from dataclasses import dataclass from typing import List, Optional, Tuple, Union, Dict from copy import deepcopy import re import math import torch import torch.utils.checkpoint from torch import nn import matplotlib.pyplot as plt from transformers import PreTrainedModel from transformers.activations import ACT2FN from transformers.cache_utils import Cache from transformers.modeling_outputs import ModelOutput from transformers.utils import ( add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings, ) from transformers.models.auto import AutoModel, AutoModelForCausalLM from .configuration_gecko import GeckoConfig logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "GeckoConfig" @dataclass class GeckoCausalLMOutputWithPast(ModelOutput): """ Base class for Llava causal language model (or autoregressive) outputs. Args: loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction). logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(torch.FloatTensor)`, *optional*): Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver """ loss: Optional[torch.FloatTensor] = None logits: torch.FloatTensor = None past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None hidden_states: Optional[Tuple[torch.FloatTensor]] = None attentions: Optional[Tuple[torch.FloatTensor]] = None image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None class GeckoPreTrainedModel(PreTrainedModel): config_class = GeckoConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["GeckoVisionAttention"] _skip_keys_device_placement = "past_key_values" _supports_flash_attn_2 = True def _init_weights(self, module): std = ( self.config.intializer_range if hasattr(self.config, "intializer_range") else self.config.text_config.initializer_range ) if hasattr(module, "class_embedding"): module.class_embedding.data.normal_(mean=0.0, std=std) if isinstance(module, (nn.Linear, nn.Conv2d)): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() @property def _supports_sdpa(self): return self.language_model._supports_sdpa class PositionalEncoding2D(nn.Module): def __init__(self, config: GeckoConfig): """ :param channels: The last dimension of the tensor you want to apply pos emb to. """ super(PositionalEncoding2D, self).__init__() if config.positional_information == "2d_before": channels = config.vision_config.hidden_size else: channels = config.text_config.hidden_size self.org_channels = channels channels = int(math.ceil(channels / 4) * 2) self.channels = channels inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2).float() / channels)) self.register_buffer("inv_freq", inv_freq) self.register_buffer("cached_penc", None, persistent=False) def get_emb(self, sin_inp): """ Gets a base embedding for one dimension with sin and cos intertwined """ emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1) return torch.flatten(emb, -2, -1) def forward(self, tensor): """ :param tensor: A 4d tensor of size (x, y, num_tokens, ch) :return: Positional Encoding Matrix of size (x, y, num_tokens, ch) """ if len(tensor.shape) != 4: raise RuntimeError("The input tensor has to be 4d!") if self.cached_penc is not None and self.cached_penc.shape == tensor.shape: return self.cached_penc self.cached_penc = None x, y, num_tokens, orig_ch = tensor.shape pos_x = torch.arange(x, device=tensor.device, dtype=self.inv_freq.dtype) pos_y = torch.arange(y, device=tensor.device, dtype=self.inv_freq.dtype) sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq) sin_inp_y = torch.einsum("i,j->ij", pos_y, self.inv_freq) emb_x = self.get_emb(sin_inp_x).unsqueeze(1) emb_y = self.get_emb(sin_inp_y) emb = torch.zeros( (x, y, self.channels * 2), device=tensor.device, dtype=tensor.dtype, ) emb[:, :, : self.channels] = emb_x emb[:, :, self.channels : 2 * self.channels] = emb_y self.cached_penc = emb[:, :, None, :orig_ch].repeat(1, 1, num_tokens, 1) return self.cached_penc class GeckoMultiModalProjector(nn.Module): def __init__(self, config: GeckoConfig): super().__init__() self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True) self.act = ACT2FN[config.projector_hidden_act] self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True) def forward(self, image_features): hidden_states = self.linear_1(image_features) hidden_states = self.act(hidden_states) hidden_states = self.linear_2(hidden_states) return hidden_states class GeckoForConditionalGeneration(GeckoPreTrainedModel): def __init__(self, config: GeckoConfig, vision_tower=None, language_model=None, multimodal_projector=None): super().__init__(config) self.vision_tower = AutoModel.from_config(config.vision_config) if vision_tower is None else vision_tower self.positional_encoding = PositionalEncoding2D(config) if '2d' in config.positional_information else None self.multi_modal_projector = GeckoMultiModalProjector(config) self.vocab_size = config.vocab_size self.language_model = AutoModelForCausalLM.from_config( config.text_config, attn_implementation=config._attn_implementation ) if language_model is None else language_model self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1 self.post_init() def load_text_encoder(self, processor): self.tokenizer = processor.tokenizer self.clip_tokenizer = processor.clip_tokenizer self.eos_token_id = [self.tokenizer.eos_token_id, self.tokenizer.convert_tokens_to_ids("<|eot_id|>")] self.encoder_type = self.config.vision_config.model_type if 'clip' in self.encoder_type: self.encoder = AutoModel.from_pretrained('openai/clip-vit-large-patch14-336', torch_dtype=self.dtype, device_map=self.device) elif 'siglip' in self.encoder_type: self.encoder = AutoModel.from_pretrained("google/siglip-so400m-patch14-384", torch_dtype=self.dtype, device_map=self.device) else: raise ValueError(f"Vision model {self.config.vision_config.model_type} is not supported.") def get_input_embeddings(self): return self.language_model.get_input_embeddings() def set_input_embeddings(self, value): self.language_model.set_input_embeddings(value) def get_output_embeddings(self): return self.language_model.get_output_embeddings() def set_output_embeddings(self, new_embeddings): self.language_model.set_output_embeddings(new_embeddings) def set_decoder(self, decoder): self.language_model.set_decoder(decoder) def get_decoder(self): return self.language_model.get_decoder() def tie_weights(self): return self.language_model.tie_weights() def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding: model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of) # update vocab size self.config.text_config.vocab_size = model_embeds.num_embeddings self.config.vocab_size = model_embeds.num_embeddings self.vocab_size = model_embeds.num_embeddings return model_embeds # def _get_highest_similarity(self, cls_token, keyword_hidden_states, top_patches): # num_patches, embed_dim = cls_token.shape # batch_size, sequence_length, hidden_size = keyword_hidden_states.shape # assert embed_dim == hidden_size, f"The embedding dimension of cls token and keyword hidden states do not match. Dimension of cls token: {embed_dim} and dimension of keyword hidden states: {hidden_size}." # keyword_hidden_states = keyword_hidden_states.squeeze(0) # # calculate the similarity between the cls token and the keyword hidden states # similarity_score = torch.matmul(cls_token, keyword_hidden_states.T) # shape: (num_patches, sequence_length) # similarity_score = similarity_score.mean(dim=1) # shape: (num_patches) # # take the index of the patch with the highest similarity score # patch_index = torch.topk(similarity_score, top_patches).indices # return patch_index # def _select_patches(self, image_features, keyword_hidden_states, top_patches=1): # selected_patches = [] # # iterate through each image # for image in image_features: # if keyword_hidden_states is not None: # # take the first token of each patch # cls_token = image[:, 0, :].squeeze(1) # # get the index of the patch with the highest similarity score # patch_index = self._get_highest_similarity(cls_token, keyword_hidden_states, top_patches) # else: # top_patches = image.shape[0] # patch_index = torch.arange(top_patches) # # select the patch with the highest similarity score # if self.multimodal_projector == 'mlp': # image = image[patch_index, 1:, :].reshape(-1, image.shape[-1]).type(self.vision_tower.dtype) # elif self.multimodal_projector == 'perceiver': # image = image[patch_index, :, :].reshape(-1, image.shape[-1]).type(self.vision_tower.dtype) # else: # raise ValueError(f"Multimodal projector {self.multimodal_projector} is not supported.") # selected_patches.append(image) # return selected_patches # shape: list with shape of num_images, each element of shape (num_tokens * num_patches_i, embed_dim) # def _input_to_vision_tower(self, pixel_values): # output = [] # for i in range(len(pixel_values)): # num_patches = pixel_values[i].shape[0] # pixel_batch_size = 2 # processed_pixel_values # def _input_to_multimodal_projector(self, selected_image_features): # output = [] # for selected_image in selected_image_features: # selected_image = self.multi_modal_projector(selected_image) # output.append(selected_image) # return output # shape: list with shape of num_images, each element of shape (num_patches_i, num_tokens, embed_dim) where i is the index of the image # def _process_keyword_input(self, keyword_input_ids, maximum_keyword_tokens=10): # self.language_model.eval() # with torch.no_grad(): # output_ids = self.language_model.generate(input_ids=keyword_input_ids, return_dict_in_generate=True, max_new_tokens=maximum_keyword_tokens) # output_ids = output_ids.sequences[:, keyword_input_ids.shape[-1]:] # self.language_model.train() # # conditions # if output_ids[0, 0:2].tolist() == [35581, 25]: # condition where the output is in the form Keyword: # keyword_ids = output_ids[:, 2:-1] # if keyword_ids[0, 0].item() == 482: # return None # return self.get_input_embeddings()(keyword_ids) # else: # output # return None def generate_keywords(self, keywords_text, criteria='template'): keywords_text = keywords_text.lstrip('\n') first_sentence = keywords_text.split('.')[0] + '.' if re.search(r'are (.+?)\.', first_sentence): objects = re.search(r'are (.+?)\.', first_sentence).group(1).split(' and ') elif re.search(r'is (.+?)\.', first_sentence): objects = [re.search(r'is (.+?)\.', first_sentence).group(1)] else: objects = [] def generate_template(object, description): if object[0] in ['a', 'e', 'i', 'o', 'u']: return f'An {object}, which {description}' else: return f'A {object}, which {description}' descriptions = [] keywords = [] for i, obj in enumerate(objects): keywords.append(obj) if criteria == 'word': descriptions.append([obj]) elif criteria == 'template': descriptions.append([f'a photo of {obj}']) elif criteria == 'description': # pattern = rf"'{obj}':(.*?)('|\Z)" # match = re.search(pattern, keywords_text, re.DOTALL) # if match: # # Extract the feature keywords_text and clean it up # feature_text = match.group(1).strip() # # Split on new lines and strip each line # feature_list = [generate_template(obj, line.strip('* ').strip()) for line in feature_text.split('\n') if line.strip()] # descriptions.append(feature_list) # The problem of the above code is that it does not work for the case where the object is not found in the text # make it more general features = re.findall(r"\* (.+)", keywords_text, re.MULTILINE) descriptions.append([generate_template(obj, feature) for feature in features[i * len(features) // len(objects): (i + 1) * len(features) // len(objects)]]) else: raise ValueError(f'invalid criteria: {criteria}') return keywords, descriptions def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels): num_images = len(image_features) num_image_tokens = torch.tensor([x.shape[0] for x in image_features], device=self.vision_tower.device, dtype=torch.int64) # total image tokens embed_dim = image_features[0].shape[-1] batch_size, sequence_length = input_ids.shape left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id)) # 1. Create a mask to know where special image tokens are special_image_token_mask = input_ids == self.config.image_token_index # num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1) # Compute the maximum embed dimension # max_embed_dim = (num_special_image_tokens.max() * (num_image_tokens - 1)) + sequence_length max_embed_dim = torch.sum(num_image_tokens) - num_images + sequence_length batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index) _, image_indices = torch.where(input_ids == self.config.image_token_index) # 2. Compute the positions where text should be written # Calculate new positions for text tokens in merged image-text sequence. # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens. # `torch.cumsum` computes how each image token shifts subsequent text token positions. # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one. image_token_mask = special_image_token_mask * 1 image_token_mask[0, image_indices] = num_image_tokens - 1 # for i, index in enumerate(image_indices): # special_image_token_mask[0, index] = num_image_tokens[i] - 1 new_token_positions = torch.cumsum((image_token_mask) + 1, -1) - 1 # new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1 nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1] if left_padding: new_token_positions += nb_image_pad[:, None] # offset for left padding text_to_overwrite = new_token_positions[batch_indices, non_image_indices] # 3. Create the full embedding, already padded to the maximum position final_embedding = torch.zeros( batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device ) final_attention_mask = torch.zeros( batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device ) if labels is not None: final_labels = torch.full( (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device ) # In case the Vision model or the Language model has been offloaded to CPU, we need to manually # set the corresponding tensors into their correct target device. target_device = inputs_embeds.device batch_indices, non_image_indices, text_to_overwrite = ( batch_indices.to(target_device), non_image_indices.to(target_device), text_to_overwrite.to(target_device), ) attention_mask = attention_mask.to(target_device) # 4. Fill the embeddings based on the mask. If we have ["hey" "", "how", "are"] # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices] final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices] if labels is not None: final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices] # 5. Fill the embeddings corresponding to the images. Anything that is still zeros needs filling image_to_overwrite = torch.all(final_embedding == 0, dim=-1) image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device) if image_to_overwrite.sum() != torch.sum(num_image_tokens): raise ValueError( f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while" f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation." ) final_embedding[image_to_overwrite] = torch.cat([image_patches for image_patches in image_features], dim=0).to(target_device) # final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device) final_attention_mask |= image_to_overwrite position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1) if labels is None: final_labels = None return final_embedding, final_attention_mask, final_labels, position_ids def forward( self, input_ids: torch.LongTensor = None, pixel_values: List[torch.FloatTensor] = None, coords: List[torch.FloatTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, keyword_prompt_input_ids: torch.LongTensor = None, vision_feature_select_strategy: Optional[str] = None, vision_feature_layer: Optional[int] = None, patch_picking_strategy: Optional[str] = None, topk: Optional[int] = None, keyword_criteria: Optional[str] = None, positional_information: Optional[str] = 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, visualize_patches: Optional[bool] = None, visualize_topk_patches: Optional[bool] = None, print_keyword: Optional[bool] = None, print_topk_patches: Optional[bool] = None, ) -> Union[Tuple, GeckoCausalLMOutputWithPast]: """ Parameters: text_inputs: Dict Output of tokenizer for text data. A dictionary containing the following keys: - input_ids: torch.LongTensor of shape (batch_size, sequence_length) - attention_mask: torch.LongTensor of shape (batch_size, sequence_length) - token_type_ids: torch.LongTensor of shape (batch_size, sequence_length) keyword_inputs: Dict Output of tokenizer for keyword data. A dictionary containing the following keys: - input_ids: torch.LongTensor of shape (batch_size, sequence_length) - attention_mask: torch.LongTensor of shape (batch_size, sequence_length) - token_type_ids: torch.LongTensor of shape (batch_size, sequence_length) image_inputs: Dict Output of ImageProcessor for image data. A dictionary containing the following keys: - pixel_values: torch.FloatTensor of shape (num_images, num_patches, num_tokens, embed_dim) - coords: List of shape (batch_size, num_images) """ # processing image and text inputs output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict vision_feature_layer = ( vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer ) vision_feature_select_strategy = ( vision_feature_select_strategy if vision_feature_select_strategy is not None else self.config.vision_feature_select_strategy ) patch_picking_strategy = patch_picking_strategy if patch_picking_strategy is not None else self.config.patch_picking_strategy topk = topk if topk is not None else self.config.topk keyword_criteria = keyword_criteria if keyword_criteria is not None else self.config.keyword_criteria positional_information = positional_information if positional_information is not None else self.config.positional_information visualize_patches = visualize_patches if visualize_patches is not None else self.config.visualize_patches visualize_topk_patches = visualize_topk_patches if visualize_topk_patches is not None else self.config.visualize_topk_patches print_keyword = print_keyword if print_keyword is not None else self.config.print_keyword print_topk_patches = print_topk_patches if print_topk_patches is not None else self.config.print_topk_patches if inputs_embeds is None: # 1. Extra the input embeddings inputs_embeds = self.get_input_embeddings()(input_ids) # 2. Merge text and images if pixel_values is not None and input_ids.shape[1] != 1: with torch.no_grad(): keyword_input_ids = self.language_model.generate(keyword_prompt_input_ids, return_dict_in_generate=True, max_new_tokens=1024, eos_token_id=self.eos_token_id) keyword_input_ids = keyword_input_ids.sequences[:, keyword_prompt_input_ids.shape[-1]:] keyword_text = self.tokenizer.decode(keyword_input_ids[0], skip_special_tokens=True) # print(keyword_text) generated_keywords, generated_descriptions = self.generate_keywords(keyword_text, criteria=keyword_criteria) all_text_features = [] for descriptions in generated_descriptions: one_text_features = [] for description in descriptions: keyword_ids = self.clip_tokenizer(description, return_tensors='pt') keyword_ids = {k: v.to(self.device) for k, v in keyword_ids.items()} text_features = self.encoder.get_text_features(**keyword_ids) one_text_features.append(text_features / text_features.norm(p=2, dim=-1, keepdim=True)) all_text_features.append(torch.cat(one_text_features, dim=0)) pixel_values = [pixel_value.to(self.vision_tower.device, dtype=self.vision_tower.dtype) for pixel_value in pixel_values] selected_image_features = [] selected_coords = [] for p, pixel_value in enumerate(pixel_values): # iterate through each image print_keyword_text = f'Keywords (criteria: {keyword_criteria}):\n' all_hidden_states = self.vision_tower(pixel_value, output_hidden_states=True).hidden_states # tuple of size (num_layers, num_patch, num_tokens, vison_embed_dim) if patch_picking_strategy == 'last_layer': hidden_states = [all_hidden_states[-1]] elif patch_picking_strategy == 'across_layers': hidden_states = deepcopy(all_hidden_states) top_patches = [0] for i, text_feature in enumerate(all_text_features): print_keyword_text += f' {i+1}: ' + "\n ".join(generated_descriptions[i]) + '\n' top_index = [] for hidden_state in hidden_states: # iterate through each layer if 'clip' in self.encoder_type: if vision_feature_select_strategy == 'cls': image_features = self.encoder.visual_projection(self.encoder.vision_model.post_layernorm(hidden_state[1:, 0, :])) # (num_patch-1, embed_dim) elif vision_feature_select_strategy == 'image_features': image_features = self.encoder.visual_projection(self.encoder.vision_model.post_layernorm(hidden_state[1:, 1:, :])) # (num_patch-1 * num_tokens, embed_dim) num_tokens = hidden_state.shape[1] - 1 elif 'siglip' in self.encoder_type: if vision_feature_select_strategy == 'cls': image_features = self.encoder.vision_model.head(self.encoder.vision_model.post_layernorm(hidden_state[1:, :, :])) # (num_patch-1, embed_dim) elif vision_feature_select_strategy == 'image_features': image_features = self.encoder.vision_model.post_layernorm(hidden_state[1:, :, :]) # (num_patch-1 * num_tokens, embed_dim) num_tokens = hidden_state.shape[1] image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True) if vision_feature_select_strategy == 'cls': similarity_score = torch.matmul(image_features, text_feature.T).mean(dim=1) # (num_patch-1) if patch_picking_strategy == 'across_layers': index = torch.topk(similarity_score, 1).indices top_index.append(index.item()+1) elif patch_picking_strategy == 'last_layer': index = torch.topk(similarity_score, math.ceil(topk / len(all_text_features))).indices + 1 # take top k patches top_index += index.tolist() elif vision_feature_select_strategy == 'image_features': image_features = image_features.flatten(0, 1) similarity_score = torch.matmul(image_features, text_feature.T).mean(dim=1) # (num_patch-1 * num_tokens) index = torch.topk(similarity_score, 100).indices # take top 100 tokens patch_index = torch.floor(index / num_tokens) # get the patch index count = torch.nn.functional.one_hot(patch_index.to(torch.int64)).sum(dim=0) # count the occurrences of each patch if patch_picking_strategy == 'across_layers': top_count = torch.topk(count, 1).indices # take top 1 top_index.append(top_count.item()+1) elif patch_picking_strategy == 'last_layer': top_count = torch.topk(count, math.ceil(topk / len(all_text_features))).indices + 1 top_index += top_count.tolist() if visualize_patches and patch_picking_strategy == 'across_layers': if 'clip' in self.encoder_type: (x, y) = (5, 5) elif 'siglip' in self.encoder_type: (x, y) = (7, 4) fig, axs = plt.subplots(y, x, figsize=(x * 2, y * 2)) fig.suptitle(f'keyword: {generated_keywords[i]}') for k, index in enumerate(top_index): axs[k // x, k % x].imshow(pixel_value[index].to(torch.float32).cpu().numpy().transpose(1, 2, 0)) axs[k // x, k % x].set_title(f'Layer {k}') axs[k // x, k % x].axis('off') plt.show() if patch_picking_strategy == 'across_layers': top_patches += torch.topk(torch.bincount(torch.tensor(top_index, dtype=torch.int64)), math.ceil(topk / len(all_text_features))).indices.to(dtype=torch.int64).tolist() elif patch_picking_strategy == 'last_layer': top_patches += top_index topk_patches = list(set(top_patches)) if visualize_topk_patches: fig, axs = plt.subplots(1, len(topk_patches), figsize=(len(topk_patches) * 2, 2)) fig.suptitle(f'top-{len(topk_patches)} patches') for m, topk_patch in enumerate(topk_patches): axs[m].imshow(pixel_value[topk_patch].to(torch.float32).cpu().numpy().transpose(1, 2, 0)) axs[m].axis('off') plt.show() if 'clip' in self.encoder_type: selected_image_features.append(all_hidden_states[vision_feature_layer][topk_patches, 1:, :]) elif 'siglip' in self.encoder_type: selected_image_features.append(all_hidden_states[vision_feature_layer][topk_patches, :, :]) selected_coords.append([coords[p][q-1] for q in topk_patches[1:]]) # if isinstance(pixel_values, list): # pixel_values = torch.cat([x for x in pixel_values if x is not None], dim=0) if print_keyword: print(print_keyword_text) multimodal_projector_features = [] for x, (selected_image_feature, selected_coord) in enumerate(zip(selected_image_features, selected_coords)): print(f'image {x+1}: {selected_coord}') if '2d' in positional_information: max_width = max(selected_coord, key= lambda x: x[0])[0] + 1 max_height = max(selected_coord, key= lambda x: x[1])[1] + 1 positional_encoding = self.positional_encoding(torch.ones((max_width, max_height, selected_image_feature.shape[1], self.positional_encoding.org_channels), dtype=self.dtype, device=self.device)) accumulate = [] for i, top_patch in enumerate(selected_image_feature): if positional_information == '2d_before' and i != 0: top_patch += positional_encoding[selected_coord[i-1][0], selected_coord[i-1][1], :, :] aligned_image_feature = self.multi_modal_projector(top_patch) if positional_information == '2d_after' and i != 0: aligned_image_feature += positional_encoding[selected_coord[i-1][0], selected_coord[i-1][1], :, :] accumulate.append(aligned_image_feature) if i == 0: accumulate.append(self.get_input_embeddings()(self.tokenizer(', ', padding=False, truncation=False, max_length=None, return_tensors='pt')['input_ids'].to(device=self.device)[0, 1:])) continue if positional_information == 'explicit': accumulate.append(self.get_input_embeddings()(self.tokenizer(f' at {str(selected_coord[i-1])}, ', padding=False, truncation=False, max_length=None, return_tensors='pt')['input_ids'].to(device=self.device)[0, 1:])) else: accumulate.append(self.get_input_embeddings()(self.tokenizer(f', ', padding=False, truncation=False, max_length=None, return_tensors='pt')['input_ids'].to(device=self.device)[0, 1:])) multimodal_projector_features.append(torch.cat(accumulate, dim=0)) # dimension of (num_selected_patch * num_tokens-1 + num_selected_patch * sep_len - 1) -> (num_selected_patch * num_tokens - 1) as sep_len = 1 assert len(selected_image_features) == len(multimodal_projector_features), f"The number of selected image features and image features do not match. Dimension of selected image features: {len(selected_image_features)} and dimension of image features: {len(multimodal_projector_features)}." # print(multimodal_projector_features[0].shape) inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features( multimodal_projector_features, inputs_embeds, input_ids, attention_mask, labels ) if labels is None: labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long) else: # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of # generation with cache if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1: # Retrieve the first layer to inspect the logits and mask out the hidden states # that are set to 0 first_layer_past_key_value = past_key_values[0][0][:, :, :, 0] # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941 batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0) # Get the target length target_seqlen = first_layer_past_key_value.shape[-1] + 1 extended_attention_mask = torch.ones( (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]), dtype=attention_mask.dtype, device=attention_mask.device, ) # Filter out only the tokens that can be un-attended, this can happen # if one uses Llava + Fused modules where the cache on the # first iteration is already big enough, or if one passes custom cache valid_indices = non_attended_tokens < extended_attention_mask.size(-1) new_batch_index = batch_index[valid_indices] new_non_attended_tokens = non_attended_tokens[valid_indices] # Zero-out the places where we don't need to attend extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0 attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1) position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1 outputs = self.language_model( attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) logits = outputs[0] batch_shift = 100 loss = None if labels is not None: # Shift so that tokens < n predict n if attention_mask is not None: shift_attention_mask = attention_mask[..., 1:] logits_shape = logits.shape labels_shape = labels.shape shift_attention_mask_shape = shift_attention_mask.shape for i in range(0, shift_attention_mask.shape[1], batch_shift): shift_logits = logits[..., i:min(i+batch_shift, logits_shape[1]-1), :][shift_attention_mask[..., i:min(i+batch_shift, shift_attention_mask_shape[1])].to(logits.device) != 0].contiguous() shift_labels = labels[..., i+1:min(i+batch_shift+1, labels_shape[1])][shift_attention_mask[..., i:min(i+batch_shift, shift_attention_mask_shape[1])].to(labels.device) != 0].contiguous() else: shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = nn.CrossEntropyLoss() loss = loss_fct( shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device) ) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return GeckoCausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, inputs_embeds=None, pixel_values=None, attention_mask=None, keyword_prompt_input_ids=None, coords=None, **kwargs ): if past_key_values is not None: if isinstance(past_key_values, Cache): cache_length = past_key_values.get_seq_length() past_length = past_key_values.seen_tokens else: cache_length = past_length = past_key_values[0][0].shape[2] # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as # input) if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard # input_ids based on the past_length. elif past_length < input_ids.shape[1]: input_ids = input_ids[:, past_length:] # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. elif self.config.image_token_index in input_ids: input_ids = input_ids[:, input_ids.shape[1] - 1 :] # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the # older attention values, as their corresponding values are not part of the input. if cache_length < past_length and attention_mask is not None: attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :] position_ids = kwargs.get("position_ids", None) if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) if past_key_values: position_ids = position_ids[:, -input_ids.shape[1] :] # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None and past_key_values is None: model_inputs = {"inputs_embeds": inputs_embeds} else: model_inputs = {"input_ids": input_ids} model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, "use_cache": kwargs.get("use_cache"), "attention_mask": attention_mask, "pixel_values": pixel_values, "keyword_prompt_input_ids": keyword_prompt_input_ids, "coords": coords, "topk": kwargs.get("topk"), "vision_feature_select_strategy": kwargs.get("vision_feature_select_strategy"), "vision_feature_layer": kwargs.get("vision_feature_layer"), "patch_picking_strategy": kwargs.get("patch_picking_strategy"), "keyword_criteria": kwargs.get("keyword_criteria"), "positional_information": kwargs.get("positional_information"), "visualize_patches": kwargs.get("visualize_patches"), "visualize_topk_patches": kwargs.get("visualize_topk_patches"), "print_keyword": kwargs.get("print_keyword"), "print_topk_patches": kwargs.get("print_topk_patches"), } ) return model_inputs def _reorder_cache(self, *args, **kwargs): return self.language_model._reorder_cache(*args, **kwargs)