832b514bbfc7ff6bcdb0d9f292.dir
+ size: 2377833
+ nfiles: 2003
+ classify@poleval:
+ cmd: PYTHONPATH=. python experiments/scripts/classify.py --dataset_name poleval
+ --output_dir data/classification/poleval
+ deps:
+ - path: data/models/poleval/
+ md5: 8f806cb1b2eb0dd097811d42e4bf9c2d.dir
+ size: 501609312
+ nfiles: 7
+ - path: data/preprocessed/poleval/
+ md5: b0ea9f0ad1dba6d3b474c0a3cedf866e.dir
+ size: 2812175
+ nfiles: 3
+ - path: experiments/scripts/classify.py
+ md5: 6fc1a6a0a11ba6cd99a8b6625a96d9f5
+ size: 1181
+ outs:
+ - path: data/classification/poleval
+ md5: f207458f9365a74672c31b5ffb2a83af.dir
+ size: 787456
+ nfiles: 2
diff --git a/dvc.yaml b/dvc.yaml
index 0455a1b..e0cd0d3 100644
--- a/dvc.yaml
+++ b/dvc.yaml
@@ -32,6 +32,7 @@ stages:
get_model:
foreach:
- enron_spam
+ - poleval
do:
wdir: .
cmd: >-
@@ -46,6 +47,7 @@ stages:
classify:
foreach:
- enron_spam
+ - poleval
- 20_news
- wiki_pl
do:
@@ -63,6 +65,7 @@ stages:
explain:
foreach:
- enron_spam
+ - poleval
- 20_news
- wiki_pl
do:
diff --git a/text_attacks/models/poleval.py b/text_attacks/models/poleval.py
index a037f8d..cc2b0c1 100644
--- a/text_attacks/models/poleval.py
+++ b/text_attacks/models/poleval.py
@@ -1,13 +1,180 @@
"""Classification model for enron_spam"""
+import os
+import pandas as pd
+from transformers import AutoConfig, BertForSequenceClassification
+from transformers import AutoTokenizer, TrainingArguments, Trainer
+from transformers import BertConfig, AutoModelForSequenceClassification
+
+import random
+import numpy as np
+import torch
+
+from tqdm import tqdm
+
+from sklearn.metrics import accuracy_score, precision_recall_fscore_support
+
+
+class Dataset(torch.utils.data.Dataset):
+ def __init__(self, encodings, labels):
+ self.encodings = encodings
+ self.labels = labels
+
+ def __getitem__(self, idx):
+ item = {k: torch.tensor(v[idx]) for k, v in self.encodings.items()}
+ item["labels"] = torch.tensor([self.labels[idx]])
+ return item
+
+ def __len__(self):
+ return len(self.labels)
+
+
+def compute_metrics(pred):
+ labels = pred.label_ids
+ preds = pred.predictions.argmax(-1)
+ precision, recall, f1, _ = precision_recall_fscore_support(
+ labels, preds, average='binary'
+ )
+ acc = accuracy_score(labels, preds)
+ return {
+ 'accuracy': acc,
+ 'f1': f1,
+ 'precision': precision,
+ 'recall': recall
+ }
+
+
+def train_model():
+ tokenizer = AutoTokenizer.from_pretrained("./data/models/wiki_pl")
+ model = AutoModelForSequenceClassification.from_pretrained(
+ "./data/models/wiki_pl", num_labels=2,
+ ignore_mismatched_sizes=True
+ )
+
+ test = pd.read_json(f"data/preprocessed/poleval/test.jsonl", lines=True)
+ train = pd.read_json(f"data/preprocessed/poleval/train.jsonl", lines=True)
+ y_test = [0 if y == "normal" else 1 for y in test["label"]]
+ y_train = [0 if y == "normal" else 1 for y in train["label"]]
+ x_test = test["text"].tolist()
+ x_train = train["text"].tolist()
+
+ train_encodings = tokenizer(
+ x_train, truncation=True, padding=True, max_length=512
+ )
+ train_dataset = Dataset(train_encodings, y_train)
+
+ test_encodings = tokenizer(
+ x_test, truncation=True, padding=True, max_length=512
+ )
+ test_dataset = Dataset(test_encodings, y_test)
+
+ training_args = TrainingArguments(
+ output_dir='./tmp',
+ num_train_epochs=100,
+ warmup_steps=100,
+ weight_decay=0.01,
+ per_device_train_batch_size=32,
+ per_device_eval_batch_size=32,
+ logging_dir='./tmp/logs',
+ logging_steps=500,
+ save_steps=500,
+ save_total_limit=10,
+ learning_rate=1e-5,
+ load_best_model_at_end=True,
+ evaluation_strategy="steps",
+ )
+ trainer = Trainer(
+ model=model,
+ args=training_args,
+ train_dataset=train_dataset,
+ eval_dataset=test_dataset,
+ compute_metrics=compute_metrics,
+ )
+ trainer.train()
+ return model, tokenizer
+
+
+def train_model_old():
+ tokenizer = AutoTokenizer.from_pretrained("bert-base-multilingual-cased")
+ config = AutoConfig.from_pretrained(
+ "mrm8488/bert-tiny-finetuned-enron-spam-detection"
+ )
+ config.update({"vocab_size": tokenizer.vocab_size})
+
+ model = BertForSequenceClassification(config)
+ test = pd.read_json(f"data/preprocessed/poleval/test.jsonl", lines=True)
+ train = pd.read_json(f"data/preprocessed/poleval/train.jsonl", lines=True)
+ y_test = [0 if y == "normal" else 1 for y in test["label"]]
+ y_train = [0 if y == "normal" else 1 for y in train["label"]]
+ x_test = test["text"].tolist()
+ x_train = train["text"].tolist()
+
+ train_encodings = tokenizer(
+ x_train, truncation=True, padding=True, max_length=512
+ )
+ train_dataset = Dataset(train_encodings, y_train)
+
+ test_encodings = tokenizer(
+ x_test, truncation=True, padding=True, max_length=512
+ )
+ test_dataset = Dataset(test_encodings, y_test)
+
+ training_args = TrainingArguments(
+ output_dir='./tmp',
+ num_train_epochs=250,
+ warmup_steps=100,
+ weight_decay=0.01,
+ logging_dir='./tmp/logs',
+ logging_steps=1000,
+ save_steps=1000,
+ save_total_limit=10,
+ load_best_model_at_end=True,
+ evaluation_strategy="steps",
+ )
+ trainer = Trainer(
+ model=model,
+ args=training_args,
+ train_dataset=train_dataset,
+ eval_dataset=test_dataset,
+ compute_metrics=compute_metrics,
+ )
+ trainer.train()
+ return model, tokenizer
def get_model_and_tokenizer():
- return None, None
+ model_path = "./data/models/poleval/"
+ if not os.path.exists(model_path + "config.json"):
+ model, tokenizer = train_model()
+ else:
+ tokenizer = AutoTokenizer.from_pretrained(model_path)
+ model = AutoModelForSequenceClassification.from_pretrained(model_path)
+ model.config.id2label = {0: "normal", 1: "hate"}
+ return model, tokenizer
-def get_classify_function():
+def get_classify_function(device="cpu"):
+ model, tokenizer = get_model_and_tokenizer()
+ model.eval()
+ model = model.to(device)
def fun(texts):
- return "dummy"
+ logits = list()
+ i = 0
+ for chunk in tqdm(
+ [texts[pos:pos + 128] for pos in range(0, len(texts), 128)]
+ ):
+ encoded_inputs = tokenizer(
+ chunk,
+ return_tensors="pt",
+ padding=True,
+ truncation=True,
+ max_length=512
+ ).to(device)
+ with torch.no_grad():
+ logits.append(model(**encoded_inputs).logits.cpu())
+ logits = torch.cat(logits, dim=0)
+ pred_y = torch.argmax(logits, dim=1).tolist()
+ pred_y = [model.config.id2label[p] for p in pred_y]
+ return pred_y
return fun
--
GitLab