# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright: # Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright: # Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import os import random os.environ["WANDB_PROJECT"] = 'radiolog_llava' import copy from dataclasses import dataclass, field import json import logging import pathlib from typing import Dict, Optional, Sequence, List, Tuple import torch from torch import Tensor from skimage import io import transformers from torchvision.transforms import Compose, Resize, ToTensor, CenterCrop, transforms from torchvision.transforms import functional as F, InterpolationMode from llava.constants import IGNORE_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN from torch.utils.data import Dataset from llava.train.llava_trainer import LLaVATrainer from llava import conversation as conversation_lib from llava.model import * from llava.mm_utils import tokenizer_image_token from PIL import Image import numpy as np local_rank = None # if LLAVA_MED # IGNORE_INDEX = -100 # DEFAULT_PAD_TOKEN = "[PAD]" # DEFAULT_EOS_TOKEN = "" # DEFAULT_BOS_TOKEN = "" # DEFAULT_UNK_TOKEN = "" # DEFAULT_IMAGE_TOKEN = "" # DEFAULT_IMAGE_PATCH_TOKEN = "" # DEFAULT_IM_START_TOKEN = "" # DEFAULT_IM_END_TOKEN = "" def rank0_print(*args): if local_rank == 0: print(*args) @dataclass class ModelArguments: model_name_or_path: Optional[str] = field(default="facebook/opt-125m") version: Optional[str] = field(default="v0") freeze_backbone: bool = field(default=False) tune_mm_mlp_adapter: bool = field(default=False) vision_tower: Optional[str] = field(default=None) mm_vision_select_layer: Optional[int] = field(default=-1) # default to the last layer pretrain_mm_mlp_adapter: Optional[str] = field(default=None) mm_projector_type: Optional[str] = field(default='linear') mm_use_im_start_end: bool = field(default=False) mm_use_im_patch_token: bool = field(default=True) mm_vision_select_feature: Optional[str] = field(default="patch") mv_type: Optional[str] = field(default='concat') @dataclass class DataArguments: data_path: str = field(default=None, metadata={"help": "Path to the training data."}) lazy_preprocess: bool = False is_multimodal: bool = False image_folder: Optional[str] = field(default=None) image_aspect_ratio: str = 'square' do_augment: bool = field(default=False) do_img_order_augment: bool = field(default=False) @dataclass class TrainingArguments(transformers.TrainingArguments): cache_dir: Optional[str] = field(default=None) optim: str = field(default="adamw_torch") remove_unused_columns: bool = field(default=False) freeze_mm_mlp_adapter: bool = field(default=False) unfreeze_n_vision_tower_layers: Optional[int] = field(default=None) mpt_attn_impl: Optional[str] = field(default="triton") model_max_length: int = field( default=512, metadata={ "help": "Maximum sequence length. Sequences will be right padded (and possibly truncated)." }, ) double_quant: bool = field( default=True, metadata={"help": "Compress the quantization statistics through double quantization."} ) quant_type: str = field( default="nf4", metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."} ) bits: int = field( default=16, metadata={"help": "How many bits to use."} ) lora_enable: bool = False lora_r: int = 64 lora_alpha: int = 16 lora_dropout: float = 0.05 lora_weight_path: str = "" lora_bias: str = "none" mm_projector_lr: Optional[float] = None group_by_modality_length: bool = field(default=False) def maybe_zero_3(param, ignore_status=False, name=None): from deepspeed import zero from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus if hasattr(param, "ds_id"): if param.ds_status == ZeroParamStatus.NOT_AVAILABLE: if not ignore_status: logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}") with zero.GatheredParameters([param]): param = param.data.detach().cpu().clone() else: param = param.detach().cpu().clone() return param # Borrowed from peft.utils.get_peft_model_state_dict def get_peft_state_maybe_zero_3(named_params, bias): if bias == "none": to_return = {k: t for k, t in named_params if "lora_" in k} elif bias == "all": to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k} elif bias == "lora_only": to_return = {} maybe_lora_bias = {} lora_bias_names = set() for k, t in named_params: if "lora_" in k: to_return[k] = t bias_name = k.split("lora_")[0] + "bias" lora_bias_names.add(bias_name) elif "bias" in k: maybe_lora_bias[k] = t for k, t in maybe_lora_bias: if bias_name in lora_bias_names: to_return[bias_name] = t else: raise NotImplementedError to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()} return to_return def get_peft_state_non_lora_maybe_zero_3_extended(model, require_grad_only=True): named_entities = list(model.named_parameters()) + list(model.named_buffers()) to_return = {k: v for k, v in named_entities if "lora_" not in k} if require_grad_only: # For buffers, requires_grad attribute does not apply, so they should be included regardless to_return = {k: v for k, v in to_return.items() if type(v) == torch.Tensor or v.requires_grad} to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()} return to_return def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match): to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)} to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()} return to_return def find_all_linear_names(model): cls = torch.nn.Linear lora_module_names = set() multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler', 'image_pooler'] 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: # needed for 16-bit lora_module_names.remove('lm_head') return list(lora_module_names) def safe_save_model_for_hf_trainer(trainer: transformers.Trainer, output_dir: str): """Collects the state dict and dump to disk.""" if getattr(trainer.args, "tune_mm_mlp_adapter", False): # Only save Adapter keys_to_match = ['mm_projector'] if getattr(trainer.args, "use_im_start_end", False): keys_to_match.extend(['embed_tokens', 'embed_in']) weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match) trainer.model.config.save_pretrained(output_dir) current_folder = output_dir.split('/')[-1] parent_folder = os.path.dirname(output_dir) if trainer.args.local_rank == 0 or trainer.args.local_rank == -1: if current_folder.startswith('checkpoint-'): mm_projector_folder = os.path.join(parent_folder, "mm_projector") os.makedirs(mm_projector_folder, exist_ok=True) torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin')) else: torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin')) return if trainer.deepspeed: torch.cuda.synchronize() trainer.save_model(output_dir) return state_dict = trainer.model.state_dict() if trainer.args.should_save: cpu_state_dict = { key: value.cpu() for key, value in state_dict.items() } del state_dict trainer._save(output_dir, state_dict=cpu_state_dict) # noqa def smart_tokenizer_and_embedding_resize( special_tokens_dict: Dict, tokenizer: transformers.PreTrainedTokenizer, model: transformers.PreTrainedModel, ): """Resize tokenizer and embedding. Note: This is the unoptimized version that may make your embedding size not be divisible by 64. """ num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict) 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 def _tokenize_fn(strings: Sequence[str], tokenizer: transformers.PreTrainedTokenizer) -> Dict: """Tokenize a list of strings.""" tokenized_list = [ tokenizer( text, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ) for text in strings ] input_ids = labels = [ tokenized.input_ids[0] for tokenized in tokenized_list ] input_ids_lens = labels_lens = [ tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item() for tokenized in tokenized_list ] return dict( input_ids=input_ids, labels=labels, input_ids_lens=input_ids_lens, labels_lens=labels_lens, ) def _mask_targets(target, tokenized_lens, speakers): # cur_idx = 0 cur_idx = tokenized_lens[0] tokenized_lens = tokenized_lens[1:] target[:cur_idx] = IGNORE_INDEX for tokenized_len, speaker in zip(tokenized_lens, speakers): if speaker == "human": target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX cur_idx += tokenized_len def _add_speaker_and_signal(header, source, get_conversation=True): """Add speaker and start/end signal on each round.""" BEGIN_SIGNAL = "### " END_SIGNAL = "\n" conversation = header for sentence in source: from_str = sentence["from"] if from_str.lower() == "human": from_str = conversation_lib.default_conversation.roles[0] elif from_str.lower() == "gpt": from_str = conversation_lib.default_conversation.roles[1] else: from_str = 'unknown' sentence["value"] = (BEGIN_SIGNAL + from_str + ": " + sentence["value"] + END_SIGNAL) if get_conversation: conversation += sentence["value"] conversation += BEGIN_SIGNAL return conversation def preprocess_multimodal( sources: Sequence[str], data_args: DataArguments ) -> Dict: is_multimodal = data_args.is_multimodal if not is_multimodal: return sources for source in sources: for sentence in source: if DEFAULT_IMAGE_TOKEN in sentence['value']: sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '').strip() sentence['value'] = DEFAULT_IMAGE_TOKEN + '\n' + sentence['value'] sentence['value'] = sentence['value'].strip() if "mmtag" in conversation_lib.default_conversation.version: sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '' + DEFAULT_IMAGE_TOKEN + '') replace_token = DEFAULT_IMAGE_TOKEN if data_args.mm_use_im_start_end: replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token) return sources def preprocess_llama_2( sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations if has_image: input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) else: input_ids = tokenizer( conversations, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ).input_ids targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_2 # Mask targets sep = "[/INST] " for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep2) cur_len = 1 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep if has_image: round_len = len(tokenizer_image_token(rou, tokenizer)) instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2 else: round_len = len(tokenizer(rou).input_ids) instruction_len = len(tokenizer(parts[0]).input_ids) - 2 target[cur_len : cur_len + instruction_len] = IGNORE_INDEX cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_v1( sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations if has_image: input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) else: input_ids = tokenizer( conversations, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ).input_ids targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.TWO # Mask targets sep = conv.sep + conv.roles[1] + ": " for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep2) cur_len = 1 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep if has_image: round_len = len(tokenizer_image_token(rou, tokenizer)) instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2 else: round_len = len(tokenizer(rou).input_ids) instruction_len = len(tokenizer(parts[0]).input_ids) - 2 if i == len(rounds) - 2: #last round, keep answer for training target[cur_len : cur_len + instruction_len] = IGNORE_INDEX else: target[cur_len : cur_len + round_len] = IGNORE_INDEX #previous rounds - mask everything cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_mpt( sources, tokenizer: transformers.PreTrainedTokenizer, ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.MPT # Mask targets sep = conv.sep + conv.roles[1] for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep) re_rounds = [conv.sep.join(rounds[:3])] # system + user + gpt for conv_idx in range(3, len(rounds), 2): re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx+2])) # user + gpt cur_len = 0 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(re_rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep round_len = len(tokenizer_image_token(rou, tokenizer)) + len(tokenizer_image_token(conv.sep, tokenizer)) instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) target[cur_len : cur_len + instruction_len] = IGNORE_INDEX cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_plain( sources: Sequence[str], tokenizer: transformers.PreTrainedTokenizer, ) -> Dict: # add end signal and concatenate together conversations = [] for source in sources: assert len(source) == 2 assert DEFAULT_IMAGE_TOKEN in source[0]['value'] source[0]['value'] = DEFAULT_IMAGE_TOKEN conversation = source[0]['value'] + source[1]['value'] + conversation_lib.default_conversation.sep conversations.append(conversation) # tokenize conversations input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations] targets = copy.deepcopy(input_ids) for target, source in zip(targets, sources): tokenized_len = len(tokenizer_image_token(source[0]['value'], tokenizer)) target[:tokenized_len] = IGNORE_INDEX return dict(input_ids=input_ids, labels=targets) def preprocess( sources: Sequence[str], tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: """ Given a list of sources, each is a conversation list. This transform: 1. Add signal '### ' at the beginning each sentence, with end signal '\n'; 2. Concatenate conversations together; 3. Tokenize the concatenated conversation; 4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX. """ if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN: return preprocess_plain(sources, tokenizer) if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2: return preprocess_llama_2(sources, tokenizer, has_image=has_image) if conversation_lib.default_conversation.version.startswith("v1"): return preprocess_v1(sources, tokenizer, has_image=has_image) if conversation_lib.default_conversation.version == "mpt": return preprocess_mpt(sources, tokenizer) # add end signal and concatenate together conversations = [] for source in sources: header = f"{conversation_lib.default_conversation.system}\n\n" conversation = _add_speaker_and_signal(header, source) conversations.append(conversation) # tokenize conversations def get_tokenize_len(prompts): return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts] if has_image: input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations] else: conversations_tokenized = _tokenize_fn(conversations, tokenizer) input_ids = conversations_tokenized["input_ids"] targets = copy.deepcopy(input_ids) for target, source in zip(targets, sources): if has_image: tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source]) else: tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"] speakers = [sentence["from"] for sentence in source] _mask_targets(target, tokenized_lens, speakers) return dict(input_ids=input_ids, labels=targets) class ExpandChannels: """ Transforms an image with one channel to an image with three channels by copying pixel intensities of the image along the 1st dimension. """ def __call__(self, data: torch.Tensor) -> torch.Tensor: """ :param data: Tensor of shape [1, H, W]. :return: Tensor with channel copied three times, shape [3, H, W]. """ if data.shape[0] != 1: raise ValueError(f"Expected input of shape [1, H, W], found {data.shape}") return torch.repeat_interleave(data, 3, dim=0) def _apply_op(img: Tensor, op_name: str, magnitude: float, interpolation: InterpolationMode, fill: Optional[List[float]]): if op_name == "ShearX": img = F.affine(img, angle=0.0, translate=[0, 0], scale=1.0, shear=[math.degrees(magnitude), 0.0], interpolation=interpolation, fill=fill) elif op_name == "ShearY": img = F.affine(img, angle=0.0, translate=[0, 0], scale=1.0, shear=[0.0, math.degrees(magnitude)], interpolation=interpolation, fill=fill) elif op_name == "TranslateX": img = F.affine(img, angle=0.0, translate=[int(magnitude), 0], scale=1.0, interpolation=interpolation, shear=[0.0, 0.0], fill=fill) elif op_name == "TranslateY": img = F.affine(img, angle=0.0, translate=[0, int(magnitude)], scale=1.0, interpolation=interpolation, shear=[0.0, 0.0], fill=fill) elif op_name == "Rotate": img = F.rotate(img, magnitude, interpolation=interpolation, fill=fill) elif op_name == "Brightness": img = F.adjust_brightness(img, 1.0 + magnitude) elif op_name == "Color": img = F.adjust_saturation(img, 1.0 + magnitude) elif op_name == "Contrast": img = F.adjust_contrast(img, 1.0 + magnitude) elif op_name == "Sharpness": img = F.adjust_sharpness(img, 1.0 + magnitude) elif op_name == "Posterize": img = F.posterize(img, int(magnitude)) elif op_name == "Solarize": img = F.solarize(img, magnitude) elif op_name == "AutoContrast": img = F.autocontrast(img) elif op_name == "Equalize": img = F.equalize(img) elif op_name == "Invert": img = F.invert(img) elif op_name == "Identity": pass else: raise ValueError("The provided operator {} is not recognized.".format(op_name)) return img class TrivialAugmentWide(torch.nn.Module): r"""Dataset-independent data-augmentation with TrivialAugment Wide, as described in `"TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentation" `. If the image is torch Tensor, it should be of type torch.uint8, and it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. If img is PIL Image, it is expected to be in mode "L" or "RGB". Args: num_magnitude_bins (int): The number of different magnitude values. interpolation (InterpolationMode): Desired interpolation enum defined by :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. fill (sequence or number, optional): Pixel fill value for the area outside the transformed image. If given a number, the value is used for all bands respectively. """ def __init__(self, num_magnitude_bins: int = 31, interpolation: InterpolationMode = InterpolationMode.NEAREST, fill: Optional[List[float]] = None, strength: float = 1.0) -> None: super().__init__() self.num_magnitude_bins = num_magnitude_bins self.interpolation = interpolation self.fill = fill self.strength = max(0.0, min(strength, 1.0)) # Ensuring strength is within [0, 1] def _augmentation_space(self, num_bins: int) -> Dict[str, Tuple[Tensor, bool]]: scale_factor = self.strength return { "Identity": (torch.tensor(0.0), False), "ShearX": (torch.linspace(0.0, 0.99 * scale_factor, num_bins), True), "ShearY": (torch.linspace(0.0, 0.99 * scale_factor, num_bins), True), "TranslateX": (torch.linspace(0.0, 32.0 * scale_factor, num_bins), True), "TranslateY": (torch.linspace(0.0, 32.0 * scale_factor, num_bins), True), "Rotate": (torch.linspace(0.0, 135.0 * scale_factor, num_bins), True), "Brightness": (torch.linspace(0.0, 0.99 * scale_factor, num_bins), True), "Color": (torch.linspace(0.0, 0.99 * scale_factor, num_bins), True), "Contrast": (torch.linspace(0.0, 0.99 * scale_factor, num_bins), True), "Sharpness": (torch.linspace(0.0, 0.99 * scale_factor, num_bins), True), #"Posterize": (8 - (torch.arange(num_bins) / ((num_bins - 1) / 6)).round().int(), False), "Solarize": (torch.linspace(256.0, 0.0, num_bins), False), "AutoContrast": (torch.tensor(0.0), False), } def forward(self, img: Tensor) -> Tensor: """ img (PIL Image or Tensor): Image to be transformed. Returns: PIL Image or Tensor: Transformed image. """ fill = self.fill if isinstance(img, Tensor): if isinstance(fill, (int, float)): fill = [float(fill)] * F.get_image_num_channels(img) elif fill is not None: fill = [float(f) for f in fill] op_meta = self._augmentation_space(self.num_magnitude_bins) op_index = int(torch.randint(len(op_meta), (1,)).item()) op_name = list(op_meta.keys())[op_index] magnitudes, signed = op_meta[op_name] magnitude = float(magnitudes[torch.randint(len(magnitudes), (1,), dtype=torch.long)].item()) \ if magnitudes.ndim > 0 else 0.0 if signed and torch.randint(2, (1,)): magnitude *= -1.0 return _apply_op(img, op_name, magnitude, interpolation=self.interpolation, fill=fill) def __repr__(self) -> str: s = self.__class__.__name__ + '(' s += 'num_magnitude_bins={num_magnitude_bins}' s += ', interpolation={interpolation}' s += ', fill={fill}' s += ')' return s.format(**self.__dict__) class LazySupervisedDataset(Dataset): """Dataset for supervised fine-tuning.""" def __init__(self, data_path: str, tokenizer: transformers.PreTrainedTokenizer, data_args: DataArguments): super(LazySupervisedDataset, self).__init__() list_data_dict = json.load(open(data_path, "r")) rank0_print("Formatting inputs...Skip in lazy mode") self.tokenizer = tokenizer self.list_data_dict = list_data_dict self.data_args = data_args self.do_img_order_augment = self.data_args.do_img_order_augment self.vis_transforms_biovil = self.create_chest_xray_transform_for_inference(512, center_crop_size=448) if self.data_args.do_augment: self.augment = TrivialAugmentWide(strength=0.5) #0.2 weak, 0.5 strong else: self.augment = None def __len__(self): return len(self.list_data_dict) def create_chest_xray_transform_for_inference(self, resize: int, center_crop_size: int) -> Compose: """ Defines the image transformation pipeline for Chest-Xray datasets. :param resize: The size to resize the image to. Linear resampling is used. Resizing is applied on the axis with smaller shape. :param center_crop_size: The size to center crop the image to. Square crop is applied. """ transforms = [Resize(resize), CenterCrop(center_crop_size), ToTensor(), ExpandChannels()] return Compose(transforms) @property def lengths(self): length_list = [] for sample in self.list_data_dict: img_tokens = 128 if 'image' in sample else 0 length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens) return length_list @property def modality_lengths(self): length_list = [] for sample in self.list_data_dict: cur_len = sum(len(conv['value'].split()) for conv in sample['conversations']) cur_len = cur_len if 'image' in sample else -cur_len length_list.append(cur_len) return length_list def remap_to_uint8(self, array: np.ndarray, percentiles=None) -> np.ndarray: """Remap values in input so the output range is :math:`[0, 255]`. Percentiles can be used to specify the range of values to remap. This is useful to discard outliers in the input data. :param array: Input array. :param percentiles: Percentiles of the input values that will be mapped to ``0`` and ``255``. Passing ``None`` is equivalent to using percentiles ``(0, 100)`` (but faster). :returns: Array with ``0`` and ``255`` as minimum and maximum values. """ array = array.astype(float) if percentiles is not None: len_percentiles = len(percentiles) if len_percentiles != 2: message = ( 'The value for percentiles should be a sequence of length 2,' f' but has length {len_percentiles}' ) raise ValueError(message) a, b = percentiles if a >= b: raise ValueError(f'Percentiles must be in ascending order, but a sequence "{percentiles}" was passed') if a < 0 or b > 100: raise ValueError(f'Percentiles must be in the range [0, 100], but a sequence "{percentiles}" was passed') cutoff: np.ndarray = np.percentile(array, percentiles) array = np.clip(array, *cutoff) array -= array.min() array /= array.max() array *= 255 return array.astype(np.uint8) def load_image_biovil(self, image_folder, image_file) -> Image.Image: """Load an image from disk. The image values are remapped to :math:`[0, 255]` and cast to 8-bit unsigned integers. :param path: Path to image. :returns: Image as ``Pillow`` ``Image``. """ # Although ITK supports JPEG and PNG, we use Pillow for consistency with older trained models if image_file.startswith('/home'): # full path path = pathlib.Path(image_file) elif image_file.startswith('files/'): # mimic-cxr path = pathlib.Path(os.path.join(image_folder, image_file)) else: path = pathlib.Path("/home/guests/chantal_pellegrini/" + image_file) #radrestruct if path.suffix in [".jpg", ".jpeg", ".png"]: image = io.imread(path) else: raise ValueError(f"Image type not supported, filename was: {path}") image = self.remap_to_uint8(image) return Image.fromarray(image).convert("L") def __getitem__(self, i) -> Dict[str, torch.Tensor]: sources = self.list_data_dict[i] if isinstance(i, int): sources = [sources] assert len(sources) == 1, "Don't know why it is wrapped to a list" # FIXME if 'image' in sources[0]: # 1 or multiple current images image_files = self.list_data_dict[i]['image'] if type(self.list_data_dict[i]['image']) == list else [self.list_data_dict[i]['image']] #convert to list image_folder = self.data_args.image_folder if self.do_img_order_augment: random.shuffle(image_files) n_images = random.randint(1,len(image_files)) image_files = image_files[:n_images] if self.data_args.vision_tower == 'biovil': images = [self.load_image_biovil(image_folder, image_file) for image_file in image_files] # augment images if self.augment is not None: images = [self.augment(image) for image in images] images = [self.vis_transforms_biovil(img) for img in images] else: processor = self.data_args.image_processor images = [Image.open(os.path.join(image_folder, image_file)).convert('RGB') for image_file in image_files] if self.data_args.image_aspect_ratio == 'pad': 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 images = [expand2square(image, tuple(int(x*255) for x in processor.image_mean)) for image in images] images = [processor.preprocess(image, return_tensors='pt')['pixel_values'][0] for image in images] else: images = [processor.preprocess(image, return_tensors='pt')['pixel_values'][0] for image in images] # stack images images = torch.stack(images, dim=0).squeeze() #stack and drop unnecessary dimension if 'prev_image' in sources[0]: # 1 or multiple previous images raise NotImplementedError("Previous image is not supported yet") prev_image_files = self.list_data_dict[i]['prev_image'] if type(self.list_data_dict[i]['prev_image']) == list else [ self.list_data_dict[i]['prev_image']] # convert to list image_folder = self.data_args.image_folder if self.data_args.vision_tower == 'biovil': prev_images = [self.load_image_biovil(image_folder, image_file) for image_file in prev_image_files] prev_images = [self.vis_transforms_biovil(img) for img in prev_images] else: processor = self.data_args.image_processor prev_images = [Image.open(os.path.join(image_folder, image_file)).convert('RGB') for image_file in prev_image_files] if self.data_args.image_aspect_ratio == 'pad': 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 prev_images = [expand2square(image, tuple(int(x * 255) for x in processor.image_mean)) for image in prev_images] prev_images = [processor.preprocess(image, return_tensors='pt')['pixel_values'][0] for image in prev_images] else: prev_images = [processor.preprocess(image, return_tensors='pt')['pixel_values'][0] for image in prev_images] # stack images prev_images = torch.stack(prev_images, dim=0).squeeze() if len(prev_images) > 0 else prev_images #stack and drop unnecessary dimension # drop images from the data sources = preprocess_multimodal( copy.deepcopy([e["conversations"] for e in sources]), self.data_args) else: sources = copy.deepcopy([e["conversations"] for e in sources]) data_dict = preprocess( sources, self.tokenizer, has_image=('image' in self.list_data_dict[i])) if isinstance(i, int): data_dict = dict(input_ids=data_dict["input_ids"][0], labels=data_dict["labels"][0]) # image exist in the data if 'image' in self.list_data_dict[i]: data_dict['image'] = images[0] if len(images) == 1 else images if 'prev_image' in self.list_data_dict[i]: data_dict['prev_image'] = prev_images[0] if len(prev_images) == 1 else prev_images if prev_images == []: data_dict['prev_image'] = None elif self.data_args.is_multimodal: # image does not exist in the data, but the model is multimodal crop_size = self.data_args.image_processor.crop_size data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width']) return data_dict @dataclass class DataCollatorForSupervisedDataset(object): """Collate examples for supervised fine-tuning.""" tokenizer: transformers.PreTrainedTokenizer def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]: input_ids, labels = tuple([instance[key] for instance in instances] for key in ("input_ids", "labels")) input_ids = torch.nn.utils.rnn.pad_sequence( input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id) labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX) input_ids = input_ids[:, :self.tokenizer.model_max_length] labels = labels[:, :self.tokenizer.model_max_length] batch = dict( input_ids=input_ids, labels=labels, attention_mask=input_ids.ne(self.tokenizer.pad_token_id), ) if 'image' in instances[0]: images = [instance['image'] for instance in instances] if not 'prev_image' in instances[0] and all(x is not None and x.shape == images[0].shape for x in images): batch['images'] = torch.stack(images) else: # extend the dimension of all images to 4 if it is only 3 (1x dimension to be treated as multi-image) images = [image.unsqueeze(0) if len(image.shape) == 3 else image for image in images] batch['images'] = images if 'prev_image' in instances[0]: prev_images = [instance['prev_image'] for instance in instances] prev_images = [image.unsqueeze(0) if len(image.shape) == 3 else image for image in prev_images] batch['prev_images'] = prev_images return batch def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args) -> Dict: """Make dataset and collator for supervised fine-tuning.""" train_dataset = LazySupervisedDataset(tokenizer=tokenizer, data_path=data_args.data_path, data_args=data_args) data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer) return dict(train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator) def train(): global local_rank parser = transformers.HfArgumentParser( (ModelArguments, DataArguments, TrainingArguments)) model_args, data_args, training_args = parser.parse_args_into_dataclasses() data_args.vision_tower = model_args.vision_tower #info for image preprocessing local_rank = training_args.local_rank compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32)) bnb_model_from_pretrained_args = {} if training_args.bits in [4, 8]: from transformers import BitsAndBytesConfig bnb_model_from_pretrained_args.update(dict( device_map={"": training_args.device}, load_in_4bit=training_args.bits == 4, load_in_8bit=training_args.bits == 8, quantization_config=BitsAndBytesConfig( load_in_4bit=training_args.bits == 4, load_in_8bit=training_args.bits == 8, llm_int8_skip_modules=["mm_projector", "image_pooler"], llm_int8_threshold=6.0, llm_int8_has_fp16_weight=False, bnb_4bit_compute_dtype=compute_dtype, bnb_4bit_use_double_quant=training_args.double_quant, bnb_4bit_quant_type=training_args.quant_type # {'fp4', 'nf4'} ) )) if model_args.vision_tower is not None: if 'mpt' in model_args.model_name_or_path: config = transformers.AutoConfig.from_pretrained(model_args.model_name_or_path, trust_remote_code=True) config.attn_config['attn_impl'] = training_args.mpt_attn_impl model = LlavaMPTForCausalLM.from_pretrained( model_args.model_name_or_path, config=config, cache_dir=training_args.cache_dir, **bnb_model_from_pretrained_args ) else: model = LlavaLlamaForCausalLM.from_pretrained( model_args.model_name_or_path, mv_type = model_args.mv_type, mm_vision_tower=model_args.vision_tower, cache_dir=training_args.cache_dir, ignore_mismatched_sizes=True, **bnb_model_from_pretrained_args ) else: model = transformers.LlamaForCausalLM.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, **bnb_model_from_pretrained_args ) model.config.use_cache = False if model_args.freeze_backbone: model.model.requires_grad_(False) if training_args.bits in [4, 8]: from peft import prepare_model_for_kbit_training model.config.torch_dtype=(torch.float32 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32)) model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing) if training_args.gradient_checkpointing: if hasattr(model, "enable_input_require_grads"): model.enable_input_require_grads() else: def make_inputs_require_grad(module, input, output): output.requires_grad_(True) model.get_input_embeddings().register_forward_hook(make_inputs_require_grad) if training_args.lora_enable: from peft import LoraConfig, get_peft_model lora_config = LoraConfig( r=training_args.lora_r, lora_alpha=training_args.lora_alpha, target_modules=find_all_linear_names(model), lora_dropout=training_args.lora_dropout, bias=training_args.lora_bias, task_type="CAUSAL_LM", ) if training_args.bits == 16: if training_args.bf16: model.to(torch.bfloat16) if training_args.fp16: model.to(torch.float16) rank0_print("Adding LoRA adapters...") model = get_peft_model(model, lora_config) if 'mpt' in model_args.model_name_or_path: tokenizer = transformers.AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, model_max_length=training_args.model_max_length, padding_side="right" ) else: tokenizer = transformers.AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, model_max_length=training_args.model_max_length, padding_side="right", use_fast=False, ) if model_args.version == "v0": if tokenizer.pad_token is None: smart_tokenizer_and_embedding_resize( special_tokens_dict=dict(pad_token="[PAD]"), tokenizer=tokenizer, model=model, ) elif model_args.version == "v0.5": tokenizer.pad_token = tokenizer.unk_token else: tokenizer.pad_token = tokenizer.unk_token if model_args.version in conversation_lib.conv_templates: conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version] else: conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1"] if model_args.vision_tower is not None: model.get_model().initialize_vision_modules( model_args=model_args, fsdp=training_args.fsdp ) vision_tower = model.get_vision_tower() vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device) data_args.image_processor = vision_tower.image_processor data_args.is_multimodal = True model.config.image_aspect_ratio = data_args.image_aspect_ratio model.config.tokenizer_padding_side = tokenizer.padding_side model.config.tokenizer_model_max_length = tokenizer.model_max_length model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter if model_args.tune_mm_mlp_adapter: model.requires_grad_(False) for p in model.get_model().mm_projector.parameters(): p.requires_grad = True model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter if training_args.freeze_mm_mlp_adapter: for p in model.get_model().mm_projector.parameters(): p.requires_grad = False # reinitialize image_pooler if model.get_model().image_pooler is not None: model.get_model().image_pooler.bert = model.get_model().image_pooler.bert.apply(model.get_model().image_pooler.bert._init_weights) if training_args.bits in [4, 8]: model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device) model.get_model().image_pooler.to(dtype=compute_dtype, device=training_args.device) model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end model.config.mm_projector_lr = training_args.mm_projector_lr training_args.use_im_start_end = model_args.mm_use_im_start_end model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer) if training_args.unfreeze_n_vision_tower_layers is not None: if data_args.vision_tower == 'biovil': print(f'Unfreezing all vision tower layers') for param in model.get_vision_tower().parameters(): param.requires_grad = True # unfreeze vit_pooler and last image encoder layer # print(f'Unfreezing partial vision tower layers') # for param in model.get_vision_tower().encoder.encoder.layer4.parameters(): # param.requires_grad = True # for param in model.get_vision_tower().encoder.encoder.fc.parameters(): # param.requires_grad = True # for param in model.get_vision_tower().encoder.vit_pooler.parameters(): # param.requires_grad = True else: print(f'Unfreezing last {training_args.unfreeze_n_vision_tower_layers} layers of vision tower') for layer in model.get_vision_tower().vision_tower.vision_model.encoder.layers[-training_args.unfreeze_n_vision_tower_layers:]: for param in layer.parameters(): param.requires_grad = True if training_args.bits in [4, 8]: from peft.tuners.lora import LoraLayer for name, module in model.named_modules(): if isinstance(module, LoraLayer): if training_args.bf16: module = module.to(torch.bfloat16) if 'norm' in name: module = module.to(torch.float32) if 'lm_head' in name or 'embed_tokens' in name: if hasattr(module, 'weight'): if training_args.bf16 and module.weight.dtype == torch.float32: module = module.to(torch.bfloat16) data_module = make_supervised_data_module(tokenizer=tokenizer, data_args=data_args) from transformers import TrainerCallback class SaveCallback(TrainerCallback): def on_save(self, args, state, control, **kwargs): print("on save") checkpoint_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(state.global_step)) if args.lora_enable: state_dict = get_peft_state_maybe_zero_3( model.named_parameters(), training_args.lora_bias ) non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3_extended( model ) print("Saving LoRA state dict...") print(args.local_rank) if args.local_rank in [-1, 0]: model.config.save_pretrained(checkpoint_dir) model.save_pretrained(checkpoint_dir, state_dict=state_dict) print(checkpoint_dir) torch.save(non_lora_state_dict, os.path.join(checkpoint_dir, 'non_lora_trainables.bin')) from llava.train.llama_patch import upcast_layer_for_flash_attention model = upcast_layer_for_flash_attention(model, torch.bfloat16) trainer = LLaVATrainer(model=model, tokenizer=tokenizer, args=training_args, callbacks=[SaveCallback()], **data_module) if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")): trainer.train(resume_from_checkpoint=True) else: trainer.train() trainer.save_state() model.config.use_cache = True if training_args.lora_enable: state_dict = get_peft_state_maybe_zero_3( model.named_parameters(), training_args.lora_bias ) non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3_extended( model ) if training_args.local_rank == 0 or training_args.local_rank == -1: model.config.save_pretrained(training_args.output_dir) model.save_pretrained(training_args.output_dir, state_dict=state_dict) torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin')) else: safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir) if __name__ == "__main__": train()