modeling_emu.py

from functools import partial
from typing import Any, List, Optional, Mapping, Callable
from collections import OrderedDict
from argparse import Namespace
import torch
from torch import nn
import torch.nn.functional as F
import torchvision.transforms as T
import PIL
import transformers
from transformers import PreTrainedModel, PreTrainedTokenizer

from .configuration_emu import EmuConfig
from .constants import *
from .modeling_llama import LlamaForCausalLM
from .visual import EVAVisionTransformer


class EmuPreTrainedModel(PreTrainedModel):
    config_class = EmuConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = False
    _no_split_modules = ["LlamaDecoderLayer", "Block"]
    _skip_keys_device_placement = "past_key_values"

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

class EmuForClsAndRegression(EmuPreTrainedModel):

    def __init__(self, config):
        super(EmuForClsAndRegression, self).__init__(config)

        self.lm = LlamaForCausalLM(config=config)

        self.lm.model.embed_tokens.padding_idx = config.pad_token_id

    def get_num_layers(self):
        return len(self.lm.model.layers)

class EmuModel(EmuPreTrainedModel):

    def __init__(self, config):
        super().__init__(config)

        vision_config = Namespace(**config.vision_config)

        self.visual = EVAVisionTransformer(
            img_size=vision_config.image_size,
            patch_size=vision_config.patch_size,
            embed_dim=vision_config.width,
            depth=vision_config.layers,
            num_heads=vision_config.width // vision_config.head_width,
            mlp_ratio=vision_config.mlp_ratio,
            qkv_bias=vision_config.qkv_bias,
            drop_path_rate=vision_config.drop_path_rate,
            norm_layer=partial(nn.LayerNorm, eps=vision_config.layer_norm_eps),
            xattn=vision_config.xattn,
            postnorm=vision_config.postnorm,
        )

        self.decoder = EmuForClsAndRegression(config)

        self.gradient_checkpointing = False
        
        self.n_query = vision_config.n_query
        self.v_query = vision_config.v_query

    @property
    def device(self):
        return next(iter(self.parameters())).device

    @property
    def dtype(self):
        return next(iter(self.parameters())).dtype

    @torch.no_grad()
    def encode_image(self, image: torch.Tensor, *, n_query=None):
        n_query = n_query if n_query is not None else self.n_query

        image_embeds = self.visual(image)
        image_embeds = image_embeds[:, 1:, :]
        b, n, c = image_embeds.shape
        sqrt_n = int(n**0.5)
        image_embeds = image_embeds.permute(0, 2, 1).view(b, c, sqrt_n, sqrt_n)

        stride = int(sqrt_n // (n_query ** 0.5))
        image_embeds = F.avg_pool2d(image_embeds, kernel_size=(stride, stride), stride=stride)
        image_embeds = image_embeds.view(b, c, -1).permute(0, 2, 1).contiguous()
        return image_embeds


class EmuForCausalLM(EmuPreTrainedModel):
    _auto_class = "AutoModelForCausalLM"

    def __init__(self, config):
        super().__init__(config)

        self.config = config
        self.model = EmuModel(config)
        # LM to EVA
        self.project_down = nn.Linear(config.hidden_size, config.d_model, bias=False)
        # EVA to LM
        self.project_up = nn.Linear(config.d_model, config.hidden_size, bias=False)

        self.n_query = self.model.n_query
        self.v_query = self.model.v_query

        self.image_placeholder = DEFAULT_IMG_TOKEN + DEFAULT_IMAGE_TOKEN * self.n_query + DEFAULT_IMG_END_TOKEN
        # temporarily borrow [gIMG] as the video frame feature placeholder.
        self.video_placeholder = DEFAULT_IMG_TOKEN + DEFAULT_gIMG_TOKEN * self.v_query + DEFAULT_IMG_END_TOKEN

    @property
    def device(self):
        return next(iter(self.parameters())).device

    @property
    def dtype(self):
        return next(iter(self.parameters())).dtype


    @torch.no_grad()
    def generate(
        self,
        input_ids,
        attention_mask,
        image: Optional[torch.Tensor] = None,
        video: Optional[torch.Tensor] = None,
        num_beams=5,
        max_new_tokens=10,
        min_len=1,
        do_sample=False,
        penalty_alpha=None,
        top_p=None,
        top_k=None,
        temperature=None,
        length_penalty=-1,
        repetition_penalty=1.0,
        **kwargs
    ):

        text_embeds = self.model.decoder.lm.model.embed_tokens(input_ids).to("cuda")
        if image is not None:
            prompt_image_embeds = self.model.encode_image(image, n_query=self.n_query)
            _, _, c = prompt_image_embeds.shape
            prompt_image_embeds = prompt_image_embeds.view(-1, c)
            prompt_image_embeds = self.project_up(prompt_image_embeds)
            image_idx = (input_ids == IMAGE)
            text_embeds[image_idx] = prompt_image_embeds.to(text_embeds.device)

        if video is not None:
            prompt_video_embeds = self.model.encode_image(video, n_query=self.v_query)
            _, _, c = prompt_video_embeds.shape
            prompt_video_embeds = prompt_video_embeds.view(-1, c)
            prompt_video_embeds = self.project_up(prompt_video_embeds)
            video_idx = (input_ids == VIDEO)
            text_embeds[video_idx] = prompt_video_embeds.to(text_embeds.device)

        outputs = self.model.decoder.lm.generate(
            inputs_embeds=text_embeds,
            attention_mask=attention_mask,
            do_sample=do_sample,
            num_beams=num_beams,
            max_new_tokens=max_new_tokens,
            min_length=min_len,
            length_penalty=length_penalty,
            repetition_penalty=repetition_penalty,
            penalty_alpha=penalty_alpha,
            top_k=top_k,
            top_p=top_p,
            temperature=temperature,
            **kwargs,
        )

        return outputs

    def prepare_image_input(self, images):
        image_size: int = self.config.vision_config['image_size']
        transform = T.Compose(
            [
                T.Resize(
                    (image_size, image_size), interpolation=T.InterpolationMode.BICUBIC
                ),
                T.ToTensor(),
                T.Normalize(OPENAI_DATASET_MEAN, OPENAI_DATASET_STD),
            ]
        )
        images = [transform(image) for image in images]
        return torch.stack(images, 0)

    def _prepare_chat_template(self, text, system_msg=""):
        text = [
            system_msg + USER_TOKEN + ": " + t + ASSISTANT_TOKEN +":"
            for t in text
        ]
        return text

    def prepare_text_input(
        self, 
        text: List[str],
        tokenizer: PreTrainedTokenizer,
        image_placeholder: str = DEFAULT_IMG_PLACEHOLDER,
        video_placeholder: str = DEFAULT_VID_PLACEHOLDER,
        ):
        text = [
            t.replace(image_placeholder, self.image_placeholder).replace(video_placeholder, self.video_placeholder)
            for t in text
        ]
        input_ids = tokenizer(text, padding="longest", return_tensors="pt")
        return input_ids
        

    def build_input_ids(
            self,
            text: List[str],
            tokenizer: PreTrainedTokenizer,
            image: Optional[List["PIL.Image"]] = None,
            video: Optional[List["PIL.Image"]] = None,
            system_msg: str = "",
            to_cuda: bool = True
        ):

        if self.config.model_version == "chat":
            text = self._prepare_chat_template(text, system_msg)

        if image is not None:
            image = self.prepare_image_input(image)
        if video is not None:
            video = self.prepare_image_input(video)
        inputs = self.prepare_text_input(text, tokenizer)
        input_ids = inputs.input_ids
        attention_mask =  inputs.attention_mask

        if to_cuda:
            input_ids = input_ids.to("cuda")
            attention_mask = attention_mask.to("cuda")
            if image is not None:
                image = image.to("cuda")
            if video is not None:
                video = video.to("cuda")


        
        return {
            'input_ids': input_ids,
            'attention_mask': attention_mask,
            'image': image,
            'video': video
        }