added mantis_eval

Author: Oscar Lo

open_flamingo/eval/eval_datasets.py

@@ -15,6 +15,7 @@
     "vizwiz",
     "textvqa",
     "gqa",
+    "mantiseval",
     "hateful_memes",
     "imagenet",
 ]
@@ -107,14 +108,26 @@ def get_img_path(self, question):
             return os.path.join(self.image_dir_path, question["image_id"])
         elif self.dataset_name == "textvqa" or self.dataset_name == "gqa":
             return os.path.join(self.image_dir_path, f"{question['image_id']}.jpg")
+        elif self.dataset_name == "mantiseval":
+            img_paths = []
+            for img_id in question['image_id']:
+                img_paths.append(os.path.join(self.image_dir_path, f"{img_id}.jpg"))
+            return img_paths
         else:
             raise Exception(f"Unknown VQA dataset {self.dataset_name}")
 
     def __getitem__(self, idx):
         question = self.questions[idx]
         img_path = self.get_img_path(question)
-        image = Image.open(img_path)
-        image.load()
+        if self.dataset_name == "mantiseval":
+            image = []
+            for path in img_path:
+                img = Image.open(path)
+                img.load()
+                image.append(img)
+        else:
+            image = Image.open(img_path)
+            image.load()
         results = {
             "image": image,
             "question": question["question"],

open_flamingo/eval/eval_models/blip.py

@@ -5,7 +5,7 @@
 
 from transformers import Blip2Processor, Blip2ForConditionalGeneration
 from eval_models.eval_model import BaseEvalModel
-from utils import unwrap_model
+from utils import unwrap_model, combine_images
 from transformers.modeling_outputs import CausalLMOutputWithPast
 
 
@@ -27,9 +27,14 @@ def required_args(self):
 
     def prepare_images(self, batch: List[List[Image.Image]]) -> torch.Tensor:
         batch_images = None
+        for i in range(len(batch)):
+            if len(batch[i]) > 1:
+                batch[i] = combine_images(batch[i])
+        """
         assert all(
             len(example) == 1 for example in batch
         ), "BLIP-2 only supports one image per example"
+        """
         for example in batch:
             if batch_images is None:
                 batch_images = self.processor.image_processor(
@@ -111,6 +116,9 @@ def get_textvqa_prompt(self, question, answer=None) -> str:
 
     def get_gqa_prompt(self, question, answer=None) -> str:
         return f"Question:{question} Short answer:{answer if answer is not None else ''}"
+    
+    def get_mantiseval_prompt(self, question, answer=None) -> str:
+        return f"Question:{question} Short answer:{answer if answer is not None else ''}"
 
     def get_coco_prompt(self, caption=None) -> str:
         return f"A photo of {caption if caption is not None else ''}"

open_flamingo/eval/eval_models/open_flamingo.py

@@ -291,6 +291,9 @@ def get_textvqa_prompt(self, question, answer=None) -> str:
     def get_gqa_prompt(self, question, answer=None) -> str:
         return f"<image>Question:{question} Short answer:{answer if answer is not None else ''}{'<|endofchunk|>' if answer is not None else ''}"
 
+    def get_mantiseval_prompt(self, question, answer=None) -> str:
+        return f"<image>Question:{question} Short answer:{answer if answer is not None else ''}{'<|endofchunk|>' if answer is not None else ''}"
+
     def get_coco_prompt(self, caption=None) -> str:
         return f"<image>Output:{caption if caption is not None else ''}{'<|endofchunk|>' if caption is not None else ''}"
 

open_flamingo/eval/evaluate.py

@@ -34,7 +34,7 @@
     HatefulMemesDataset,
 )
 from ok_vqa_utils import postprocess_ok_vqa_generation
-from vqa_metric import compute_vqa_accuracy, postprocess_vqa_generation
+from vqa_metric import compute_vqa_accuracy, postprocess_vqa_generation, compute_mantis_accuracy
 
 parser = argparse.ArgumentParser()
 parser.add_argument(
@@ -152,29 +152,30 @@
     default=None,
 )
 
-## VQAV2, OK-VQA, VizWiz, TextVQA, GQA Datasets
-for task in ['vqav2', 'okvqa', 'vizwiz', 'textvqa', 'gqa']:
+## VQAV2, OK-VQA, VizWiz, TextVQA, GQA, Mantis-Eval Datasets
+for task in ['vqav2', 'okvqa', 'vizwiz', 'textvqa', 'gqa', 'mantiseval']:
     parser.add_argument(
-        f"--{task}_image_dir_path" if task=='gqa' or task=='textvqa' else f"--{task}_train_image_dir_path",
+        f"--{task}_image_dir_path" if task=='gqa' or task=='textvqa' or task=='mantiseval' else f"--{task}_train_image_dir_path",
         type=str,
         default=None,
     )
-    if task!='gqa' and task!='textvqa':
+    if task != 'mantiseval':
+        if task!='gqa' and task!='textvqa':
+            parser.add_argument(
+                f"--{task}_test_image_dir_path",
+                type=str,
+                default=None,
+            )
         parser.add_argument(
-            f"--{task}_test_image_dir_path",
+            f"--{task}_train_questions_json_path",
+            type=str,
+            default=None,
+        )
+        parser.add_argument(
+            f"--{task}_train_annotations_json_path",
             type=str,
             default=None,
         )
-    parser.add_argument(
-        f"--{task}_train_questions_json_path",
-        type=str,
-        default=None,
-    )
-    parser.add_argument(
-        f"--{task}_train_annotations_json_path",
-        type=str,
-        default=None,
-    )
     parser.add_argument(
         f"--{task}_test_questions_json_path",
         type=str,
@@ -315,7 +316,7 @@ def main():
                         }
                     )
 
-    for vqa_task in ["okvqa", "vqav2", "vizwiz", "textvqa", "gqa"]:
+    for vqa_task in ["okvqa", "vqav2", "vizwiz", "textvqa", "gqa", "mantiseval"]:
         if var_args[f"eval_{vqa_task}"]:
             print(f"Evaluating on {vqa_task}...")
 
@@ -601,16 +602,16 @@ def evaluate_vqa(
         float: accuracy score
     """
     var_args = vars(args)
-    for task in ["okvqa", "vqav2", "vizwiz", "textvqa", "gqa"]:
+    for task in ["okvqa", "vqav2", "vizwiz", "textvqa", "gqa", "mantiseval"]:
         if dataset_name == task:
             task = task
-            train_image_dir_path = var_args[f"{task}_train_image_dir_path" if task!="textvqa" and task!="gqa" else f"{task}_image_dir_path"]
-            train_questions_json_path = var_args[f"{task}_train_questions_json_path"]
-            train_annotations_json_path = var_args[f"{task}_train_annotations_json_path"]
-            test_image_dir_path = var_args[f"{task}_test_image_dir_path" if task!="textvqa" and task!="gqa" else f"{task}_image_dir_path"]
+            train_image_dir_path = var_args[f"{task}_train_image_dir_path" if task!="textvqa" and task!="gqa" and task!="mantiseval" else f"{task}_image_dir_path"]
+            train_questions_json_path = var_args[f"{task}_train_questions_json_path"] if task!="mantiseval" else var_args[f"{task}_test_questions_json_path"]
+            train_annotations_json_path = var_args[f"{task}_train_annotations_json_path"] if task!="mantiseval" else var_args[f"{task}_test_annotations_json_path"]
+            test_image_dir_path = var_args[f"{task}_test_image_dir_path" if task!="textvqa" and task!="gqa" and task!="mantiseval" else f"{task}_image_dir_path"]
             test_questions_json_path = var_args[f"{task}_test_questions_json_path"]
             test_annotations_json_path = var_args[f"{task}_test_annotations_json_path"]
-    if dataset_name not in ["okvqa", "vqav2", "vizwiz", "textvqa", "gqa"]:
+    if dataset_name not in ["okvqa", "vqav2", "vizwiz", "textvqa", "gqa", "mantiseval"]:
         raise ValueError(f"Unsupported dataset: {dataset_name}")
 
     train_dataset = VQADataset(
@@ -675,7 +676,10 @@ def evaluate_vqa(
                 context_images = [x["image"] for x in batch_demo_samples[i]]
             else:
                 context_images = []
-            batch_images.append(context_images + [batch["image"][i]])
+            if dataset_name == "mantiseval":
+                batch_images.append(context_images + batch["image"][i])
+            else:
+                batch_images.append(context_images + [batch["image"][i]])
 
             context_text = "".join(
                 [
@@ -703,7 +707,7 @@ def evaluate_vqa(
             num_beams=num_beams,
             length_penalty=length_penalty,
         )
-
+        
         process_function = (
             postprocess_ok_vqa_generation
             if dataset_name == "okvqa"
@@ -732,11 +736,17 @@ def evaluate_vqa(
         f.write(json.dumps(all_predictions, indent=4))
 
     if test_annotations_json_path is not None:
-        acc = compute_vqa_accuracy(
-            f"{dataset_name}results_{random_uuid}.json",
-            test_questions_json_path,
-            test_annotations_json_path,
-        )
+        if dataset_name == "mantiseval":
+            acc = compute_mantis_accuracy(
+                f"{dataset_name}results_{random_uuid}.json",
+                test_annotations_json_path,
+            )
+        else:
+            acc = compute_vqa_accuracy(
+                f"{dataset_name}results_{random_uuid}.json",
+                test_questions_json_path,
+                test_annotations_json_path,
+            )
         # delete the temporary file
         os.remove(f"{dataset_name}results_{random_uuid}.json")
 

open_flamingo/eval/utils.py

@@ -3,6 +3,7 @@
 import random
 import torch.nn as nn
 from contextlib import suppress
+from PIL import Image
 
 
 def random_seed(seed=42, rank=0):
@@ -122,3 +123,25 @@ def get_autocast(precision):
         return lambda: torch.cuda.amp.autocast(dtype=torch.bfloat16)
     else:
         return suppress
+
+def combine_images(images):
+    img_heights, _ = zip(*(img.size for img in images))
+    avg_height = sum(img_heights) // len(img_heights)
+    for i, img in enumerate(images):
+        images[i] = img.resize((int(img.size[0] * avg_height / img.size[1]), avg_height))
+    resized_heights, resized_widths = zip(*(img.size for img in images))
+    total_width = sum(resized_widths)
+    max_height = max(resized_heights)
+    new_img = Image.new("RGB", (total_width + 10 * (len(images) - 1), max_height))
+    x_offset = 0
+    for i, img in enumerate(images):
+        if i > 0:
+            new_img.paste(Image.new("RGB", (1, max_height), (0, 0, 0)), (x_offset, 0))
+            x_offset += 1
+            new_img.paste(Image.new("RGB", (8, max_height), (255, 255, 255)), (x_offset, 0))
+            x_offset += 8
+            new_img.paste(Image.new("RGB", (1, max_height), (0, 0, 0)), (x_offset, 0))
+            x_offset += 1
+        new_img.paste(img, (x_offset, 0))
+        x_offset += img.size[0]
+    return new_img

open_flamingo/eval/vqa_metric.py

@@ -553,6 +553,26 @@ def compute_vqa_accuracy(result_json_path, question_json_path, annotation_json_p
 
     return vqaEval.accuracy["overall"]
 
+def compute_mantis_accuracy(result_json_path, annotation_json_path):
+    dataset = json.load(open(annotation_json_path, "r"))
+    gt_ans = {}
+    for ann in dataset["annotations"]:
+        gt_ans[ann["question_id"]] = {"answer": ann["answers"][0]["answer"], "type": ann["question_type"]}
+    results = json.load(open(result_json_path, "r"))
+    assert type(results) == list, "results is not an array of objects"
+    correct = 0
+    for res in results:
+        res_ans = res["answer"].lower().strip('()\n ')
+        if gt_ans[res["question_id"]]["type"] == "multi-choice":
+            if len(res_ans) > 1:
+                for c in res_ans:
+                    if c.isalpha():
+                        res_ans = c
+                        break
+        if res_ans == gt_ans[res["question_id"]]["answer"].lower().strip('()\n '):
+            correct+=1
+    acc = correct / len(results)
+    return acc
 
 def postprocess_vqa_generation(predictions):
     answer = re.split("Question|Answer|Short", predictions, 1)[0]