diff --git a/M-MMMU/README.md b/M-MMMU/README.md deleted file mode 100644 index f883830c..00000000 --- a/M-MMMU/README.md +++ /dev/null @@ -1,58 +0,0 @@ -# Multilingual MMMU -Multilingual MMMU: Multilingual Massive Multi-task Multimodality Understanding - -## Instruction -Multilingual MMMU is a benchmark that aims to evaluate the performance of multimodal foundation models in understanding multilingual multimodal data across multiple tasks. It provides a standardized evaluation framework and dataset for researchers to compare and analyze the general understandability of different multimodal models in handling various knowledge tasks in multilingual cases. - -We are looking for high school or college exam questions that have both image and text in the question: -- The question needs to contain both image and text. -- The question needs to be highly precise, and the answer needs to be easy to verify, e.g., a multiple-choice answer, a number, a list of number, or a boolean variable, etc. -- The topic covers Art, Business, Medical, Humanities, Science and Engineering. See the full list of the category in the following. -- You can either find questions from the Internet, or Textbook or come up with your own question (if you are an expert in that field). - -## Category -| Category | Translation Count | Annotation Count | -|--------------------------------------- |------------------- |------------------ | -| **Art and Design** | | | -| - Art | | | -| - Art_Theory | | | -| - Design | | | -| - Music | | | -| | | | -| **Business** | | | -| - Accounting | | | -| - Economics | | | -| - Finance | | | -| - Manage | | | -| - Marketing | | | -| | | | -| **Health and Medicine** | | | -| - Basic_Medical_Science | | | -| - Clinical_Medicine | | | -| - Diagnostics_and_Laboratory_Medicine | | | -| - Pharmacy | | | -| - Public_Health | | | -| | | | -| **Humanities and Social Science** | | | -| - History | | | -| - Literature | | | -| - Psychology | | | -| - Sociology | | | -| | | | -| **Science** | | | -| - Biology | | | -| - Chemistry | | | -| - Geography | | | -| - Math | | | -| - Physics | | | -| | | | -| **Tech and Engineering** | | | -| - Agriculture | | | -| - Architecture_and_Engineering | | | -| - Computer_Science | | | -| - Electronics | | | -| - Energy_and_Power | | | -| - Materials | | | -| - Mechanical_Engineering | | | - -## \ No newline at end of file diff --git a/lmms_eval/models/__init__.py b/lmms_eval/models/__init__.py index 86711a9c..7e28a9bf 100755 --- a/lmms_eval/models/__init__.py +++ b/lmms_eval/models/__init__.py @@ -12,7 +12,6 @@ logger.add(sys.stdout, level="WARNING") AVAILABLE_MODELS = { - "cambrian1": "Cambrian1", "llava": "Llava", "qwen_vl": "Qwen_VL", "fuyu": "Fuyu", diff --git a/lmms_eval/models/cambrian1.py b/lmms_eval/models/cambrian1.py deleted file mode 100644 index 06648482..00000000 --- a/lmms_eval/models/cambrian1.py +++ /dev/null @@ -1,420 +0,0 @@ -import torch - -torch.backends.cuda.matmul.allow_tf32 = True - - -import copy -from tqdm import tqdm -from datetime import timedelta - -from lmms_eval import utils -from lmms_eval.api.instance import Instance -from lmms_eval.api.model import lmms -from lmms_eval.api.registry import register_model -from lmms_eval.utils import stop_sequences_criteria - -from accelerate import Accelerator, DistributedType, InitProcessGroupKwargs -from accelerate.state import AcceleratorState -from typing import List, Optional, Union, Tuple -from packaging import version -import warnings - -warnings.filterwarnings("ignore") - -from loguru import logger as eval_logger - -try: - from cambrian.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN - from cambrian.conversation import conv_templates - from cambrian.model.builder import load_pretrained_model - from cambrian.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path -except Exception as e: - eval_logger.debug("Cambrian is not installed. Please head over https://github.com/cambrian-mllm/cambrian/ and install Cambrian to use this model.\nError: %s" % e) - -if version.parse(torch.__version__) >= version.parse("2.1.2"): - best_fit_attn_implementation = "sdpa" -else: - best_fit_attn_implementation = "eager" - -# Cambrian model names and their corresponding conversation templates -# So that we can use the correct conversation template for the model -conv_modes = { - "cambrian-phi3-3b": "phi3", - "cambrian-8b": "llama_3", - "cambrian-13b": "vicuna_v1", - "cambrian-34b": "chatml_direct", -} - -@register_model("cambrian1") -class Cambrian1(lmms): - """ - Cambrian1 Model - """ - - def __init__( - self, - pretrained: str = "nyu-visionx/cambrian-8b", - truncation: Optional[bool] = True, - device: Optional[str] = "cuda:0", - batch_size: Optional[Union[int, str]] = 1, - model_name=None, - attn_implementation=best_fit_attn_implementation, - device_map="cuda:0", - conv_template="llama_3", - use_cache=True, - tie_weights: bool = True, - truncate_context=False, # whether to truncate the context in generation, set it False - customized_config=None, # ends in json - **kwargs, - ) -> None: - super().__init__() - # Do not use kwargs for now - assert kwargs == {}, f"Unexpected kwargs: {kwargs}" - - accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52)) - accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs]) - self.accelerator = accelerator - if accelerator.num_processes > 1: - self._device = torch.device(f"cuda:{accelerator.local_process_index}") - self.device_map = f"cuda:{accelerator.local_process_index}" - elif accelerator.num_processes == 1 and device_map == "auto": - self._device = torch.device(device) - self.device_map = device_map - else: - self._device = torch.device(f"cuda:{accelerator.local_process_index}") - self.device_map = f"cuda:{accelerator.local_process_index}" - - cambrian_model_args = {} - if customized_config is not None: - cambrian_model_args["customized_config"] = customized_config - if attn_implementation is not None: - cambrian_model_args["attn_implementation"] = attn_implementation - if "use_flash_attention_2" in kwargs: - cambrian_model_args["use_flash_attention_2"] = kwargs["use_flash_attention_2"] - model_name = model_name if model_name is not None else get_model_name_from_path(pretrained) - try: - # Try to load the model with the multimodal argument - self._tokenizer, self._model, self._image_processor, self._max_length = load_pretrained_model(pretrained, None, model_name, device_map=self.device_map, **cambrian_model_args) - except TypeError: - # Load the model without cambrian_model_args - self._tokenizer, self._model, self._image_processor, self._max_length = load_pretrained_model(pretrained, None, model_name, device_map=self.device_map) - - self._config = self._model.config - self.model.eval() - if tie_weights: - self.model.tie_weights() - - self.truncation = truncation - batch_size = int(batch_size) - assert batch_size == 1, f"Batch size should be 1 for Cambrian, but got {batch_size}." - self.batch_size_per_gpu = batch_size - self.conv_template = conv_modes[model_name] if model_name in conv_modes else conv_template - self.use_cache = use_cache - self.truncate_context = truncate_context - if accelerator.num_processes > 1: - assert accelerator.distributed_type in [DistributedType.FSDP, DistributedType.MULTI_GPU, DistributedType.DEEPSPEED], "Unsupported distributed type provided. Only DDP and FSDP are supported." - # If you want to use DistributedType.DEEPSPEED, you have to run accelerate config before using the model - # Also, you have to select zero stage 0 (equivalent to DDP) in order to make the prepare model works - # I tried to set different parameters in the kwargs to let default zero 2 stage works, but it didn't work. - if accelerator.distributed_type == DistributedType.DEEPSPEED: - kwargs = { - "train_micro_batch_size_per_gpu": self.batch_size_per_gpu, - "train_batch_size": self.batch_size_per_gpu * accelerator.num_processes, - } - AcceleratorState().deepspeed_plugin.deepspeed_config_process(must_match=True, **kwargs) - eval_logger.info("Detected that you are using DistributedType.DEEPSPEED. Make sure you run `accelerate config` and set zero stage to 0") - - if accelerator.distributed_type == DistributedType.FSDP or accelerator.distributed_type == DistributedType.DEEPSPEED: - self._model = accelerator.prepare(self.model) - else: - self._model = accelerator.prepare_model(self.model, evaluation_mode=True) - self.accelerator = accelerator - if self.accelerator.is_local_main_process: - eval_logger.info(f"Using {accelerator.num_processes} devices with data parallelism") - self._rank = self.accelerator.local_process_index - self._world_size = self.accelerator.num_processes - elif accelerator.num_processes == 1 and device_map == "auto": - eval_logger.info(f"Using {accelerator.num_processes} devices with tensor parallelism") - self._rank = 0 - self._word_size = 1 - else: - eval_logger.info(f"Using single device: {self._device}") - self.model.to(self._device) - self._rank = 0 - self._world_size = 1 - - @property - def config(self): - # return the associated transformers.AutoConfig for the given pretrained model. - return self._config - - @property - def tokenizer(self): - return self._tokenizer - - @property - def model(self): - # returns the model, unwrapping it if using Accelerate - if hasattr(self, "accelerator"): - return self.accelerator.unwrap_model(self._model) - else: - return self._model - - @property - def eot_token_id(self): - # we use EOT because end of *text* is more accurate for what we're doing than end of *sentence* - return self.tokenizer.eos_token_id - - @property - def max_length(self): - return self._max_length - - def pad_sequence(self, input_ids, batch_first, padding_value): - if self.tokenizer.padding_side == "left": - input_ids = [torch.flip(_input_ids, [0]) for _input_ids in input_ids] - input_ids = torch.nn.utils.rnn.pad_sequence(input_ids, batch_first=batch_first, padding_value=padding_value) - if self.tokenizer.padding_side == "left": - input_ids = torch.flip(input_ids, [1]) - return input_ids - - @property - def batch_size(self): - return self.batch_size_per_gpu - - @property - def device(self): - return self._device - - @property - def rank(self): - return self._rank - - @property - def world_size(self): - return self._world_size - - def tok_encode(self, string: str, left_truncate_len=None, add_special_tokens=None) -> List[int]: - """ """ - add_special_tokens = False if add_special_tokens is None else add_special_tokens - encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens) - # left-truncate the encoded context to be at most `left_truncate_len` tokens long - if left_truncate_len: - encoding = encoding[-left_truncate_len:] - return encoding - - def tok_decode(self, tokens): - try: - return self.tokenizer.decode(tokens) - except: - return self.tokenizer.decode([tokens]) - - def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]: - # TODO - res = [] - pbar = tqdm(total=len(requests), disable=(self.rank != 0), desc="Model Responding") - - for contexts, doc_to_target, doc_to_visual, doc_id, task, split in [reg.args for reg in requests]: - # encode, pad, and truncate contexts for this batch - if type(doc_to_target) == str: - continuation = doc_to_target - else: - continuation = doc_to_target(self.task_dict[task][split][doc_id]) - visuals = [doc_to_visual(self.task_dict[task][split][doc_id])] - # If there are more than one image in the sample, take the first image - # Cambrian only supports one image at a time - for idx, visual in enumerate(visuals): - if len(visual) > 1: - visuals[idx] = [visual[0]] - - visuals = self.flatten(visuals) - image_sizes = [[visual.size[0], visual.size[1]] for visual in visuals] - if visuals: - image = process_images(visuals, self._image_processor, self._config) - if type(image) is list: - image = [_image.to(dtype=torch.float16, device=self.device) for _image in image] - else: - image = image.to(dtype=torch.float16, device=self.device) - else: - image = None - - prompts_input = contexts[0] if isinstance(contexts, list) else contexts - - if image is not None and len(image) != 0 and DEFAULT_IMAGE_TOKEN not in prompts_input: - """ - Three senarios: - 1. No image, and there for, no image token should be added. - 2. image token is already specified in the context, so we don't need to add it. - 3. image token is not specified in the context and there is image inputs, so we need to add it. In this case, we add the image token at the beginning of the context and add a new line. - """ - if self.model.config.mm_use_im_start_end: - image_tokens = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN - else: - image_tokens = DEFAULT_IMAGE_TOKEN - prompts_input = image_tokens + "\n" + (contexts[0] if isinstance(contexts, list) else contexts) - - conv = conv_templates[self.conv_template].copy() - conv.append_message(conv.roles[0], prompts_input) - conv.append_message(conv.roles[1], None) - prompt = conv.get_prompt() - pad_token_id = self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id - contxt_id = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device) - # Add the answer of the second role - conv.messages[1][1] = continuation - - prompt = conv.get_prompt() - input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device) - labels = input_ids.clone() - # Context part no need to calculate for loss - labels[0, : contxt_id.shape[1]] = -100 - with torch.inference_mode(): - outputs = self.model(input_ids=input_ids, labels=labels, images=image, use_cache=True, image_sizes=image_sizes) - loss = outputs["loss"] - # loss = torch.exp(loss) - logits = outputs["logits"] - greedy_tokens = logits.argmax(dim=-1) - cont_toks = input_ids[:, contxt_id.shape[1] :] # [1, seq] - greedy_tokens = greedy_tokens[:, contxt_id.shape[1] : input_ids.shape[1]] # [1, seq] - max_equal = (greedy_tokens == cont_toks).all() - res.append((float(loss.item()), bool(max_equal))) - pbar.update(1) - pbar.close() - return res - - def flatten(self, input): - new_list = [] - for i in input: - for j in i: - new_list.append(j) - return new_list - - def generate_until(self, requests: List[Instance]) -> List[str]: - res = [] - - def _collate(x): - # the negative sign on len(toks) sorts descending - this has a few advantages: - # - time estimates will always be over not underestimates, which is more useful for planning - # - to know the size of a batch when going through the list, you know the first one is always the batch - # padded context length. this is useful to simplify the batching logic and more importantly to make - # automatic adaptive batches much much easier to implement - # - any OOMs will happen right away rather than near the end - toks = self.tok_encode(x[0]) - return -len(toks), x[0] - - # we group requests by their generation_kwargs, - # so that we don't try to execute e.g. greedy sampling and temp=0.8 sampling - # in the same batch. - re_ords = utils.Collator([reg.args for reg in requests], _collate, grouping=True) - chunks = re_ords.get_batched(n=self.batch_size, batch_fn=None) - num_iters = len(requests) // self.batch_size if len(requests) % self.batch_size == 0 else len(requests) // self.batch_size + 1 - pbar = tqdm(total=num_iters, disable=(self.rank != 0), desc="Model Responding") - for chunk in chunks: - contexts, all_gen_kwargs, doc_to_visual, doc_id, task, split = zip(*chunk) - task = task[0] - split = split[0] - batched_visuals = [doc_to_visual[0](self.task_dict[task][split][ids]) for ids in doc_id] # [B, N] - # If there are more than one image in the sample, take the first image - # Cambrian only supports one image at a time - for idx, visual in enumerate(batched_visuals): - if len(visual) > 1: - batched_visuals[idx] = [visual[0]] - flattened_visuals = self.flatten(batched_visuals) # [B*N] - # we assume all gen kwargs in the batch are the same - # this is safe to assume because the `grouper` object ensures it. - gen_kwargs = all_gen_kwargs[0] - - # Set default values for until and max_new_tokens - until = [self.tok_decode(self.eot_token_id)] - - # Update values from gen_kwargs if present - if "until" in gen_kwargs: - until = gen_kwargs.pop("until") - if isinstance(until, str): - until = [until] - elif not isinstance(until, list): - raise ValueError(f"Expected `gen_kwargs['until']` to be of type Union[str,list] but got {type(until)}") - - if "image_aspect_ratio" in gen_kwargs.keys() and "image_aspect_ratio" not in self._config.__dict__: - # here we should pop it out of gen_kwargs so that it doesn't get passed to the model for next step of generation - self._config.image_aspect_ratio = gen_kwargs.pop("image_aspect_ratio") - eval_logger.info(f"Setting image aspect ratio: {self._config.image_aspect_ratio}") - # encode, pad, and truncate contexts for this batch - if flattened_visuals: - image_tensor = process_images(flattened_visuals, self._image_processor, self._config) - if type(image_tensor) is list: - image_tensor = [_image.to(dtype=torch.float16, device=self.device) for _image in image_tensor] - else: - image_tensor = image_tensor.to(dtype=torch.float16, device=self.device) - else: - image_tensor = None - - # prompts_input = contexts[0] - - question_input = [] - - for visual, context in zip(batched_visuals, contexts): - if image_tensor is not None and len(image_tensor) != 0 and DEFAULT_IMAGE_TOKEN not in context: - """ - Three senarios: - 1. No image, and there for, no image token should be added. - 2. image token is already specified in the context, so we don't need to add it. - 3. image token is not specified in the context and there is image inputs, so we need to add it. In this case, we add the image token at the beginning of the context and add a new line. - """ - if self.model.config.mm_use_im_start_end: - image_tokens = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN - else: - image_tokens = DEFAULT_IMAGE_TOKEN - question = image_tokens + "\n" + context - else: - question = context - - conv = conv_templates[self.conv_template].copy() - conv.append_message(conv.roles[0], question) - conv.append_message(conv.roles[1], None) - prompt_question = conv.get_prompt() - question_input.append(prompt_question) - - # preconfigure gen_kwargs with defaults - gen_kwargs["image_sizes"] = [flattened_visuals[idx].size for idx in range(len(flattened_visuals))] - if "max_new_tokens" not in gen_kwargs: - gen_kwargs["max_new_tokens"] = 1024 - if "temperature" not in gen_kwargs: - gen_kwargs["temperature"] = 0 - if "top_p" not in gen_kwargs: - gen_kwargs["top_p"] = None - if "num_beams" not in gen_kwargs: - gen_kwargs["num_beams"] = 1 - - input_ids_list = [tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt") for prompt in question_input] - pad_token_ids = self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id - input_ids = self.pad_sequence(input_ids_list, batch_first=True, padding_value=pad_token_ids).to(self.device) - attention_masks = input_ids.ne(pad_token_ids).to(self.device) - # These steps are not in Cambrian's original code, but are necessary for generation to work - try: - cont = self.model.generate( - input_ids, - attention_mask=attention_masks, - pad_token_id=pad_token_ids, - images=image_tensor, - image_sizes=gen_kwargs["image_sizes"], - do_sample=True if gen_kwargs["temperature"] > 0 else False, - temperature=gen_kwargs["temperature"], - top_p=gen_kwargs["top_p"], - num_beams=gen_kwargs["num_beams"], - max_new_tokens=gen_kwargs["max_new_tokens"], - use_cache=self.use_cache, - ) - text_outputs = self.tokenizer.batch_decode(cont, skip_special_tokens=True) - except Exception as e: - raise e - eval_logger.error(f"Error {e} in generating") - cont = "" - text_outputs = [""] - - res.extend(text_outputs) - self.cache_hook.add_partial("generate_until", (context, gen_kwargs), text_outputs) - pbar.update(1) - # reorder this group of results back to original unsorted form - res = re_ords.get_original(res) - - pbar.close() - return res \ No newline at end of file