Add files via upload

datasets/ocr_dataset.py

@@ -0,0 +1,22 @@
+from torch.utils.data import Dataset
+import os
+class ocrDataset(Dataset):
+    def __init__(
+        self,
+        image_dir_path= "./data/ocr",
+        dataset_name = "ct80"
+    ):
+        self.image_dir_path = image_dir_path
+        self.dataset_name = dataset_name
+        file_path = os.path.join(image_dir_path, f'{dataset_name}/test_label.txt')
+        file = open(file_path, "r")
+        self.lines = file.readlines()
+    def __len__(self):
+        return len(self.lines)
+    def __getitem__(self, idx):
+        image_id = self.lines[idx].split()[0]
+        img_path = os.path.join(self.image_dir_path,f'{self.dataset_name}/{image_id}')
+        answers = self.lines[idx].split()[1]
+        return {
+            "image_path": img_path,
+            "gt_answers": answers}

datasets/vqa_dataset.py

@@ -0,0 +1,103 @@
+from torch.utils.data import Dataset
+import os
+import json
+class textVQADataset(Dataset):
+    def __init__(
+        self,
+        image_dir_path= "./data/textVQA/train_images",
+        ann_path= "./data/textVQA/TextVQA_0.5.1_val.json"
+    ):
+        
+        self.data = json.load(open(ann_path, "r"))["data"]
+        self.image_dir_path = image_dir_path
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        question = self.data[idx]['question']
+        answers = self.data[idx]['answers']
+        img_path = os.path.join(self.image_dir_path, f"{self.data[idx]['image_id']}.jpg")
+        return {
+            "image_path": img_path,
+            "question": question,
+            "gt_answers": answers}
+
+class docVQADataset(Dataset):
+    def __init__(
+        self,
+        image_dir_path= "./data/docVQA/val",
+        ann_path= "./data/docVQA/val/val_v1.0.json",
+    ):
+        
+        self.data = json.load(open(ann_path, "r"))["data"]
+        self.image_dir_path = image_dir_path
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        question = self.data[idx]['question']
+        answers = self.data[idx]['answers']
+        img_path = os.path.join(self.image_dir_path, self.data[idx]['image'])
+        return {
+            "image_path": img_path,
+            "question": question,
+            "gt_answers": answers}
+
+class ocrVQADataset(Dataset):
+    def __init__(
+        self,
+        image_dir_path= "./data/ocrVQA/images",
+        ann_path= "./data/ocrVQA/dataset.json",
+    ):
+        self.image_list = []
+        self.question_list = []
+        self.answer_list = []
+        dataset = json.load(open(ann_path, "r"))
+        import pdb;pdb.set_trace()
+        for idx, data in enumerate(dataset):
+            questions =  dataset[data]['questions']
+            for index, question in enumerate(questions):
+                image_file = os.path.join(image_dir_path, f'{data}.jpg')
+                gt_answers = dataset[data]['answers'][index]
+                self.image_list.append(image_file)
+                self.answer_list.append(gt_answers)
+                self.question_list.append(question)
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        question = self.question_list[idx]
+        answers = self.answer_list[idx]
+        img_path = self.image_list[idx]
+        return {
+            "image_path": img_path,
+            "question": question,
+            "gt_answers": answers}
+
+class STVQADataset(Dataset):
+    def __init__(
+        self,
+        image_dir_path= "./data/STVQA",
+        ann_path= "./data/STVQA/train_task_3.json",
+    ):
+        self.image_list = []
+        self.question_list = []
+        self.answer_list = []
+        data = json.load(open(ann_path, "r"))
+        for i in range(len(data)):
+            image_path = image_dir_path+'/'+data['data'][i]['dataset']+'/'+data['data'][i]['file_name']
+            self.image_list.append(image_path)
+            self.answer_list.append(data['data'][i]['answers'])
+            self.question_list.append(data['data'][i]['question'])
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        question = self.question_list[idx]
+        answers = self.answer_list[idx]
+        img_path = self.image_list[idx]
+        return {
+            "image_path": img_path,
+            "question": question,
+            "gt_answers": answers}

eval.py

@@ -0,0 +1,440 @@
+import argparse
+#from models.BLIP2.BLIP2 import BLIP2
+import more_itertools
+from tqdm import tqdm
+import datetime
+import os
+import json
+import re
+from datasets.vqa_dataset import textVQADataset, docVQADataset, ocrVQADataset, STVQADataset
+from datasets.ocr_dataset import ocrDataset
+from models.lavis.lavis import lavis
+import torch
+import numpy as np
+def get_model(args):
+    if args.model_name=='BLIP2':
+        #model = BLIP2(args.BLIP2_model_path, args.device)
+        model = lavis(args.BLIP2_model_name, args.BLIP2_model_type, args.device)
+    #elif args.model_name=='mPLUG-Owl':
+    #   model = 
+    return model
+def has_word(sentence, word):
+    pattern = r"\b" + re.escape(word) + r"\b"
+    match = re.search(pattern, sentence)
+    if match:
+        return True
+    else:
+        return False
+def remove_special_chars(s):
+    pattern = r"[^a-zA-Z0-9\s]"
+    s = re.sub(pattern, "", s)
+    return s
+class VQAEval:
+    def __init__(self):
+        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, answer, gt_answers):
+        answer = answer.replace("\n", " ")
+        answer = answer.replace("\t", " ")
+        answer = answer.strip()
+        answer = self.processPunctuation(answer)
+        answer = self.processDigitArticle(answer)
+        if type(gt_answers)==list:
+            for i in range(len(gt_answers)):
+                gt_answers[i] = gt_answers[i].replace("\n", " ")
+                gt_answers[i] = gt_answers[i].replace("\t", " ")
+                gt_answers[i] = gt_answers[i].strip()
+                gt_answers[i] = self.processPunctuation(gt_answers[i])
+                gt_answers[i] = self.processDigitArticle(gt_answers[i])
+                if has_word(answer, gt_answers[i]):
+                    return 1
+            return 0
+        else:
+            gt_answers = gt_answers.replace("\n", " ")
+            gt_answers= gt_answers.replace("\t", " ")
+            gt_answers = gt_answers.strip()
+            gt_answers = self.processPunctuation(gt_answers)
+            gt_answers = self.processDigitArticle(gt_answers)
+            if has_word(answer, gt_answers[i]):
+                return 1
+            else:
+                return 0
+
+    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 evaluate_VQA(
+    model,
+    dataset,
+    model_name,
+    dataset_name,
+    time,
+    batch_size=1,
+    answer_path='./answers'
+):
+    predictions=[]
+    for batch in more_itertools.chunked(
+        tqdm(dataset, desc="Running inference"), batch_size
+    ):
+        batch = batch[0]
+        output = model.generate(image=batch['image_path'], question=batch['question'])
+        answer_dict={'question':batch['question'], 'answer':output, 
+        'gt_answers':batch['gt_answers'], 'image_path':batch['image_path'],
+        'model_name':model_name}
+        predictions.append(answer_dict)
+    answer_dir = os.path.join(answer_path, time)
+    os.makedirs(answer_dir, exist_ok=True)
+    answer_path = os.path.join(answer_dir, f"{dataset_name}.json")
+    with open(answer_path, "w") as f:
+        f.write(json.dumps(predictions, indent=4))
+    eval = VQAEval()
+    correct = 0
+    num = 0
+    with open(answer_path, 'r') as f:
+        dict = json.load(f)
+        for i in range(len(dict)):
+            gt_answers = dict[i]['gt_answers']
+            answer = dict[i]['answer']
+            if eval.evaluate(answer,gt_answers)==1:
+                correct+=1
+            num+=1
+    print(f'{dataset_name}:{float(correct)/num}')
+    return float(correct)/num
+def evaluate_OCR(
+    model,
+    dataset,
+    model_name,
+    dataset_name,
+    time,
+    question='what is written in the image?',
+    batch_size=1,
+    answer_path='./answers'
+):
+    predictions=[]
+    for batch in more_itertools.chunked(
+        tqdm(dataset, desc="Running inference"), batch_size
+    ):
+        batch = batch[0]
+        output = model.generate(image=batch['image_path'], question=question)
+        answer_dict={'question':question, 'answer':output, 
+        'gt_answers':batch['gt_answers'], 'image_path':batch['image_path'],
+        'model_name':model_name}
+        predictions.append(answer_dict)
+    answer_dir = os.path.join(answer_path, time)
+    os.makedirs(answer_dir, exist_ok=True)
+    answer_path = os.path.join(answer_dir, f"{dataset_name}.json")
+    with open(answer_path, "w") as f:
+        f.write(json.dumps(predictions, indent=4))
+    correct = 0
+    num = 0
+    with open(answer_path, 'r') as f:
+        dict = json.load(f)
+        for i in range(len(dict)):
+            gt_answers = dict[i]['gt_answers']
+            answer = dict[i]['answer']
+            gt_answers = remove_special_chars(gt_answers).lower()
+            answer = remove_special_chars(answer).lower()
+            if has_word(answer, gt_answers):
+                correct+=1
+            num+=1
+    print(f'{dataset_name}:{float(correct)/num}')
+    return float(correct)/num
+            
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Demo")
+    #OCR datasets
+    parser.add_argument("--ocr_dir_path", type=str, default="./data")
+    parser.add_argument("--ocr_dataset_name", type=str, default="IIIT5K svt IC13_857 IC15_1811 svtp ct80 cocotext ctw totaltext HOST WOST WordArt")
+    #textVQA
+    parser.add_argument("--textVQA_image_dir_path", type=str, default="./data/textVQA/train_images")
+    parser.add_argument("--textVQA_ann_path", type=str, default="./data/textVQA/TextVQA_0.5.1_val.json")
+
+    #docVQA
+    parser.add_argument("--docVQA_image_dir_path", type=str, default="./data/docVQA/val")
+    parser.add_argument("--docVQA_ann_path", type=str, default="./data/docVQA/val/val_v1.0.json")
+
+    #ocrVQA
+    parser.add_argument("--ocrVQA_image_dir_path", type=str, default="./data/ocrVQA/images")
+    parser.add_argument("--ocrVQA_ann_path", type=str, default="./data/ocrVQA/dataset.json")
+
+    #STVQA
+    parser.add_argument("--STVQA_image_dir_path", type=str, default="./data/STVQA")
+    parser.add_argument("--STVQA_ann_path", type=str, default="./data/STVQA/train_task_3.json")
+
+    #result_path
+    parser.add_argument("--answer_path", type=str, default="./answers")
+
+    parser.add_argument(
+        "--eval_textVQA",
+        action="store_true",
+        default=False,
+        help="Whether to evaluate on textVQA."
+    )
+    parser.add_argument(
+        "--eval_docVQA",
+        action="store_true",
+        default=False,
+        help="Whether to evaluate on docVQA."
+    )
+    parser.add_argument(
+        "--eval_ocrVQA",
+        action="store_true",
+        default=False,
+        help="Whether to evaluate on ocrVQA."
+    )
+    parser.add_argument(
+        "--eval_STVQA",
+        action="store_true",
+        default=False,
+        help="Whether to evaluate on STVQA."
+    )
+    parser.add_argument(
+        "--eval_ocr",
+        action="store_true",
+        default=False,
+        help="Whether to evaluate on ocr."
+    )
+    #BLIP2
+    #parser.add_argument("--BLIP2_model_path", type=str, default="/home/zhangli/GPT4/BLIP2-flant5")
+    parser.add_argument("--BLIP2_model_name", type=str, default="blip2_opt")#blip2_t5  blip2_opt blip2_vicuna_instruct
+    parser.add_argument("--BLIP2_model_type", type=str, default="pretrain_opt6.7b")#pretrain_flant5xxl pretrain_opt6.7b vicuna13b
+
+
+    parser.add_argument("--model_name", type=str, default="BLIP2")#mPLUG,miniGPT4,LLaVA
+    parser.add_argument("--device", type=str, default="cuda:2")
+    args = parser.parse_args()
+    return args
+
+def main(args):
+    np.random.seed(0)
+    max_sample_num = 5000
+    model = get_model(args)
+    '''ocr_dataset_name=['IIIT5K','svt','IC13_857','IC15_1811','svtp','ct80',
+                  'cocotext','ctw','totaltext','HOST','WOST','WordArt']'''
+    ocr_dataset_name = args.ocr_dataset_name.split()
+    result = {}
+    time = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
+    if args.eval_textVQA:
+        dataset = textVQADataset(args.textVQA_image_dir_path, args.textVQA_ann_path)
+        acc = evaluate_VQA(model, dataset, args.model_name, 'textVQA', time)
+        result['textVQA'] = acc
+    if args.eval_docVQA:
+        dataset = docVQADataset(args.docVQA_image_dir_path, args.docVQA_ann_path)
+        acc = evaluate_VQA(model, dataset, args.model_name, 'docVQA', time)
+        result['docVQA'] = acc
+
+    if args.eval_ocrVQA:
+        dataset = ocrVQADataset(args.ocrVQA_image_dir_path, args.ocrVQA_ann_path)
+        random_indices = np.random.choice(
+            len(dataset), max_sample_num, replace=False
+        )
+        dataset = torch.utils.data.Subset(dataset,random_indices)
+        acc = evaluate_VQA(model, dataset, args.model_name, 'ocrVQA', time)
+        result['ocrVQA'] = acc
+    
+    if args.eval_STVQA:
+        dataset = STVQADataset(args.STVQA_image_dir_path, args.STVQA_ann_path)
+        random_indices = np.random.choice(
+            len(dataset), max_sample_num, replace=False
+        )
+        dataset = torch.utils.data.Subset(dataset,random_indices)
+        acc = evaluate_VQA(model, dataset, args.model_name, 'STVQA', time)
+        result['STVQA'] = acc
+
+    if args.eval_ocr:
+        for i in range(len(ocr_dataset_name)):
+            dataset = ocrDataset(args.ocr_dir_path, ocr_dataset_name[i])
+            acc = evaluate_OCR(model, dataset, args.model_name, ocr_dataset_name[i], time)
+            result[ocr_dataset_name[i]] = acc
+    result_path = os.path.join(os.path.join(args.answer_path, time), 'result.json')
+    with open(result_path, "w") as f:
+        f.write(json.dumps(result, indent=4))
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

models/BLIP2/BLIP2.py

@@ -0,0 +1,21 @@
+from transformers import Blip2Processor, Blip2ForConditionalGeneration
+import torch
+from PIL import Image
+from ..process import pad_image
+#There are some issues with the Hugging Face version of the BLIP2-opt model.
+class BLIP2:
+    def __init__(self, model_path, device = "cuda") -> None:
+        self.processor = Blip2Processor.from_pretrained(model_path)
+        self.model = Blip2ForConditionalGeneration.from_pretrained(
+        model_path, torch_dtype=torch.float16).to(device)
+        self.model.eval()
+        self.device = device
+    def generate(self, image, question, pad=True):
+        prompt =f'Question: {question} Answer:'
+        image = Image.open(image)
+        if pad:
+            image = pad_image(image, (224,224))
+        inputs = self.processor(images=image, text=prompt, return_tensors="pt").to(self.device, torch.float16)
+        generated_ids = self.model.generate(**inputs)
+        generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
+        return generated_text

models/LLaVA/LLaVA.py

@@ -0,0 +1,47 @@
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+from transformers import CLIPVisionModel, CLIPImageProcessor, StoppingCriteria
+import torch
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+def disable_torch_init():
+    """
+    Disable the redundant torch default initialization to accelerate model creation.
+    """
+    setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
+    setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
+def patch_config(config):
+    patch_dict = {
+        "use_mm_proj": True,
+        "mm_vision_tower": "openai/clip-vit-large-patch14",
+        "mm_hidden_size": 1024
+    }
+    cfg = AutoConfig.from_pretrained(config)
+    if not hasattr(cfg, "mm_vision_tower"):
+        print(f'`mm_vision_tower` not found in `{config}`, applying patch and save to disk.')
+        for k, v in patch_dict.items():
+            setattr(cfg, k, v)
+        cfg.save_pretrained(config)
+class LLaVA:
+    def __init__(self, model_path) -> None:
+        tokenizer = AutoTokenizer.from_pretrained(model_path)
+        patch_config(model_path)
+        model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16).cuda()
+        image_processor = CLIPImageProcessor.from_pretrained(model.config.mm_vision_tower, torch_dtype=torch.float16)
+
+        mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
+        tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        if mm_use_im_start_end:
+            tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+
+        vision_tower = model.model.vision_tower[0]
+        vision_tower.to(device='cuda', dtype=torch.float16)
+        vision_config = vision_tower.config
+        vision_config.im_patch_token = tokenizer.convert_tokens_to_ids([DEFAULT_IMAGE_PATCH_TOKEN])[0]
+        vision_config.use_im_start_end = mm_use_im_start_end
+        if mm_use_im_start_end:
+            vision_config.im_start_token, vision_config.im_end_token = tokenizer.convert_tokens_to_ids([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN])
+        image_token_len = (vision_config.image_size // vision_config.patch_size) ** 2
+    def generate(self, image, question):
+        

models/lavis/lavis.py

@@ -0,0 +1,20 @@
+import torch
+from PIL import Image
+from lavis.models import load_model_and_preprocess
+from ..process import pad_image
+class lavis:
+    def __init__(self, model_name, model_type, device) -> None:
+        model, vis_processors, txt_processors = load_model_and_preprocess(name = model_name, model_type = model_type, is_eval=True, device=device)
+        self.model = model
+        self.vis_processors = vis_processors
+        self.txt_processors = txt_processors
+        self.device = device
+    def generate(self, image, question, pad=True):
+        prompt = f'Question: {question} Short answer:'
+        image = Image.open(image).convert("RGB")
+        if pad:
+            image = pad_image(image, (224,224))
+        image = self.vis_processors["eval"](image).unsqueeze(0).to(self.device)
+        prompt = self.txt_processors["eval"](prompt)
+        answer = self.model.predict_answers(samples={"image": image, "text_input": prompt}, inference_method="generate", max_len=32)[0]
+        return answer

models/mPLUG_owl/clip/configuration_clip.py

@@ -0,0 +1,454 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" CLIP model configuration"""
+
+import copy
+import os
+from collections import OrderedDict
+from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
+
+
+if TYPE_CHECKING:
+    from transformers.processing_utils import ProcessorMixin
+    from transformers.utils import TensorType
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "openai/clip-vit-base-patch32": "https://huggingface.co/openai/clip-vit-base-patch32/resolve/main/config.json",
+    # See all CLIP models at https://huggingface.co/models?filter=clip
+}
+
+
+class CLIPTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPTextModel`]. It is used to instantiate a CLIP
+    text encoder according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the text encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 49408):
+            Vocabulary size of the CLIP text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`CLIPModel`].
+        hidden_size (`int`, *optional*, defaults to 512):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        max_position_embeddings (`int`, *optional*, defaults to 77):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPTextConfig, CLIPTextModel
+
+    >>> # Initializing a CLIPTextConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPTextConfig()
+
+    >>> # Initializing a CLIPTextModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "clip_text_model"
+
+    def __init__(
+        self,
+        vocab_size=49408,
+        hidden_size=512,
+        intermediate_size=2048,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=8,
+        max_position_embeddings=77,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "clip":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class CLIPVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPVisionModel`]. It is used to instantiate a
+    CLIP vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPVisionConfig, CLIPVisionModel
+
+    >>> # Initializing a CLIPVisionConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPVisionConfig()
+
+    >>> # Initializing a CLIPVisionModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "clip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class CLIPConfig(PretrainedConfig):
+    r"""
+    [`CLIPConfig`] is the configuration class to store the configuration of a [`CLIPModel`]. It is used to instantiate
+    a CLIP model according to the specified arguments, defining the text model and vision model configs. Instantiating
+    a configuration with the defaults will yield a similar configuration to that of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The inital value of the *logit_scale* paramter. Default is used as per the original CLIP implementation.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPConfig, CLIPModel
+
+    >>> # Initializing a CLIPConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPConfig()
+
+    >>> # Initializing a CLIPModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a CLIPConfig from a CLIPTextConfig and a CLIPVisionConfig
+    >>> from transformers import CLIPTextConfig, CLIPVisionConfig
+
+    >>> # Initializing a CLIPText and CLIPVision configuration
+    >>> config_text = CLIPTextConfig()
+    >>> config_vision = CLIPVisionConfig()
+
+    >>> config = CLIPConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "clip"
+    is_composition = True
+
+    def __init__(
+        self, text_config=None, vision_config=None, projection_dim=512, logit_scale_init_value=2.6592, **kwargs
+    ):
+        # If `_config_dict` exist, we use them for the backward compatibility.
+        # We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot
+        # of confusion!).
+        text_config_dict = kwargs.pop("text_config_dict", None)
+        vision_config_dict = kwargs.pop("vision_config_dict", None)
+
+        super().__init__(**kwargs)
+
+        # Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in
+        # `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most
+        # cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.
+        if text_config_dict is not None:
+            if text_config is None:
+                text_config = {}
+
+            # This is the complete result when using `text_config_dict`.
+            _text_config_dict = CLIPTextConfig(**text_config_dict).to_dict()
+
+            # Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.
+            for key, value in _text_config_dict.items():
+                if key in text_config and value != text_config[key] and key not in ["transformers_version"]:
+                    # If specified in `text_config_dict`
+                    if key in text_config_dict:
+                        message = (
+                            f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "
+                            f'The value `text_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The "
+                            f'value `text_config["{key}"]` will be overriden.'
+                        )
+                    logger.warning(message)
+
+            # Update all values in `text_config` with the ones in `_text_config_dict`.
+            text_config.update(_text_config_dict)
+
+        if vision_config_dict is not None:
+            if vision_config is None:
+                vision_config = {}
+
+            # This is the complete result when using `vision_config_dict`.
+            _vision_config_dict = CLIPVisionConfig(**vision_config_dict).to_dict()
+            # convert keys to string instead of integer
+            if "id2label" in _vision_config_dict:
+                _vision_config_dict["id2label"] = {
+                    str(key): value for key, value in _vision_config_dict["id2label"].items()
+                }
+
+            # Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.
+            for key, value in _vision_config_dict.items():
+                if key in vision_config and value != vision_config[key] and key not in ["transformers_version"]:
+                    # If specified in `vision_config_dict`
+                    if key in vision_config_dict:
+                        message = (
+                            f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "
+                            f'values. The value `vision_config_dict["{key}"]` will be used instead.'
+                        )
+                    # If inferred from default argument values (just to be super careful)
+                    else:
+                        message = (
+                            f"`vision_config_dict` is provided which will be used to initialize `CLIPVisionConfig`. "
+                            f'The value `vision_config["{key}"]` will be overriden.'
+                        )
+                    logger.warning(message)
+
+            # Update all values in `vision_config` with the ones in `_vision_config_dict`.
+            vision_config.update(_vision_config_dict)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("`text_config` is `None`. Initializing the `CLIPTextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `CLIPVisionConfig` with default values.")
+
+        self.text_config = CLIPTextConfig(**text_config)
+        self.vision_config = CLIPVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: CLIPTextConfig, vision_config: CLIPVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`CLIPConfig`] (or a derived class) from clip text model configuration and clip vision model
+        configuration.
+
+        Returns:
+            [`CLIPConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["text_config"] = self.text_config.to_dict()
+        output["vision_config"] = self.vision_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output
+
+
+class CLIPOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("input_ids", {0: "batch", 1: "sequence"}),
+                ("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),
+                ("attention_mask", {0: "batch", 1: "sequence"}),
+            ]
+        )
+
+    @property
+    def outputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("logits_per_image", {0: "batch"}),
+                ("logits_per_text", {0: "batch"}),
+                ("text_embeds", {0: "batch"}),
+                ("image_embeds", {0: "batch"}),
+            ]
+        )
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-4
+
+    def generate_dummy_inputs(
+        self,
+        processor: "ProcessorMixin",
+        batch_size: int = -1,
+        seq_length: int = -1,
+        framework: Optional["TensorType"] = None,
+    ) -> Mapping[str, Any]:
+        text_input_dict = super().generate_dummy_inputs(
+            processor.tokenizer, batch_size=batch_size, seq_length=seq_length, framework=framework
+        )
+        image_input_dict = super().generate_dummy_inputs(
+            processor.feature_extractor, batch_size=batch_size, framework=framework
+        )
+        return {**text_input_dict, **image_input_dict}
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 14

models/mPLUG_owl/mPLUG.py

@@ -0,0 +1,44 @@
+import argparse
+import json
+import torch
+from transformers.models.llama.configuration_llama import LlamaConfig
+from mplug_owl.configuration_mplug_owl import mPLUG_OwlConfig
+from mplug_owl.modeling_mplug_owl import mPLUG_OwlForConditionalGeneration
+from transformers.models.llama.tokenization_llama import LlamaTokenizer
+from mplug_owl.modeling_mplug_owl import ImageProcessor
+from mplug_owl.tokenize_utils import tokenize_prompts
+class mPLUG:
+    def __init__(self, checkpoint_path=None, tokenizer_path=None) -> None:
+        config = mPLUG_OwlConfig()
+        self.model = mPLUG_OwlForConditionalGeneration(config=config).to(torch.bfloat16)
+        self.model.eval()
+
+        if checkpoint_path is not None:
+            tmp_ckpt = torch.load(
+                checkpoint_path, map_location='cpu')
+            msg = self.model.load_state_dict(tmp_ckpt, strict=False)
+            print(msg)
+
+        assert tokenizer_path is not None
+        self.tokenizer = LlamaTokenizer(
+            tokenizer_path, pad_token='<unk>', add_bos_token=False)
+        self.img_processor = ImageProcessor()
+    def generate(self, image, question, max_length=512, top_k=1, do_sample=True, **generate_kwargs):
+        prompts = [
+        f'''The following is a conversation between a curious human and AI assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
+        Human: <image>
+        Human: {question}
+        AI: ''']
+        tokens_to_generate = 0
+        add_BOS = True
+        context_tokens_tensor, context_length_tensorm, attention_mask = tokenize_prompts(
+            prompts=prompts, tokens_to_generate=tokens_to_generate, add_BOS=add_BOS, tokenizer=self.tokenizer, ignore_dist=True)
+        images = self.img_processor(image).to(torch.bfloat16).cuda()
+        context_tokens_tensor = context_tokens_tensor.cuda()
+        self.model.eval()
+        with torch.no_grad():
+            res = self.model.generate(input_ids=context_tokens_tensor, pixel_values=images,
+                                attention_mask=attention_mask, max_lengt=max_length,top_k=top_k,do_sample=do_sample,**generate_kwargs)
+        sentence = self.tokenizer.decode(res.tolist()[0], skip_special_tokens=True)
+        return sentence
+    

models/mPLUG_owl/mplug_owl/configuration_mplug_owl.py

@@ -0,0 +1,154 @@
+# coding=utf-8
+# Copyright 2023 Alibaba Inc. and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import copy
+import os
+from typing import Union
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.models.auto import CONFIG_MAPPING
+from transformers.models.auto.modeling_auto import \
+    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class mPLUG_OwlVisualAbstractorConfig(PretrainedConfig):
+
+    model_type = "mPLUG_OwlVisualAbstractor"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=1024,
+        num_hidden_layers=6,
+        num_attention_heads=8,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        classifier_dropout=None,
+        cross_attention_frequency=2,
+        encoder_hidden_size=1024,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.classifier_dropout = classifier_dropout
+        self.cross_attention_frequency = cross_attention_frequency
+        self.encoder_hidden_size = encoder_hidden_size
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs)
+
+        if config_dict.get("model_type") == "mplug_owl":
+            config_dict = config_dict["abstractor_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class mPLUG_OwlConfig(PretrainedConfig):
+
+    model_type = "mplug_owl"
+    is_composition = True
+
+    def __init__(self, vision_config=None, visual_abstractor_config=None, text_config=None, num_query_tokens=64, **kwargs):
+        super().__init__(**kwargs)
+        from clip.configuration_clip import CLIPVisionConfig
+        if vision_config is None:
+            # By defalt we use openai-clip large patch14
+
+            vision_config = CLIPVisionConfig(
+                **vision_config_dict, layer_norm_eps=1e-6).to_dict()
+            logger.info(
+                "vision_config is None.")
+
+        if visual_abstractor_config is None:
+            visual_abstractor_config = {}
+            logger.info(
+                "abstractor_config is None. ")
+
+        if text_config is None:
+            # we use LLAMA 7b by default
+            from transformers.models.llama.configuration_llama import \
+                LlamaConfig
+            text_config = LlamaConfig(pad_token_id=2).to_dict()
+            logger.info("text_config is None.")
+
+        self.vision_config = CLIPVisionConfig(**vision_config)
+        self.visual_abstractor_config = mPLUG_OwlVisualAbstractorConfig(
+            **visual_abstractor_config)
+        self.visual_abstractor_config.layer_norm_eps = 1e-6
+        text_model_type = text_config["model_type"] if "model_type" in text_config else "opt"
+        self.text_config = CONFIG_MAPPING[text_model_type](**text_config)
+
+        self.tie_word_embeddings = self.text_config.tie_word_embeddings
+        self.is_encoder_decoder = self.text_config.is_encoder_decoder
+
+        self.num_query_tokens = num_query_tokens
+        self.visual_abstractor_config.encoder_hidden_size = self.vision_config.hidden_size
+        self.use_decoder_only_language_model = self.text_config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+        self.initializer_factor = 1.0
+        self.initializer_range = 0.02
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["vision_config"] = self.vision_config.to_dict()
+        output["abstractor_config"] = self.visual_abstractor_config.to_dict()
+        output["text_config"] = self.text_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output
+
+
+vision_config_dict = {
+    "hidden_size": 1024,
+    "intermediate_size": 4096,
+    "num_attention_heads": 8,
+    "num_hidden_layers": 24,
+    "patch_size": 14,
+    "projection_dim": 768}

models/mPLUG_owl/mplug_owl/tokenize_utils.py

@@ -0,0 +1,171 @@
+# coding=utf-8
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tokenization utilities."""
+
+
+import re
+
+import torch
+from icecream import ic
+
+
+def detokenize_generations(tokens_gpu_tensor,
+                           lengths_gpu_tensor,
+                           return_segments, tokenizer):
+    """Detokenize the generated tokens."""
+
+    prompts_plus_generations = []
+    if return_segments:
+        prompts_plus_generations_segments = []
+
+    tokens = tokens_gpu_tensor.cpu().numpy().tolist()
+    lengths = lengths_gpu_tensor.cpu().numpy().tolist()
+    for sequence_tokens, length in zip(tokens, lengths):
+        sequence_tokens = sequence_tokens[:length]
+        prompts_plus_generations.append(
+            tokenizer.detokenize(sequence_tokens))
+        if return_segments:
+            from tokenizers.decoders import Metaspace
+            if hasattr(tokenizer, 'tokenizer'):
+                if isinstance(tokenizer.tokenizer.decoder, Metaspace):
+                    words = tokenizer.tokenizer.decode(sequence_tokens)
+                else:
+                    words = []
+                    for token in sequence_tokens:
+                        word = tokenizer.tokenizer.decoder[token]
+                        word = bytearray(
+                            [tokenizer.tokenizer.byte_decoder[c] for c in word]).decode(
+                                'utf-8', errors='replace')
+                        words.append(word)
+                prompts_plus_generations_segments.append(words)
+            else:
+                words = tokenizer.detokenize(sequence_tokens)
+                # else:
+                #     words = []
+                #     for token in sequence_tokens:
+                #         word = tokenizer.tokenizer.decoder[token]
+                #         word = bytearray(
+                #             [tokenizer.tokenizer.byte_decoder[c] for c in word]).decode(
+                #                 'utf-8', errors='replace')
+                #         words.append(word)
+                prompts_plus_generations_segments.append(words)
+
+    if return_segments:
+        return tokens, prompts_plus_generations, \
+            prompts_plus_generations_segments
+
+    return tokens, prompts_plus_generations
+
+
+def tokenize_prompts(prompts=None, tokens_to_generate=None,
+                     add_BOS=None, rank=0, tokenizer=None, ignore_dist=False):
+    """Tokenize prompts and make them avaiable on all ranks."""
+
+    # On all ranks set to None so we can pass them to functions
+    sizes_list = None
+    prompts_tokens_cuda_long_tensor = None
+    prompts_length_cuda_long_tensor = None
+
+    # On the specified rank, build the above.
+    attention_mask = None
+    if ignore_dist or torch.distributed.get_rank() == rank:
+        assert prompts is not None
+        assert tokens_to_generate is not None
+        # Tensor of tokens padded and their unpadded length.
+        prompts_tokens_cuda_long_tensor, prompts_length_cuda_long_tensor, attention_mask = \
+            _tokenize_prompts_and_batch(
+                prompts, tokens_to_generate, add_BOS, tokenizer)
+        # We need the sizes of these tensors for the boradcast
+        sizes_list = [prompts_tokens_cuda_long_tensor.size(0),  # Batch size
+                      prompts_tokens_cuda_long_tensor.size(1)]  # Sequence lenght
+
+    return prompts_tokens_cuda_long_tensor, prompts_length_cuda_long_tensor, attention_mask
+
+
+def _tokenize_prompts_and_batch(prompts, tokens_to_generate, add_BOS, tokenizer):
+    """Given a set of prompts and number of tokens to generate:
+        - tokenize prompts
+        - set the sequence length to be the max of length of prompts
+          plus the number of tokens we would like to generate
+        - pad all the sequences to this length so we can convert them
+          into a 2D tensor.
+    """
+
+    # Tokenize all the prompts.
+    # if add_BOS:
+    #     prompts_tokens = [[tokenizer.bos] + tokenizer.tokenize(prompt)
+    #                       for prompt in prompts]
+    # else:
+    #     prompts_tokens = [tokenizer.tokenize(prompt) for prompt in prompts]
+
+    prompts_tokens = [_tokenize_prompt(
+        prompt, tokenizer, add_BOS) for prompt in prompts]
+
+    # Now we have a list of list of tokens which each list has a different
+    # size. We want to extend this list to:
+    #   - incorporate the tokens that need to be generated
+    #   - make all the sequences equal length.
+    # Get the prompts length.
+    prompts_length = [len(prompt_tokens) for prompt_tokens in prompts_tokens]
+    # Get the max prompts length.
+    max_prompt_len = max(prompts_length)
+    # Number of tokens in the each sample of the batch.
+    samples_length = max_prompt_len + tokens_to_generate
+    # Now update the list of list to be of the same size: samples_length.
+    for prompt_tokens, prompt_length in zip(prompts_tokens, prompts_length):
+        padding_size = samples_length - prompt_length
+        prompt_tokens.extend([tokenizer.eos_token_id] * padding_size)
+
+    # Now we are in a structured format, we can convert to tensors.
+    prompts_tokens_tensor = torch.LongTensor(prompts_tokens)
+    prompts_length_tensor = torch.LongTensor(prompts_length)
+    attention_mask = torch.zeros(prompts_tokens_tensor.shape[:2])
+    for i, l in enumerate(prompts_length_tensor):
+        attention_mask[i, :l] = 1
+    return prompts_tokens_tensor, prompts_length_tensor, attention_mask
+
+
+def _tokenize_prompt(prompt, tokenizer, add_BOS=False, media_info={'<image>': 65}):
+    media_tokens = {k: -int(i+1) for i, k in enumerate(media_info.keys())}
+    media_lengths = media_info.copy()
+
+    if add_BOS:
+        prompt_chunk = [tokenizer.bos_token_id]
+    else:
+        prompt_chunk = []
+
+    # Pure Text
+    if all([media_token not in prompt for media_token in media_tokens.keys()]):
+        enc_chunk = prompt_chunk + \
+            tokenizer(prompt, add_special_tokens=False)['input_ids']
+
+    # Multi-Modal Text
+    else:
+        enc_chunk = prompt_chunk
+        pattern = '|'.join(map(re.escape, list(media_tokens.keys())))
+        chunk_strs = re.split(f'({pattern})', prompt)
+        chunk_strs = [x for x in chunk_strs if len(x) > 0]
+        for idx, chunk_str in enumerate(chunk_strs):
+            if chunk_str in media_tokens:
+                enc_chunk += [media_tokens[chunk_str]] * \
+                    media_lengths[chunk_str]
+            else:
+                tmp_chunk = tokenizer(chunk_str, add_special_tokens=False)[
+                    'input_ids']
+                # if idx < len(chunk_strs) - 1: # Last chunk should not have eos
+                #     tmp_chunk += [tokenizer.eod_id]
+                enc_chunk += tmp_chunk
+    return enc_chunk

models/process.py

@@ -0,0 +1,29 @@
+import PIL
+def pad_image(image, target_size):
+ 
+    """
+    :param image: input image
+    :param target_size: a tuple (num,num)
+    :return: new image
+    """
+ 
+    iw, ih = image.size  # 原始图像的尺寸
+    w, h = target_size  # 目标图像的尺寸
+ 
+    scale = min(w / iw, h / ih)  # 转换的最小比例
+ 
+    # 保证长或宽，至少一个符合目标图像的尺寸 0.5保证四舍五入
+    nw = int(iw * scale+0.5)
+    nh = int(ih * scale+0.5)
+
+    w += 128
+    h += 128
+ 
+
+    image = image.resize((nw, nh), PIL.Image.BICUBIC)  # 更改图像尺寸，双立法插值效果很好
+    #image.show()
+    new_image = PIL.Image.new('RGB', (w, h), (0, 0, 0))  # 生成黑色图像
+    # // 为整数除法，计算图像的位置
+    new_image.paste(image, ((w - nw) // 2, (h - nh) // 2))  # 将图像填充为中间图像，两侧为黑色的样式
+
+    return new_image