Merge pull request #69 from hunterheiden/hsh/new_task/WebSRC

[New Task] WebSRC (multimodal Q&A on web screenshots)

Author: Luodian

README.md

@@ -260,6 +260,9 @@ We also provide the raw data exported from Weights & Biases for the detailed res
 - VQAv2 (vqav2)
   - VQAv2 Validation (vqav2_val)
   - VQAv2 Test (vqav2_test)
+- WebSRC (websrc)
+  - WebSRC Validation (websrc_val)
+  - WebSRC Test (websrc_test)
 
 ## Datasets to be added and tested
 - TallyQA (tallyqa)

lmms_eval/tasks/websrc/README.md

@@ -0,0 +1,51 @@
+# WebSRC
+
+## Paper 
+
+Title: WebSRC: A Dataset for Web-Based Structural Reading Comprehension
+
+Abstract: https://arxiv.org/abs/2101.09465
+
+Homepage: https://x-lance.github.io/WebSRC/#
+
+WebSRC is a dataset for web-based structural reading comprehension.
+Its full train/dev/test split contains over 400k questions across 6.4k webpages. 
+This version of the dataset does not contain OCR or original HTML, it simply treats WebSRC as a image-and-text-based multimodal Q&A benchmark on webpage screenshots.
+
+## Citation
+
+```bibtex
+@inproceedings{chen2021websrc,
+  title={WebSRC: A Dataset for Web-Based Structural Reading Comprehension},
+  author={Chen, Xingyu and Zhao, Zihan and Chen, Lu and Ji, Jiabao and Zhang, Danyang and Luo, Ao and Xiong, Yuxuan and Yu, Kai},
+  booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
+  pages={4173--4185},
+  year={2021}
+}
+```
+
+## Groups & Tasks
+
+### Groups 
+
+- `websrc`: Evaluates `websrc-val` and generates a submission file for `websrc-test`.
+
+### Tasks
+
+- `websrc-val`: Given a question and a web page, predict the answer.
+- `websrc-test`: Given a question and a web page, predict the answer. Ground truth is not provided for this task.
+
+## Metrics
+
+This task uses SQUAD-style evaluation metrics, of which F1 score over tokens is used. 
+The orignal paper also uses Exact Match (EM) score, but this is not implemented here as that metric is more conducive for Encoder-only extraction models.
+
+### F1 Score
+
+F1 Score is the harmonic mean of precision and recall.
+We calculate precision and recall at the token level, then compute the F1 score as normal using these values.
+
+### Test Submission
+
+When evaluaing on the test split, a prediction JSON will be compiled instead of metrics computed. 
+Instructions for submission are available on the [WebSRC homepage](https://x-lance.github.io/WebSRC/#) and in their [Original GitHub Repo](https://github.com/X-LANCE/WebSRC-Baseline#obtain-test-result).

lmms_eval/tasks/websrc/utils.py

@@ -0,0 +1,148 @@
+from collections import defaultdict
+import re
+import ast
+import base64
+import io
+import random
+import numpy as np
+import os
+import json
+import logging
+from PIL import Image
+
+from lmms_eval.tasks._task_utils.file_utils import generate_submission_file
+
+lmms_logger = logging.getLogger("lmms-eval")
+
+OPEN_ENDED_PROMPT = "Answer the question using a single word or phrase."
+
+
+def construct_prompt(doc):
+    question = doc["question"]
+    question = f"{OPEN_ENDED_PROMPT}\n{question}"
+    return question
+
+
+def websrc_doc_to_text(doc):
+    question = construct_prompt(doc)
+    return question
+
+
+def websrc_doc_to_visual(doc):
+    img_bs64 = doc["image"]
+    img = Image.open(io.BytesIO(base64.b64decode(img_bs64)))
+    del doc['image']
+    return [img]
+
+
+def websrc_process_results(doc, results):
+    pred = results[0]
+    parsed_pred = pred
+    id = doc["page_id"]
+    websrc_ans = {"id": id, "domain": doc['domain'], "parsed_pred": parsed_pred}
+    if "answer" in doc:
+        websrc_ans["answer"] = doc["answer"]
+
+    if 'id' in doc:
+        websrc_ans['question_id'] = doc['id']
+
+    return {
+        "websrc_squad_f1": websrc_ans,
+        "submission": {
+            websrc_ans['question_id']: pred,
+        },
+    }
+
+
+def websrc_test_aggregate_results_for_submission(results, args):
+    path = generate_submission_file("websrc_test_for_submission.json", args)
+    with open(path, "w") as f:
+        out = {}
+        for result in results:
+            out.update(result)
+        json.dump(out, f, indent=4)
+    lmms_logger.info(f"Results saved to {path}.")
+
+
+def websrc_aggregate_results(results):
+    evaluation_result = {}
+
+    # Group results by domain
+    subset_to_eval_samples = defaultdict(list)
+    for result in results:
+        subset_to_eval_samples[result["domain"]].append(result)
+
+    # Evaluate each domain
+    for subset, sub_eval_samples in subset_to_eval_samples.items():
+        judge_dict, metric_dict = evaluate_websrc(sub_eval_samples)
+        metric_dict.update({"num_example": len(sub_eval_samples)})
+        evaluation_result[subset] = metric_dict
+
+    # Aggregate results for all domains
+    printable_results = {}
+    for domain in DOMAINS:
+        if domain not in evaluation_result:
+            continue
+        printable_results[domain] = {
+            "num": int(evaluation_result[domain]["num_example"]),
+            "f1": round(evaluation_result[domain]["f1"], 3),
+        }
+    all_ins_f1 = np.sum([cat_results["f1"] * cat_results["num_example"] for cat_results in evaluation_result.values()]) / sum(
+        [cat_results["num_example"] for cat_results in evaluation_result.values()]
+    )
+    printable_results["Overall"] = {
+        "num": sum([cat_results["num_example"] for cat_results in evaluation_result.values()]),
+        "f1": round(all_ins_f1, 3),
+    }
+    print(printable_results)
+    return printable_results["Overall"]["f1"]
+
+
+##################
+# Helper functions written by official MMMU repo.
+##################
+DOMAINS = [
+    'auto',
+    'book',
+    'camera',
+    'game',
+    'jobs',
+    'movie',
+    'phone',
+    'restaurant',
+    'sports',
+    'university',
+    'hotel',
+]
+
+
+def evaluate_websrc(samples):
+
+    def _normalize_str(string):
+        # lower it
+        string = string.lower()
+
+        # strip non-alphanumeric characters
+        string = re.sub(r"[^a-zA-Z0-9]", "", string)
+
+        # strip leading and trailing whitespaces
+        string = string.strip()
+        
+        return string
+
+    judge_list = []
+    for sample in samples:
+        gold_i = set(_normalize_str(sample["answer"]))
+        pred_i = set(_normalize_str( sample["parsed_pred"]))
+        if len(pred_i) == 0:
+            judge_list.append(0.0)
+            continue
+
+        comm_i = gold_i.intersection(pred_i)
+        prec_i = len(comm_i) / len(pred_i)
+        rec_i = len(comm_i) / len(gold_i)
+        f1_i = 2 * prec_i * rec_i / (prec_i + rec_i) if prec_i + rec_i > 0 else 0
+        judge_list.append(f1_i)
+
+    f1 = np.mean(judge_list)
+    return judge_list, {"f1": f1}

lmms_eval/tasks/websrc/websrc.yaml

@@ -0,0 +1,4 @@
+group: websrc
+task:
+- websrc_val
+- websrc_test

lmms_eval/tasks/websrc/websrc_test.yaml

@@ -0,0 +1,19 @@
+dataset_path: rootsautomation/websrc-test
+task: "websrc_test"
+test_split: test
+output_type: generate_until
+doc_to_visual: !function utils.websrc_doc_to_visual
+doc_to_text: !function utils.websrc_doc_to_text
+doc_to_target: "answer"
+# The return value of process_results will be used by metrics
+process_results: !function utils.websrc_process_results
+# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
+generation_kwargs:
+  max_new_tokens: 16
+  image_aspect_ratio: pad
+metric_list:
+  - metric: submission
+    aggregation: !function utils.websrc_test_aggregate_results_for_submission
+    higher_is_better: true
+metadata:
+  - version: 0.0

lmms_eval/tasks/websrc/websrc_val.yaml

@@ -0,0 +1,19 @@
+dataset_path: rootsautomation/websrc
+task: "websrc_val"
+test_split: dev
+output_type: generate_until
+doc_to_visual: !function utils.websrc_doc_to_visual
+doc_to_text: !function utils.websrc_doc_to_text
+doc_to_target: "answer"
+# The return value of process_results will be used by metrics
+process_results: !function utils.websrc_process_results
+# Note that the metric name can be either a registed metric function (such as the case for GQA) or a key name returned by process_results
+generation_kwargs:
+  max_new_tokens: 16
+  image_aspect_ratio: pad
+metric_list:
+  - metric: websrc_squad_f1
+    aggregation: !function utils.websrc_aggregate_results
+    higher_is_better: true
+metadata:
+  - version: 0.0