EveryDream2trainer/caption_cog.py

"""
Copyright [2022-2023] Victor C Hall

Licensed under the GNU Affero General Public License;
You may not use this code except in compliance with the License.
You may obtain a copy of the License at

    https://www.gnu.org/licenses/agpl-3.0.en.html

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 io
import argparse
import time
import json
import logging
import re
from typing import TYPE_CHECKING, Generator, Optional, List, Tuple, Dict, Any

import torch
from torchvision import transforms

from PIL import Image
import PIL.ImageOps as ImageOps
from pynvml import *

from transformers import AutoModelForCausalLM, LlamaTokenizer, PreTrainedTokenizer, BitsAndBytesConfig
from transformers.modeling_outputs import BaseModelOutputWithPast
from colorama import Fore, Style

from plugins.caption_plugins import load_prompt_alteration_plugin
from utils.patch_cog import patch_cog
from data.generators import image_path_generator, SUPPORTED_EXT

try:
    from moai.load_moai import prepare_moai
except ImportError:
    print("moai not found, skipping")

Image.MAX_IMAGE_PIXELS = 715827880*4 # expand the size limit

IMAGE_SIZE: int = 490
PATCH_SIZE: int = 14

patch_cog() # fixes inv_freq key error with cogvlm, quantization, and newer transformers revisions

def build_conversation_input_ids(
            tokenizer: PreTrainedTokenizer,
            *,
            query: str,
            history: Optional[List[Tuple[str, str]]] = None,
            images: Optional[List[Image.Image]] = None,
            starts_with: Optional[str] = None,
    ):
        # based on https://huggingface.co/THUDM/cogvlm-chat-hf/blob/main/modeling_cogvlm.py
        image_size: int = IMAGE_SIZE
        patch_size: int = PATCH_SIZE
        assert images is None or len(images) <= 1, f"not support multi images by now."
        history = history or []

        text = f"Question: {query} Answer: "
        text += starts_with if starts_with is not None else ""

        input_ids = [tokenizer.bos_token_id]
        token_type_ids = [0]
        if images is not None and len(images) == 1:
            # vision
            transform = transforms.Compose(
                [
                    transforms.Resize(
                        (image_size, image_size), interpolation=transforms.InterpolationMode.BICUBIC
                    ),
                    transforms.ToTensor(),
                    transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
                ]
            )
            images = [transform(images[0])]
            vision_token_num = (image_size // patch_size) * (image_size // patch_size) + 2
            input_ids += [tokenizer.pad_token_id] * vision_token_num
            token_type_ids += [1] * vision_token_num
        text_ids = tokenizer.encode(text, add_special_tokens=False)

        input_ids += text_ids
        token_type_ids += [0] * len(text_ids)
        attention_mask = [1] * len(input_ids)

        return {
            "input_ids": torch.tensor(input_ids, dtype=torch.long),
            "token_type_ids": torch.tensor(token_type_ids, dtype=torch.long),
            "attention_mask": torch.tensor(attention_mask, dtype=torch.long),
            "images": images,
        }

def get_gpu_memory_map():
    nvmlInit()
    handle = nvmlDeviceGetHandleByIndex(0)
    info = nvmlDeviceGetMemoryInfo(handle)
    nvmlShutdown()
    return info.used/1024/1024

def save_params(args, gen_kwargs):
    save_path = os.path.join(args.image_dir, "caption_cog_params.txt")
    args_dict = {
        "args": vars(args),
        "gen_kwargs": gen_kwargs,
    }
    pretty_print = json.dumps(args_dict, indent=4)
    with open(save_path, "w") as f:
        f.write(pretty_print)

def create_bnb_config():
    return BitsAndBytesConfig(
        bnb_4bit_compute_dtype="float32",
        bnb_4bit_quant_type= "fp4",
        bnb_4bit_use_double_quant=False,
        llm_int8_enable_fp32_cpu_offload=False,
        llm_int8_has_fp16_weight=False,
        llm_int8_skip_modules=None,
        llm_int8_threshold= 6.0,
        load_in_4bit=True,
        load_in_8bit=False,
        quant_method="bitsandbytes"
    )

class MoaiManager:
    def __init__(self, model_name: str):
        self.model_name = model_name
        self.moai_model = None
        self.moai_processor = None
        self.seg_model = None
        self.seg_processor = None
        self.od_model = None
        self.od_processor = None
        self.sgg_model = None
        self.ocr_model = None

    def load_model(self, bits: int=4, grad_ckpt: bool=False, lora: bool=False, dtype: str="fp16"):
        moai_model, moai_processor, seg_model, seg_processor, od_model, od_processor, sgg_model, ocr_model \
            = prepare_moai(moai_path=self.model_name, bits=bits, grad_ckpt=grad_ckpt, lora=lora, dtype=dtype)
        self.moai_model = moai_model
        self.moai_processor = moai_processor
        self.seg_model = seg_model
        self.seg_processor = seg_processor
        self.od_model = od_model
        self.od_processor = od_processor
        self.sgg_model = sgg_model
        self.ocr_model = ocr_model

        return moai_model, moai_processor

    def get_inputs(self, image: Image.Image, prompt: str):
        moai_inputs = self.moai_model.demo_process(image=image, 
                                    prompt=prompt, 
                                    processor=self.moai_processor,
                                    seg_model=self.seg_model,
                                    seg_processor=self.seg_processor,
                                    od_model=self.od_model,
                                    od_processor=self.od_processor,
                                    sgg_model=self.sgg_model,
                                    ocr_model=self.ocr_model,
                                    device="cuda:0")
        return moai_inputs

    def __call__(self, moai_inputs, do_sample=True, temperature=0.9, top_p=0.95, max_new_tokens=256, use_cache=True) -> Any:
        with torch.inference_mode():
            generate_ids = self.moai_model.generate(**moai_inputs, do_sample=do_sample, temperature=temperature, top_p=top_p, max_new_tokens=max_new_tokens, use_cache=use_cache)
        answer = self.moai_processor.batch_decode(generate_ids, skip_special_tokens=True)[0].split("[U")[0]
        return answer

class CogVLMManager:
    def __init__(self, model_name: str):
        self.model_name = model_name
        self.tokenizer = None
        self.model = None

    def load_model(self):
        self.tokenizer = LlamaTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
        self.model = AutoModelForCausalLM.from_pretrained(
            self.model_name,
            torch_dtype=torch.bfloat16,
            low_cpu_mem_usage=True,
            trust_remote_code=True,
            quantization_config=create_bnb_config()
        )
        return self.model, self.tokenizer

    def get_inputs(self, prompt: str, history: List[Tuple[str, str]], images: List[Image.Image], starts_with: str):
        return build_conversation_input_ids(self.tokenizer, query=prompt, history=history, images=images, starts_with=starts_with)

    def get_gen_kwargs(self, args):
        gen_kwargs = {
            "max_length": args.max_length,
            "do_sample": args.top_k is not None or args.top_p is not None or args.temp is not None or False,
            "length_penalty": args.length_penalty,
            "num_beams": args.num_beams,
            "temperature": args.temp,
            "top_k": args.top_k,
            "top_p": args.top_p,
            "repetition_penalty": args.repetition_penalty,
            "no_repeat_ngram_size": args.no_repeat_ngram_size,
            "min_new_tokens": args.min_new_tokens,
            "max_new_tokens": args.max_new_tokens,
            "length_penalty": args.length_penalty,
        }
        print(gen_kwargs)
        if args.max_new_tokens is not None:
            logging.info(f"** max_new_tokens set to {args.max_new_tokens}, ignoring max_length")
            del gen_kwargs["max_length"]

        if not gen_kwargs["do_sample"]:
            logging.info(f"** Using greedy sampling")
            del gen_kwargs["top_k"]
            del gen_kwargs["top_p"]
            del gen_kwargs["temperature"]
        else:
            logging.info(f"** Sampling enabled")
        return gen_kwargs

def model_manager_factory(model_name: str):
    if "moai" in model_name:
        return MoaiManager(model_name)
    else:
        return CogVLMManager(model_name)

def main(args):
    prompt_plugin_fn = load_prompt_alteration_plugin(args.prompt_plugin, args=args)
    model_manager = model_manager_factory(args.model)

    model, tokenizer = model_manager.load_model()

    args.append = args.append or ""
    if len(args.append) > 0:
        args.append = " " + args.append.strip()

    gen_kwargs = model_manager.get_gen_kwargs(args)

    force_words_ids = None
    if args.force_words is not None:
        force_words = args.force_words.split(",") if args.force_words is not None else []
        logging.info(f"** force_words: {Fore.LIGHTGREEN_EX}{force_words}{Style.RESET_ALL}")
        force_words_ids = tokenizer(force_words, add_special_tokens=False)["input_ids"] if force_words else []

    bad_words_ids = None
    if args.bad_words is not None:
        bad_words = args.bad_words.split(",") if args.bad_words is not None else []
        logging.info(f"** bad_words: {Fore.LIGHTGREEN_EX}{bad_words}{Style.RESET_ALL}")
        bad_words_ids = tokenizer(bad_words, add_special_tokens=False)["input_ids"] if bad_words else []

    logging.info(f"** gen_kwargs: \n{Fore.LIGHTGREEN_EX}{gen_kwargs}{Style.RESET_ALL}")

    save_params(args, gen_kwargs)

    total_start_time = time.time()
    i_processed = 0

    starts_with = args.starts_with.strip() if args.starts_with is not None else ""

    for i, image_path in enumerate(image_path_generator(args.image_dir, do_recurse=not args.no_recurse)):
        candidate_caption_path = image_path.replace(os.path.splitext(image_path)[-1], ".txt")

        if args.no_overwrite and os.path.exists(candidate_caption_path):
            logging.warning(f"Skipping {image_path}, caption already exists.")
            continue

        cap_start_time = time.time()
        image = Image.open(image_path)

        try:
            image = image.convert("RGB")
            image = ImageOps.exif_transpose(image)
        except Exception as e:
            logging.warning(f"Non-fatal error processing {image_path}: {e}")
            continue
        
        logging.debug(f" __ Prompt before plugin: {Fore.LIGHTGREEN_EX}{args.prompt}{Style.RESET_ALL}")
        prompt = prompt_plugin_fn(image_path, args=args)
        logging.debug(f" __ Modified prompt after plugin: {Fore.LIGHTGREEN_EX}{prompt}{Style.RESET_ALL}")

        inputs = build_conversation_input_ids(tokenizer, query=prompt, history=[], images=[image], starts_with=args.starts_with)  # chat mode

        inputs = {
            "input_ids": inputs["input_ids"].unsqueeze(0).to("cuda"),
            "token_type_ids": inputs['token_type_ids'].unsqueeze(0).to("cuda"),
            "attention_mask": inputs['attention_mask'].unsqueeze(0).to("cuda"),
            "images": [[inputs["images"][0].to("cuda").to(torch.bfloat16)] for _ in range(args.num_beams)],
            "output_hidden_states": True,
            "return_dict": True
        }

        # print(f"** inputs type: {type(inputs)}") # dict
        # print(f"** inputs len: {len(inputs)}") # 4
        # print(f"** inputs keys: {inputs.keys()}") # dict_keys(['input_ids', 'token_type_ids', 'attention_mask', 'images'])
        # print(f"** inputs['images'] shape: {inputs['images'].shape}") # list has no shape
        # print(f"** image_path: {image_path}")

        with torch.no_grad():
            #input_decoded = tokenizer.decode(inputs["input_ids"][0], skip_special_tokens=True)
            #logging.debug(f"inputs decoded: {input_decoded}")
            #print(f"calling generate with input shapes: {inputs['input_ids'].shape}, {inputs['token_type_ids'].shape}, {inputs['attention_mask'].shape}, {inputs['images'][0][0].shape}")
            #calling generate with input shapes: torch.Size([1, 1352]), torch.Size([1, 1352]), torch.Size([1, 1352]), torch.Size([3, 490, 490])
            outputs = model.generate(**inputs, **gen_kwargs, force_words_ids=force_words_ids, bad_words_ids=bad_words_ids)
            print(f"type of outputs: {type(outputs)}, outputs shape: {outputs.shape}") 
            #print(f"type of hidden_states: {type(hidden_states)}, outputs shape: {hidden_states.shape}")

            len_inputs = inputs['input_ids'].shape[1]
            outputs_without_prompt = outputs[:, len_inputs:]

            caption = tokenizer.decode(outputs_without_prompt[0], skip_special_tokens=True)
            if not args.remove_starts_with:
                # deal with caption starting with comma, etc
                if not re.match(r"^\W", caption):
                    caption = starts_with + " " + caption
                else:
                    caption = starts_with + caption

            caption += args.append

            with open(candidate_caption_path, "w") as f:
                f.write(caption)
            vram_gb = get_gpu_memory_map()
            elapsed_time = time.time() - cap_start_time
            logging.info(f"n:{i:05}, VRAM: {Fore.LIGHTYELLOW_EX}{vram_gb:0.1f} GB{Style.RESET_ALL}, elapsed: {Fore.LIGHTYELLOW_EX}{elapsed_time:0.1f}{Style.RESET_ALL} sec, Captioned {Fore.LIGHTYELLOW_EX}{image_path}{Style.RESET_ALL}: ")
            logging.info(f"{Fore.LIGHTCYAN_EX}{caption}{Style.RESET_ALL}")
            i_processed += 1

    if i_processed == 0:
        logging.info(f"** No images found in {args.image_dir} with extension in {SUPPORTED_EXT} OR no images left to caption (did you use --no_overwrite?)")
        exit(1)

    total_elapsed_time = time.time() - total_start_time
    avg_time = total_elapsed_time / i_processed
    hh_mm_ss = time.strftime("%H:%M:%S", time.gmtime(total_elapsed_time))
    logging.info(f"** Done captioning {args.image_dir} with prompt '{prompt}', total elapsed: {hh_mm_ss} (hh_mm_ss), avg: {avg_time:0.1f} sec/image")


def configure_logging(args: argparse.Namespace):
    level = logging.INFO if not args.debug else logging.DEBUG
    filemode = "a" if args.append_log else "w"
    logging.basicConfig(filename="caption_cog.log", 
                        level=level, 
                        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
                        filemode=filemode)

    console = logging.StreamHandler()
    console.setLevel(level)
    console.setFormatter(logging.Formatter('%(message)s'))
    logging.getLogger('').addHandler(console)

EXAMPLES = """ex.
Basic example:
  python caption_cog.py --image_dir /mnt/mydata/kyrie/ --prompt 'Describe this image in detail, including the subject matter and medium of the artwork.'

Use probabilistic sampling by using any of top_k, top_p, or temp:
  python caption_cog.py --image_dir \"c:/users/chadley/my documents/pictures\" --prompt \"What is this?\" --top_p 0.9 
  
Use beam search and probabilistic sampling:
  python caption_cog.py --image_dir \"c:/users/chadley/my documents/pictures\" --prompt \"Write a description.\" --max_new_tokens 75 --num_beams 4 --temp 0.9 --top_k 3 --top_p 0.9 --repetition_penalty 1.0 --no_repeat_ngram_size 0 --min_new_tokens 5

Force "cat" and "dog" and disallow the word "depicts":
  python caption_cog.py --image_dir /mnt/lcl/nvme/mldata/test --num_beams 3 --force_words "cat,dog" --bad_words "depicts"

Use a lot of beams and try to control the length with length_penalty:
  python caption_cog.py --image_dir /mnt/lcl/nvme/mldata/test --num_beams 8 --length_penalty 0.8 --prompt "Write a single sentence description."

Notes:
  1. Setting top_k, top_p, or temp enables probabilistic sampling (aka "do_sample"), otherwise greedy sampling is used.
    a. num_beams 1 and do_sample false uses "greedy decoding"
    b. num_beams 1 and do_sample true uses "multinomial sampling"
    c. num_beams > 1 and do_sample true uses "beam-search multinomial sampling"
    d. num_beams > 1 and do_sample false uses "beam-search decoding"
  2. Max_length and max_new_tokens are mutually exclusive.  If max_new_tokens is set, max_length is ignored.  Default is max_length 2048 if nothing set.
    Using Max may abruptly end caption, consider modifying prompt or use length_penalty instead.

Find more info on the Huggingface Transformers documentation: https://huggingface.co/docs/transformers/main_classes/text_generation
Parameters definitions and use map directly to their API.
"""

DESCRIPTION = f"** {Fore.LIGHTBLUE_EX}CogVLM captioning script{Style.RESET_ALL} **\n Use --help for usage."

if __name__ == "__main__":
    argparser = argparse.ArgumentParser()
    argparser.add_argument("--debug", action="store_true", help="Enable debug logging")
    argparser.add_argument("--disable_4bit", action="store_true", help="Disables 4bit inference for compatibility or experimentation. Bad for VRAM, fallback is bf16.")
    argparser.add_argument("--temp", type=float, default=None, help="Temperature for sampling")
    argparser.add_argument("--num_beams", type=int, default=2, help="Number of beams for beam search, default 1 (off)")
    argparser.add_argument("--top_k", type=int, default=None, help="Top-k, filter k highest probability tokens before sampling")
    argparser.add_argument("--top_p", type=float, default=None, help="Top-p, for sampling, selects from top tokens with cumulative probability >= p")
    argparser.add_argument("--repetition_penalty", type=float, default=1.0, help="Repetition penalty")
    argparser.add_argument("--no_repeat_ngram_size", type=int, default=0, help="No repetition n-gram size")
    argparser.add_argument("--min_new_tokens", type=int, default=5, help="Minimum number of tokens in returned caption.")
    argparser.add_argument("--max_new_tokens", type=int, default=None, help="Maximum number of tokens in returned caption.")
    argparser.add_argument("--max_length", type=int, default=2048, help="Alternate to max_new_tokens, limits context.")
    argparser.add_argument("--length_penalty", type=float, default=1.0, help="Length penalty, lower values encourage shorter captions.")
    argparser.add_argument("--prompt", type=str, default="Write a description.", help="Prompt that will guide captioning")
    argparser.add_argument("--image_dir", type=str, default=None, help="Path to folder of images to caption")
    argparser.add_argument("--no_overwrite", action="store_true", help="Skips captioning images that already have a caption file.")
    argparser.add_argument("--force_words", type=str, default=None, help="Forces the model to include these words in the caption, use CSV format.")
    argparser.add_argument("--bad_words", type=str, default=None, help="Words that will not be allowed, use CSV format.")
    argparser.add_argument("--append", type=str, default=None, help="Extra string to append to all captions. ex. 'painted by John Doe'")
    argparser.add_argument("--no_recurse", action="store_true", help="Do not recurse into subdirectories.")
    argparser.add_argument("--prompt_plugin", type=str, default=None, help="Function name to modify prompt, edit code to add plugins.")
    argparser.add_argument("--starts_with", type=str, default=None, help="Force start words on the output caption.")
    argparser.add_argument("--remove_starts_with", action="store_true", help="Removes the starts_with words from the output caption.")
    argparser.add_argument("--append_log", action="store_true", help="Sets logging to append mode.")
    argparser.add_argument("--model", type=str, default="THUDM/cogvlm-chat-hf", help="Model to use for captioning.")
    args, unknown_args = argparser.parse_known_args()
    
    configure_logging(args)

    unknown_args_dict = {}
    for i in range(0, len(unknown_args), 2):
        key = unknown_args[i].lstrip('-')  # Remove the leading '--'
        value = unknown_args[i + 1]
        unknown_args_dict[key] = value
        setattr(args, key, value)  # Add each unknown argument to the args namespace

    logging.info(f"** Unknown args have been added to args for plugins: {Fore.LIGHTGREEN_EX}{unknown_args_dict}{Style.RESET_ALL}")

    print(DESCRIPTION)
    print(EXAMPLES)

    if args.image_dir is None:
        logging.error(f"** {Fore.RED}Error: image_dir is required.{Style.RESET_ALL}")
        exit(1)

    if not os.path.exists(args.image_dir):
        logging.error(f"** {Fore.RED}Error: image_dir {args.image_dir} does not exist.{Style.RESET_ALL}")
        exit(1)

    startprint = f"** Running: {args.image_dir} with prompt '{args.prompt}"
    if args.starts_with is not None:
        startprint += f" {args.starts_with}'"
    else:
        startprint += "'"
    startprint += f" <caption>"
    if args.append is not None:
        startprint += f", and appending: {args.append}"
    logging.info(startprint)

    main(args)