import cv2 import numpy as np from PIL import Image from concurrent.futures import ThreadPoolExecutor from config.configu import * from models.model import * from models.similarity import * from sklearn.cluster import KMeans from utils.utils import * import warnings from typing import Any, List, Optional, Tuple, Union import torch import random import torch.utils.checkpoint import transformers from torch import nn from torch.nn import CrossEntropyLoss from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM, LlamaTokenizer) from transformers.modeling_outputs import CausalLMOutputWithPast from transformers.modeling_utils import PreTrainedModel from transformers.utils import ModelOutput, logging from .configuration_internvl_chat import InternVLChatConfig from .conversation import get_conv_template from .modeling_intern_vit import InternVisionModel from .modeling_internlm2 import InternLM2ForCausalLM logger = logging.get_logger(__name__) def coord_transform(box,return_4=True): if return_4: return [box[0][0],box[0][1],box[1][0],box[1][1]] else: return [[box[0],box[1]],[box[2],box[3]]] def insert_zeros(input_ids, attention_mask, num_zeros=5): device = input_ids.device # 获取原始设备 input_ids = input_ids.cpu().clone() # 将张量移到 CPU 并克隆 attention_mask = attention_mask.cpu().clone() # 将张量移到 CPU 并克隆 for _ in range(num_zeros): # 随机选择插入位置 insert_pos = random.randint(0, input_ids.size(1)) # 在 input_ids 中插入 0 input_ids = torch.cat((input_ids[:, :insert_pos], torch.tensor([[0]]), input_ids[:, insert_pos:]), dim=1) # 在 attention_mask 中插入 1 attention_mask = torch.cat((attention_mask[:, :insert_pos], torch.tensor([[1]]), attention_mask[:, insert_pos:]), dim=1) # 将张量移回原始设备 input_ids = input_ids.to(device) attention_mask = attention_mask.to(device) return input_ids, attention_mask def add_Gaussian_noise(input_embeds, rate=1e-1): device = input_embeds.device input_embeds = input_embeds.cpu().clone() mean = input_embeds.mean() std = input_embeds.std() noise = torch.randn(input_embeds.size()) * std + mean noisy_input_embeds = input_embeds + rate * noise noisy_input_embeds = noisy_input_embeds.to(device) noisy_input_embeds = noisy_input_embeds.to(torch.bfloat16) return noisy_input_embeds def version_cmp(v1, v2, op='eq'): import operator from packaging import version op_func = getattr(operator, op) return op_func(version.parse(v1), version.parse(v2)) def most_frequent_rgb(image_array): """找一张图片中最frequent的rgb,用于填充mask""" # Flatten the image array to a 2D array where each row is an RGB tuple pixels = image_array.reshape(-1, image_array.shape[-1]) # Use np.unique with return_counts to find unique rows and their counts unique_pixels, counts = np.unique(pixels, axis=0, return_counts=True) # Find the index of the most frequent pixel most_frequent_index = np.argmax(counts) # Get the most frequent pixel and its count most_frequent_pixel = unique_pixels[most_frequent_index] frequency = counts[most_frequent_index] return most_frequent_pixel, frequency def most_frequent_rgb_fast(image_array): """快速查找图片中最频繁的RGB值，不返回频率""" # 将RGB每个通道的值映射为一个唯一的整数，形如 R * 256^2 + G * 256 + B flattened = image_array.reshape(-1, 3) rgb_ints = flattened[:, 0] * 256**2 + flattened[:, 1] * 256 + flattened[:, 2] # 使用np.bincount统计每个唯一RGB组合出现的次数 counts = np.bincount(rgb_ints) # 找到出现次数最多的那个整数 most_frequent_index = np.argmax(counts) # 将整数转换回RGB值 r = (most_frequent_index // 256**2) % 256 g = (most_frequent_index // 256) % 256 b = most_frequent_index % 256 return (r, g, b) def mask_area(image_array,coords,color): """对一张图片在框定的一系列box进行mask""" # Define the bounding box (x1, y1, x2, y2) #color=average_rgb(modified_image) for coord in coords: x1, y1, x2, y2 = coord image_array[y1:y2, x1:x2] =color # 255 for white in an RGB image return image_array class InternVLChatModel(PreTrainedModel): config_class = InternVLChatConfig main_input_name = 'pixel_values' _supports_flash_attn_2 = True _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'InternLM2DecoderLayer'] def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None): super().__init__(config) assert version_cmp(transformers.__version__, '4.36.2', 'ge') image_size = config.force_image_size or config.vision_config.image_size patch_size = config.vision_config.patch_size self.patch_size = patch_size self.select_layer = config.select_layer self.template = config.template ##TODO change the number of img tokens self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2)) #self.num_image_token = 3 self.downsample_ratio = config.downsample_ratio self.ps_version = config.ps_version self.mu_sigma=torch.load(NORM_PARAMS_PATH)['weight'] self.mu=self.mu_sigma[:,0].reshape((-1,1)) self.sigma=self.mu_sigma[:,1].reshape((-1,1)) #[vocab_size, 1] self.normed_emb,self.mu_sigma=self.load_normed_tok_embeddings(load_checkboard=True) self.resampler=load_perceiver_resampler_2(PERCEIVER_CHECKPOINT,num_layers=4) self.sorter=load_orderformer(ORDERFORMER_CHECKPOINT) logger.info(f'num_image_token: {self.num_image_token}') logger.info(f'ps_version: {self.ps_version}') # print('vision_model', vision_model) # print('language_model', language_model) # print('config.llm_config.architectures[0]', config.llm_config.architectures[0]) if vision_model is not None: self.vision_model = vision_model else: self.vision_model = InternVisionModel(config.vision_config) if language_model is not None: self.language_model = language_model else: if config.llm_config.architectures[0] == 'LlamaForCausalLM': self.language_model = LlamaForCausalLM(config.llm_config) elif config.llm_config.architectures[0] == 'InternLM2ForCausalLM': self.language_model = InternLM2ForCausalLM(config.llm_config) else: raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.') vit_hidden_size = config.vision_config.hidden_size llm_hidden_size = config.llm_config.hidden_size self.mlp1 = nn.Sequential( nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2), nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size), nn.GELU(), nn.Linear(llm_hidden_size, llm_hidden_size) ) self.img_context_token_id = None self.conv_template = get_conv_template(self.template) self.system_message = self.conv_template.system_message def load_normed_tok_embeddings(self,vocab_size=92553, llm_hidden_size=4096,load_checkboard=False): tok_embeddings = nn.Embedding(vocab_size, llm_hidden_size, padding_idx=2).to_empty(device=torch.device('cuda')).to(torch.bfloat16) tok_embeddings.load_state_dict(torch.load(NORM_TOK_EMBEDDING_PATH, weights_only=True, map_location="cpu")) if load_checkboard: checkboard_norm=torch.load(NORM_PARAMS_PATH) # (voc_size, 2) mu sigma pred * sigma + mu (逐行) return tok_embeddings,checkboard_norm['weight'] return tok_embeddings def forward( self, pixel_values: torch.FloatTensor, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, image_flags: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: return_dict = return_dict if return_dict is not None else self.config.use_return_dict image_flags = image_flags.squeeze(-1) input_embeds = self.language_model.get_input_embeddings()(input_ids) vit_embeds = self.extract_feature(pixel_values) vit_embeds = vit_embeds[image_flags == 1] vit_batch_size = pixel_values.shape[0] B, N, C = input_embeds.shape input_embeds = input_embeds.reshape(B * N, C) if torch.distributed.get_rank() == 0: print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}') input_ids = input_ids.reshape(B * N) selected = (input_ids == self.img_context_token_id) try: input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C) except Exception as e: vit_embeds = vit_embeds.reshape(-1, C) print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, ' f'vit_embeds.shape={vit_embeds.shape}') n_token = selected.sum() input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token] input_embeds = input_embeds.reshape(B, N, C) outputs = self.language_model( inputs_embeds=input_embeds, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) logits = outputs.logits loss = None if labels is not None: # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) loss = loss_fct(shift_logits, shift_labels) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def pixel_shuffle(self, x, scale_factor=0.5): n, w, h, c = x.size() # N, W, H, C --> N, W, H * scale, C // scale x = x.view(n, w, int(h * scale_factor), int(c / scale_factor)) # N, W, H * scale, C // scale --> N, H * scale, W, C // scale x = x.permute(0, 2, 1, 3).contiguous() # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2) x = x.view(n, int(h * scale_factor), int(w * scale_factor), int(c / (scale_factor * scale_factor))) if self.ps_version == 'v1': warnings.warn("In ps_version 'v1', the height and width have not been swapped back, " 'which results in a transposed image.') else: x = x.permute(0, 2, 1, 3).contiguous() return x def extract_feature(self, pixel_values): if self.select_layer == -1: vit_embeds = self.vision_model( pixel_values=pixel_values, output_hidden_states=False, return_dict=True).last_hidden_state else: vit_embeds = self.vision_model( pixel_values=pixel_values, output_hidden_states=True, return_dict=True).hidden_states[self.select_layer] vit_embeds = vit_embeds[:, 1:, :] h = w = int(vit_embeds.shape[1] ** 0.5) vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1) vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio) vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1]) vit_embeds = self.mlp1(vit_embeds) return vit_embeds @torch.no_grad() def calli_align(self,img_path,detect_model, drop_zero = False, use_hard_vector_quant=False,save_path=None,verbose=False): def dynamic_read(img_path,mode='c'): # 如果是字符串类型（文件路径），用 cv2 读取 if isinstance(img_path, str): img = cv2.imread(img_path) if img is None: try: img = Image.open(img_path).convert("RGB") img = np.array(img) except: raise ValueError(f"Image at path {img_path} could not be loaded.") # 如果是 PIL.Image.Image 类型，将其转为 cv2 格式 elif isinstance(img_path, Image.Image): img = np.array(img_path) # PIL 转 numpy 数组 # 因为 OpenCV 是 BGR，需要将 RGB 转为 BGR img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR) else: raise TypeError(f"Unsupported image type: {type(img_path)}") if mode=='i': img=Image.fromarray(img).convert("RGB") return img import time def iterative_only_boxes(model,jpg_path): image = dynamic_read(jpg_path) image_array = np.array(image) h, w, channels = image.shape boxes=[] color=most_frequent_rgb_fast(image_array) while True: res=model(image_array,verbose=False)[0] to_be_masked=[] for box in res.boxes: xyxy = box.xyxy.squeeze().tolist() x1, y1, x2, y2 = int(xyxy[0]), int(xyxy[1]), int(xyxy[2]), int(xyxy[3]) to_be_masked.append([x1,y1,x2,y2]) boxes.extend(to_be_masked) if len(to_be_masked)>250: image_array=mask_area(image_array,to_be_masked,color) else: break boxes=[[[max(item[0],0),max(item[1],0)],[min(item[2],w),min(item[3],h)]]for item in boxes] i=0 length=len(boxes) while i0.8: rm = boxes[j] boxes.remove(rm) if j (group_0[0][1][0] - group_0[0][0][0]):# and len(group_1) > 0.8*len(group_0): # 1 为正文，0为落款 g1_hs = np.array([x[1][1]-x[0][1] for x in group_1]).mean() thr1 = 1*( group_1[-1][1][0] - group_1[-1][0][0]) thr2 = 0.8*g1_hs #luokuan_mean_area = np.array([(ele[1][0] - ele[0][0])*(ele[1][1] - ele[0][1]) for ele in group_0]).mean() new_0 = [] for ele in group_0: if (ele[1][0] - ele[0][0]) >= thr1 or (ele[1][1] - ele[0][1]) >= thr2 or (areas.min()/(ele[1][0] - ele[0][0])*(ele[1][1] - ele[0][1]) <= 1/5 and areas.mean() / ((ele[1][0] - ele[0][0])*(ele[1][1] - ele[0][1])) <= 1.3): group_1.append(ele) else: new_0.append(ele) grouped_luokuan = merge_boxes(new_0.copy()) final_ = [] for ele in new_0: if ele in grouped_luokuan: group_1.append(ele) else: final_.append(ele) group_0 = final_ elif (group_0[0][1][0] - group_0[0][0][0]) > (group_1[0][1][0] - group_1[0][0][0]):# and len(group_0) > 0.8*len(group_1): g0_hs = np.array([x[1][1]-x[0][1] for x in group_0]).mean() thr1 = 1*( group_0[-1][1][0] - group_0[-1][0][0]) thr2 = 0.8*g0_hs #luokuan_mean_area = np.array([(ele[1][0] - ele[0][0])*(ele[1][1] - ele[0][1]) for ele in group_1]).mean() new_1 = [] for ele in group_1: if (ele[1][0] - ele[0][0]) >= thr1 or (ele[1][1] - ele[0][1]) >= thr2 or (areas.min()/(ele[1][0] - ele[0][0])*(ele[1][1] - ele[0][1]) <= 1/5 and areas.mean() / ((ele[1][0] - ele[0][0])*(ele[1][1] - ele[0][1])) <=1.3): group_0.append(ele) else: new_1.append(ele) grouped_luokuan = merge_boxes(new_1.copy()) final_ = [] for ele in new_1: if ele in grouped_luokuan: group_0.append(ele) else: final_.append(ele) group_1 = final_ return group_0,group_1 def toint(lst): if len(lst)==2: return [[int(lst[0][0]),int(lst[0][1])],[int(lst[1][0]),int(lst[1][1])]] else: return [int(lst[0]),int(lst[1]),int(lst[2]),int(lst[3])] img = dynamic_read(jpg_path) h, w, channels = img.shape normalized_boxes=[[[item[0][0]/w,item[0][1]/h],[item[1][0]/w,item[1][1]/h]] for item in boxes] S=np.array([(item[0][0]-item[1][0])*(item[0][1]-item[1][1]) for item in normalized_boxes]) # print(np.max(S)-np.min(S),h,w) # print(boxes) # print(normalized_boxes) coef_var=np.std(S)/np.mean(S) boxes2class=None col2class=None if coef_var>0.66 and S.min()/S.mean() <= 1/8: boxes1,boxes2=kmeans_boxes(normalized_boxes) boxes1=[[[item[0][0]*w,item[0][1]*h],[item[1][0]*w,item[1][1]*h]] for item in boxes1] boxes2=[[[item[0][0]*w,item[0][1]*h],[item[1][0]*w,item[1][1]*h]] for item in boxes2] columns1=merge_boxes(boxes1.copy()) columns2=merge_boxes(boxes2.copy()) columns=columns1+columns2 boxes2class={1:[toint(item) for item in boxes1],2:[toint(item) for item in boxes2]} col2class={1:[toint(item) for item in columns1],2:[toint(item) for item in columns2]} #[[481.3252033886607, 1185.3073037637248], [748.9909909909909, 1616.216216216216]] else: columns=merge_boxes(boxes.copy()) results={"imageHeight":h,"imageWidth":w,"shapes":[{"points":toint(col)} for col in columns], "boxes2class":boxes2class,"col2class":col2class} #print("saving results...") # if verbose: # frame = dynamic_read(jpg_path) # name=jpg_path.split("/")[-1] # os.makedirs("./detect_boxes_char2col/result_merge", exist_ok=True) # for i,box in enumerate(results['shapes']): # xyxy = box['points'] # x1, y1, x2, y2 = int(xyxy[0][0]), int(xyxy[0][1]), int(xyxy[1][0]), int(xyxy[1][1]) # colo = (255,0,0) # cv2.rectangle(frame, (x1, y1), (x2, y2), thickness=2,color=colo,lineType=cv2.LINE_AA) # # put labels # if boxes2class is not None: # if xyxy in col2class[1]: # cv2.putText(frame, str(1), ((x1+x2)//2, (y1+y2)//2), cv2.FONT_HERSHEY_SIMPLEX, 1.5, colo, thickness=2, lineType=cv2.LINE_AA) # elif xyxy in col2class[2]: # cv2.putText(frame, str(2), ((x1+x2)//2, (y1+y2)//2), cv2.FONT_HERSHEY_SIMPLEX, 1.5, (0, 30, 235), thickness=2, lineType=cv2.LINE_AA) # #cv2.putText(frame, str(i+1), ((x1+x2)//2, (y1+y2)//2), cv2.FONT_HERSHEY_SIMPLEX, 1.5, colo, thickness=2, lineType=cv2.LINE_AA) # cv2.imwrite("./detect_boxes_char2col/result_merge"+name,frame) return results def sort_boxes(jpg,detector,model,thres=0.8): boxes=iterative_only_boxes(detector,jpg) data=char2col_with_kmeans(jpg,boxes,verbose=False) res=model.predict(data,jpg) final_results=[] for idx,col in res.items(): lst=[] for item in boxes: ratio=calculate_iou(col,[item[0][0],item[0][1],item[1][0],item[1][1]],mini=True) if ratio>=thres: lst.append([item[0][0],item[0][1],item[1][0],item[1][1]]) lst=sorted(lst, key=lambda item: (item[1]+item[3])/2) final_results.extend(lst) #print(len(boxes),len(res),len(final_results)) return final_results if img_path is None: return None,None st=time.time() boxes=sort_boxes(img_path,detect_model,self.sorter) ed=time.time() if verbose: print(f"YOLO+Orderformer {ed-st:.2f}s") if save_path!=None: frame = dynamic_read(img_path) name=img_path.split("/")[-1] for i,box in enumerate(boxes): xyxy = box x1, y1, x2, y2 = int(xyxy[0]), int(xyxy[1]), int(xyxy[2]), int(xyxy[3]) colo = (255,0,0) cv2.rectangle(frame, (x1, y1), (x2, y2), thickness=2,color=colo,lineType=cv2.LINE_AA) # put labels cv2.putText(frame, str(i+1), ((x1+x2)//2, (y1+y2)//2), cv2.FONT_HERSHEY_SIMPLEX, 1.5, colo, thickness=2, lineType=cv2.LINE_AA) print(save_path+"oredered_result_"+name) cv2.imwrite(save_path+"oredered_result_"+name,frame) st=time.time() pixel_values=[] img=np.array(dynamic_read(img_path,mode='i').convert("RGB")) for xyxy in boxes: x1, y1, x2, y2 = int(xyxy[0]), int(xyxy[1]), int(xyxy[2]), int(xyxy[3]) sub_img=Image.fromarray(img[y1:y2,x1:x2]) pixel_values.append(load_image_2(sub_img).to(torch.bfloat16).cuda()) ed1=time.time() results=torch.cat(pixel_values) image_embeddings=self.extract_feature(results) ed2=time.time() output=self.resampler(image_embeddings) ed3=time.time() #TODO 可以indices转换回去 outs=vq_cos_sim(self.normed_emb,output, use_hard_vector_quant) #(B, 3) #如果use_vq的话现在改成dynamic: 对于max cos_sim小于等于thresh的，使用向量量化进行替换 ed4=time.time() if verbose: print(f"Get pixel values {ed1-st:.2f}s") print(f"extract feat {ed2-ed1:.2f}s") print(f"Resampler forward {ed3-ed2:.2f}") print(f"vq cos sim {ed4-ed3:.2f}s") if use_hard_vector_quant: indices, cos_sim_values = outs #### DEFINE THRESH!!! thresh = 0.5 else: indices = outs if use_hard_vector_quant: print("Dynamic vector quantization...") below_mask = (cos_sim_values <= thresh).to(torch.bfloat16).unsqueeze(-1) output = output * (1-below_mask) + self.normed_emb.weight[indices] * below_mask flattened_output = output.view(-1, output.shape[-1]) flattened_indices = indices.view(-1) if drop_zero: filtered_indices=flattened_indices[flattened_indices!=0] filtered_output=flattened_output[flattened_indices!=0] sigma_flat = self.sigma[filtered_indices] # 形状 (183 * 3, 1) mu_flat = self.mu[filtered_indices] sigma_flat = sigma_flat.expand(-1, filtered_output.shape[-1]) mu_flat = mu_flat.expand(-1, filtered_output.shape[-1]) back_to_origin_flat = filtered_output * sigma_flat + mu_flat else: sigma_flat = self.sigma[flattened_indices] mu_flat = self.mu[flattened_indices] sigma_flat = sigma_flat.expand(-1, flattened_output.shape[-1]) mu_flat = mu_flat.expand(-1, flattened_output.shape[-1]) back_to_origin_flat = flattened_output * sigma_flat + mu_flat return back_to_origin_flat, indices def find_coordinates(self,text): import re numbers = re.findall(r'\d+', text) numbers = [int(num) for num in numbers] # 如果需要浮点数，可以用 float() return numbers def chat_ocr(self, tokenizer, detect_model,img_path, questions, generation_config, num_patches_list=None, history=None, return_history=False, IMG_START_TOKEN='', IMG_END_TOKEN='', IMG_CONTEXT_TOKEN='', ALIGNED_TOKEN="[UNUSED_TOKEN_140]",verbose=False, image_counts=None,batch=False, use_p=True, drop_zero=False, hard_vq=False, repetition_penalty=1.5,region_wise=False): pixel_values = None if img_path is not None: try: if region_wise: img=np.array(Image.open(img_path).convert("RGB")) coord=self.find_coordinates(questions) x1,x2,y1,y2=coord sub_img=Image.fromarray(img[y1:y2,x1:x2]) questions="输出图片中所有文字:" pixel_values=load_image(sub_img).to(torch.bfloat16).to(torch.device("cuda")) else: pixel_values=load_image(img_path).to(torch.bfloat16).to(torch.device("cuda")) except: raise FileNotFoundError if use_p: import time st=time.time() if region_wise: try: out_tokens, indices =self.calli_align(sub_img,detect_model, drop_zero = drop_zero, use_hard_vector_quant=hard_vq,verbose=verbose) except: return "检测失败" else: out_tokens, indices =self.calli_align(img_path,detect_model, drop_zero = drop_zero, use_hard_vector_quant=hard_vq,verbose=verbose) #,tokenizer=tokenizer) if verbose: print(f"Calli Align: {time.time()-st:.2f}s") # 删掉多余0 # indices 备用，因为我们也想未来看仅使用calliAlign效果 if pixel_values is None: question=questions if pixel_values is not None and '' not in questions: question = '\n' + questions #question = questions elif history is None and pixel_values is None: question=questions elif '' in questions: question=questions if history is None and use_p and '[UNUSED_TOKEN_140]' not in question: question =question+'[UNUSED_TOKEN_140]'*out_tokens.shape[0] if num_patches_list is None: num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else [] assert pixel_values is None or len(pixel_values) == sum(num_patches_list) img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) self.img_context_token_id = img_context_token_id template = get_conv_template(self.template) template.system_message = self.system_message eos_token_id = tokenizer.convert_tokens_to_ids(template.sep) history = [] if history is None else history for (old_question, old_answer) in history: template.append_message(template.roles[0], old_question) template.append_message(template.roles[1], old_answer) template.append_message(template.roles[0], question) template.append_message(template.roles[1], None) query = template.get_prompt() for num_patches in num_patches_list: image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN query = query.replace('', image_tokens, 1) model_inputs = tokenizer(query, return_tensors='pt') input_ids = model_inputs['input_ids'].cuda() attention_mask = model_inputs['attention_mask'].cuda() generation_config['eos_token_id'] = eos_token_id if use_p: generation_output = self.generate_ocr( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, reference_embeds=out_tokens, repetition_penalty=repetition_penalty, **generation_config ) else: generation_output = self.generate_ocr( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, repetition_penalty=repetition_penalty, **generation_config ) response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0] response = response.split(template.sep)[0].strip() history.append((question, response)) if return_history: return response, history else: query_to_print = query.replace(IMG_CONTEXT_TOKEN, '') query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '') return response def dynamic_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None, history=None, return_history=False, IMG_START_TOKEN='', IMG_END_TOKEN='', IMG_CONTEXT_TOKEN='', verbose=False, image_counts=None,batch=False,use_p=True): if use_p: self.num_image_token=3 if batch: assert isinstance(questions,list) and len(questions)>0 and isinstance(questions[0],str) if history is not None or return_history: print('Now multi-turn chat is not supported in batch_chat.') raise NotImplementedError if image_counts is not None: num_patches_list = image_counts print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.') img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) self.img_context_token_id = img_context_token_id if verbose and pixel_values is not None: image_bs = pixel_values.shape[0] print(f'dynamic ViT batch size: {image_bs}') queries = [] for idx, num_patches in enumerate(num_patches_list): question = questions[idx] if pixel_values is not None and '' not in question: question = '\n' + question template = get_conv_template(self.template) template.append_message(template.roles[0], question) template.append_message(template.roles[1], None) query = template.get_prompt() image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN query = query.replace('', image_tokens, 1) queries.append(query) # print(query) tokenizer.padding_side = 'left' model_inputs = tokenizer(queries, return_tensors='pt', padding=True) input_ids = model_inputs['input_ids'].cuda() attention_mask = model_inputs['attention_mask'].cuda() eos_token_id = tokenizer.convert_tokens_to_ids(template.sep) generation_config['eos_token_id'] = eos_token_id if use_p: generation_output = self.generate( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, **generation_config ) else: generation_output = self.generate_origin( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, **generation_config ) responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True) responses = [response.split(template.sep)[0].strip() for response in responses] return responses else: assert isinstance(questions,str) if num_patches_list is None: num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else [] assert pixel_values is None or len(pixel_values) == sum(num_patches_list) img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) self.img_context_token_id = img_context_token_id template = get_conv_template(self.template) template.system_message = self.system_message eos_token_id = tokenizer.convert_tokens_to_ids(template.sep) history = [] if history is None else history for (old_question, old_answer) in history: template.append_message(template.roles[0], old_question) template.append_message(template.roles[1], old_answer) template.append_message(template.roles[0], questions) template.append_message(template.roles[1], None) query = template.get_prompt() if verbose and pixel_values is not None: image_bs = pixel_values.shape[0] print(f'dynamic ViT batch size: {image_bs}') # print('num_image_token', self.num_image_token) # print('num_patches_list', num_patches_list) query=f"""<|im_start|>system你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。<|im_end|>\n<|im_start|>user{questions}""" query = query+'' for num_patches in num_patches_list: #image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN image_tokens = IMG_CONTEXT_TOKEN * self.num_image_token #print('tokens_num', len(image_tokens)) query = query.replace('', image_tokens, 1) query+="<|im_end|>\n<|im_start|>assistant" # print(query) model_inputs = tokenizer(query, return_tensors='pt') input_ids = model_inputs['input_ids'].cuda() attention_mask = model_inputs['attention_mask'].cuda() generation_config['eos_token_id'] = eos_token_id if use_p: generation_output = self.generate( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, **generation_config ) else: generation_output = self.generate_origin( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, **generation_config ) response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0] response = response.split(template.sep)[0].strip() history.append((questions, response)) if return_history: return response, history else: query_to_print = query.replace(IMG_CONTEXT_TOKEN, '') query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '') if verbose: print(query_to_print, response) return response def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None, history=None, return_history=False, IMG_START_TOKEN='', IMG_END_TOKEN='', IMG_CONTEXT_TOKEN='', verbose=False, image_counts=None): if history is not None or return_history: print('Now multi-turn chat is not supported in batch_chat.') raise NotImplementedError if image_counts is not None: num_patches_list = image_counts print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.') img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) self.img_context_token_id = img_context_token_id if verbose and pixel_values is not None: image_bs = pixel_values.shape[0] print(f'dynamic ViT batch size: {image_bs}') queries = [] for idx, num_patches in enumerate(num_patches_list): question = questions[idx] if pixel_values is not None and '' not in question: question = '\n' + question template = get_conv_template(self.template) template.append_message(template.roles[0], question) template.append_message(template.roles[1], None) query = template.get_prompt() image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN query = query.replace('', image_tokens, 1) queries.append(query) # print(query) tokenizer.padding_side = 'left' model_inputs = tokenizer(queries, return_tensors='pt', padding=True) input_ids = model_inputs['input_ids'].cuda() attention_mask = model_inputs['attention_mask'].cuda() eos_token_id = tokenizer.convert_tokens_to_ids(template.sep) generation_config['eos_token_id'] = eos_token_id generation_output = self.generate_origin( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, **generation_config ) responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True) responses = [response.split(template.sep)[0].strip() for response in responses] return responses #When call internvl,this func is called def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False, num_patches_list=None, IMG_START_TOKEN='', IMG_END_TOKEN='', IMG_CONTEXT_TOKEN='', verbose=False): #self.num_image_token=3 # original_question = question if history is None and pixel_values is not None and '' not in question: question = '\n' + question if num_patches_list is None: num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else [] assert pixel_values is None or len(pixel_values) == sum(num_patches_list) img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN) self.img_context_token_id = img_context_token_id template = get_conv_template(self.template) template.system_message = self.system_message eos_token_id = tokenizer.convert_tokens_to_ids(template.sep) history = [] if history is None else history for (old_question, old_answer) in history: template.append_message(template.roles[0], old_question) template.append_message(template.roles[1], old_answer) template.append_message(template.roles[0], question) template.append_message(template.roles[1], None) query = template.get_prompt() if verbose and pixel_values is not None: image_bs = pixel_values.shape[0] for num_patches in num_patches_list: image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN query = query.replace('', image_tokens, 1) print(num_patches,self.num_image_token) print(pixel_values.shape[0]) model_inputs = tokenizer(query, return_tensors='pt') input_ids = model_inputs['input_ids'].cuda() attention_mask = model_inputs['attention_mask'].cuda() generation_config['eos_token_id'] = eos_token_id generation_output = self.generate_origin( pixel_values=pixel_values, input_ids=input_ids, attention_mask=attention_mask, **generation_config ) response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0] response = response.split(template.sep)[0].strip() history.append((question, response)) if return_history: return response, history else: query_to_print = query.replace(IMG_CONTEXT_TOKEN, '') query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '') if verbose: print(query_to_print, response) return response @torch.no_grad() def generate_origin( self, pixel_values: Optional[torch.FloatTensor] = None, input_ids: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.LongTensor] = None, visual_features: Optional[torch.FloatTensor] = None, generation_config: Optional[GenerationConfig] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, **generate_kwargs, ) -> torch.LongTensor: assert self.img_context_token_id is not None if pixel_values is not None: if visual_features is not None: vit_embeds = visual_features else: vit_embeds = self.extract_feature(pixel_values) input_embeds = self.language_model.get_input_embeddings()(input_ids) B, N, C = input_embeds.shape input_embeds = input_embeds.reshape(B * N, C) input_ids = input_ids.reshape(B * N) selected = (input_ids == self.img_context_token_id) assert selected.sum() != 0 input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device) print("ID: ",self.img_context_token_id) input_embeds = input_embeds.reshape(B, N, C) else: input_embeds = self.language_model.get_input_embeddings()(input_ids) outputs = self.language_model.generate( inputs_embeds=input_embeds, attention_mask=attention_mask, generation_config=generation_config, output_hidden_states=output_hidden_states, return_dict=return_dict, use_cache=True, **generate_kwargs, ) return outputs @torch.no_grad() def generate_ocr( self, pixel_values: Optional[torch.FloatTensor] = None, input_ids: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.LongTensor] = None, visual_features: Optional[torch.FloatTensor] = None, generation_config: Optional[GenerationConfig] = None, reference_embeds=None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, repetition_penalty=1.5, **generate_kwargs, ) -> torch.LongTensor: assert self.img_context_token_id is not None if pixel_values is not None: if visual_features is not None: vit_embeds = visual_features else: vit_embeds = self.extract_feature(pixel_values) input_embeds = self.language_model.get_input_embeddings()(input_ids) B, N, C = input_embeds.shape input_embeds = input_embeds.reshape(B * N, C) input_ids = input_ids.reshape(B * N) selected = (input_ids == self.img_context_token_id) assert selected.sum() != 0 input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device) if reference_embeds is not None: selected = (input_ids == 92537) assert selected.sum() != 0 input_embeds[selected] =reference_embeds.reshape(-1, C).to(input_embeds.device) input_embeds = input_embeds.reshape(B, N, C) else: input_embeds = self.language_model.get_input_embeddings()(input_ids) outputs = self.language_model.generate( inputs_embeds=input_embeds, attention_mask=attention_mask, generation_config=generation_config, output_hidden_states=output_hidden_states, return_dict=return_dict, use_cache=True, repetition_penalty=repetition_penalty, **generate_kwargs, ) return outputs @torch.no_grad() def generate( self, pixel_values: Optional[torch.FloatTensor] = None, input_ids: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.LongTensor] = None, visual_features: Optional[torch.FloatTensor] = None, generation_config: Optional[GenerationConfig] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, **generate_kwargs, ) -> torch.LongTensor: assert self.img_context_token_id is not None if pixel_values is not None: if visual_features is not None: vit_embeds = visual_features else: vit_embeds = self.extract_feature(pixel_values) input_embeds = self.language_model.get_input_embeddings()(input_ids) vit_embeds = self.resampler(vit_embeds) mu=self.mu_sigma[:,0].reshape((-1,1)) sigma=self.mu_sigma[:,1].reshape((-1,1)) indices=vq_cos_sim(self.normed_emb,vit_embeds).reshape((-1,)) vit_embeds=vit_embeds.reshape((-1,vit_embeds.shape[-1]))*sigma[indices][:]+mu[indices][:] B, N, C = input_embeds.shape input_embeds = input_embeds.reshape(B * N, C) input_ids = input_ids.reshape(B * N) selected = (input_ids == self.img_context_token_id) assert selected.sum() != 0 input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device) input_embeds = input_embeds.reshape(B, N, C) else: input_embeds = self.language_model.get_input_embeddings()(input_ids) outputs = self.language_model.generate( inputs_embeds=input_embeds, attention_mask=attention_mask, generation_config=generation_config, output_hidden_states=output_hidden_states, return_dict=return_dict, use_cache=True, **generate_kwargs, ) return outputs