Add WCAG classification training pipeline
Browse files- .gitignore +15 -0
- load_dataset.py +15 -0
- load_model.py +25 -0
- predict.py +49 -0
- test.csv +11 -0
- tokenize_dataset.py +52 -0
- train.csv +81 -0
- train_model.py +136 -0
- validation.csv +11 -0
.gitignore
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Python virtual environment
|
| 2 |
+
.venv/
|
| 3 |
+
|
| 4 |
+
# Python cache
|
| 5 |
+
__pycache__/
|
| 6 |
+
*.pyc
|
| 7 |
+
|
| 8 |
+
# Hugging Face training output
|
| 9 |
+
results/
|
| 10 |
+
|
| 11 |
+
# Generated trained model
|
| 12 |
+
wcag-classifier/
|
| 13 |
+
|
| 14 |
+
# Hugging Face / dataset caches
|
| 15 |
+
.cache/
|
load_dataset.py
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from datasets import load_dataset
|
| 2 |
+
|
| 3 |
+
dataset = load_dataset(
|
| 4 |
+
"csv",
|
| 5 |
+
data_files={
|
| 6 |
+
"train": "train.csv",
|
| 7 |
+
"validation": "validation.csv",
|
| 8 |
+
"test": "test.csv",
|
| 9 |
+
}
|
| 10 |
+
)
|
| 11 |
+
|
| 12 |
+
print(dataset)
|
| 13 |
+
|
| 14 |
+
print("\nFirst training example:")
|
| 15 |
+
print(dataset["train"][0])
|
load_model.py
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import AutoModelForSequenceClassification
|
| 2 |
+
|
| 3 |
+
MODEL_NAME = "distilbert-base-uncased"
|
| 4 |
+
|
| 5 |
+
id2label = {
|
| 6 |
+
0: "1.3.1 Info and Relationships",
|
| 7 |
+
1: "2.1.1 Keyboard",
|
| 8 |
+
2: "2.4.3 Focus Order",
|
| 9 |
+
3: "2.4.7 Focus Visible",
|
| 10 |
+
4: "4.1.2 Name, Role, Value",
|
| 11 |
+
}
|
| 12 |
+
|
| 13 |
+
label2id = {
|
| 14 |
+
label: id
|
| 15 |
+
for id, label in id2label.items()
|
| 16 |
+
}
|
| 17 |
+
|
| 18 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
| 19 |
+
MODEL_NAME,
|
| 20 |
+
num_labels=5,
|
| 21 |
+
id2label=id2label,
|
| 22 |
+
label2id=label2id,
|
| 23 |
+
)
|
| 24 |
+
|
| 25 |
+
print(model.config)
|
predict.py
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
|
| 3 |
+
from transformers import (
|
| 4 |
+
AutoModelForSequenceClassification,
|
| 5 |
+
AutoTokenizer,
|
| 6 |
+
)
|
| 7 |
+
|
| 8 |
+
MODEL_PATH = Path(__file__).resolve().parent / "wcag-classifier"
|
| 9 |
+
|
| 10 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
|
| 11 |
+
|
| 12 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
| 13 |
+
MODEL_PATH
|
| 14 |
+
)
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
def predict(text):
|
| 18 |
+
inputs = tokenizer(
|
| 19 |
+
text,
|
| 20 |
+
return_tensors="pt",
|
| 21 |
+
truncation=True,
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
with torch.no_grad():
|
| 25 |
+
outputs = model(**inputs)
|
| 26 |
+
|
| 27 |
+
probabilities = torch.softmax(
|
| 28 |
+
outputs.logits,
|
| 29 |
+
dim=-1
|
| 30 |
+
)[0]
|
| 31 |
+
|
| 32 |
+
predicted_id = torch.argmax(probabilities).item()
|
| 33 |
+
|
| 34 |
+
predicted_label = model.config.id2label[predicted_id]
|
| 35 |
+
confidence = probabilities[predicted_id].item()
|
| 36 |
+
|
| 37 |
+
return predicted_label, confidence
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
while True:
|
| 41 |
+
text = input("\nDescribe an accessibility issue (or 'quit'): ")
|
| 42 |
+
|
| 43 |
+
if text.lower() == "quit":
|
| 44 |
+
break
|
| 45 |
+
|
| 46 |
+
label, confidence = predict(text)
|
| 47 |
+
|
| 48 |
+
print(f"\nPrediction: {label}")
|
| 49 |
+
print(f"Confidence: {confidence:.1%}")
|
test.csv
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
text,label
|
| 2 |
+
Focus jumps from the navigation directly to the footer,2.4.3 Focus Order
|
| 3 |
+
Dropdown options cannot be selected using the keyboard,2.1.1 Keyboard
|
| 4 |
+
Users navigating with Tab cannot see which link is currently active,2.4.7 Focus Visible
|
| 5 |
+
The focus indicator disappears against the control background,2.4.7 Focus Visible
|
| 6 |
+
Custom switch does not expose its on or off state,"4.1.2 Name, Role, Value"
|
| 7 |
+
Image button has no programmatically determinable name,"4.1.2 Name, Role, Value"
|
| 8 |
+
Custom slider cannot be adjusted using the arrow keys,2.1.1 Keyboard
|
| 9 |
+
Section titles are presented using bold text instead of heading markup,1.3.1 Info and Relationships
|
| 10 |
+
After deleting an item focus moves to the page header instead of the next relevant control,2.4.3 Focus Order
|
| 11 |
+
A list of navigation links is visually presented as a list but is not marked up as one,1.3.1 Info and Relationships
|
tokenize_dataset.py
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from datasets import load_dataset
|
| 2 |
+
from transformers import AutoTokenizer
|
| 3 |
+
|
| 4 |
+
MODEL_NAME = "distilbert-base-uncased"
|
| 5 |
+
|
| 6 |
+
LABELS = [
|
| 7 |
+
"1.3.1 Info and Relationships",
|
| 8 |
+
"2.1.1 Keyboard",
|
| 9 |
+
"2.4.3 Focus Order",
|
| 10 |
+
"2.4.7 Focus Visible",
|
| 11 |
+
"4.1.2 Name, Role, Value",
|
| 12 |
+
]
|
| 13 |
+
|
| 14 |
+
label2id = {label: index for index, label in enumerate(LABELS)}
|
| 15 |
+
id2label = {index: label for index, label in enumerate(LABELS)}
|
| 16 |
+
|
| 17 |
+
dataset = load_dataset(
|
| 18 |
+
"csv",
|
| 19 |
+
data_files={
|
| 20 |
+
"train": "train.csv",
|
| 21 |
+
"validation": "validation.csv",
|
| 22 |
+
"test": "test.csv",
|
| 23 |
+
}
|
| 24 |
+
)
|
| 25 |
+
|
| 26 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def preprocess(example):
|
| 30 |
+
encoded = tokenizer(
|
| 31 |
+
example["text"],
|
| 32 |
+
truncation=True
|
| 33 |
+
)
|
| 34 |
+
|
| 35 |
+
encoded["label"] = label2id[example["label"]]
|
| 36 |
+
|
| 37 |
+
return encoded
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
tokenized_dataset = dataset.map(preprocess)
|
| 41 |
+
|
| 42 |
+
print(tokenized_dataset)
|
| 43 |
+
|
| 44 |
+
print("\nOriginal example:")
|
| 45 |
+
print(dataset["train"][0])
|
| 46 |
+
|
| 47 |
+
print("\nTokenized example:")
|
| 48 |
+
print(tokenized_dataset["train"][0])
|
| 49 |
+
|
| 50 |
+
print("\nLabels:")
|
| 51 |
+
for label, id in label2id.items():
|
| 52 |
+
print(f"{id}: {label}")
|
train.csv
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
text,label
|
| 2 |
+
Custom radio control does not expose its selected state,"4.1.2 Name, Role, Value"
|
| 3 |
+
Navigation items show no visual change when they receive keyboard focus,2.4.7 Focus Visible
|
| 4 |
+
Tabbing from the search field moves focus to an unrelated control further down the page,2.4.3 Focus Order
|
| 5 |
+
The custom checkbox has no visible focus styling,2.4.7 Focus Visible
|
| 6 |
+
The submit button does not display an outline when focused,2.4.7 Focus Visible
|
| 7 |
+
Keyboard users cannot operate the carousel controls,2.1.1 Keyboard
|
| 8 |
+
Keyboard focus on the accordion header is not visually apparent,2.4.7 Focus Visible
|
| 9 |
+
Keyboard focus on the menu trigger is visually indistinguishable from its default state,2.4.7 Focus Visible
|
| 10 |
+
Tab order does not follow the visual order of the page,2.4.3 Focus Order
|
| 11 |
+
Keyboard users cannot close the dialog,2.1.1 Keyboard
|
| 12 |
+
Table header relationships are lost when the content is accessed with a screen reader,1.3.1 Info and Relationships
|
| 13 |
+
Form controls are visually grouped under a heading but have no programmatic group relationship,1.3.1 Info and Relationships
|
| 14 |
+
Opening the menu places focus on content outside the menu,2.4.3 Focus Order
|
| 15 |
+
Close button is announced without an accessible name,"4.1.2 Name, Role, Value"
|
| 16 |
+
Keyboard focus is not visible,2.4.7 Focus Visible
|
| 17 |
+
The relationship between a form section heading and its controls is only conveyed visually,1.3.1 Info and Relationships
|
| 18 |
+
Custom control does not respond to keyboard input,2.1.1 Keyboard
|
| 19 |
+
Menu cannot be opened without using a mouse,2.1.1 Keyboard
|
| 20 |
+
Custom slider does not expose its current value to assistive technology,"4.1.2 Name, Role, Value"
|
| 21 |
+
The focus outline has been removed from links,2.4.7 Focus Visible
|
| 22 |
+
User must use a mouse to activate the control,2.1.1 Keyboard
|
| 23 |
+
Drag and drop control provides no keyboard method of operation,2.1.1 Keyboard
|
| 24 |
+
When tabbing through the form there is no indication of the active field,2.4.7 Focus Visible
|
| 25 |
+
The selected navigation item and the keyboard-focused navigation item look identical,2.4.7 Focus Visible
|
| 26 |
+
Table headers are not programmatically associated with their cells,1.3.1 Info and Relationships
|
| 27 |
+
Date picker can only be operated using a mouse,2.1.1 Keyboard
|
| 28 |
+
Navigation control can be clicked with a mouse but cannot be activated from the keyboard,2.1.1 Keyboard
|
| 29 |
+
Combobox does not communicate its role to screen reader users,"4.1.2 Name, Role, Value"
|
| 30 |
+
Interactive icon cannot receive keyboard input,2.1.1 Keyboard
|
| 31 |
+
Content that appears as a numbered list is not programmatically identified as a list,1.3.1 Info and Relationships
|
| 32 |
+
Tabbing to the search button produces no visible focus indicator,2.4.7 Focus Visible
|
| 33 |
+
Text is styled to look like a heading but assistive technology identifies it as ordinary text,1.3.1 Info and Relationships
|
| 34 |
+
Keyboard navigation moves between controls in an illogical sequence,2.4.3 Focus Order
|
| 35 |
+
Toggle button does not communicate whether it is pressed,"4.1.2 Name, Role, Value"
|
| 36 |
+
Menu button does not expose whether the menu is currently open,"4.1.2 Name, Role, Value"
|
| 37 |
+
The currently focused control cannot be distinguished from surrounding controls,2.4.7 Focus Visible
|
| 38 |
+
Interactive icons provide no visible indication when they receive focus,2.4.7 Focus Visible
|
| 39 |
+
Focus unexpectedly jumps back to the top of the page,2.4.3 Focus Order
|
| 40 |
+
User cannot select an item from the list using the keyboard,2.1.1 Keyboard
|
| 41 |
+
Custom dropdown does not expose whether it is expanded or collapsed,"4.1.2 Name, Role, Value"
|
| 42 |
+
Tab sequence does not match the logical reading order,2.4.3 Focus Order
|
| 43 |
+
Visually grouped checkboxes are not contained within a programmatically identified group,1.3.1 Info and Relationships
|
| 44 |
+
Related form fields are positioned together visually but their relationship cannot be determined programmatically,1.3.1 Info and Relationships
|
| 45 |
+
The carousel controls receive keyboard focus without displaying a focus indicator,2.4.7 Focus Visible
|
| 46 |
+
Keyboard users encounter footer links before completing the form,2.4.3 Focus Order
|
| 47 |
+
Custom checkbox does not expose its checked state to assistive technology,"4.1.2 Name, Role, Value"
|
| 48 |
+
Focus styling is missing from the custom dropdown,2.4.7 Focus Visible
|
| 49 |
+
Button cannot be activated using the Enter or Space key,2.1.1 Keyboard
|
| 50 |
+
Tooltip trigger cannot be activated from the keyboard,2.1.1 Keyboard
|
| 51 |
+
Focus moves to controls that are visually positioned later in the interface,2.4.3 Focus Order
|
| 52 |
+
Focus moves outside the modal while the dialog is still open,2.4.3 Focus Order
|
| 53 |
+
Interactive icon is announced without its purpose being identified,"4.1.2 Name, Role, Value"
|
| 54 |
+
After closing the dialog focus does not return to the control that opened it,2.4.3 Focus Order
|
| 55 |
+
Interactive chart controls cannot be operated without a pointing device,2.1.1 Keyboard
|
| 56 |
+
Items in the custom listbox cannot be selected without a mouse,2.1.1 Keyboard
|
| 57 |
+
Keyboard users cannot visually identify which form field currently has focus,2.4.7 Focus Visible
|
| 58 |
+
Focus reaches background page controls before the active dialog controls,2.4.3 Focus Order
|
| 59 |
+
Search icon is exposed as a button but has no accessible name,"4.1.2 Name, Role, Value"
|
| 60 |
+
Focus moves through form fields in an unexpected sequence,2.4.3 Focus Order
|
| 61 |
+
Form fields are visually grouped but the relationship is not available to assistive technology,1.3.1 Info and Relationships
|
| 62 |
+
Screen reader identifies the interactive element as plain text instead of a button,"4.1.2 Name, Role, Value"
|
| 63 |
+
Screen reader does not announce the current value of the custom range control,"4.1.2 Name, Role, Value"
|
| 64 |
+
Radio buttons are visually grouped but are not programmatically associated with their group label,1.3.1 Info and Relationships
|
| 65 |
+
Navigation submenu cannot be expanded using the keyboard,2.1.1 Keyboard
|
| 66 |
+
Opening the modal does not move focus into the dialog,2.4.3 Focus Order
|
| 67 |
+
Table cells are not associated with the appropriate row headers,1.3.1 Info and Relationships
|
| 68 |
+
After submitting the form focus moves to an unrelated navigation item,2.4.3 Focus Order
|
| 69 |
+
Button has no accessible name,"4.1.2 Name, Role, Value"
|
| 70 |
+
Focus moves to the footer before reaching the main content,2.4.3 Focus Order
|
| 71 |
+
Screen reader does not announce the state of the checkbox,"4.1.2 Name, Role, Value"
|
| 72 |
+
Data is arranged visually in rows and columns but no table semantics are provided,1.3.1 Info and Relationships
|
| 73 |
+
There is no visual indication when a navigation link receives keyboard focus,2.4.7 Focus Visible
|
| 74 |
+
Headings are visually styled but are not marked up as headings,1.3.1 Info and Relationships
|
| 75 |
+
Heading structure does not reflect the page hierarchy,1.3.1 Info and Relationships
|
| 76 |
+
The visual hierarchy of the content is not represented in the page markup,1.3.1 Info and Relationships
|
| 77 |
+
Expandable navigation control does not communicate its expanded state,"4.1.2 Name, Role, Value"
|
| 78 |
+
Screen reader does not identify the custom control as a button,"4.1.2 Name, Role, Value"
|
| 79 |
+
Keyboard focus moves behind the open dialog,2.4.3 Focus Order
|
| 80 |
+
A data table does not identify its column headers programmatically,1.3.1 Info and Relationships
|
| 81 |
+
Accordion sections cannot be expanded using keyboard controls,2.1.1 Keyboard
|
train_model.py
ADDED
|
@@ -0,0 +1,136 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
|
| 3 |
+
from datasets import load_dataset
|
| 4 |
+
from sklearn.metrics import accuracy_score, precision_recall_fscore_support
|
| 5 |
+
from transformers import (
|
| 6 |
+
AutoModelForSequenceClassification,
|
| 7 |
+
AutoTokenizer,
|
| 8 |
+
DataCollatorWithPadding,
|
| 9 |
+
Trainer,
|
| 10 |
+
TrainingArguments,
|
| 11 |
+
)
|
| 12 |
+
|
| 13 |
+
MODEL_NAME = "distilbert-base-uncased"
|
| 14 |
+
|
| 15 |
+
LABELS = [
|
| 16 |
+
"1.3.1 Info and Relationships",
|
| 17 |
+
"2.1.1 Keyboard",
|
| 18 |
+
"2.4.3 Focus Order",
|
| 19 |
+
"2.4.7 Focus Visible",
|
| 20 |
+
"4.1.2 Name, Role, Value",
|
| 21 |
+
]
|
| 22 |
+
|
| 23 |
+
label2id = {label: i for i, label in enumerate(LABELS)}
|
| 24 |
+
id2label = {i: label for i, label in enumerate(LABELS)}
|
| 25 |
+
|
| 26 |
+
# Load our CSV files
|
| 27 |
+
dataset = load_dataset(
|
| 28 |
+
"csv",
|
| 29 |
+
data_files={
|
| 30 |
+
"train": "train.csv",
|
| 31 |
+
"validation": "validation.csv",
|
| 32 |
+
"test": "test.csv",
|
| 33 |
+
},
|
| 34 |
+
)
|
| 35 |
+
|
| 36 |
+
# Load DistilBERT's tokenizer
|
| 37 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def preprocess(example):
|
| 41 |
+
encoded = tokenizer(
|
| 42 |
+
example["text"],
|
| 43 |
+
truncation=True,
|
| 44 |
+
)
|
| 45 |
+
|
| 46 |
+
encoded["label_id"] = label2id[example["label"]]
|
| 47 |
+
|
| 48 |
+
return encoded
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
tokenized_dataset = dataset.map(
|
| 52 |
+
preprocess,
|
| 53 |
+
remove_columns=["text", "label"],
|
| 54 |
+
)
|
| 55 |
+
|
| 56 |
+
tokenized_dataset = tokenized_dataset.rename_column(
|
| 57 |
+
"label_id",
|
| 58 |
+
"labels",
|
| 59 |
+
)
|
| 60 |
+
|
| 61 |
+
# Pads each batch to the longest sentence in that batch
|
| 62 |
+
data_collator = DataCollatorWithPadding(
|
| 63 |
+
tokenizer=tokenizer
|
| 64 |
+
)
|
| 65 |
+
|
| 66 |
+
# Load pretrained DistilBERT with our new 5-class head
|
| 67 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
| 68 |
+
MODEL_NAME,
|
| 69 |
+
num_labels=len(LABELS),
|
| 70 |
+
id2label=id2label,
|
| 71 |
+
label2id=label2id,
|
| 72 |
+
)
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def compute_metrics(eval_pred):
|
| 76 |
+
logits, labels = eval_pred
|
| 77 |
+
|
| 78 |
+
predictions = np.argmax(logits, axis=-1)
|
| 79 |
+
|
| 80 |
+
precision, recall, f1, _ = precision_recall_fscore_support(
|
| 81 |
+
labels,
|
| 82 |
+
predictions,
|
| 83 |
+
average="weighted",
|
| 84 |
+
zero_division=0,
|
| 85 |
+
)
|
| 86 |
+
|
| 87 |
+
accuracy = accuracy_score(
|
| 88 |
+
labels,
|
| 89 |
+
predictions,
|
| 90 |
+
)
|
| 91 |
+
|
| 92 |
+
return {
|
| 93 |
+
"accuracy": accuracy,
|
| 94 |
+
"precision": precision,
|
| 95 |
+
"recall": recall,
|
| 96 |
+
"f1": f1,
|
| 97 |
+
}
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
training_args = TrainingArguments(
|
| 101 |
+
output_dir="./results",
|
| 102 |
+
num_train_epochs=5,
|
| 103 |
+
per_device_train_batch_size=8,
|
| 104 |
+
per_device_eval_batch_size=8,
|
| 105 |
+
learning_rate=2e-5,
|
| 106 |
+
eval_strategy="epoch",
|
| 107 |
+
save_strategy="epoch",
|
| 108 |
+
logging_strategy="epoch",
|
| 109 |
+
load_best_model_at_end=True,
|
| 110 |
+
metric_for_best_model="f1",
|
| 111 |
+
)
|
| 112 |
+
|
| 113 |
+
trainer = Trainer(
|
| 114 |
+
model=model,
|
| 115 |
+
args=training_args,
|
| 116 |
+
train_dataset=tokenized_dataset["train"],
|
| 117 |
+
eval_dataset=tokenized_dataset["validation"],
|
| 118 |
+
data_collator=data_collator,
|
| 119 |
+
compute_metrics=compute_metrics,
|
| 120 |
+
)
|
| 121 |
+
|
| 122 |
+
print("\nStarting training...\n")
|
| 123 |
+
|
| 124 |
+
trainer.train()
|
| 125 |
+
|
| 126 |
+
print("\nEvaluating final model on test data...\n")
|
| 127 |
+
|
| 128 |
+
test_results = trainer.evaluate(
|
| 129 |
+
tokenized_dataset["test"]
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
print(test_results)
|
| 133 |
+
|
| 134 |
+
# Save our finished model locally
|
| 135 |
+
trainer.save_model("./wcag-classifier")
|
| 136 |
+
tokenizer.save_pretrained("./wcag-classifier")
|
validation.csv
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
text,label
|
| 2 |
+
Nested headings do not represent the structure shown visually on the page,1.3.1 Info and Relationships
|
| 3 |
+
Icon button does not provide an accessible name,"4.1.2 Name, Role, Value"
|
| 4 |
+
Focus is programmatically moved to the close button but there is no visual indication,2.4.7 Focus Visible
|
| 5 |
+
Modal cannot be opened using the keyboard,2.1.1 Keyboard
|
| 6 |
+
Required fields are indicated by visual formatting that is not programmatically available,1.3.1 Info and Relationships
|
| 7 |
+
Keyboard navigation reaches hidden content before visible controls,2.4.3 Focus Order
|
| 8 |
+
Tabbing through the page skips to the sidebar before the main content,2.4.3 Focus Order
|
| 9 |
+
Custom toggle responds to mouse clicks but not keyboard activation,2.1.1 Keyboard
|
| 10 |
+
Accordion header does not communicate whether its panel is expanded or collapsed,"4.1.2 Name, Role, Value"
|
| 11 |
+
Focused buttons cannot be visually distinguished from unfocused buttons,2.4.7 Focus Visible
|