ola_vlm/train/sherlock_dsg_train.py

# 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 os
import copy
from dataclasses import dataclass, field
import json
import logging
import pathlib
from typing import Dict, Optional, Sequence, List

import numpy as np
import torch

import transformers
import tokenizers

from ola_vlm.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
from torch.utils.data import Dataset
from ola_vlm.train.llava_trainer import LLaVATrainer

from ola_vlm import conversation as conversation_lib
from ola_vlm.model import *
from ola_vlm.mm_utils import tokenizer_image_token

from PIL import Image, ImageFile
from transformers import set_seed

set_seed(42)

# Enable loading of truncated images
ImageFile.LOAD_TRUNCATED_IMAGES = True


local_rank = None


def rank0_print(*args):
    if local_rank == 0:
        print(*args)


from packaging import version
IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse('0.14')


@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)
    unfreeze_mm_vision_tower: bool = field(default=False)
    unfreeze_whole_model: bool = field(default=False)
    use_s2: bool = field(default=False)
    s2_scales: Optional[str] = field(default="336,1008")
    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_patch_merge_type: Optional[str] = field(default='flat')
    mm_vision_select_feature: Optional[str] = field(default="patch")

    attn_mask_type: Optional[str] = field(default="causal")

    contrastive_loss_weight: Optional[float] = field(default=0.1)
    
    # visual interpretors
    image_generator: Optional[str] = field(default="stabilityai/stable-diffusion-2-1-unclip")
    image_segmentor: Optional[str] = field(default="shi-labs/oneformer_coco_swin_large") # sam_vit_l_0b3195.pth
    depth_estimator: Optional[str] = field(default="depth_anything_v2_vitl.pth")

    mode: Optional[str] = field(default="depth-seg-gen")
    num_task_tokens: Optional[int] = 0
    task_token_format: Optional[str] = "expand_emb"
    sample_tokens: Optional[bool] = False
    pass_text_to_aux: Optional[bool] = False

    # dinov2
    use_dinov2: Optional[bool] = False
    dinov2_model: Optional[str] = "/mnt/projects4jw/jiteshjain_sherlock/dinov2-large-res336"
    dinov2_dim: Optional[str] = 1024
    dinov2_layers: Optional[str] = "8-12"
    dinov2_loss_weight: Optional[float] = 0.25
    
    use_contrastive: Optional[bool] = True
    use_ce: Optional[bool] = False
    layer_indices: Optional[str] = "d8-14_s10-16_g12-18"
    loss_weights: Optional[str] = "d0.5_s0.5_g0.5"

    # gen
    img_head_depth: Optional[int] = 1
    img_head_dim_head: Optional[int] = 32
    img_head_num_heads: Optional[int] = 4
    img_head_num_tokens: Optional[int] = 1
    img_head_output_dim: Optional[int] = 1024
    img_head_ff_mult: Optional[int] = 1

    # seg
    seg_head_depth: Optional[int] = 1
    seg_head_dim_head: Optional[int] = 32
    seg_head_num_heads: Optional[int] = 4
    seg_head_num_tokens: Optional[int] = 576
    seg_head_output_dim: Optional[int] = 1536 # 256
    seg_head_ff_mult: Optional[int] = 1
    seg_teacher: Optional[str] = "oneformer" # "sam"

    # depth
    depth_head_depth: Optional[int] = 1
    depth_head_dim_head: Optional[int] = 32
    depth_head_num_heads: Optional[int] = 4
    depth_head_num_tokens: Optional[int] = 576
    depth_head_output_dim: Optional[int] = 1024
    depth_head_ff_mult: Optional[int] = 1
    use_intermediate_depth: Optional[bool] = False

    freeze_task_token: Optional[bool] = field(default=False)
    freeze_aux_heads: Optional[bool] = field(default=False)
    use_reference_model: Optional[bool] = field(default=False)

@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)
    depth_folder: Optional[str] = field(default=None)
    unclip_folder: Optional[str] = field(default=None)
    seg_folder: Optional[str] = field(default=None)
    image_aspect_ratio: str = 'square'
    use_cost: bool = 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)
    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
    mm_vision_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(named_params, require_grad_only=True):
    to_return = {k: t for k, t in named_params if "lora_" not in k}
    if require_grad_only:
        to_return = {k: t for k, t in to_return.items() if t.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']
    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'))

    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, '<Image>' + DEFAULT_IMAGE_TOKEN + '</Image>')
            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_phi_3(
    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.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 = 1
        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

            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 > 0:
                round_len -= 2
                instruction_len -= 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_llama_3(
    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.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])]
        for conv_idx in range(3, len(rounds), 2):
            re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx + 2]))
        cur_len = 1
        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

            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 > 0:
            #     round_len -= 1
            #     instruction_len -= 1
            
            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_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 != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14:
                round_len -= 1
                instruction_len -= 1

            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_qwen(
        sources, 
        tokenizer: transformers.PreTrainedTokenizer, 
        has_image: bool = False, 
        system_message: str = "You are a helpful assistant."
    ) -> Dict:
    # roles = {"human": "<|im_start|>user", "gpt": "<|im_start|>assistant"}
    roles = {"human": "user", "gpt": "assistant"}

    # Add image tokens to tokenizer as a special tokens
    # Use a deepcopy of tokenizer so that we don't modify on the tokenizer
    tokenizer = copy.deepcopy(tokenizer)
    # When there is actually an image, we add the image tokens as a special token
    if has_image:
        tokenizer.add_tokens(["<image>"], special_tokens=True)

    image_token_index = tokenizer.convert_tokens_to_ids("<image>")
    im_start, im_end = tokenizer.additional_special_tokens_ids
    # unmask_tokens = ["<|im_start|>", "<|im_start|>", "\n"]
    unmask_tokens_idx =  [198, im_start, im_end]
    nl_tokens = tokenizer("\n").input_ids

    # Reset Qwen chat templates so that it won't include system message every time we apply
    chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
    tokenizer.chat_template = chat_template

    # _system = tokenizer("system").input_ids + nl_tokens
    # _user = tokenizer("user").input_ids + nl_tokens
    # _assistant = tokenizer("assistant").input_ids + nl_tokens

    # Apply prompt templates
    input_ids, targets = [], []
    for i, source in enumerate(sources):
        if roles[source[0]["from"]] != roles["human"]:
            source = source[1:]

        input_id, target = [], []

        # New version, use apply chat template
        # Build system message for each sentence
        input_id += tokenizer.apply_chat_template([{"role" : "system", "content" : system_message}])
        target += [IGNORE_INDEX] * len(input_id)

        for conv in source:
            # Make sure llava data can load
            try:
                role = conv["role"]
                content = conv["content"]
            except:
                role = conv["from"]
                content = conv["value"]

            role =  roles.get(role, role)
            
            conv = [{"role" : role, "content" : content}]
            encode_id = tokenizer.apply_chat_template(conv)
            input_id += encode_id
            if role in ["user", "system"]:
                target += [IGNORE_INDEX] * len(encode_id)
            else:
                target += encode_id
                    
        assert len(input_id) == len(target), f"{len(input_id)} != {len(target)}"
        for idx, encode_id in enumerate(input_id):
            if encode_id in unmask_tokens_idx:
                target[idx] = encode_id
            if encode_id == image_token_index:
                input_id[idx] = IMAGE_TOKEN_INDEX
        input_ids.append(input_id)
        targets.append(target)
    input_ids = torch.tensor(input_ids, dtype=torch.long)
    targets = torch.tensor(targets, dtype=torch.long)

    return dict(
        input_ids=input_ids,  # tensor(bs x seq_len)
        labels=targets,  # tensor(bs x seq_len)
    )


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.version == "llama3":
        return preprocess_llama_3(sources, tokenizer, has_image=has_image)
    if conversation_lib.default_conversation.version == "phi3":
        return preprocess_phi_3(sources, tokenizer, has_image=has_image)
    if conversation_lib.default_conversation.version == "qwen":
        return preprocess_qwen(sources, tokenizer, has_image=has_image)
    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)
    # 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)


def read_jsonl(path):
    list_data_dict = []

    with open(path, "r") as file:
        for line in file:
            d = json.loads(line)
            list_data_dict.append(d)
    return list_data_dict


def _obtain_seg_texts(file_path):
    def _remove_specific_word(text, word_to_remove):
        import re
        tokens = re.findall(r'\b\w+\b|[,.]', text)
        result_tokens = []
        word_found = False

        for i, token in enumerate(tokens):
            if token == word_to_remove:
                if not word_found:
                    # Keep the first occurrence and mark it as found
                    result_tokens.append(token)
                    word_found = True
                else:
                    # Remove any preceding punctuation if it's just before this word
                    if i > 0 and tokens[i-1] in {',', '.'}:
                        result_tokens.pop()
            else:
                result_tokens.append(token)

        # Join tokens and clean up spaces before punctuation
        result_text = ' '.join(result_tokens)
        result_text = re.sub(r'\s([,.](?:\s|$))', r'\1', result_text)
        return result_text

    with open(file_path) as f:
        lines = f.readlines()
    
    seg_labels = {}
    for line in lines:
        key = line.split("<IMG>")[1].strip("\n")
        label = line.split("<IMG>")[2].strip("\n")
        label = _remove_specific_word(label, "wall")
        label = _remove_specific_word(label, "window")
        seg_labels[key] = label
    
    return seg_labels

from ola_vlm.ola_utils import PANOPTIC_QUESTIONS, SEMANTIC_QUESTIONS, INSTANCE_QUESTIONS
import random
def get_object_data_split(data_args):
    list_data_dict = []
    for bucket in ["train"]:
            panoptic_labels = _obtain_seg_texts(os.path.join(data_args.image_folder, "coco", "panoptic.txt"))
            semantic_labels = _obtain_seg_texts(os.path.join(data_args.image_folder, "coco", "semantic.txt"))
            instance_labels = _obtain_seg_texts(os.path.join(data_args.image_folder, "coco", "instance.txt"))

            for key in panoptic_labels.keys():
                assert key in semantic_labels.keys() and key in instance_labels.keys(), "Instance, semantic, and panoptic labels should have the same keys."
                prob_task = np.random.uniform(0,1.)
                question_prob = np.random.uniform(0,1.)
                if prob_task < 0.33:
                    answer = semantic_labels[key]
                    if question_prob > 0.90:
                        question = "What objects can be seen in the image?"
                    else:
                        question = random.choice(SEMANTIC_QUESTIONS)
                elif prob_task < 0.66:
                    answer = instance_labels[key]
                    if question_prob > 0.90:
                        question = "What objects can be seen in the image?"
                    else:
                        question = random.choice(INSTANCE_QUESTIONS)
                else:
                    answer = panoptic_labels[key]
                    if question_prob > 0.90:
                        question = "What objects can be seen in the image?"
                    else:
                        question = random.choice(PANOPTIC_QUESTIONS)

                question += "\n<image>"
                conversations = [ 
                        {
                            "from": "human", 
                            "value": question
                        }, 
                        {
                            "from": "gpt",
                            "value": answer
                        },
                    ]
                list_data_dict.append(
                        {
                            "conversations": conversations,
                            "image": "coco/" + bucket + "2017/" + key,
                        }
                    )
                    
    random.shuffle(list_data_dict)
    return list_data_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__()
        if "jsonl" in data_path:
            list_data_dict = read_jsonl(data_path)
        else:
            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

        if data_args.use_cost:
            cost_list_data = get_object_data_split(data_args)
            self.list_data_dict.extend(cost_list_data)

    def __len__(self):
        return len(self.list_data_dict)

    @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 __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]:
            image_file = self.list_data_dict[i]['image']
            image_folder = self.data_args.image_folder
            processor = self.data_args.image_processor
            try:
                crop_size = self.data_args.image_processor.crop_size
            except:
                crop_size = self.data_args.image_processor.size

            try:
                image = Image.open(os.path.join(image_folder, image_file)).convert('RGB')
                pil_image = Image.open(os.path.join(image_folder, image_file)).convert('RGB')
            except Exception as e:
                from icecream import ic
                ic("----------------------------------")
                ic("OS ERROROROROROROROROROROOR")
                ic("OS ERROROROROROROROROROROOR")
                ic(image_file)
                ic(e)
                ic("OS ERROROROROROROROROROROOR")
                ic("OS ERROROROROROROROROROROOR")
                ic("===================================")
                return self.__getitem__(0)
            
            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
                image = expand2square(image, tuple(int(x*255) for x in processor.image_mean))
                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
            else:
                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
            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'] = image
            data_dict['pil_image'] = pil_image
            data_dict['seg_mask'] = 1
            data_dict['depth_mask'] = 1
            data_dict['gen_mask'] = 1
        elif self.data_args.is_multimodal:
            try:
                crop_size = self.data_args.image_processor.crop_size
            except:
                crop_size = self.data_args.image_processor.size
            data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width'])
            data_dict['pil_image'] = Image.new('RGB', (crop_size['width'], crop_size['height']), color='black')
            data_dict['seg_mask'] = 0
            data_dict['depth_mask'] = 0
            data_dict['gen_mask'] = 0
        
        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 all(x is not None and x.shape == images[0].shape for x in images):
                batch['images'] = torch.stack(images)
            else:
                batch['images'] = images
        
        if 'pil_image' in instances[0]:
            pil_images = [instance['pil_image'] for instance in instances]
            batch['pil_images'] = pil_images

            seg_mask = [instance['seg_mask'] for instance in instances]
            batch['seg_mask'] = torch.tensor(seg_mask)
            
            depth_mask = [instance['depth_mask'] for instance in instances]
            batch['depth_mask'] = torch.tensor(depth_mask)
            
            gen_mask = [instance['gen_mask'] for instance in instances]
            batch['gen_mask'] = torch.tensor(gen_mask)
        
        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 add_special_tokens(
    special_tokens: List,
    tokenizer: transformers.PreTrainedTokenizer,
    model: transformers.PreTrainedModel,
):
    """Resize tokenizer and embedding.

    Initialize new token embeddings to follow the distribution of existing embeddings.
    """
    # Add special tokens to tokenizer
    num_new_tokens = tokenizer.add_tokens(special_tokens, special_tokens=True)
    # Resize the token embeddings in the model
    model.resize_token_embeddings(len(tokenizer))

    if num_new_tokens > 0:
        # Get input embeddings and compute global mean and std over all dimensions
        input_embeddings = model.get_input_embeddings().weight.data
        input_mean, input_std = input_embeddings.mean(), input_embeddings.std()

        # Initialize new input embeddings with the same distribution as existing ones
        input_embeddings[-num_new_tokens:] = torch.nn.init.normal_(
            torch.empty(num_new_tokens, input_embeddings.size(1)),
            mean=input_mean.item(),
            std=input_std.item()
        )

        # Check if model has output embeddings and initialize them similarly
        if model.get_output_embeddings() is not None:
            output_embeddings = model.get_output_embeddings().weight.data
            output_mean, output_std = output_embeddings.mean(), output_embeddings.std()

            # Initialize new output embeddings with the same distribution as existing ones
            output_embeddings[-num_new_tokens:] = torch.nn.init.normal_(
                torch.empty(num_new_tokens, output_embeddings.size(1)),
                mean=output_mean.item(),
                std=output_std.item()
            )
            
def train(attn_implementation=None):
    global local_rank

    parser = transformers.HfArgumentParser(
        (ModelArguments, DataArguments, TrainingArguments))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
    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"],
                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 'phi' in model_args.model_name_or_path.lower():
            model = OlaLlavaPhi3ForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **bnb_model_from_pretrained_args
            )
        elif 'qwen' in model_args.model_name_or_path.lower():
            model = OlaLlavaQwenForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **bnb_model_from_pretrained_args
            )
        else:
            model = OlaLlavaLlamaForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **bnb_model_from_pretrained_args
            )
    else:
        if 'phi' in model_args.model_name_or_path.lower():
            model = transformers.Phi3ForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **bnb_model_from_pretrained_args
            )
        elif 'qwen2' in model_args.model_name_or_path.lower():
            model = transformers.Qwen2ForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **bnb_model_from_pretrained_args
            )
        else:
            model = transformers.LlamaForCausalLM.from_pretrained(
                model_args.model_name_or_path,
                cache_dir=training_args.cache_dir,
                attn_implementation=attn_implementation,
                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
                **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 "qwen" in model_args.model_name_or_path.lower():
        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 tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.unk_token
    if tokenizer.pad_token_id is None:
        smart_tokenizer_and_embedding_resize(
                special_tokens_dict=dict(pad_token="<pad>"),
                tokenizer=tokenizer,
                model=model,
            )

    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["llava_phi_3"]

    if "sherlock" in model_args.model_name_or_path:
        vision_tower = model.get_vision_tower()

        if vision_tower is None:
            model.get_model().initialize_vision_modules(
                model_args=model_args,
                fsdp=training_args.fsdp
            )
            vision_tower = model.get_vision_tower()

        if not vision_tower.is_loaded:
            vision_tower.load_model()
        vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device)

    elif 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_grid_pinpoints = [[336,672], [672,336], [672,672], [1008,336], [336,1008]]        
    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

    if training_args.bits in [4, 8]:
        model.get_model().mm_projector.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 model_args.unfreeze_mm_vision_tower:
        model.requires_grad_(False)
        for p in model.get_model().mm_projector.parameters():
            p.requires_grad = True
        model.get_model().vision_tower.requires_grad_(True)
    else:
        model.get_model().vision_tower.requires_grad_(False)

    model.config.use_s2 = model_args.use_s2
    model.config.s2_scales = model_args.s2_scales

    if "sherlock" not in model_args.model_name_or_path.split("/")[-1]:
        aux_mode = model_args.mode
        model.config.aux_mode = model_args.mode
        model.config.contrastive_loss_weight = model_args.contrastive_loss_weight
        model.config.num_task_tokens = model_args.num_task_tokens
        model.config.task_token_format = model_args.task_token_format
        model.config.pass_text_to_aux = model_args.pass_text_to_aux
        model.config.use_contrastive = model_args.use_contrastive
        model.config.use_ce = model_args.use_ce

        layer_indices = model_args.layer_indices

        pattern = r'[a-zA-Z]\d+(?:-\d+)?'

        import re
        # Extract matching substrings from each string
        matches = re.findall(pattern, layer_indices)

        depth_layer_indices = "0"
        seg_layer_indices = "0"
        img_layer_indices = "0"

        for match in matches:
            if match.startswith('d'):
                depth_layer_indices = match[1:]
            elif match.startswith('s'):
                seg_layer_indices = match[1:]
            elif match.startswith('g'):
                img_layer_indices = match[1:]
        
        loss_weights = model_args.loss_weights

        pattern = r'[a-zA-Z]\d+\.\d+'
        matches = re.findall(pattern, loss_weights)

        img_loss_weight = 0.5
        seg_loss_weight = 0.5
        depth_loss_weight = 0.5

        for match in matches:
            if match.startswith('d'):
                depth_loss_weight = float(match[1:])
            elif match.startswith('s'):
                seg_loss_weight = float(match[1:])
            elif match.startswith('g'):
                img_loss_weight = float(match[1:])
        
        model.config.image_gen = {
            "depth": model_args.img_head_depth,
            "dim_head": model_args.img_head_dim_head,
            "num_heads": model_args.img_head_num_heads,
            "num_tokens": model_args.img_head_num_tokens,
            "output_dim": model_args.img_head_output_dim,
            "ff_mult": model_args.img_head_ff_mult,
            "img_layer_indices": img_layer_indices,
            "img_loss_weight": img_loss_weight,
        }
        model.config.image_generator = model_args.image_generator

        model.config.image_seg = {
            "depth": model_args.seg_head_depth,
            "dim_head": model_args.seg_head_dim_head,
            "num_heads": model_args.seg_head_num_heads,
            "num_tokens": model_args.seg_head_num_tokens,
            "output_dim": model_args.seg_head_output_dim,
            "ff_mult": model_args.seg_head_ff_mult,
            "seg_layer_indices": seg_layer_indices,
            "seg_loss_weight": seg_loss_weight,
            "seg_teacher": model_args.seg_teacher,
        }
        model.config.image_segmentor = model_args.image_segmentor

        model.config.image_depth = {
            "depth": model_args.depth_head_depth,
            "dim_head": model_args.depth_head_dim_head,
            "num_heads": model_args.depth_head_num_heads,
            "num_tokens": model_args.depth_head_num_tokens,
            "output_dim": model_args.depth_head_output_dim,
            "ff_mult": model_args.depth_head_ff_mult,
            "depth_layer_indices": depth_layer_indices,
            "depth_loss_weight": depth_loss_weight,
            "use_intermediate_depth": model_args.use_intermediate_depth,
        }
        model.config.depth_estimator = model_args.depth_estimator
        model.config.sample_tokens = model_args.sample_tokens
        num_task_tokens = model_args.num_task_tokens

        if model_args.use_dinov2:
            model.config.dinov2_feats = {
                "model": model_args.dinov2_model,
                "dinov2_layer_indices": model_args.dinov2_layers,
                "dim": model_args.dinov2_dim,
                "dinov2_loss_weight": model_args.dinov2_loss_weight,
            }

        model.config.num_task_tokens = model_args.num_task_tokens
        model.config.task_token_format = model_args.task_token_format
        if model_args.num_task_tokens > 0:
            if model_args.task_token_format == "text":
                if "depth" in aux_mode:
                    special_depth_tokens = [f"<depth_{i}>" for i in range(num_task_tokens)]
                    special_depth_tokens_str = "".join(special_depth_tokens)
                    add_special_tokens(
                        special_tokens=special_depth_tokens,
                        tokenizer=tokenizer,
                        model=model,
                    )
                    model.config.depth_tokens = tokenizer(special_depth_tokens_str).input_ids[1:]
                if "seg" in aux_mode:
                    special_seg_tokens = [f"<seg_{i}>" for i in range(num_task_tokens)]
                    special_seg_tokens_str = "".join(special_seg_tokens)
                    add_special_tokens(
                        special_tokens=special_seg_tokens,
                        tokenizer=tokenizer,
                        model=model,
                    )
                    model.config.seg_tokens = tokenizer(special_seg_tokens_str).input_ids[1:]
                if "gen" in aux_mode:
                    special_gen_tokens = [f"<gen_{i}>" for i in range(num_task_tokens)]
                    special_gen_tokens_str = "".join(special_gen_tokens)
                    add_special_tokens(
                        special_tokens=special_gen_tokens,
                        tokenizer=tokenizer,
                        model=model,
                    )
                    model.config.gen_tokens = tokenizer(special_gen_tokens_str).input_ids[1:]

            model.get_model().initialize_special_tokens(model.config)

        model.init_heads(model.config)
        model.init_target_models(model.config)
    elif model_args.unfreeze_whole_model:
        model.requires_grad_(True)
    elif model_args.unfreeze_mm_vision_tower:
        if "depth" in model_args.mode:
            for p in model.image_depth_heads.parameters():
                p.requires_grad = True
        if "gen" in model_args.mode:
            for p in model.image_gen_heads.parameters():
                p.requires_grad = True
        if "seg" in model_args.mode:
            for p in model.image_seg_heads.parameters():
                p.requires_grad = True
        if "emb" in model.config.task_token_format and model.config.num_task_tokens > 0:
            if "gen" in aux_mode:
                model.get_model().special_gen_tokens.requires_grad_(True)
            if "seg" in aux_mode:
                model.get_model().special_seg_tokens.requires_grad_(True)
            if "depth" in aux_mode:
                model.get_model().special_depth_tokens.requires_grad_(True)
    elif not model_args.tune_mm_mlp_adapter:
         if "emb" in model.config.task_token_format and model.config.num_task_tokens > 0:
            if "gen" in model.config.aux_mode:
                model.get_model().special_gen_tokens.requires_grad_(False)
            if "seg" in model.config.aux_mode:
                model.get_model().special_seg_tokens.requires_grad_(False)
            if "depth" in model.config.aux_mode:
                model.get_model().special_depth_tokens.requires_grad_(False)

    
    loss_weights = model_args.loss_weights

    import re
    pattern = r'[a-zA-Z]\d+\.\d+'
    matches = re.findall(pattern, loss_weights)

    img_loss_weight = 0.5
    seg_loss_weight = 0.5
    depth_loss_weight = 0.5

    for match in matches:
        if match.startswith('d'):
            depth_loss_weight = float(match[1:])
        elif match.startswith('s'):
            seg_loss_weight = float(match[1:])
        elif match.startswith('g'):
            img_loss_weight = float(match[1:])
    
    model.config.image_seg["seg_loss_weight"] = seg_loss_weight
    model.config.image_gen["img_loss_weight"] = img_loss_weight
    model.config.image_depth["depth_loss_weight"] = depth_loss_weight
    
    if model_args.use_reference_model:
        model.init_reference_model()
    
    for name, p in model.named_parameters():
        if "sam." in name or "da_v2_head." in name or "dinov2_model." in name or "gen_encoder." in name or "dav2_backbone." in name or "oneformer." in name:
            p.requires_grad = False

    model.img_gen_loss_weight = img_loss_weight
    model.img_seg_loss_weight = seg_loss_weight
    model.img_depth_loss_weight = depth_loss_weight

    if model_args.num_task_tokens > 0:
        if "emb" in model.config.task_token_format and model_args.freeze_task_token:
            if "gen" in model.config.aux_mode:
                model.get_model().special_gen_tokens.requires_grad_(False)
            if "seg" in model.config.aux_mode:
                model.get_model().special_seg_tokens.requires_grad_(False)
            if "depth" in model.config.aux_mode:
                model.get_model().special_depth_tokens.requires_grad_(False)
        else:
            if "gen" in model.config.aux_mode:
                model.get_model().special_gen_tokens.requires_grad_(True)
            if "seg" in model.config.aux_mode:
                model.get_model().special_seg_tokens.requires_grad_(True)
            if "depth" in model.config.aux_mode:
                model.get_model().special_depth_tokens.requires_grad_(True)

    if model_args.freeze_aux_heads:
        model.get_model().vision_tower.requires_grad_(False)
        if "depth" in model.config.aux_mode:
            for p in model.image_depth_heads.parameters():
                p.requires_grad = False
            model.depth_logit_scale.requires_grad_(False)
        if "gen" in model.config.aux_mode:
            for p in model.image_gen_heads.parameters():
                p.requires_grad = False
            model.gen_logit_scale.requires_grad_(False)
        if "seg" in model.config.aux_mode:
            for p in model.image_seg_heads.parameters():
                p.requires_grad = False
            model.seg_logit_scale.requires_grad_(False)
        
    import torch.distributed as dist
    from icecream import ic
    if dist.get_rank() == 0:
        gen_heads = 0
        depth_heads = 0
        seg_heads = 0
        for n, p in model.named_parameters():
            if p.requires_grad:
                if "gen_head" in n:
                    gen_heads += p.numel()
                elif "depth_head" in n:
                    depth_heads += p.numel()
                elif "seg_head" in n:
                    seg_heads += p.numel()
                ic(n)
        ic(depth_heads, gen_heads, seg_heads)

    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)
    trainer = LLaVATrainer(model=model,
                    tokenizer=tokenizer,
                    args=training_args,
                    **data_module)
    
    print('starting training...', local_rank)

    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(
            model.named_parameters()
        )
        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()