个人实现的反向文心（无需训练的AI看图说话，你不心动？）

本文介绍ZeroCap的中文Paddle迁移实现，这是一个零样本图像描述模型。项目用Ernie-VIL替换原论文的CLIP，GPT采用中文版，涉及GPTChineseTokenizer、GPTLMHeadModel等模型。代码包含安装库、模型初始化、定义相关函数及生成文本等内容，还展示了效果示例，适合想了解深度学习Image Caption的新手，可参考相关B站科普视频。

ZeroCap:zero shot的image caption模型paddle迁移实现（中文版）

论文：ZeroCap: Zero-Shot Image-to-Text Generation for Visual-Semantic Arithmetic ，2024.3.31

代码: https://github.com/YoadTew/zero-shot-image-to-text

英文Paddle实现(对于zerpcap论文个人讲解也可参考该AI Studio项目) : ZeroCap

对深度学习 Image Caption 什么都不太了解 但是很想去了解学习 的同学可以看看我做的这个B站视频:

用大白话讲Paper之Image caption ZeroCap，科普视频，不想学的请直接划走

这个视频我真的尽可能用大白话去把一个个概念用最朴素的语言讲了出来，麻烦各位看官动动你们的小手为我可怜的视频增加点播放量，谢谢大家

超级 超级 推荐新手观看上面这个B站视频！！！！

本项目使用以下模型：

GPTChineseTokenizer,GPTLMHeadModel,ErnieViLProcessor, ErnieViLModel. Ernie-VIL替换原论文英文版的CLIP,然后GPT使用中文版

from paddlenlp.transformers import GPTChineseTokenizer,GPTLMHeadModelfrom paddlenlp.transformers import ErnieViLProcessor, ErnieViLModel

效果展示：

output: 这张图片讲的故事是猫躲帐

output: 这张图片讲的故事是小狗Jack在草地上.

output: 这张图片讲的故事是梅西在运球的时候

#安装库!pip install --upgrade pip!pip uninstall paddlenlp -y!pip install paddlenlp==2.4.1!pip install regex!pip install fastcore

import paddlenlp.transformers.clip as clipfrom paddlenlp.transformers import GPTLMHeadModel as GPT2LMHeadModel,GPTTokenizer as GPT2Tokenizerimport paddleimport paddle.nn as nnfrom paddlenlp.transformers import CLIPProcessor, CLIPModelfrom PIL import Imageimport numpy as npimport collectionsfrom paddlenlp.data import Padfrom paddlenlp.transformers import ErnieViLTokenizer#测试ErnieViLTokenizer使用tokenizer = ErnieViLTokenizer.from_pretrained('ernie_vil-2.0-base-zh')print(tokenizer('我爱你宝贝'))

[2024-01-29 03:28:12,985] [    INFO] - Downloading https://bj.bcebos.com/paddlenlp/models/transformers/ernie_vil/ernie_vil-2.0-base-zh/vocab.txt and saved to /home/aistudio/.paddlenlp/models/ernie_vil-2.0-base-zh[2024-01-29 03:28:12,989] [    INFO] - Downloading vocab.txt from https://bj.bcebos.com/paddlenlp/models/transformers/ernie_vil/ernie_vil-2.0-base-zh/vocab.txt100%|██████████| 182k/182k [00:00<00:00, 3.23MB/s][2024-01-29 03:28:13,209] [    INFO] - tokenizer config file saved in /home/aistudio/.paddlenlp/models/ernie_vil-2.0-base-zh/tokenizer_config.json[2024-01-29 03:28:13,212] [    INFO] - Special tokens file saved in /home/aistudio/.paddlenlp/models/ernie_vil-2.0-base-zh/special_tokens_map.json

{'input_ids': [1, 75, 329, 226, 707, 1358, 2]}

import paddlenlp.transformers.clip as clipfrom paddlenlp.transformers import GPTChineseTokenizer,GPTLMHeadModelimport paddle# from paddlenlp.transformers import ErnieForGenerationimport paddle.nn as nnfrom paddlenlp.transformers import ErnieViLProcessor, ErnieViLModelfrom PIL import Imageimport numpy as npimport collectionsfrom paddlenlp.data import Padfrom paddlenlp.transformers import ErnieViLTokenizertokenizer = ErnieViLTokenizer.from_pretrained('ernie_vil-2.0-base-zh')print(tokenizer('我爱你宝贝'))from fastcore.all import *@patch_to(GPTChineseTokenizer)def convert_ids_to_tokens(self, ids, skip_special_tokens=False):    if not isinstance(ids, (list, tuple)):        return self._convert_id_to_token(ids)    tokens = [self._convert_id_to_token(_id) for _id in ids]    if skip_special_tokens:        return [            token for token in tokens            if token not in self.all_special_tokens        ]    return tokensdef add_context0(x, y):    return (x[0] + y[0], x[1] + y[1])def add_context(x, y):    a = x.__class__(k=(x.k+y[0]),v = (x.v+y[1]))    return afrom datetime import datetimeimport sysdef log_info(text, verbose=True):    if verbose:        dt_string = datetime.now().strftime("%d/%m/%Y %H:%M:%S")        print(f'{dt_string} | {text}')        sys.stdout.flush()# def add_context(x, y):#     # print(x.__class__,x)#     Cache = collections.namedtuple("Cache","k,v")#     a = Cache(k=(x.k+y[0]).mean(axis = 2,keepdim = True),v = (x.v+y[1]).mean(axis = 2,keepdim = True))#     return aclass CLIPTextGenerator:    def __init__(self,                 seed=0,                 lm_model='gpt-2',                 forbidden_tokens_file_path='./forbidden_tokens.npy',                 clip_checkpoints='./clip_checkpoints',                 target_seq_length=50,                 reset_context_delta=True,                 num_iterations=5,                 clip_loss_temperature=0.01,                 clip_scale=1.,                 ce_scale=0.2,#0.2                 stepsize=0.3,                 grad_norm_factor=0.9,                 fusion_factor=0.99,                 repetition_penalty=1.,                 end_token='。',                 end_factor=1.01,                 forbidden_factor=20,                 **kwargs):        # set Random seed        paddle.seed(seed)        np.random.seed(seed)        # Initialize Language model        self.context_prefix = ''        self.lm_tokenizer = GPTChineseTokenizer.from_pretrained('gpt-cpm-large-cn')        self.lm_model = GPTLMHeadModel.from_pretrained("gpt-cpm-large-cn")        # self.context_prefix = self.lm_tokenizer.bos_token        self.lm_model.eval()        self.forbidden_tokens = np.load(forbidden_tokens_file_path)    #     # Freeze LM weights        for param in self.lm_model.parameters():            param.requires_grad = False    #     # Initialize CLIP        self.clip = ErnieViLModel.from_pretrained("ernie_vil-2.0-base-zh")        self.clip_preprocess = ErnieViLProcessor.from_pretrained("ernie_vil-2.0-base-zh")    #     # convert_models_to_fp32(self.clip)    #     # Init arguments        self.target_seq_length = target_seq_length        self.reset_context_delta = reset_context_delta        self.num_iterations = num_iterations        self.clip_loss_temperature = clip_loss_temperature        self.clip_scale = clip_scale        self.ce_scale = ce_scale        self.stepsize = stepsize        self.grad_norm_factor = grad_norm_factor        self.fusion_factor = fusion_factor        self.repetition_penalty = repetition_penalty        # self.end_token = self.lm_tokenizer.encode(end_token)[0]        self.end_token = self.lm_tokenizer.encode(end_token)["input_ids"][0]        self.end_factor = end_factor        self.ef_idx = 1        self.forbidden_factor = forbidden_factor    def get_img_feature(self, img_path, weights = None):        imgs = [Image.open(x) for x in img_path]        # print("imgs",imgs)        clip_imgs = [self.clip_preprocess(images = x,return_tensors="pd")["pixel_values"] for x in imgs]        with paddle.no_grad():            image_fts = [self.clip.get_image_features(x) for x in clip_imgs]            # print("image_fts",image_fts)            if weights is not None:                image_features = sum([x * weights[i] for i, x in enumerate(image_fts)])            else:                image_features = sum(image_fts)            image_features = image_features / image_features.norm(axis=-1, keepdim=True)            return image_features.detach()    def get_txt_features(self, text):        # print("text",text)        clip_texts = tokenizer(text)        clip_texts = Pad(pad_val=0)(clip_texts["input_ids"])        clip_texts = paddle.to_tensor(clip_texts)        # clip_texts = clip.tokenize(text)        # print("clip_text",clip_texts)        with paddle.no_grad():            text_features = self.clip.get_text_features(clip_texts)            text_features = text_features / text_features.norm(axis=-1, keepdim=True)        return text_features.detach()    def run(self, image_features, cond_text, beam_size):        self.image_features = image_features        context_tokens = self.lm_tokenizer.encode(self.context_prefix + cond_text)        # print("context_tokens0",context_tokens)        output_tokens, output_text = self.generate_text(context_tokens["input_ids"], beam_size)        return output_text    def generate_text(self, context_tokens, beam_size):        context_tokens = paddle.to_tensor(context_tokens).unsqueeze(0)        print("158context_tokens",context_tokens)        gen_tokens = None        scores = None        seq_lengths = paddle.ones([beam_size])        is_stopped = paddle.zeros([beam_size], dtype=paddle.bool)        for i in range(self.target_seq_length):            # print("146行")            probs = self.get_next_probs(i, context_tokens)            logits = probs.log()            if scores is None:                scores, next_tokens = logits.topk(beam_size, -1)                context_tokens = context_tokens.expand([beam_size, *context_tokens.shape[1:]])                # print(next_tokens.shape)                next_tokens, scores = next_tokens.transpose([1, 0]), scores.squeeze(0)                if gen_tokens is None:                    gen_tokens = next_tokens                else:                    gen_tokens = gen_tokens.expand(beam_size, *gen_tokens.shape[1:])                    gen_tokens = paddle.concat((gen_tokens, next_tokens), axis=1)            else:                # print("logits",logits.shape)                # print("is_stopped",is_stopped)                # print("i",i)                logits[is_stopped] = -float(np.inf)                logits[is_stopped, 0] = 0                scores_sum = scores[:, None] + logits                seq_lengths[~is_stopped] += 1                scores_sum_average = scores_sum / seq_lengths[:, None]                scores_sum_average, next_tokens = scores_sum_average.reshape([-1]).topk(                    beam_size, -1)                next_tokens_source = next_tokens // scores_sum.shape[1]                seq_lengths = seq_lengths[next_tokens_source]                next_tokens = next_tokens % scores_sum.shape[1]                next_tokens = next_tokens.unsqueeze(1)                gen_tokens = gen_tokens[next_tokens_source]                gen_tokens = paddle.concat((gen_tokens, next_tokens), axis=-1)                context_tokens = context_tokens[next_tokens_source]                scores = scores_sum_average * seq_lengths                # print("is_stopped",is_stopped,"next_tokens_source",next_tokens_source)                # is_stopped = is_stopped[next_tokens_source]                is_stopped = is_stopped[list(map(int,list(next_tokens_source.numpy())))]            context_tokens = paddle.concat((context_tokens, next_tokens), axis=1)            # print("next_tokens",next_tokens)            # print("is_stopped",is_stopped)            temp_a = next_tokens.equal(paddle.full_like(next_tokens,self.end_token)).astype("float32").squeeze()            # print(temp_a)            # is_stopped = is_stopped + next_tokens.equal(self.end_token).astype("float32").squeeze()            # is_stopped = is_stopped + next_tokens.equal(paddle.full_like(next_tokens,self.end_token)).astype("float32").squeeze()            is_stopped = paddle.any(paddle.stack([is_stopped.astype("float32"),temp_a],axis=0).astype("bool"),axis=0)            # print("is_stopped",is_stopped)                        ####            tmp_scores = scores / seq_lengths            tmp_output_list = gen_tokens.numpy()            tmp_output_texts = [self.lm_tokenizer.convert_ids_to_string(list(map(int,list(tmp_output)))) for tmp_output, tmp_length in zip(tmp_output_list, seq_lengths)]            tmp_order = tmp_scores.argsort(descending=True)            tmp_output_texts = [tmp_output_texts[i] + ' %% ' + str(tmp_scores[i].numpy()) for i in tmp_order]            log_info(tmp_output_texts, verbose=True)            ####            if is_stopped.all():                break        scores = scores / seq_lengths        output_list = gen_tokens.numpy()        output_texts = [            self.lm_tokenizer.convert_ids_to_string(list(map(int,list(output[: int(length)]))))            for output, length in zip(output_list, seq_lengths)        ]        order = scores.argsort(descending=True)        output_texts = [output_texts[i] for i in order]        return context_tokens, output_texts    def get_next_probs(self, i, context_tokens):        last_token = context_tokens[:, -1:]        if self.reset_context_delta and context_tokens.shape[1] > 1:            # print(context_tokens[:, :-1])            # print(self.lm_model(context_tokens[:, :-1],use_cache=True))            context = self.lm_model(context_tokens[:, :-1],use_cache=True)[1] #得到k v            # print("context180",context)        # Logits of LM with unshifted context        logits_before_shift = self.lm_model(context_tokens)        # print("220row",logits_before_shift.shape)        logits_before_shift = logits_before_shift[:, -1, :]        probs_before_shift = nn.functional.softmax(logits_before_shift, axis=-1).detach()        if context:            context = self.shift_context(i, context, last_token, context_tokens, probs_before_shift)        lm_output = self.lm_model(last_token, cache=context,use_cache=True)        logits, past = (            lm_output[0],            lm_output[1],        )        logits = logits[:, -1, :]        # logits = self.update_special_tokens_logits(context_tokens, i, logits)        probs = nn.functional.softmax(logits, axis=-1)        probs = (probs ** self.fusion_factor) * (probs_before_shift ** (1 - self.fusion_factor))        probs = probs / probs.sum()        return probs    def shift_context(self, i, context, last_token, context_tokens, probs_before_shift):        context_delta = [tuple([np.zeros(x.shape).astype("float32") for x in p]) for p in context]        # context_delta = [ for p in context]        window_mask = paddle.ones_like(context[0][0])        for i in range(self.num_iterations):            curr_shift = [tuple([paddle.to_tensor(x,stop_gradient = False) for x in p_]) for p_ in                            context_delta]            # for p0, p1 in curr_shift:            #     # p0.retain_grad()            #     # p1.retain_grad()            # print("context220",len(cond_text),context)            shifted_context = list(map(add_context, context, curr_shift))            # print(last_token,len(shifted_context))            # print(shifted_context)            shifted_outputs = self.lm_model(last_token, cache=shifted_context,use_cache = True)            # logits = shifted_outputs["logits"][:, -1, :]            logits = shifted_outputs[0][:, -1, :]            probs = nn.functional.softmax(logits, axis=-1)            loss = 0.0            # CLIP LOSS            clip_loss, clip_losses = self.clip_loss(probs, context_tokens)            loss += self.clip_scale * clip_loss            # CE/Fluency loss            if isinstance(self.ce_scale,float):                ce_loss = self.ce_scale * ((probs * probs.log()) - (probs * probs_before_shift.log())).sum(-1)            else:                a = self.ce_scale[0]                b = self.ce_scale[1]                ce_loss = (b - (b-a)/self.num_iterations*i) * ((probs * probs.log()) - (probs * probs_before_shift.log())).sum(-1)            loss += ce_loss.sum()            loss.backward()            # print("loss finish")            # ---------- Weights ----------            combined_scores_k = -(ce_loss)            combined_scores_c = -(self.clip_scale * paddle.stack(clip_losses)).squeeze(1)            # print(295,"combined_scores_k",combined_scores_k.shape,"combined_scores_c",combined_scores_c.shape)            # minmax            if combined_scores_k.shape[0] == 1:                tmp_weights_c = tmp_weights_k = paddle.ones(combined_scores_k.shape)            else:                tmp_weights_k = ((combined_scores_k - combined_scores_k.min())) / (combined_scores_k.max() - combined_scores_k.min())                tmp_weights_c = ((combined_scores_c - combined_scores_c.min())) / (combined_scores_c.max() - combined_scores_c.min())            # print(tmp_weights_k)            tmp_weights = 0.5 * tmp_weights_k + 0.5 * tmp_weights_c            # print("305 tmp_weights",tmp_weights.shape)                        tmp_weights = tmp_weights.reshape([tmp_weights.shape[0], 1, 1, 1])            factor = 1            # --------- Specific Gen ---------            sep_grads = None            for b in range(context_tokens.shape[0]):                tmp_sep_norms = [[(paddle.norm(x.grad[b:(b + 1)] * window_mask[b:(b + 1)]) + 1e-15) for x in p_]                                    for p_ in curr_shift]                # normalize gradients                tmp_grad = [tuple([-self.stepsize * factor * (                        x.grad[b:(b + 1)] * window_mask[b:(b + 1)] / tmp_sep_norms[i][                    j] ** self.grad_norm_factor).numpy()                                    for j, x in enumerate(p_)])                            for i, p_ in enumerate(curr_shift)]                if sep_grads is None:                    sep_grads = tmp_grad                else:                    for l_index in range(len(sep_grads)):                        sep_grads[l_index] = list(sep_grads[l_index])                        for k_index in range(len(sep_grads[0])):                            sep_grads[l_index][k_index] = np.concatenate(                                (sep_grads[l_index][k_index], tmp_grad[l_index][k_index]), axis=0)                        sep_grads[l_index] = tuple(sep_grads[l_index])            final_grads = sep_grads            # --------- update context ---------            context_delta = list(map(add_context0, final_grads, context_delta))            # print("curr_shift",len(curr_shift),curr_shift[0])            for p0, p1 in curr_shift:                # print(p0.grad)                p0.stop_gradient = True                p1.stop_gradient = True                p0.grad.zero_()                p1.grad.zero_()                p0.stop_gradient = False                p1.stop_gradient = False            # with paddle.no_grad():            #     for p0, p1 in curr_shift:            #         p0.grad.zero_()            #         p1.grad.zero_()                        new_context = []            for p0, p1 in context:                # new_context.append((p0.detach(), p1.detach()))                new_context.append(shifted_outputs[1][0].__class__(p0.detach(), p1.detach()))            context = new_context        context_delta = [tuple([paddle.to_tensor(x,stop_gradient = False) for x in p_])                            for p_ in context_delta]        context = list(map(add_context, context, context_delta))        new_context = []        for p0, p1 in context:            p0 = p0.detach()            p0.stop_gradient = False            p1 = p1.detach()            p1.stop_gradient = False            # new_context.append((p0, p1))            new_context.append(shifted_outputs[1][0].__class__(p0, p1))        context = new_context        return context    def update_special_tokens_logits(self, context_tokens, i, logits):        for beam_id in range(context_tokens.shape[0]):            for token_idx in set(context_tokens[beam_id][-4:].tolist()):                factor = self.repetition_penalty if logits[beam_id, token_idx] > 0 else (1 / self.repetition_penalty)                logits[beam_id, token_idx] /= factor            if i >= self.ef_idx:                factor = self.end_factor if logits[beam_id, self.end_token] > 0 else (1 / self.end_factor)                logits[beam_id, self.end_token] *= factor            if i == 0:                start_factor = 1.6                factor = start_factor if logits[beam_id, self.end_token] > 0 else (1 / start_factor)                logits[beam_id, self.end_token] /= factor            for token_idx in list(self.forbidden_tokens):                factor = self.forbidden_factor if logits[beam_id, token_idx] > 0 else (1 / self.forbidden_factor)                logits[beam_id, token_idx] /= factor        return logits    def clip_loss(self, probs, context_tokens):        for p_ in self.clip.text_model.parameters():            if p_.grad is not None:                p_.grad.data.zero_()        top_size = 512        _, top_indices = probs.topk(top_size, -1)        # print("417",context_tokens[0])        # print("418",self.lm_tokenizer.decode(context_tokens[0]))        # prefix_texts = [self.lm_tokenizer.decode(x).replace(self.lm_tokenizer.bos_token, '') for x in context_tokens]        prefix_texts = [self.lm_tokenizer.convert_ids_to_string(list(map(int,list(x.numpy())))) for x in context_tokens]        # print(422,prefix_texts)        clip_loss = 0        losses = []        for idx_p in range(probs.shape[0]):            top_texts = []            prefix_text = prefix_texts[idx_p]            for x in top_indices[idx_p]:                top_texts.append(prefix_text + self.lm_tokenizer.convert_ids_to_string(list(map(int,list(x.numpy())))))            text_features = self.get_txt_features(top_texts)            with paddle.no_grad():                similiraties = (self.image_features @ text_features.T)                target_probs = nn.functional.softmax(similiraties / self.clip_loss_temperature, axis=-1).detach()                # print("target_probs",target_probs)                target_probs = target_probs.astype(paddle.float32)            target = paddle.zeros_like(probs[idx_p])            target.stop_gradient = True            # print("target_probs",target_probs.shape,target_probs.stop_gradient)            # print("top_indices",top_indices.stop_gradient,"idx_p",idx_p)            # print("target",target.stop_gradient)            target[top_indices[idx_p]] = target_probs[0]            target = target.unsqueeze(0)            cur_clip_loss = paddle.sum(-(target * paddle.log(probs[idx_p:(idx_p + 1)])))            clip_loss += cur_clip_loss            losses.append(cur_clip_loss)        return clip_loss, losses    # text_generator = CLIPTextGenerator()# image_features = text_generator.get_img_feature(["微信图片_20241026225709.webp"])# cond_text = "Image of a"# beam_size = 5# text_generator.run(image_features,cond_text,beam_size)

[2024-01-29 03:28:13,274] [    INFO] - Already cached /home/aistudio/.paddlenlp/models/ernie_vil-2.0-base-zh/vocab.txt[2024-01-29 03:28:13,302] [    INFO] - tokenizer config file saved in /home/aistudio/.paddlenlp/models/ernie_vil-2.0-base-zh/tokenizer_config.json[2024-01-29 03:28:13,305] [    INFO] - Special tokens file saved in /home/aistudio/.paddlenlp/models/ernie_vil-2.0-base-zh/special_tokens_map.json

{'input_ids': [1, 75, 329, 226, 707, 1358, 2]}

image_path = "微信图片_20241026225709.webp" #请修改对应图片路径text_generator = CLIPTextGenerator(ce_scale=0.2)image_features = text_generator.get_img_feature([image_path])#输入图片路径地址cond_text = "这张图片讲的故事是"# cond_text = "这张图片描述的是（详细描述）："beam_size = 5captions = text_generator.run(image_features,cond_text,beam_size)encoded_captions = [text_generator.clip.get_text_features(paddle.to_tensor(tokenizer(c)["input_ids"]).unsqueeze(0)) for c in captions]encoded_captions = [x / x.norm(axis=-1, keepdim=True) for x in encoded_captions]best_clip_idx = (paddle.concat(encoded_captions) @ image_features.t()).squeeze().argmax().item()print(captions)print('best clip:', cond_text + captions[best_clip_idx])

你是一名 AI 行业编辑，请围绕下面这条热点输出一份资讯解读：
热点：个人实现的反向文心（无需训练的AI看图说话，你不心动？）要求：
1. 先用一句话解释这条热点在讲什么
2. 再总结它为什么重要
3. 说明会影响哪些 AI 产品或内容方向
4. 最后给出 3 个适合资讯站使用的标题