diff --git a/dataset/README_MINIGPTv2_FINETUNE.md b/dataset/README_MINIGPTv2_FINETUNE.md index 5a539f4..b29138d 100644 --- a/dataset/README_MINIGPTv2_FINETUNE.md +++ b/dataset/README_MINIGPTv2_FINETUNE.md @@ -15,7 +15,7 @@ RefCOCOg | annotations AOK-VQA | annotations OCR-VQA | annotations -GQA | images    annotations +GQA | images    annotations Filtered flickr-30k | annotations Multi-task conversation | annotations Filtered unnatural instruction | annotations diff --git a/eval_configs/minigptv2_benchmark_evaluation.yaml b/eval_configs/minigptv2_benchmark_evaluation.yaml new file mode 100644 index 0000000..c0e3a26 --- /dev/null +++ b/eval_configs/minigptv2_benchmark_evaluation.yaml @@ -0,0 +1,79 @@ +model: + arch: minigpt_v2 + model_type: pretrain + max_txt_len: 500 + end_sym: "" + low_resource: False + prompt_template: '[INST] {} [/INST]' + llama_model: "" + ckpt: "" + lora_r: 64 + lora_alpha: 16 + + +datasets: + cc_sbu_align: + vis_processor: + train: + name: "blip2_image_eval" + image_size: 448 + text_processor: + train: + name: "blip_caption" + +evaluation_datasets: + refcoco: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + refcocog: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + refcoco+: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + gqa: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + okvqa: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + vizwiz: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + iconvqa: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + vsr: + eval_file_path: cambridgeltl/vsr_zeroshot + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 10 + hm: + eval_file_path: /path/to/eval/annotation/path + img_path: /path/to/eval/image/path + max_new_tokens: 20 + batch_size: 100 + +run: + task: image_text_pretrain + name: minigptv2_evaluation + save_path: /path/to/save/folder_path + + + + + diff --git a/eval_configs/minigptv2_eval.yaml b/eval_configs/minigptv2_eval.yaml index 0479f2a..46ac2fa 100644 --- a/eval_configs/minigptv2_eval.yaml +++ b/eval_configs/minigptv2_eval.yaml @@ -1,11 +1,11 @@ model: arch: minigpt_v2 model_type: pretrain - max_txt_len: 160 + max_txt_len: 500 end_sym: "" low_resource: True prompt_template: '[INST] {} [/INST]' - ckpt: 'please set this value to the path of pretrained checkpoint' + ckpt: "please set this value to the path of pretrained checkpoint" lora_r: 64 lora_alpha: 16 diff --git a/eval_scripts/EVAL_README.md b/eval_scripts/EVAL_README.md new file mode 100644 index 0000000..d12e436 --- /dev/null +++ b/eval_scripts/EVAL_README.md @@ -0,0 +1,104 @@ +## Evaluation Instruction for MiniGPT-v2 + +### Data preparation +Images download +Image source | Download path +--- | :---: +OKVQA| annotations    images +gqa | annotations    images +hateful meme | images and annotations +iconqa | images and annotation +vizwiz | images and annotation +RefCOCO | annotations +RefCOCO+ | annotations +RefCOCOg | annotations + +### Evaluation dataset structure + +``` +${MINIGPTv2_EVALUATION_DATASET} +├── gqa +│ └── test_balanced_questions.json +│ ├── testdev_balanced_questions.json +│ ├── gqa_images +├── hateful_meme +│ └── hm_images +│ ├── dev.jsonl +├── iconvqa +│ └── iconvqa_images +│ ├── choose_text_val.json +├── vizwiz +│ └── vizwiz_images +│ ├── val.json +├── vsr +│ └── vsr_images +├── okvqa +│ ├── okvqa_test_split.json +│ ├── mscoco_val2014_annotations_clean.json +│ ├── OpenEnded_mscoco_val2014_questions_clean.json +├── refcoco +│ └── instances.json +│ ├── refs(google).p +│ ├── refs(unc).p +├── refcoco+ +│ └── instances.json +│ ├── refs(unc).p +├── refercocog +│ └── instances.json +│ ├── refs(google).p +│ ├── refs(und).p +... +``` + + +### environment setup + +``` +export PYTHONPATH=$PYTHONPATH:/path/to/directory/of/MiniGPT-4 +``` + +### config file setup + +Set **llama_model** to the path of LLaMA model. +Set **ckpt** to the path of our pretrained model. +Set **eval_file_path** to the path of the annotation files for each evaluation data. +Set **img_path** to the img_path for each evaluation dataset. +Set **save_path** to the save_path for evch evaluation dataset. + +in [minigpt4/eval_configs/minigptv2_benchmark_evaluation.yaml](../minigpt4/eval_configs/minigptv2_benchmark_evaluation.yaml) + + + + +### start evalauting RefCOCO, RefCOCO+, RefCOCOg +port=port_number +cfg_path=/path/to/eval_configs/minigptv2_benchmark_evaluation.yaml + +dataset names: +| refcoco | refcoco+ | refcocog | +| ------- | -------- | -------- | + +``` +torchrun --master-port ${port} --nproc_per_node 1 eval_ref.py \ + --cfg-path ${cfg_path} --dataset refcoco,refcoco+,refcocog --resample +``` + + +### start evaluating visual question answering + +port=port_number +cfg_path=/path/to/eval_configs/minigptv2_benchmark_evaluation.yaml + +dataset names: +| okvqa | vizwiz | iconvqa | gqa | vsr | hm | +| ------- | -------- | -------- |-------- | -------- | -------- | + + +``` +torchrun --master-port ${port} --nproc_per_node 1 eval_vqa.py \ + --cfg-path ${cfg_path} --dataset okvqa,vizwiz,iconvqa,gqa,vsr,hm +``` + + + + diff --git a/eval_scripts/eval.sh b/eval_scripts/eval.sh new file mode 100644 index 0000000..3bebd23 --- /dev/null +++ b/eval_scripts/eval.sh @@ -0,0 +1,25 @@ +#!/bin/bash --login + +export PYTHONPATH=$PYTHONPATH:$(pwd) + +cfg_path=eval_configs/minigpt4_llama2_eval.yaml +CKPT=YOUR_CKPT_PATH +NAME=EXP_NAME +IMG_PATH=YOUR_IMG_PATH +EVAL_FILE_PATH=YOUR_EVAL_FILE_PATH + +# torchrun --nproc_per_node 1 eval_scripts/eval_ref.py --name ${NAME} \ +# --cfg-path ${cfg_path} \ +# --ckpt ${CKPT} --dataset refcoco,refcoco+,refcocog --lora_r 64 --lora_alpha 16 \ +# --batch_size 64 --max_new_tokens 20 --resample --img_path ${IMG_PATH} --eval_file_path ${EVAL_FILE_PATH} + +# torchrun --nproc_per_node 1 eval_scripts/eval_vqa.py --name ${NAME} \ +# --cfg-path ${cfg_path} \ +# --ckpt ${CKPT} --split val,test --dataset okvqa,vizwiz,iconqa,gqa,vsr,hm --lora_r 64 --lora_alpha 16 \ +# --batch_size 32 --max_new_tokens 20 --resample + + +torchrun --nproc_per_node 1 eval_scripts/eval_vqa.py --name ${NAME} \ + --cfg-path ${cfg_path} \ + --ckpt ${CKPT} --split val,test --dataset okvqa --lora_r 64 --lora_alpha 16 \ + --batch_size 32 --max_new_tokens 20 \ No newline at end of file diff --git a/eval_scripts/eval_ref.py b/eval_scripts/eval_ref.py new file mode 100644 index 0000000..28d55a2 --- /dev/null +++ b/eval_scripts/eval_ref.py @@ -0,0 +1,128 @@ +import os +import re +import json +import argparse +from collections import defaultdict +import random +import numpy as np +from PIL import Image +from tqdm import tqdm +import torch +from torch.utils.data import DataLoader +from minigpt4.common.config import Config +from minigpt4.common.eval_utils import prepare_texts, init_model, eval_parser, computeIoU +from minigpt4.conversation.conversation import CONV_VISION_minigptv2 + +from minigpt4.datasets.datasets.coco_caption import RefCOCOEvalData + +def list_of_str(arg): + return list(map(str, arg.split(','))) + +parser = eval_parser() +parser.add_argument("--dataset", type=list_of_str, default='refcoco', help="dataset to evaluate") +parser.add_argument("--res", type=float, default=100.0, help="resolution used in refcoco") +parser.add_argument("--resample", action='store_true', help="resolution used in refcoco") +args = parser.parse_args() + +cfg = Config(args) + +eval_dict = {'refcoco': ['val','testA','testB'], + 'refcoco+': ['val','testA','testB'], + 'refcocog': ['val','test']} + + +model, vis_processor = init_model(args) +model.eval() +CONV_VISION = CONV_VISION_minigptv2 +conv_temp = CONV_VISION.copy() +conv_temp.system = "" + +# +model.eval() +save_path = cfg.run_cfg.save_path + + + +for dataset in args.dataset: + for split in eval_dict[dataset]: + + eval_file_path = cfg.evaluation_datasets_cfg[dataset]["eval_file_path"] + img_path = cfg.evaluation_datasets_cfg[dataset]["img_path"] + batch_size = cfg.evaluation_datasets_cfg[dataset]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg[dataset]["max_new_tokens"] + + with open(os.path.join(eval_file_path,f"{dataset}/{dataset}_{split}.json"), 'r') as f: + refcoco = json.load(f) + + data = RefCOCOEvalData(refcoco, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + minigpt4_predict = defaultdict(list) + resamples = [] + + for images, questions, img_ids in tqdm(eval_dataloader): + texts = prepare_texts(questions, conv_temp) # warp the texts with conversation template + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False) + for answer, img_id, question in zip(answers, img_ids, questions): + answer = answer.replace("","").replace(" ","").strip() + pattern = r'\{<\d{1,3}><\d{1,3}><\d{1,3}><\d{1,3}>\}' + if re.match(pattern, answer): + minigpt4_predict[img_id].append(answer) + else: + resamples.append({'img_id': img_id, 'sents': [question.replace('[refer] give me the location of','').strip()]}) + if args.resample: + for i in range(20): + data = RefCOCOEvalData(resamples, vis_processor, img_path) + resamples = [] + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + for images, questions, img_ids in tqdm(eval_dataloader): + texts = prepare_texts(questions, conv_temp) # warp the texts with conversation template + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False) + for answer, img_id, question in zip(answers, img_ids, questions): + answer = answer.replace("","").replace(" ","").strip() + pattern = r'\{<\d{1,3}><\d{1,3}><\d{1,3}><\d{1,3}>\}' + if re.match(pattern, answer) or i == 4: + minigpt4_predict[img_id].append(answer) + else: + resamples.append({'img_id': img_id, 'sents': [question.replace('[refer] give me the location of','').strip()]}) + + if len(resamples) == 0: + break + + file_save_path = os.path.join(save_path,f"{args.dataset}_{split}.json") + with open(file_save_path,'w') as f: + json.dump(minigpt4_predict, f) + + count=0 + total=len(refcoco) + res=args.res + refcoco_dict = defaultdict() + for item in refcoco: + refcoco_dict[item['img_id']] = item + for img_id in refcoco_dict: + item = refcoco_dict[img_id] + bbox = item['bbox'] + outputs = minigpt4_predict[img_id] + for output in outputs: + try: + integers = re.findall(r'\d+', output) + pred_bbox = [int(num) for num in integers] + height = item['height'] + width = item['width'] + pred_bbox[0] = pred_bbox[0] / res * width + pred_bbox[1] = pred_bbox[1] / res * height + pred_bbox[2] = pred_bbox[2] / res * width + pred_bbox[3] = pred_bbox[3] / res * height + + gt_bbox = [0,0,0,0] + gt_bbox[0] = bbox[0] + gt_bbox[1] = bbox[1] + gt_bbox[2] = bbox[0] + bbox[2] + gt_bbox[3] = bbox[1] + bbox[3] + + iou_score = computeIoU(pred_bbox, gt_bbox) + if iou_score > 0.5: + count+=1 + except: + continue + + print(f'{dataset} {split}:', count / total * 100, flush=True) diff --git a/eval_scripts/eval_vqa.py b/eval_scripts/eval_vqa.py new file mode 100644 index 0000000..75f0d91 --- /dev/null +++ b/eval_scripts/eval_vqa.py @@ -0,0 +1,252 @@ +import os +import re +import json +import argparse +from collections import defaultdict + +import numpy as np +from PIL import Image +from tqdm import tqdm +import torch +from torch.utils.data import DataLoader +from datasets import load_dataset + + +from minigpt4.datasets.datasets.vqa_datasets import OKVQAEvalData,VizWizEvalData,IconQAEvalData,GQAEvalData,VSREvalData,HMEvalData +from minigpt4.common.vqa_tools.VQA.PythonHelperTools.vqaTools.vqa import VQA +from minigpt4.common.vqa_tools.VQA.PythonEvaluationTools.vqaEvaluation.vqaEval import VQAEval + +from minigpt4.common.eval_utils import prepare_texts, init_model, eval_parser +from minigpt4.conversation.conversation import CONV_VISION_minigptv2 +from minigpt4.common.config import Config + + +def list_of_str(arg): + return list(map(str, arg.split(','))) + +parser = eval_parser() +parser.add_argument("--dataset", type=list_of_str, default='refcoco', help="dataset to evaluate") +args = parser.parse_args() +cfg = Config(args) + + + +model, vis_processor = init_model(args) +conv_temp = CONV_VISION_minigptv2.copy() +conv_temp.system = "" +model.eval() +save_path = cfg.run_cfg.save_path + + +if 'okvqa' in args.dataset: + + eval_file_path = cfg.evaluation_datasets_cfg["okvqa"]["eval_file_path"] + img_path = cfg.evaluation_datasets_cfg["okvqa"]["img_path"] + batch_size = cfg.evaluation_datasets_cfg["okvqa"]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg["okvqa"]["max_new_tokens"] + + + evaluation_annntation_path = os.path.join(eval_file_path, "okvqa_test_split.json") + with open(evaluation_annntation_path) as f: + ok_vqa_test_split = json.load(f) + + data = OKVQAEvalData(ok_vqa_test_split, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + minigpt4_predict = [] + + for images, questions, question_ids, img_ids in eval_dataloader: + texts = prepare_texts(questions, conv_temp) # warp the texts with conversation template + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False) + + for answer, question_id, question, img_id in zip(answers, question_ids, questions, img_ids): + result = dict() + answer = answer.lower().replace('','').strip() + result['answer'] = answer + result['question_id'] = int(question_id) + minigpt4_predict.append(result) + + file_save_path= os.path.join(save_path,"okvqa.json") + with open(file_save_path,'w') as f: + json.dump(minigpt4_predict, f) + + annFile = os.path.join(eval_file_path,"mscoco_val2014_annotations_clean.json") + quesFile = os.path.join(eval_file_path,"OpenEnded_mscoco_val2014_questions_clean.json" ) + + vqa = VQA(annFile, quesFile) + vqaRes = vqa.loadRes(file_save_path, quesFile) + + vqaEval = VQAEval(vqa, vqaRes, n=2) + vqaEval.evaluate() + print ("Overall OKVQA Accuracy is: %.02f\n" %(vqaEval.accuracy['overall']), flush=True) + +if 'vizwiz' in args.dataset: + + eval_file_path = cfg.evaluation_datasets_cfg["vizwiz"]["eval_file_path"] + img_path = cfg.evaluation_datasets_cfg["vizwiz"]["img_path"] + batch_size = cfg.evaluation_datasets_cfg["vizwiz"]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg["vizwiz"]["max_new_tokens"] + + vizwiz = json.load(open(eval_file_path, 'r')) + + data = VizWizEvalData(vizwiz, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + minigpt4_predict = [] + total_acc = [] + for images, texts, gt_answers in tqdm(eval_dataloader): + texts = prepare_texts(texts, conv_temp) # warp the texts with conversation template + with torch.no_grad(): + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False,repetition_penalty=1.0) + + for answer, gt_answer in zip(answers, gt_answers): + result = dict() + result['answer'] = answer.replace('','').strip() + minigpt4_predict.append(result) + count=0 + gt_answer = gt_answer.split('_') + for gt in gt_answer: + if gt.lower() == answer.lower(): + count += 1 + acc = min(count/3.0, 1.0) + total_acc.append(acc) + + file_save_path = os.path.join(save_path, "vizwiz.json") + with open(file_save_path,'w') as f: + json.dump(minigpt4_predict, f) + print('vizwiz Acc: ', np.average(total_acc)* 100.0, flush=True) + + +if 'iconvqa' in args.dataset: + + eval_file_path = cfg.evaluation_datasets_cfg["iconvqa"]["eval_file_path"] + img_path = cfg.evaluation_datasets_cfg["iconvqa"]["img_path"] + batch_size = cfg.evaluation_datasets_cfg["iconvqa"]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg["iconvqa"]["max_new_tokens"] + + iconqa_text_val = json.load(open(eval_file_path,"r")) + + data = IconQAEvalData(iconqa_text_val, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + + count = 0 + for images, texts, candidates, answers in tqdm(eval_dataloader): + candidates = [candidate.split('_') for candidate in candidates] + num_cand = [len(candidate) for candidate in candidates] + for candidate in candidates: + candidate.extend(['none'] * (max(num_cand) - len(candidate))) + candidates = [list(x) for x in zip(*candidates)] + instructions = ["[INST] {} [/INST]".format(text) for text in texts] + answer_ranks = model.multi_select(images, instructions, candidates, num_cand=num_cand) + for idx, answer in enumerate(answers): + if answer_ranks[idx][0] == answer: + count += 1 + + print('iconqa Acc: ', count / len(iconqa_text_val) * 100.0, flush=True) + + +if 'gqa' in args.dataset: + + eval_file_path = cfg.evaluation_datasets_cfg["gqa"]["eval_file_path"] + img_path = cfg.evaluation_datasets_cfg["gqa"]["img_path"] + batch_size = cfg.evaluation_datasets_cfg["gqa"]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg["gqa"]["max_new_tokens"] + + gqa = json.load(open(eval_file_path)) + data = GQAEvalData(gqa, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + count=0 + total=0 + minigpt4_predict = [] + for images, texts, labels in tqdm(eval_dataloader): + texts = prepare_texts(texts, conv_temp) # warp the texts with conversation template + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False) + + for answer, label in zip(answers, labels): + result = dict() + result['pred'] = answer.lower().replace('','').strip() + result['gt'] = label + minigpt4_predict.append(result) + if answer.lower() == label: + count+=1 + total+=1 + print('gqa val:', count / total * 100, flush=True) + + file_save_path = os.path.join(save_path, "gqa.json") + with open(file_save_path,'w') as f: + json.dump(minigpt4_predict, f) + +if 'vsr' in args.dataset: + + img_path = cfg.evaluation_datasets_cfg["vsr"]["img_path"] + batch_size = cfg.evaluation_datasets_cfg["vsr"]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg["vsr"]["max_new_tokens"] + + annotation = load_dataset("cambridgeltl/vsr_zeroshot", split='test') + data = VSREvalData(annotation, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + count=0 + total=0 + + minigpt4_predict = [] + + for images, texts, labels in tqdm(eval_dataloader): + texts = prepare_texts(texts, conv_temp) # warp the texts with conversation template + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False) + + for answer, label in zip(answers, labels): + result = dict() + result['pred'] = answer.replace('','').strip() + result['gt'] = label + minigpt4_predict.append(result) + if answer.lower() == label.lower(): + count+=1 + total+=1 + print('vsr test:', count / total * 100, flush=True) + file_save_path = os.path.join(save_path,"vsr.json") + with open(file_save_path,'w') as f: + json.dump(minigpt4_predict, f) + +if 'hm' in args.dataset: + + eval_file_path = cfg.evaluation_datasets_cfg["hm"]["eval_file_path"] + img_path = cfg.evaluation_datasets_cfg["hm"]["img_path"] + batch_size = cfg.evaluation_datasets_cfg["hm"]["batch_size"] + max_new_tokens = cfg.evaluation_datasets_cfg["hm"]["max_new_tokens"] + + annotation = [] + with open(eval_file_path, 'r') as jsonl_file: + for line in jsonl_file: + json_obj = json.loads(line) + annotation.append(json_obj) + + data = HMEvalData(annotation, vis_processor, img_path) + eval_dataloader = DataLoader(data, batch_size=batch_size, shuffle=False) + count=0 + total=0 + + minigpt4_predict = [] + + for images, texts, labels in tqdm(eval_dataloader): + texts = prepare_texts(texts, conv_temp) # warp the texts with conversation template + + answers = model.generate(images, texts, max_new_tokens=max_new_tokens, do_sample=False) + + for answer, label in zip(answers, labels): + result = dict() + if answer.lower().strip() =="yes": + answer=1 + elif answer.lower().strip()=="no": + answer=0 + else: + print("non-matching answer",answer) + + result['pred'] = answer + result['gt'] = int(label) + minigpt4_predict.append(result) + if answer == label: + count+=1 + total+=1 + + print('hm val:', count / total * 100, flush=True) + file_save_path = os.path.join(save_path, "hm.json") + with open(file_save_path,'w') as f: + json.dump(minigpt4_predict, f) diff --git a/minigpt4/common/config.py b/minigpt4/common/config.py index e184b1f..5fe7239 100644 --- a/minigpt4/common/config.py +++ b/minigpt4/common/config.py @@ -29,6 +29,7 @@ class Config: runner_config = self.build_runner_config(config) model_config = self.build_model_config(config, **user_config) dataset_config = self.build_dataset_config(config) + evaluation_dataset_config = self.build_evaluation_dataset_config(config) # Validate the user-provided runner configuration # model and dataset configuration are supposed to be validated by the respective classes @@ -37,7 +38,7 @@ class Config: # Override the default configuration with user options. self.config = OmegaConf.merge( - runner_config, model_config, dataset_config, user_config + runner_config, model_config, dataset_config,evaluation_dataset_config, user_config ) def _validate_runner_config(self, runner_config): @@ -111,6 +112,28 @@ class Config: return dataset_config + + @staticmethod + def build_evaluation_dataset_config(config): + datasets = config.get("evaluation_datasets", None) + if datasets is None: + raise KeyError( + "Expecting 'datasets' as the root key for dataset configuration." + ) + + dataset_config = OmegaConf.create() + + for dataset_name in datasets: + builder_cls = registry.get_builder_class(dataset_name) + + # hierarchy override, customized config > default config + dataset_config = OmegaConf.merge( + dataset_config, + {"evaluation_datasets": {dataset_name: config["evaluation_datasets"][dataset_name]}}, + ) + + return dataset_config + def _convert_to_dot_list(self, opts): if opts is None: opts = [] @@ -136,6 +159,10 @@ class Config: def datasets_cfg(self): return self.config.datasets + @property + def evaluation_datasets_cfg(self): + return self.config.evaluation_datasets + @property def model_cfg(self): return self.config.model diff --git a/minigpt4/common/eval_utils.py b/minigpt4/common/eval_utils.py new file mode 100644 index 0000000..3087d2a --- /dev/null +++ b/minigpt4/common/eval_utils.py @@ -0,0 +1,76 @@ +import argparse +import numpy as np +from nltk.translate.bleu_score import sentence_bleu + +from minigpt4.common.registry import registry +from minigpt4.common.config import Config + +# imports modules for registration +from minigpt4.datasets.builders import * +from minigpt4.models import * +from minigpt4.processors import * +from minigpt4.runners import * +from minigpt4.tasks import * + + + +def eval_parser(): + parser = argparse.ArgumentParser(description="Demo") + parser.add_argument("--cfg-path", required=True, help="path to configuration file.") + parser.add_argument("--name", type=str, default='A2', help="evaluation name") + parser.add_argument("--ckpt", type=str, help="path to configuration file.") + parser.add_argument("--eval_opt", type=str, default='all', help="path to configuration file.") + parser.add_argument("--max_new_tokens", type=int, default=10, help="max number of generated tokens") + parser.add_argument("--batch_size", type=int, default=32) + parser.add_argument("--lora_r", type=int, default=64, help="lora rank of the model") + parser.add_argument("--lora_alpha", type=int, default=16, help="lora alpha") + parser.add_argument( + "--options", + nargs="+", + help="override some settings in the used config, the key-value pair " + "in xxx=yyy format will be merged into config file (deprecate), " + "change to --cfg-options instead.", + ) + return parser + + +def prepare_texts(texts, conv_temp): + convs = [conv_temp.copy() for _ in range(len(texts))] + [conv.append_message( + conv.roles[0], ' {}'.format(text)) for conv, text in zip(convs, texts)] + [conv.append_message(conv.roles[1], None) for conv in convs] + texts = [conv.get_prompt() for conv in convs] + return texts + + +def init_model(args): + print('Initialization Model') + cfg = Config(args) + # cfg.model_cfg.ckpt = args.ckpt + # cfg.model_cfg.lora_r = args.lora_r + # cfg.model_cfg.lora_alpha = args.lora_alpha + + model_config = cfg.model_cfg + model_cls = registry.get_model_class(model_config.arch) + model = model_cls.from_config(model_config).to('cuda:0') + +# import pudb; pudb.set_trace() + key = list(cfg.datasets_cfg.keys())[0] + vis_processor_cfg = cfg.datasets_cfg.get(key).vis_processor.train + vis_processor = registry.get_processor_class(vis_processor_cfg.name).from_config(vis_processor_cfg) + print('Initialization Finished') + return model, vis_processor + +def computeIoU(bbox1, bbox2): + x1, y1, x2, y2 = bbox1 + x3, y3, x4, y4 = bbox2 + intersection_x1 = max(x1, x3) + intersection_y1 = max(y1, y3) + intersection_x2 = min(x2, x4) + intersection_y2 = min(y2, y4) + intersection_area = max(0, intersection_x2 - intersection_x1 + 1) * max(0, intersection_y2 - intersection_y1 + 1) + bbox1_area = (x2 - x1 + 1) * (y2 - y1 + 1) + bbox2_area = (x4 - x3 + 1) * (y4 - y3 + 1) + union_area = bbox1_area + bbox2_area - intersection_area + iou = intersection_area / union_area + return iou diff --git a/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvalDemo.py b/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvalDemo.py new file mode 100644 index 0000000..07ca21d --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvalDemo.py @@ -0,0 +1,89 @@ +# coding: utf-8 + +import sys +dataDir = '../../VQA' +sys.path.insert(0, '%s/PythonHelperTools/vqaTools' %(dataDir)) +from vqa import VQA +from vqaEvaluation.vqaEval import VQAEval +import matplotlib.pyplot as plt +import skimage.io as io +import json +import random +import os + +# set up file names and paths +versionType ='v2_' # this should be '' when using VQA v2.0 dataset +taskType ='OpenEnded' # 'OpenEnded' only for v2.0. 'OpenEnded' or 'MultipleChoice' for v1.0 +dataType ='mscoco' # 'mscoco' only for v1.0. 'mscoco' for real and 'abstract_v002' for abstract for v1.0. +dataSubType ='train2014' +annFile ='%s/Annotations/%s%s_%s_annotations.json'%(dataDir, versionType, dataType, dataSubType) +quesFile ='%s/Questions/%s%s_%s_%s_questions.json'%(dataDir, versionType, taskType, dataType, dataSubType) +imgDir ='%s/Images/%s/%s/' %(dataDir, dataType, dataSubType) +resultType ='fake' +fileTypes = ['results', 'accuracy', 'evalQA', 'evalQuesType', 'evalAnsType'] + +# An example result json file has been provided in './Results' folder. + +[resFile, accuracyFile, evalQAFile, evalQuesTypeFile, evalAnsTypeFile] = ['%s/Results/%s%s_%s_%s_%s_%s.json'%(dataDir, versionType, taskType, dataType, dataSubType, \ +resultType, fileType) for fileType in fileTypes] + +# create vqa object and vqaRes object +vqa = VQA(annFile, quesFile) +vqaRes = vqa.loadRes(resFile, quesFile) + +# create vqaEval object by taking vqa and vqaRes +vqaEval = VQAEval(vqa, vqaRes, n=2) #n is precision of accuracy (number of places after decimal), default is 2 + +# evaluate results +""" +If you have a list of question ids on which you would like to evaluate your results, pass it as a list to below function +By default it uses all the question ids in annotation file +""" +vqaEval.evaluate() + +# print accuracies +print "\n" +print "Overall Accuracy is: %.02f\n" %(vqaEval.accuracy['overall']) +print "Per Question Type Accuracy is the following:" +for quesType in vqaEval.accuracy['perQuestionType']: + print "%s : %.02f" %(quesType, vqaEval.accuracy['perQuestionType'][quesType]) +print "\n" +print "Per Answer Type Accuracy is the following:" +for ansType in vqaEval.accuracy['perAnswerType']: + print "%s : %.02f" %(ansType, vqaEval.accuracy['perAnswerType'][ansType]) +print "\n" +# demo how to use evalQA to retrieve low score result +evals = [quesId for quesId in vqaEval.evalQA if vqaEval.evalQA[quesId]<35] #35 is per question percentage accuracy +if len(evals) > 0: + print 'ground truth answers' + randomEval = random.choice(evals) + randomAnn = vqa.loadQA(randomEval) + vqa.showQA(randomAnn) + + print '\n' + print 'generated answer (accuracy %.02f)'%(vqaEval.evalQA[randomEval]) + ann = vqaRes.loadQA(randomEval)[0] + print "Answer: %s\n" %(ann['answer']) + + imgId = randomAnn[0]['image_id'] + imgFilename = 'COCO_' + dataSubType + '_'+ str(imgId).zfill(12) + '.jpg' + if os.path.isfile(imgDir + imgFilename): + I = io.imread(imgDir + imgFilename) + plt.imshow(I) + plt.axis('off') + plt.show() + +# plot accuracy for various question types +plt.bar(range(len(vqaEval.accuracy['perQuestionType'])), vqaEval.accuracy['perQuestionType'].values(), align='center') +plt.xticks(range(len(vqaEval.accuracy['perQuestionType'])), vqaEval.accuracy['perQuestionType'].keys(), rotation='0',fontsize=10) +plt.title('Per Question Type Accuracy', fontsize=10) +plt.xlabel('Question Types', fontsize=10) +plt.ylabel('Accuracy', fontsize=10) +plt.show() + +# save evaluation results to ./Results folder +json.dump(vqaEval.accuracy, open(accuracyFile, 'w')) +json.dump(vqaEval.evalQA, open(evalQAFile, 'w')) +json.dump(vqaEval.evalQuesType, open(evalQuesTypeFile, 'w')) +json.dump(vqaEval.evalAnsType, open(evalAnsTypeFile, 'w')) + diff --git a/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvaluation/__init__.py b/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvaluation/__init__.py new file mode 100644 index 0000000..148424d --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvaluation/__init__.py @@ -0,0 +1 @@ +author='aagrawal' diff --git a/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvaluation/vqaEval.py b/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvaluation/vqaEval.py new file mode 100644 index 0000000..8a65604 --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/PythonEvaluationTools/vqaEvaluation/vqaEval.py @@ -0,0 +1,192 @@ +# coding=utf-8 + +__author__='aagrawal' + +import re +# This code is based on the code written by Tsung-Yi Lin for MSCOCO Python API available at the following link: +# (https://github.com/tylin/coco-caption/blob/master/pycocoevalcap/eval.py). +import sys + + +class VQAEval: + def __init__(self, vqa, vqaRes, n=2): + self.n = n + self.accuracy = {} + self.evalQA = {} + self.evalQuesType = {} + self.evalAnsType = {} + self.vqa = vqa + self.vqaRes = vqaRes + self.params = {'question_id': vqa.getQuesIds()} + self.contractions = {"aint": "ain't", "arent": "aren't", "cant": "can't", "couldve": "could've", "couldnt": "couldn't", \ + "couldn'tve": "couldn't've", "couldnt've": "couldn't've", "didnt": "didn't", "doesnt": "doesn't", "dont": "don't", "hadnt": "hadn't", \ + "hadnt've": "hadn't've", "hadn'tve": "hadn't've", "hasnt": "hasn't", "havent": "haven't", "hed": "he'd", "hed've": "he'd've", \ + "he'dve": "he'd've", "hes": "he's", "howd": "how'd", "howll": "how'll", "hows": "how's", "Id've": "I'd've", "I'dve": "I'd've", \ + "Im": "I'm", "Ive": "I've", "isnt": "isn't", "itd": "it'd", "itd've": "it'd've", "it'dve": "it'd've", "itll": "it'll", "let's": "let's", \ + "maam": "ma'am", "mightnt": "mightn't", "mightnt've": "mightn't've", "mightn'tve": "mightn't've", "mightve": "might've", \ + "mustnt": "mustn't", "mustve": "must've", "neednt": "needn't", "notve": "not've", "oclock": "o'clock", "oughtnt": "oughtn't", \ + "ow's'at": "'ow's'at", "'ows'at": "'ow's'at", "'ow'sat": "'ow's'at", "shant": "shan't", "shed've": "she'd've", "she'dve": "she'd've", \ + "she's": "she's", "shouldve": "should've", "shouldnt": "shouldn't", "shouldnt've": "shouldn't've", "shouldn'tve": "shouldn't've", \ + "somebody'd": "somebodyd", "somebodyd've": "somebody'd've", "somebody'dve": "somebody'd've", "somebodyll": "somebody'll", \ + "somebodys": "somebody's", "someoned": "someone'd", "someoned've": "someone'd've", "someone'dve": "someone'd've", \ + "someonell": "someone'll", "someones": "someone's", "somethingd": "something'd", "somethingd've": "something'd've", \ + "something'dve": "something'd've", "somethingll": "something'll", "thats": "that's", "thered": "there'd", "thered've": "there'd've", \ + "there'dve": "there'd've", "therere": "there're", "theres": "there's", "theyd": "they'd", "theyd've": "they'd've", \ + "they'dve": "they'd've", "theyll": "they'll", "theyre": "they're", "theyve": "they've", "twas": "'twas", "wasnt": "wasn't", \ + "wed've": "we'd've", "we'dve": "we'd've", "weve": "we've", "werent": "weren't", "whatll": "what'll", "whatre": "what're", \ + "whats": "what's", "whatve": "what've", "whens": "when's", "whered": "where'd", "wheres": "where's", "whereve": "where've", \ + "whod": "who'd", "whod've": "who'd've", "who'dve": "who'd've", "wholl": "who'll", "whos": "who's", "whove": "who've", "whyll": "why'll", \ + "whyre": "why're", "whys": "why's", "wont": "won't", "wouldve": "would've", "wouldnt": "wouldn't", "wouldnt've": "wouldn't've", \ + "wouldn'tve": "wouldn't've", "yall": "y'all", "yall'll": "y'all'll", "y'allll": "y'all'll", "yall'd've": "y'all'd've", \ + "y'alld've": "y'all'd've", "y'all'dve": "y'all'd've", "youd": "you'd", "youd've": "you'd've", "you'dve": "you'd've", \ + "youll": "you'll", "youre": "you're", "youve": "you've"} + self.manualMap = { 'none': '0', + 'zero': '0', + 'one': '1', + 'two': '2', + 'three': '3', + 'four': '4', + 'five': '5', + 'six': '6', + 'seven': '7', + 'eight': '8', + 'nine': '9', + 'ten': '10' + } + self.articles = ['a', + 'an', + 'the' + ] + + + self.periodStrip = re.compile("(?!<=\d)(\.)(?!\d)") + self.commaStrip = re.compile("(\d)(\,)(\d)") + self.punct = [';', r"/", '[', ']', '"', '{', '}', + '(', ')', '=', '+', '\\', '_', '-', + '>', '<', '@', '`', ',', '?', '!'] + + + def evaluate(self, quesIds=None): + if quesIds == None: + quesIds = [quesId for quesId in self.params['question_id']] + gts = {} + res = {} + for quesId in quesIds: + gts[quesId] = self.vqa.qa[quesId] + res[quesId] = self.vqaRes.qa[quesId] + + # ================================================= + # Compute accuracy + # ================================================= + accQA = [] + accQuesType = {} + accAnsType = {} + # print "computing accuracy" + step = 0 + for quesId in quesIds: + for ansDic in gts[quesId]['answers']: + ansDic['answer'] = ansDic['answer'].replace('\n', ' ') + ansDic['answer'] = ansDic['answer'].replace('\t', ' ') + ansDic['answer'] = ansDic['answer'].strip() + resAns = res[quesId]['answer'] + resAns = resAns.replace('\n', ' ') + resAns = resAns.replace('\t', ' ') + resAns = resAns.strip() + gtAcc = [] + gtAnswers = [ans['answer'] for ans in gts[quesId]['answers']] + + if len(set(gtAnswers)) > 1: + for ansDic in gts[quesId]['answers']: + ansDic['answer'] = self.processPunctuation(ansDic['answer']) + ansDic['answer'] = self.processDigitArticle(ansDic['answer']) + resAns = self.processPunctuation(resAns) + resAns = self.processDigitArticle(resAns) + + for gtAnsDatum in gts[quesId]['answers']: + otherGTAns = [item for item in gts[quesId]['answers'] if item!=gtAnsDatum] + matchingAns = [item for item in otherGTAns if item['answer'].lower()==resAns.lower()] + acc = min(1, float(len(matchingAns))/3) + gtAcc.append(acc) + quesType = gts[quesId]['question_type'] + ansType = gts[quesId]['answer_type'] + avgGTAcc = float(sum(gtAcc))/len(gtAcc) + accQA.append(avgGTAcc) + if quesType not in accQuesType: + accQuesType[quesType] = [] + accQuesType[quesType].append(avgGTAcc) + if ansType not in accAnsType: + accAnsType[ansType] = [] + accAnsType[ansType].append(avgGTAcc) + self.setEvalQA(quesId, avgGTAcc) + self.setEvalQuesType(quesId, quesType, avgGTAcc) + self.setEvalAnsType(quesId, ansType, avgGTAcc) + if step%100 == 0: + self.updateProgress(step/float(len(quesIds))) + step = step + 1 + + self.setAccuracy(accQA, accQuesType, accAnsType) + # print "Done computing accuracy" + + def processPunctuation(self, inText): + outText = inText + for p in self.punct: + if (p + ' ' in inText or ' ' + p in inText) or (re.search(self.commaStrip, inText) != None): + outText = outText.replace(p, '') + else: + outText = outText.replace(p, ' ') + outText = self.periodStrip.sub("", + outText, + re.UNICODE) + return outText + + def processDigitArticle(self, inText): + outText = [] + tempText = inText.lower().split() + for word in tempText: + word = self.manualMap.setdefault(word, word) + if word not in self.articles: + outText.append(word) + else: + pass + for wordId, word in enumerate(outText): + if word in self.contractions: + outText[wordId] = self.contractions[word] + outText = ' '.join(outText) + return outText + + def setAccuracy(self, accQA, accQuesType, accAnsType): + self.accuracy['overall'] = round(100*float(sum(accQA))/len(accQA), self.n) + self.accuracy['perQuestionType'] = {quesType: round(100*float(sum(accQuesType[quesType]))/len(accQuesType[quesType]), self.n) for quesType in accQuesType} + self.accuracy['perAnswerType'] = {ansType: round(100*float(sum(accAnsType[ansType]))/len(accAnsType[ansType]), self.n) for ansType in accAnsType} + + def setEvalQA(self, quesId, acc): + self.evalQA[quesId] = round(100*acc, self.n) + + def setEvalQuesType(self, quesId, quesType, acc): + if quesType not in self.evalQuesType: + self.evalQuesType[quesType] = {} + self.evalQuesType[quesType][quesId] = round(100*acc, self.n) + + def setEvalAnsType(self, quesId, ansType, acc): + if ansType not in self.evalAnsType: + self.evalAnsType[ansType] = {} + self.evalAnsType[ansType][quesId] = round(100*acc, self.n) + + def updateProgress(self, progress): + barLength = 20 + status = "" + if isinstance(progress, int): + progress = float(progress) + if not isinstance(progress, float): + progress = 0 + status = "error: progress var must be float\r\n" + if progress < 0: + progress = 0 + status = "Halt...\r\n" + if progress >= 1: + progress = 1 + status = "Done...\r\n" + block = int(round(barLength*progress)) + text = "\rFinshed Percent: [{0}] {1}% {2}".format( "#"*block + "-"*(barLength-block), int(progress*100), status) + sys.stdout.write(text) + sys.stdout.flush() diff --git a/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaDemo.py b/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaDemo.py new file mode 100644 index 0000000..406b596 --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaDemo.py @@ -0,0 +1,73 @@ +# coding: utf-8 + +from vqaTools.vqa import VQA +import random +import skimage.io as io +import matplotlib.pyplot as plt +import os + +dataDir ='../../VQA' +versionType ='v2_' # this should be '' when using VQA v2.0 dataset +taskType ='OpenEnded' # 'OpenEnded' only for v2.0. 'OpenEnded' or 'MultipleChoice' for v1.0 +dataType ='mscoco' # 'mscoco' only for v1.0. 'mscoco' for real and 'abstract_v002' for abstract for v1.0. +dataSubType ='train2014' +annFile ='%s/Annotations/%s%s_%s_annotations.json'%(dataDir, versionType, dataType, dataSubType) +quesFile ='%s/Questions/%s%s_%s_%s_questions.json'%(dataDir, versionType, taskType, dataType, dataSubType) +imgDir = '%s/Images/%s/%s/' %(dataDir, dataType, dataSubType) + +# initialize VQA api for QA annotations +vqa=VQA(annFile, quesFile) + +# load and display QA annotations for given question types +""" +All possible quesTypes for abstract and mscoco has been provided in respective text files in ../QuestionTypes/ folder. +""" +annIds = vqa.getQuesIds(quesTypes='how many'); +anns = vqa.loadQA(annIds) +randomAnn = random.choice(anns) +vqa.showQA([randomAnn]) +imgId = randomAnn['image_id'] +imgFilename = 'COCO_' + dataSubType + '_'+ str(imgId).zfill(12) + '.jpg' +if os.path.isfile(imgDir + imgFilename): + I = io.imread(imgDir + imgFilename) + plt.imshow(I) + plt.axis('off') + plt.show() + +# load and display QA annotations for given answer types +""" +ansTypes can be one of the following +yes/no +number +other +""" +annIds = vqa.getQuesIds(ansTypes='yes/no'); +anns = vqa.loadQA(annIds) +randomAnn = random.choice(anns) +vqa.showQA([randomAnn]) +imgId = randomAnn['image_id'] +imgFilename = 'COCO_' + dataSubType + '_'+ str(imgId).zfill(12) + '.jpg' +if os.path.isfile(imgDir + imgFilename): + I = io.imread(imgDir + imgFilename) + plt.imshow(I) + plt.axis('off') + plt.show() + +# load and display QA annotations for given images +""" +Usage: vqa.getImgIds(quesIds=[], quesTypes=[], ansTypes=[]) +Above method can be used to retrieve imageIds for given question Ids or given question types or given answer types. +""" +ids = vqa.getImgIds() +annIds = vqa.getQuesIds(imgIds=random.sample(ids,5)); +anns = vqa.loadQA(annIds) +randomAnn = random.choice(anns) +vqa.showQA([randomAnn]) +imgId = randomAnn['image_id'] +imgFilename = 'COCO_' + dataSubType + '_'+ str(imgId).zfill(12) + '.jpg' +if os.path.isfile(imgDir + imgFilename): + I = io.imread(imgDir + imgFilename) + plt.imshow(I) + plt.axis('off') + plt.show() + diff --git a/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaTools/__init__.py b/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaTools/__init__.py new file mode 100644 index 0000000..072d8d9 --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaTools/__init__.py @@ -0,0 +1 @@ +__author__ = 'aagrawal' diff --git a/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaTools/vqa.py b/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaTools/vqa.py new file mode 100644 index 0000000..4f76961 --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/PythonHelperTools/vqaTools/vqa.py @@ -0,0 +1,179 @@ +__author__ = 'aagrawal' +__version__ = '0.9' + +# Interface for accessing the VQA dataset. + +# This code is based on the code written by Tsung-Yi Lin for MSCOCO Python API available at the following link: +# (https://github.com/pdollar/coco/blob/master/PythonAPI/pycocotools/coco.py). + +# The following functions are defined: +# VQA - VQA class that loads VQA annotation file and prepares data structures. +# getQuesIds - Get question ids that satisfy given filter conditions. +# getImgIds - Get image ids that satisfy given filter conditions. +# loadQA - Load questions and answers with the specified question ids. +# showQA - Display the specified questions and answers. +# loadRes - Load result file and create result object. + +# Help on each function can be accessed by: "help(COCO.function)" + +import json +import datetime +import copy + + +class VQA: + def __init__(self, annotation_file=None, question_file=None): + """ + Constructor of VQA helper class for reading and visualizing questions and answers. + :param annotation_file (str): location of VQA annotation file + :return: + """ + # load dataset + self.dataset = {} + self.questions = {} + self.qa = {} + self.qqa = {} + self.imgToQA = {} + if not annotation_file == None and not question_file == None: + # print 'loading VQA annotations and questions into memory...' + time_t = datetime.datetime.utcnow() + dataset = json.load(open(annotation_file, 'r')) + questions = json.load(open(question_file, 'r')) + # print datetime.datetime.utcnow() - time_t + self.dataset = dataset + self.questions = questions + self.createIndex() + + def createIndex(self): + imgToQA = {ann['image_id']: [] for ann in self.dataset['annotations']} + qa = {ann['question_id']: [] for ann in self.dataset['annotations']} + qqa = {ann['question_id']: [] for ann in self.dataset['annotations']} + for ann in self.dataset['annotations']: + imgToQA[ann['image_id']] += [ann] + qa[ann['question_id']] = ann + for ques in self.questions['questions']: + qqa[ques['question_id']] = ques + # print 'index created!' + + # create class members + self.qa = qa + self.qqa = qqa + self.imgToQA = imgToQA + + def info(self): + """ + Print information about the VQA annotation file. + :return: + """ + + # for key, value in self.datset['info'].items(): + # print '%s: %s'%(key, value) + + def getQuesIds(self, imgIds=[], quesTypes=[], ansTypes=[]): + """ + Get question ids that satisfy given filter conditions. default skips that filter + :param imgIds (int array) : get question ids for given imgs + quesTypes (str array) : get question ids for given question types + ansTypes (str array) : get question ids for given answer types + :return: ids (int array) : integer array of question ids + """ + imgIds = imgIds if type(imgIds) == list else [imgIds] + quesTypes = quesTypes if type(quesTypes) == list else [quesTypes] + ansTypes = ansTypes if type(ansTypes) == list else [ansTypes] + + if len(imgIds) == len(quesTypes) == len(ansTypes) == 0: + anns = self.dataset['annotations'] + else: + if not len(imgIds) == 0: + anns = sum([self.imgToQA[imgId] for imgId in imgIds if imgId in self.imgToQA], []) + else: + anns = self.dataset['annotations'] + anns = anns if len(quesTypes) == 0 else [ann for ann in anns if ann['question_type'] in quesTypes] + anns = anns if len(ansTypes) == 0 else [ann for ann in anns if ann['answer_type'] in ansTypes] + ids = [ann['question_id'] for ann in anns] + return ids + + def getImgIds(self, quesIds=[], quesTypes=[], ansTypes=[]): + """ + Get image ids that satisfy given filter conditions. default skips that filter + :param quesIds (int array) : get image ids for given question ids + quesTypes (str array) : get image ids for given question types + ansTypes (str array) : get image ids for given answer types + :return: ids (int array) : integer array of image ids + """ + quesIds = quesIds if type(quesIds) == list else [quesIds] + quesTypes = quesTypes if type(quesTypes) == list else [quesTypes] + ansTypes = ansTypes if type(ansTypes) == list else [ansTypes] + + if len(quesIds) == len(quesTypes) == len(ansTypes) == 0: + anns = self.dataset['annotations'] + else: + if not len(quesIds) == 0: + anns = sum([self.qa[quesId] for quesId in quesIds if quesId in self.qa], []) + else: + anns = self.dataset['annotations'] + anns = anns if len(quesTypes) == 0 else [ann for ann in anns if ann['question_type'] in quesTypes] + anns = anns if len(ansTypes) == 0 else [ann for ann in anns if ann['answer_type'] in ansTypes] + ids = [ann['image_id'] for ann in anns] + return ids + + def loadQA(self, ids=[]): + """ + Load questions and answers with the specified question ids. + :param ids (int array) : integer ids specifying question ids + :return: qa (object array) : loaded qa objects + """ + if type(ids) == list: + return [self.qa[id] for id in ids] + elif type(ids) == int: + return [self.qa[ids]] + + def showQA(self, anns): + """ + Display the specified annotations. + :param anns (array of object): annotations to display + :return: None + """ + if len(anns) == 0: + return 0 + for ann in anns: + quesId = ann['question_id'] + print("Question: %s" % (self.qqa[quesId]['question'])) + for ans in ann['answers']: + print("Answer %d: %s" % (ans['answer_id'], ans['answer'])) + + def loadRes(self, resFile, quesFile): + """ + Load result file and return a result object. + :param resFile (str) : file name of result file + :return: res (obj) : result api object + """ + res = VQA() + res.questions = json.load(open(quesFile)) + res.dataset['info'] = copy.deepcopy(self.questions['info']) + res.dataset['task_type'] = copy.deepcopy(self.questions['task_type']) + res.dataset['data_type'] = copy.deepcopy(self.questions['data_type']) + res.dataset['data_subtype'] = copy.deepcopy(self.questions['data_subtype']) + res.dataset['license'] = copy.deepcopy(self.questions['license']) + + # print 'Loading and preparing results... ' + time_t = datetime.datetime.utcnow() + anns = json.load(open(resFile)) + assert type(anns) == list, 'results is not an array of objects' + annsQuesIds = [ann['question_id'] for ann in anns] + assert set(annsQuesIds) == set(self.getQuesIds()), \ + 'Results do not correspond to current VQA set. Either the results do not have predictions for all question ids in annotation file or there is atleast one question id that does not belong to the question ids in the annotation file.' + for ann in anns: + quesId = ann['question_id'] + if res.dataset['task_type'] == 'Multiple Choice': + assert ann['answer'] in self.qqa[quesId][ + 'multiple_choices'], 'predicted answer is not one of the multiple choices' + qaAnn = self.qa[quesId] + ann['image_id'] = qaAnn['image_id'] + ann['question_type'] = qaAnn['question_type'] + ann['answer_type'] = qaAnn['answer_type'] + # print 'DONE (t=%0.2fs)'%((datetime.datetime.utcnow() - time_t).total_seconds()) + + res.dataset['annotations'] = anns + res.createIndex() + return res diff --git a/minigpt4/common/vqa_tools/VQA/QuestionTypes/abstract_v002_question_types.txt b/minigpt4/common/vqa_tools/VQA/QuestionTypes/abstract_v002_question_types.txt new file mode 100644 index 0000000..44304fc --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/QuestionTypes/abstract_v002_question_types.txt @@ -0,0 +1,81 @@ +how many +what color is the +is the +where is the +what +what is +are the +what is the +is there a +does the +is the woman +is the man +what is on the +is it +is the girl +is the boy +is the dog +are they +who is +what kind of +what color are the +what is in the +what is the man +is there +what is the woman +what are the +what is the boy +are there +what is the girl +is this +how +which +how many people are +is the cat +why is the +are +will the +what type of +what is the dog +do +is she +does +do the +is +is the baby +are there any +is the lady +can +what animal is +where are the +is the sun +what are they +did the +what is the cat +what is the lady +how many clouds are +is that +is the little girl +is he +are these +how many trees are +how many pillows +are the people +why +is the young +how many windows are +is this a +what is the little +is the tv +how many animals are +who +how many pictures +how many plants are +how many birds are +what color is +what is the baby +is anyone +what color +how many bushes +is the old man +none of the above diff --git a/minigpt4/common/vqa_tools/VQA/QuestionTypes/mscoco_question_types.txt b/minigpt4/common/vqa_tools/VQA/QuestionTypes/mscoco_question_types.txt new file mode 100644 index 0000000..9559050 --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/QuestionTypes/mscoco_question_types.txt @@ -0,0 +1,65 @@ +how many +is the +what +what color is the +what is the +is this +is this a +what is +are the +what kind of +is there a +what type of +is it +what are the +where is the +is there +does the +what color are the +are these +are there +which +is +what is the man +is the man +are +how +does this +what is on the +what does the +how many people are +what is in the +what is this +do +what are +are they +what time +what sport is +are there any +is he +what color is +why +where are the +what color +who is +what animal is +is the woman +is this an +do you +how many people are in +what room is +has +is this person +what is the woman +can you +why is the +is the person +what is the color of the +what is the person +could +was +is that a +what number is +what is the name +what brand +none of the above diff --git a/minigpt4/common/vqa_tools/VQA/README.md b/minigpt4/common/vqa_tools/VQA/README.md new file mode 100644 index 0000000..439d59d --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/README.md @@ -0,0 +1,80 @@ +Python API and Evaluation Code for v2.0 and v1.0 releases of the VQA dataset. +=================== +## VQA v2.0 release ## +This release consists of +- Real + - 82,783 MS COCO training images, 40,504 MS COCO validation images and 81,434 MS COCO testing images (images are obtained from [MS COCO website] (http://mscoco.org/dataset/#download)) + - 443,757 questions for training, 214,354 questions for validation and 447,793 questions for testing + - 4,437,570 answers for training and 2,143,540 answers for validation (10 per question) + +There is only one type of task +- Open-ended task + +## VQA v1.0 release ## +This release consists of +- Real + - 82,783 MS COCO training images, 40,504 MS COCO validation images and 81,434 MS COCO testing images (images are obtained from [MS COCO website] (http://mscoco.org/dataset/#download)) + - 248,349 questions for training, 121,512 questions for validation and 244,302 questions for testing (3 per image) + - 2,483,490 answers for training and 1,215,120 answers for validation (10 per question) +- Abstract + - 20,000 training images, 10,000 validation images and 20,000 MS COCO testing images + - 60,000 questions for training, 30,000 questions for validation and 60,000 questions for testing (3 per image) + - 600,000 answers for training and 300,000 answers for validation (10 per question) + +There are two types of tasks +- Open-ended task +- Multiple-choice task (18 choices per question) + +## Requirements ## +- python 2.7 +- scikit-image (visit [this page](http://scikit-image.org/docs/dev/install.html) for installation) +- matplotlib (visit [this page](http://matplotlib.org/users/installing.html) for installation) + +## Files ## +./Questions +- For v2.0, download the question files from the [VQA download page](http://www.visualqa.org/download.html), extract them and place in this folder. +- For v1.0, both real and abstract, question files can be found on the [VQA v1 download page](http://www.visualqa.org/vqa_v1_download.html). +- Question files from Beta v0.9 release (123,287 MSCOCO train and val images, 369,861 questions, 3,698,610 answers) can be found below + - [training question files](http://visualqa.org/data/mscoco/prev_rel/Beta_v0.9/Questions_Train_mscoco.zip) + - [validation question files](http://visualqa.org/data/mscoco/prev_rel/Beta_v0.9/Questions_Val_mscoco.zip) +- Question files from Beta v0.1 release (10k MSCOCO images, 30k questions, 300k answers) can be found [here](http://visualqa.org/data/mscoco/prev_rel/Beta_v0.1/Questions_Train_mscoco.zip). + +./Annotations +- For v2.0, download the annotations files from the [VQA download page](http://www.visualqa.org/download.html), extract them and place in this folder. +- For v1.0, for both real and abstract, annotation files can be found on the [VQA v1 download page](http://www.visualqa.org/vqa_v1_download.html). +- Annotation files from Beta v0.9 release (123,287 MSCOCO train and val images, 369,861 questions, 3,698,610 answers) can be found below + - [training annotation files](http://visualqa.org/data/mscoco/prev_rel/Beta_v0.9/Annotations_Train_mscoco.zip) + - [validation annotation files](http://visualqa.org/data/mscoco/prev_rel/Beta_v0.9/Annotations_Val_mscoco.zip) +- Annotation files from Beta v0.1 release (10k MSCOCO images, 30k questions, 300k answers) can be found [here](http://visualqa.org/data/mscoco/prev_rel/Beta_v0.1/Annotations_Train_mscoco.zip). + +./Images +- For real, create a directory with name mscoco inside this directory. For each of train, val and test, create directories with names train2014, val2014 and test2015 respectively inside mscoco directory, download respective images from [MS COCO website](http://mscoco.org/dataset/#download) and place them in respective folders. +- For abstract, create a directory with name abstract_v002 inside this directory. For each of train, val and test, create directories with names train2015, val2015 and test2015 respectively inside abstract_v002 directory, download respective images from [VQA download page](http://www.visualqa.org/download.html) and place them in respective folders. + +./PythonHelperTools +- This directory contains the Python API to read and visualize the VQA dataset +- vqaDemo.py (demo script) +- vqaTools (API to read and visualize data) + +./PythonEvaluationTools +- This directory contains the Python evaluation code +- vqaEvalDemo.py (evaluation demo script) +- vqaEvaluation (evaluation code) + +./Results +- OpenEnded_mscoco_train2014_fake_results.json (an example of a fake results file for v1.0 to run the demo) +- Visit [VQA evaluation page] (http://visualqa.org/evaluation) for more details. + +./QuestionTypes +- This directory contains the following lists of question types for both real and abstract questions (question types are unchanged from v1.0 to v2.0). In a list, if there are question types of length n+k and length n with the same first n words, then the question type of length n does not include questions that belong to the question type of length n+k. +- mscoco_question_types.txt +- abstract_v002_question_types.txt + +## References ## +- [VQA: Visual Question Answering](http://visualqa.org/) +- [Microsoft COCO](http://mscoco.org/) + +## Developers ## +- Aishwarya Agrawal (Virginia Tech) +- Code for API is based on [MSCOCO API code](https://github.com/pdollar/coco). +- The format of the code for evaluation is based on [MSCOCO evaluation code](https://github.com/tylin/coco-caption). diff --git a/minigpt4/common/vqa_tools/VQA/license.txt b/minigpt4/common/vqa_tools/VQA/license.txt new file mode 100644 index 0000000..f87c06b --- /dev/null +++ b/minigpt4/common/vqa_tools/VQA/license.txt @@ -0,0 +1,30 @@ +Copyright (c) 2014, Aishwarya Agrawal +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE +FOR +ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +The views and conclusions contained in the software and documentation are +those +of the authors and should not be interpreted as representing official +policies, +either expressed or implied, of the FreeBSD Project. diff --git a/minigpt4/common/vqa_tools/__init__.py b/minigpt4/common/vqa_tools/__init__.py new file mode 100644 index 0000000..9b98da8 --- /dev/null +++ b/minigpt4/common/vqa_tools/__init__.py @@ -0,0 +1,8 @@ +""" + Copyright (c) 2022, salesforce.com, inc. + All rights reserved. + SPDX-License-Identifier: BSD-3-Clause + For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause +""" + +__author__ = "aagrawal" diff --git a/minigpt4/common/vqa_tools/vqa.py b/minigpt4/common/vqa_tools/vqa.py new file mode 100644 index 0000000..a386b90 --- /dev/null +++ b/minigpt4/common/vqa_tools/vqa.py @@ -0,0 +1,211 @@ +""" + Copyright (c) 2022, salesforce.com, inc. + All rights reserved. + SPDX-License-Identifier: BSD-3-Clause + For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause +""" + +__author__ = "aagrawal" +__version__ = "0.9" + +# Interface for accessing the VQA dataset. + +# This code is based on the code written by Tsung-Yi Lin for MSCOCO Python API available at the following link: +# (https://github.com/pdollar/coco/blob/master/PythonAPI/pycocotools/coco.py). + +# The following functions are defined: +# VQA - VQA class that loads VQA annotation file and prepares data structures. +# getQuesIds - Get question ids that satisfy given filter conditions. +# getImgIds - Get image ids that satisfy given filter conditions. +# loadQA - Load questions and answers with the specified question ids. +# showQA - Display the specified questions and answers. +# loadRes - Load result file and create result object. + +# Help on each function can be accessed by: "help(COCO.function)" + +import json +import datetime +import copy + + +class VQA: + def __init__(self, annotation_file=None, question_file=None): + """ + Constructor of VQA helper class for reading and visualizing questions and answers. + :param annotation_file (str): location of VQA annotation file + :return: + """ + # load dataset + self.dataset = {} + self.questions = {} + self.qa = {} + self.qqa = {} + self.imgToQA = {} + if not annotation_file == None and not question_file == None: + print("loading VQA annotations and questions into memory...") + time_t = datetime.datetime.utcnow() + dataset = json.load(open(annotation_file, "r")) + questions = json.load(open(question_file, "r")) + self.dataset = dataset + self.questions = questions + self.createIndex() + + def createIndex(self): + # create index + print("creating index...") + imgToQA = {ann["image_id"]: [] for ann in self.dataset["annotations"]} + qa = {ann["question_id"]: [] for ann in self.dataset["annotations"]} + qqa = {ann["question_id"]: [] for ann in self.dataset["annotations"]} + for ann in self.dataset["annotations"]: + imgToQA[ann["image_id"]] += [ann] + qa[ann["question_id"]] = ann + for ques in self.questions["questions"]: + qqa[ques["question_id"]] = ques + print("index created!") + + # create class members + self.qa = qa + self.qqa = qqa + self.imgToQA = imgToQA + + def info(self): + """ + Print information about the VQA annotation file. + :return: + """ + for key, value in self.datset["info"].items(): + print("%s: %s" % (key, value)) + + def getQuesIds(self, imgIds=[], quesTypes=[], ansTypes=[]): + """ + Get question ids that satisfy given filter conditions. default skips that filter + :param imgIds (int array) : get question ids for given imgs + quesTypes (str array) : get question ids for given question types + ansTypes (str array) : get question ids for given answer types + :return: ids (int array) : integer array of question ids + """ + imgIds = imgIds if type(imgIds) == list else [imgIds] + quesTypes = quesTypes if type(quesTypes) == list else [quesTypes] + ansTypes = ansTypes if type(ansTypes) == list else [ansTypes] + + if len(imgIds) == len(quesTypes) == len(ansTypes) == 0: + anns = self.dataset["annotations"] + else: + if not len(imgIds) == 0: + anns = sum( + [self.imgToQA[imgId] for imgId in imgIds if imgId in self.imgToQA], + [], + ) + else: + anns = self.dataset["annotations"] + anns = ( + anns + if len(quesTypes) == 0 + else [ann for ann in anns if ann["question_type"] in quesTypes] + ) + anns = ( + anns + if len(ansTypes) == 0 + else [ann for ann in anns if ann["answer_type"] in ansTypes] + ) + ids = [ann["question_id"] for ann in anns] + return ids + + def getImgIds(self, quesIds=[], quesTypes=[], ansTypes=[]): + """ + Get image ids that satisfy given filter conditions. default skips that filter + :param quesIds (int array) : get image ids for given question ids + quesTypes (str array) : get image ids for given question types + ansTypes (str array) : get image ids for given answer types + :return: ids (int array) : integer array of image ids + """ + quesIds = quesIds if type(quesIds) == list else [quesIds] + quesTypes = quesTypes if type(quesTypes) == list else [quesTypes] + ansTypes = ansTypes if type(ansTypes) == list else [ansTypes] + + if len(quesIds) == len(quesTypes) == len(ansTypes) == 0: + anns = self.dataset["annotations"] + else: + if not len(quesIds) == 0: + anns = sum( + [self.qa[quesId] for quesId in quesIds if quesId in self.qa], [] + ) + else: + anns = self.dataset["annotations"] + anns = ( + anns + if len(quesTypes) == 0 + else [ann for ann in anns if ann["question_type"] in quesTypes] + ) + anns = ( + anns + if len(ansTypes) == 0 + else [ann for ann in anns if ann["answer_type"] in ansTypes] + ) + ids = [ann["image_id"] for ann in anns] + return ids + + def loadQA(self, ids=[]): + """ + Load questions and answers with the specified question ids. + :param ids (int array) : integer ids specifying question ids + :return: qa (object array) : loaded qa objects + """ + if type(ids) == list: + return [self.qa[id] for id in ids] + elif type(ids) == int: + return [self.qa[ids]] + + def showQA(self, anns): + """ + Display the specified annotations. + :param anns (array of object): annotations to display + :return: None + """ + if len(anns) == 0: + return 0 + for ann in anns: + quesId = ann["question_id"] + print("Question: %s" % (self.qqa[quesId]["question"])) + for ans in ann["answers"]: + print("Answer %d: %s" % (ans["answer_id"], ans["answer"])) + + def loadRes(self, resFile, quesFile): + """ + Load result file and return a result object. + :param resFile (str) : file name of result file + :return: res (obj) : result api object + """ + res = VQA() + res.questions = json.load(open(quesFile)) + res.dataset["info"] = copy.deepcopy(self.questions["info"]) + res.dataset["task_type"] = copy.deepcopy(self.questions["task_type"]) + res.dataset["data_type"] = copy.deepcopy(self.questions["data_type"]) + res.dataset["data_subtype"] = copy.deepcopy(self.questions["data_subtype"]) + res.dataset["license"] = copy.deepcopy(self.questions["license"]) + + print("Loading and preparing results... ") + time_t = datetime.datetime.utcnow() + anns = json.load(open(resFile)) + assert type(anns) == list, "results is not an array of objects" + annsQuesIds = [ann["question_id"] for ann in anns] + assert set(annsQuesIds) == set( + self.getQuesIds() + ), "Results do not correspond to current VQA set. Either the results do not have predictions for all question ids in annotation file or there is atleast one question id that does not belong to the question ids in the annotation file." + for ann in anns: + quesId = ann["question_id"] + if res.dataset["task_type"] == "Multiple Choice": + assert ( + ann["answer"] in self.qqa[quesId]["multiple_choices"] + ), "predicted answer is not one of the multiple choices" + qaAnn = self.qa[quesId] + ann["image_id"] = qaAnn["image_id"] + ann["question_type"] = qaAnn["question_type"] + ann["answer_type"] = qaAnn["answer_type"] + print( + "DONE (t=%0.2fs)" % ((datetime.datetime.utcnow() - time_t).total_seconds()) + ) + + res.dataset["annotations"] = anns + res.createIndex() + return res diff --git a/minigpt4/common/vqa_tools/vqa_eval.py b/minigpt4/common/vqa_tools/vqa_eval.py new file mode 100644 index 0000000..ee808b3 --- /dev/null +++ b/minigpt4/common/vqa_tools/vqa_eval.py @@ -0,0 +1,324 @@ +""" + Copyright (c) 2022, salesforce.com, inc. + All rights reserved. + SPDX-License-Identifier: BSD-3-Clause + For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause +""" + +# coding=utf-8 + +__author__ = "aagrawal" + +# This code is based on the code written by Tsung-Yi Lin for MSCOCO Python API available at the following link: +# (https://github.com/tylin/coco-caption/blob/master/pycocoevalcap/eval.py). +import sys +import re + + +class VQAEval: + def __init__(self, vqa=None, vqaRes=None, n=2): + self.n = n + self.accuracy = {} + self.evalQA = {} + self.evalQuesType = {} + self.evalAnsType = {} + self.vqa = vqa + self.vqaRes = vqaRes + if vqa is not None: + self.params = {"question_id": vqa.getQuesIds()} + self.contractions = { + "aint": "ain't", + "arent": "aren't", + "cant": "can't", + "couldve": "could've", + "couldnt": "couldn't", + "couldn'tve": "couldn't've", + "couldnt've": "couldn't've", + "didnt": "didn't", + "doesnt": "doesn't", + "dont": "don't", + "hadnt": "hadn't", + "hadnt've": "hadn't've", + "hadn'tve": "hadn't've", + "hasnt": "hasn't", + "havent": "haven't", + "hed": "he'd", + "hed've": "he'd've", + "he'dve": "he'd've", + "hes": "he's", + "howd": "how'd", + "howll": "how'll", + "hows": "how's", + "Id've": "I'd've", + "I'dve": "I'd've", + "Im": "I'm", + "Ive": "I've", + "isnt": "isn't", + "itd": "it'd", + "itd've": "it'd've", + "it'dve": "it'd've", + "itll": "it'll", + "let's": "let's", + "maam": "ma'am", + "mightnt": "mightn't", + "mightnt've": "mightn't've", + "mightn'tve": "mightn't've", + "mightve": "might've", + "mustnt": "mustn't", + "mustve": "must've", + "neednt": "needn't", + "notve": "not've", + "oclock": "o'clock", + "oughtnt": "oughtn't", + "ow's'at": "'ow's'at", + "'ows'at": "'ow's'at", + "'ow'sat": "'ow's'at", + "shant": "shan't", + "shed've": "she'd've", + "she'dve": "she'd've", + "she's": "she's", + "shouldve": "should've", + "shouldnt": "shouldn't", + "shouldnt've": "shouldn't've", + "shouldn'tve": "shouldn't've", + "somebody'd": "somebodyd", + "somebodyd've": "somebody'd've", + "somebody'dve": "somebody'd've", + "somebodyll": "somebody'll", + "somebodys": "somebody's", + "someoned": "someone'd", + "someoned've": "someone'd've", + "someone'dve": "someone'd've", + "someonell": "someone'll", + "someones": "someone's", + "somethingd": "something'd", + "somethingd've": "something'd've", + "something'dve": "something'd've", + "somethingll": "something'll", + "thats": "that's", + "thered": "there'd", + "thered've": "there'd've", + "there'dve": "there'd've", + "therere": "there're", + "theres": "there's", + "theyd": "they'd", + "theyd've": "they'd've", + "they'dve": "they'd've", + "theyll": "they'll", + "theyre": "they're", + "theyve": "they've", + "twas": "'twas", + "wasnt": "wasn't", + "wed've": "we'd've", + "we'dve": "we'd've", + "weve": "we've", + "werent": "weren't", + "whatll": "what'll", + "whatre": "what're", + "whats": "what's", + "whatve": "what've", + "whens": "when's", + "whered": "where'd", + "wheres": "where's", + "whereve": "where've", + "whod": "who'd", + "whod've": "who'd've", + "who'dve": "who'd've", + "wholl": "who'll", + "whos": "who's", + "whove": "who've", + "whyll": "why'll", + "whyre": "why're", + "whys": "why's", + "wont": "won't", + "wouldve": "would've", + "wouldnt": "wouldn't", + "wouldnt've": "wouldn't've", + "wouldn'tve": "wouldn't've", + "yall": "y'all", + "yall'll": "y'all'll", + "y'allll": "y'all'll", + "yall'd've": "y'all'd've", + "y'alld've": "y'all'd've", + "y'all'dve": "y'all'd've", + "youd": "you'd", + "youd've": "you'd've", + "you'dve": "you'd've", + "youll": "you'll", + "youre": "you're", + "youve": "you've", + } + self.manualMap = { + "none": "0", + "zero": "0", + "one": "1", + "two": "2", + "three": "3", + "four": "4", + "five": "5", + "six": "6", + "seven": "7", + "eight": "8", + "nine": "9", + "ten": "10", + } + self.articles = ["a", "an", "the"] + + self.periodStrip = re.compile("(?!<=\d)(\.)(?!\d)") + self.commaStrip = re.compile("(\d)(,)(\d)") + self.punct = [ + ";", + r"/", + "[", + "]", + '"', + "{", + "}", + "(", + ")", + "=", + "+", + "\\", + "_", + "-", + ">", + "<", + "@", + "`", + ",", + "?", + "!", + ] + + def evaluate(self, quesIds=None): + if quesIds == None: + quesIds = [quesId for quesId in self.params["question_id"]] + gts = {} + res = {} + for quesId in quesIds: + gts[quesId] = self.vqa.qa[quesId] + res[quesId] = self.vqaRes.qa[quesId] + + # ================================================= + # Compute accuracy + # ================================================= + accQA = [] + accQuesType = {} + accAnsType = {} + print("computing accuracy") + step = 0 + for quesId in quesIds: + resAns = res[quesId]["answer"] + resAns = resAns.replace("\n", " ") + resAns = resAns.replace("\t", " ") + resAns = resAns.strip() + resAns = self.processPunctuation(resAns) + resAns = self.processDigitArticle(resAns) + gtAcc = [] + gtAnswers = [ans["answer"] for ans in gts[quesId]["answers"]] + if len(set(gtAnswers)) > 1: + for ansDic in gts[quesId]["answers"]: + ansDic["answer"] = self.processPunctuation(ansDic["answer"]) + for gtAnsDatum in gts[quesId]["answers"]: + otherGTAns = [ + item for item in gts[quesId]["answers"] if item != gtAnsDatum + ] + matchingAns = [item for item in otherGTAns if item["answer"] == resAns] + acc = min(1, float(len(matchingAns)) / 3) + gtAcc.append(acc) + quesType = gts[quesId]["question_type"] + ansType = gts[quesId]["answer_type"] + avgGTAcc = float(sum(gtAcc)) / len(gtAcc) + accQA.append(avgGTAcc) + if quesType not in accQuesType: + accQuesType[quesType] = [] + accQuesType[quesType].append(avgGTAcc) + if ansType not in accAnsType: + accAnsType[ansType] = [] + accAnsType[ansType].append(avgGTAcc) + self.setEvalQA(quesId, avgGTAcc) + self.setEvalQuesType(quesId, quesType, avgGTAcc) + self.setEvalAnsType(quesId, ansType, avgGTAcc) + if step % 100 == 0: + self.updateProgress(step / float(len(quesIds))) + step = step + 1 + + self.setAccuracy(accQA, accQuesType, accAnsType) + print("Done computing accuracy") + + def processPunctuation(self, inText): + outText = inText + for p in self.punct: + if (p + " " in inText or " " + p in inText) or ( + re.search(self.commaStrip, inText) != None + ): + outText = outText.replace(p, "") + else: + outText = outText.replace(p, " ") + outText = self.periodStrip.sub("", outText, re.UNICODE) + return outText + + def processDigitArticle(self, inText): + outText = [] + tempText = inText.lower().split() + for word in tempText: + word = self.manualMap.setdefault(word, word) + if word not in self.articles: + outText.append(word) + else: + pass + for wordId, word in enumerate(outText): + if word in self.contractions: + outText[wordId] = self.contractions[word] + outText = " ".join(outText) + return outText + + def setAccuracy(self, accQA, accQuesType, accAnsType): + self.accuracy["overall"] = round(100 * float(sum(accQA)) / len(accQA), self.n) + self.accuracy["perQuestionType"] = { + quesType: round( + 100 * float(sum(accQuesType[quesType])) / len(accQuesType[quesType]), + self.n, + ) + for quesType in accQuesType + } + self.accuracy["perAnswerType"] = { + ansType: round( + 100 * float(sum(accAnsType[ansType])) / len(accAnsType[ansType]), self.n + ) + for ansType in accAnsType + } + + def setEvalQA(self, quesId, acc): + self.evalQA[quesId] = round(100 * acc, self.n) + + def setEvalQuesType(self, quesId, quesType, acc): + if quesType not in self.evalQuesType: + self.evalQuesType[quesType] = {} + self.evalQuesType[quesType][quesId] = round(100 * acc, self.n) + + def setEvalAnsType(self, quesId, ansType, acc): + if ansType not in self.evalAnsType: + self.evalAnsType[ansType] = {} + self.evalAnsType[ansType][quesId] = round(100 * acc, self.n) + + def updateProgress(self, progress): + barLength = 20 + status = "" + if isinstance(progress, int): + progress = float(progress) + if not isinstance(progress, float): + progress = 0 + status = "error: progress var must be float\r\n" + if progress < 0: + progress = 0 + status = "Halt...\r\n" + if progress >= 1: + progress = 1 + status = "Done...\r\n" + block = int(round(barLength * progress)) + text = "\rFinshed Percent: [{0}] {1}% {2}".format( + "#" * block + "-" * (barLength - block), int(progress * 100), status + ) + sys.stdout.write(text) + sys.stdout.flush() diff --git a/minigpt4/conversation/conversation.py b/minigpt4/conversation/conversation.py index 83c8bee..829e034 100644 --- a/minigpt4/conversation/conversation.py +++ b/minigpt4/conversation/conversation.py @@ -127,7 +127,14 @@ CONV_VISION_LLama2 = Conversation( sep="", ) - +CONV_VISION_minigptv2 = Conversation( + system="", + roles=("[INST] ", " [/INST]"), + messages=[], + offset=2, + sep_style=SeparatorStyle.SINGLE, + sep="", +) class Chat: def __init__(self, model, vis_processor, device='cuda:0', stopping_criteria=None): diff --git a/minigpt4/datasets/datasets/coco_caption.py b/minigpt4/datasets/datasets/coco_caption.py index 76f86e4..2cafd93 100755 --- a/minigpt4/datasets/datasets/coco_caption.py +++ b/minigpt4/datasets/datasets/coco_caption.py @@ -91,7 +91,7 @@ class RefCOCOEvalData(torch.utils.data.Dataset): image_path = os.path.join(self.root_path, f'{img_id[:27]}.jpg') image = Image.open(image_path).convert('RGB') image = self.vis_processor(image) - question = f"[refer] tell me the location of {sent}?" + question = f"[refer] give me the location of {sent}" return image, question, img_id class EvalCaptionData(torch.utils.data.Dataset): diff --git a/minigpt4/datasets/datasets/llava_dataset.py b/minigpt4/datasets/datasets/llava_dataset.py index 2766189..e12d51b 100755 --- a/minigpt4/datasets/datasets/llava_dataset.py +++ b/minigpt4/datasets/datasets/llava_dataset.py @@ -3,7 +3,6 @@ import json import pickle import random import time -# import iterto import numpy as np from PIL import Image import skimage.io as io diff --git a/minigpt4/datasets/datasets/vqa_datasets.py b/minigpt4/datasets/datasets/vqa_datasets.py index 5cdc0fa..a8df3cc 100755 --- a/minigpt4/datasets/datasets/vqa_datasets.py +++ b/minigpt4/datasets/datasets/vqa_datasets.py @@ -16,30 +16,6 @@ class VQADataset(BaseDataset): def __init__(self, vis_processor, text_processor, vis_root, ann_paths): super().__init__(vis_processor, text_processor, vis_root, ann_paths) - # def collater(self, samples): - # image_list, question_list, answer_list, weight_list = [], [], [], [] - - # num_answers = [] - - # for sample in samples: - # image_list.append(sample["image"]) - # question_list.append(sample["question"]) - - # weight_list.extend(sample["weights"]) - - # answers = sample["answer"] - - # answer_list.extend(answers) - # num_answers.append(len(answers)) - - # return { - # "image": torch.stack(image_list, dim=0), - # "text_input": question_list, - # "answer": answer_list, - # "weight": torch.Tensor(weight_list), - # "n_answers": torch.LongTensor(num_answers), - # } - class VQAEvalDataset(BaseDataset): def __init__(self, vis_processor, text_processor, vis_root, ann_paths): @@ -65,7 +41,6 @@ class OKVQAEvalData(torch.utils.data.Dataset): image = Image.open(image_path).convert('RGB') image = self.vis_processor(image) question = f"[vqa] Based on the image, respond to this question with a short answer: {question}" - # question = f"[vqa] {question} " return image, question, question_id, img_id class VizWizEvalData(torch.utils.data.Dataset): @@ -86,54 +61,9 @@ class VizWizEvalData(torch.utils.data.Dataset): image_path = os.path.join(self.root_path, img_id) image = Image.open(image_path).convert('RGB') image = self.vis_processor(image) - # question = f"[vqa] Based on the image, respond to this question with a short answer: {question} " - question = f"[vqa] Based on the image, respond to this question with a short answer: {question} and reply 'unanswerable' if you could not answer it" + question = f"[vqa] The question is '{question}' Based on the image, answer the question with a single word or phrase. and reply 'unanswerable' when the provided information is insufficient" return image, question, answers -class AOKVQADAEvalData(torch.utils.data.Dataset): - def __init__(self, loaded_data, vis_processor, root_path): - self.loaded_data = loaded_data - self.root_path = root_path - self.vis_processor = vis_processor - - def __len__(self): - return len(self.loaded_data) - - def __getitem__(self, idx): - data = self.loaded_data[idx] - img_file = data['image'] - question = data['question'] - question_id = data['question_id'] - image_path = os.path.join(self.root_path, img_file) - image = Image.open(image_path).convert('RGB') - image = self.vis_processor(image) - question = f"[vqa] Based on the image, respond to this question with a short answer: {question}" - # question = f"[vqa] {question} " - return image, question, question_id - -class AOKVQAMCEvalData(torch.utils.data.Dataset): - def __init__(self, loaded_data, vis_processor, root_path): - self.loaded_data = loaded_data - self.root_path = root_path - self.vis_processor = vis_processor - - def __len__(self): - return len(self.loaded_data) - - def __getitem__(self, idx): - data = self.loaded_data[idx] - img_file = data['image'] - question = data['question'] - question_id = data['question_id'] - image_path = os.path.join(self.root_path, img_file) - image = Image.open(image_path).convert('RGB') - image = self.vis_processor(image).half().cuda() - candidates=data['choices'] - # question = f"Given this image, choose one correct answer from {candidates} for this question: {question} " - question = f"[vqa] Based on the image, respond to this question with a short answer: {question}" - # question = f"[vqa] {question} " - return image, question, question_id, candidates - class IconQAEvalData(torch.utils.data.Dataset): def __init__(self, loaded_data, vis_processor, root_path): self.loaded_data = loaded_data @@ -152,9 +82,7 @@ class IconQAEvalData(torch.utils.data.Dataset): image = self.vis_processor(image).half().cuda() candidates = '_'.join(data['choices']) answer = data['answer'] - # question = f"Given this image, choose one correct answer from {candidates} for this question: {question} " question = f"[vqa] Based on the image, respond to this question with a short answer: {question}" - # question = f"[vqa] {question} " return image, question, candidates, answer class GQAEvalData(torch.utils.data.Dataset): @@ -173,9 +101,7 @@ class GQAEvalData(torch.utils.data.Dataset): image = Image.open(image_path).convert("RGB") image = self.vis_processor(image) question = ann["question"] - # question = f'Question: {question} Short answer: ' question = f"[vqa] Based on the image, respond to this question with a short answer: {question}" - # question = f"[vqa] {question} " labels = ann["answer"] return image, question, labels @@ -196,7 +122,7 @@ class HMEvalData(torch.utils.data.Dataset): image = Image.open(image_path).convert("RGB") image = self.vis_processor(image) question = ann["text"] - question = f"[vqa] This is an image with: \"{question}\" written on it. Is it hateful? Answer:" + question = f"This is an image writting '{question}'. Is this image hateful? Answer yes or no. Answer:" labels = ann["label"] return image, question, labels @@ -217,7 +143,6 @@ class VSREvalData(torch.utils.data.Dataset): image = self.vis_processor(image) question = ann["caption"] question = f'[vqa] Based on the image, is this statement true or false? {question}' - question_id = ann["image"].split('.')[0] labels = 'true' if ann["label"] == 1 else 'false' return image, question, labels \ No newline at end of file diff --git a/minigpt4/models/minigpt_base.py b/minigpt4/models/minigpt_base.py index 4ded50f..66c8c80 100644 --- a/minigpt4/models/minigpt_base.py +++ b/minigpt4/models/minigpt_base.py @@ -11,8 +11,6 @@ from transformers import StoppingCriteria, StoppingCriteriaList from minigpt4.conversation.conversation import StoppingCriteriaSub - - class MiniGPTBase(BaseModel): """ Base class for MiniGPT-4 and MiniGPT-v2 @@ -368,9 +366,18 @@ class MiniGPTBase(BaseModel): min_length=min_length, top_p=top_p, repetition_penalty=repetition_penalty, - stopping_criteria=stopping_criteria, + # stopping_criteria=stopping_criteria, ) + # with self.maybe_autocast(): + # outputs = self.llama_model.generate( + # inputs_embeds=embs, + # attention_mask=attn_mask, + # max_new_tokens=max_new_tokens, + # num_beams=num_beams, + # do_sample=do_sample, + # # stopping_criteria=stopping_criteria, + # ) answers = [] for output_token in outputs: if output_token[0] == 0: @@ -400,4 +407,4 @@ class MiniGPTBase(BaseModel): for i in range(all_losses.shape[0]): all_losses[i, num_cand[i]:] = 9999 output_class_ranks = torch.argsort(all_losses, dim=-1) - return output_class_ranks.tolist() \ No newline at end of file + return output_class_ranks.tolist()