| import torch |
| import transformers |
| import tokenizers |
| import os, sys |
| from dataclasses import dataclass, field |
| import argparse |
| from PIL import Image |
|
|
| def expand2square(pil_img, background_color): |
| width, height = pil_img.size |
| if width == height: |
| return pil_img |
| elif width > height: |
| result = Image.new(pil_img.mode, (width, width), background_color) |
| result.paste(pil_img, (0, (width - height) // 2)) |
| return result |
| else: |
| result = Image.new(pil_img.mode, (height, height), background_color) |
| result.paste(pil_img, ((height - width) // 2, 0)) |
| return result |
|
|
| def find_all_linear_names(model): |
| cls = torch.nn.Linear |
| lora_module_names = set() |
| multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler'] |
| for name, module in model.named_modules(): |
| if any(mm_keyword in name for mm_keyword in multimodal_keywords): |
| continue |
| if isinstance(module, cls): |
| names = name.split('.') |
| lora_module_names.add(names[0] if len(names) == 1 else names[-1]) |
|
|
| if 'lm_head' in lora_module_names: |
| lora_module_names.remove('lm_head') |
| return list(lora_module_names) |
|
|
| def add_special_tokens_and_resize_model(tokenizer, model, vq_idx_nums=8192): |
| if len(tokenizer.additional_special_tokens) != 0: |
| return tokenizer.additional_special_tokens |
| index_tokens = [f"<idx_{i}>" for i in range(vq_idx_nums)] |
| special_tokens = { |
| 'additional_special_tokens': ['<start_index>'] + index_tokens + ['<end_index>'] + ['<pixel_newline>'] |
| } |
| num_new_tokens = tokenizer.add_special_tokens(special_tokens) |
| model.resize_token_embeddings(len(tokenizer)) |
| if num_new_tokens > 0: |
| input_embeddings = model.get_input_embeddings().weight.data |
| output_embeddings = model.get_output_embeddings().weight.data |
|
|
| input_embeddings_avg = input_embeddings[:-num_new_tokens].mean( |
| dim=0, keepdim=True) |
| output_embeddings_avg = output_embeddings[:-num_new_tokens].mean( |
| dim=0, keepdim=True) |
|
|
| input_embeddings[-num_new_tokens:] = input_embeddings_avg |
| output_embeddings[-num_new_tokens:] = output_embeddings_avg |
| |
| return num_new_tokens |
|
|
| com_vision_args = argparse.Namespace( |
| freeze_backbone=False, |
| mm_patch_merge_type='flat', |
| mm_projector_type='mlp2x_gelu', |
| mm_use_im_patch_token=False, |
| mm_use_im_start_end=False, |
| mm_vision_select_feature='patch', |
| mm_vision_select_layer=-2, |
| model_name_or_path=None, |
| pretrain_mm_mlp_adapter=None, |
| tune_mm_mlp_adapter=False, |
| version=None, |
| vision_tower=None |
| ) |
|
|
| gen_vision_args = argparse.Namespace( |
| freeze_backbone=False, |
| mm_patch_merge_type='flat', |
| mm_projector_type='mlp2x_gelu', |
| mm_use_im_patch_token=False, |
| mm_use_im_start_end=False, |
| mm_vision_select_feature='patch', |
| mm_vision_select_layer=1, |
| model_name_or_path=None, |
| pretrain_mm_mlp_adapter=None, |
| tune_mm_mlp_adapter=False, |
| version=None, |
| vision_tower=None |
| ) |
|
|
| def load_weights(model, hlora_path, fusion_layer_path=None): |
| hlora_weights = torch.load(hlora_path) |
| hlora_unexpected_keys = model.load_state_dict(hlora_weights, strict=False)[1] |
| if hlora_unexpected_keys: |
| print(f"Warning: Unexpected keys in hlora checkpoint: {hlora_unexpected_keys}") |
|
|
| if fusion_layer_path: |
| fusion_layer_weights = torch.load(fusion_layer_path) |
| fusion_layer_unexpected_keys = model.load_state_dict(fusion_layer_weights, strict=False)[1] |
| if fusion_layer_unexpected_keys: |
| print(f"Warning: Unexpected keys in fusion_layer checkpoint: {fusion_layer_unexpected_keys}") |
|
|
| return model |
