[Init] Add app file

README.md

@@ -1,13 +1,151 @@
 ---
-title: VideoChatGPT
-emoji: ⚡
-colorFrom: gray
-colorTo: red
+title: Ask-Anything:ChatGPT with Video Understanding
+emoji: movie_camera
+colorFrom: green
+colorTo: green
 sdk: gradio
-sdk_version: 3.29.0
+python_version: 3.8.16
 app_file: app.py
 pinned: false
-license: cc-by-sa-4.0
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🦜 VideoChat [[paper](https://arxiv.org/abs/2305.06355)]
+
+![images](assert/framework.png)
+In this study, we initiate an exploration into video understanding by introducing VideoChat, an **end-to-end chat-centric video understanding system**. It integrates video foundation models and large language models via a learnable neural interface, excelling in **spatiotemporal reasoning, event localization, and causal relationship inference**. To instructively tune this system, we propose a **video-centric instruction dataset**, composed of thousands of videos matched with detailed descriptions and conversations. This dataset emphasizes **spatiotemporal reasoning and causal relationships**, providing a valuable asset for training chat-centric video understanding systems. Preliminary qualitative experiments reveal our system’s potential across a broad spectrum of video applications and set the standard for future research.
+
+
+# :fire: Updates
+- **2023/05/11**: Release the 🦜**VideoChat V1**, which can **handle both image and video understanding!**
+    - [Model](https://drive.google.com/file/d/1BqmWHWCZBPkhTNWDAq0IfGpbkKLz9C0V/view?usp=share_link) and [Data](https://github.com/OpenGVLab/InternVideo/blob/main/Data/instruction_data.md).
+    - 🧑‍💻 *Online demo is Preparing*.
+    - 🧑‍🔧 *Tuning scripts are cleaning*.
+
+# :hourglass_flowing_sand: Schedule
+
+- [x] Small-scale video instuction data and tuning
+- [x] Instruction tuning on BLIP+UniFormerV2+Vicuna
+- [ ] Large-scale and complex video instuction data
+- [ ] Instruction tuning on strong video foundation model
+- [ ] User-friendly interactions with longer videos
+- [ ] ...
+
+# :speech_balloon: Example
+
+<div align="center">
+<b>
+  <font size="4">Comparison with ChatGPT, MiniGPT-4, LLaVA and mPLUG-Owl. </font>
+  <br>
+  <font size="4" color="red">Our VideoChat can handle both image and video understanding well!</font>
+</b>
+</div>
+<div align="center">
+<img src="assert/comparison.png" width="90%">
+</div>
+
+<div align="center">
+  <font size="4">
+	<a href="https://pjlab-gvm-data.oss-cn-shanghai.aliyuncs.com/papers/media/jesse_dance.mp4">[Video]</a> <b>Why the video is funny?</b>
+  </font>
+</div>
+<div align="center">
+<img src="assert/humor.png" width="50%">
+</div>
+
+<div align="center">
+  <font size="4">
+	<a href="https://pjlab-gvm-data.oss-cn-shanghai.aliyuncs.com/papers/media/jp_dance.mp4">[Video]</a> <b>Spatial perception</b>
+  </font>
+</div>
+<div align="center">
+<img src="assert/spatial.png" width="50%">
+</div>
+
+<div align="center">
+  <font size="4">
+	<a href="https://pjlab-gvm-data.oss-cn-shanghai.aliyuncs.com/papers/media/car_accident.mp4">[Video]</a> <b>Temporal perception</b>
+  </font>
+</div>
+<div align="center">
+<img src="assert/temporal.png" width="50%">
+</div>
+
+<div align="center">
+  <font size="4">
+	<a href="https://pjlab-gvm-data.oss-cn-shanghai.aliyuncs.com/papers/media/idol_dancing.mp4">[Video]</a> <b>Multi-turn conversation</b>
+  </font>
+</div>
+<div align="center">
+<img src="assert/multi_turn.png" width="50%">
+</div>
+
+<div align="center">
+  <font size="4">
+	<b>Image understanding</b>
+  </font>
+</div>
+<div align="center">
+<img src="assert/image.png" width="100%">
+</div>
+
+# :running: Usage
+
+- Prepare the envirment.
+    ```shell
+    pip install -r requirements.txt
+    ```
+    
+- Download [BLIP2](https://huggingface.co/docs/transformers/main/model_doc/blip-2) model:
+    - ViT: `wget https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth`
+    - QFormer: `wget https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth`
+    - Change the `vit_model_path` and `q_former_model_path` in [config.json](./configs/config.json).
+    
+- Download [StabelVicuna](https://huggingface.co/CarperAI/stable-vicuna-13b-delta) model:
+    - LLAMA: Download it from the [original repo](https://github.com/facebookresearch/llama) or [hugging face](https://huggingface.co/decapoda-research/llama-13b-hf).
+    - If you download LLAMA from the original repo, please process it via the following command:
+    ```shell
+    # convert_llama_weights_to_hf is copied from transformers
+    python src/transformers/models/llama/convert_llama_weights_to_hf.py \
+      --input_dir /path/to/downloaded/llama/weights \
+      --model_size 7B --output_dir /output/path
+    ```
+    - Download [StableVicuna-13b-deelta](https://huggingface.co/CarperAI/stable-vicuna-13b-delta) and process it:
+    ```shell
+    # fastchat v0.1.10
+    python3 apply_delta.py \
+      --base /path/to/model_weights/llama-13b \
+      --target stable-vicuna-13b \
+      --delta CarperAI/stable-vicuna-13b-delta
+    ```
+    - Change the `llama_model_path` in [config.json](./configs/config.json).
+    
+- Download [VideoChat](https://drive.google.com/file/d/1BqmWHWCZBPkhTNWDAq0IfGpbkKLz9C0V/view?usp=share_link) model:
+  
+    - Change the `videochat_model_path` in [config.json](./configs/config.json).
+    
+- Running demo with Gradio:
+    ```shell
+    python demo.py
+    ```
+    
+- Another demo on Jupyter Notebook can found in [demo.ipynb](demo.ipynb)
+
+
+# :page_facing_up: Citation
+
+If you find this project useful in your research, please consider cite:
+```BibTeX
+@article{2023videochat,
+  title={VideoChat: Chat-Centric Video Understanding},
+  author={KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao},
+  journal={arXiv preprint arXiv:2305.06355},
+  year={2023}
+}
+```
+
+# :thumbsup: Acknowledgement
+
+Thanks to the open source of the following projects:
+
+[InternVideo](https://github.com/OpenGVLab/InternVideo), [UniFormerV2](https://github.com/OpenGVLab/UniFormerV2), [MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4), [LLaVA](https://github.com/haotian-liu/LLaVA), [BLIP2](https://huggingface.co/docs/transformers/main/model_doc/blip-2), [StableLM](https://github.com/Stability-AI/StableLM).

configs/config.json

@@ -0,0 +1,28 @@
+{
+  "model": {
+    "vit_model": "eva_clip_g",
+    "vit_model_path": "model/eva_vit_g.pth",
+    "q_former_model_path": "model/blip2_pretrained_flant5xxl.pth",
+    "llama_model_path": "model/stable-vicuna-13b",
+    "videochat_model_path": "model/videochat.pth",
+    "img_size": 224,
+    "num_query_token": 32,
+    "drop_path_rate": 0.0,
+    "use_grad_checkpoint": false,
+    "vit_precision": "fp32",
+    "freeze_vit": true,
+    "freeze_mhra": false,
+    "freeze_qformer": true,
+    "low_resource": false,
+    "max_txt_len": 320,
+    "temporal_downsample": false,
+    "no_lmhra": true,
+    "double_lmhra": false,
+    "lmhra_reduction": 2.0,
+    "gmhra_layers": 8,
+    "gmhra_drop_path_rate": 0.0,
+    "gmhra_dropout": 0.5,
+    "extra_num_query_token": 64
+  },
+  "device": "cuda"
+}

conversation.py

@@ -0,0 +1,193 @@
+from PIL import Image
+
+import torch
+from transformers import StoppingCriteria, StoppingCriteriaList
+
+from enum import auto, Enum
+
+import numpy as np
+from decord import VideoReader, cpu
+import torchvision.transforms as T
+from models.video_transformers import (
+    GroupNormalize, GroupScale, GroupCenterCrop, 
+    Stack, ToTorchFormatTensor
+)
+from torchvision.transforms.functional import InterpolationMode
+from transformers import LlamaTokenizer, LlamaConfig
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+
+class SeparatorStyle(Enum):
+    """Different separator style."""
+    SINGLE = auto()
+    TWO = auto()
+
+
+def get_prompt(conv):
+    ret = conv.system + conv.sep
+    for role, message in conv.messages:
+        if message:
+            ret += role + ": " + message + conv.sep
+        else:
+            ret += role + ":"
+    return ret
+
+
+class StoppingCriteriaSub(StoppingCriteria):
+    def __init__(self, stops=[], encounters=1):
+        super().__init__()
+        self.stops = stops
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
+        for stop in self.stops:
+            if torch.all((stop == input_ids[0][-len(stop):])).item():
+                return True
+        return False
+
+
+class Chat:
+    def __init__(self, model, device='cuda:0'):
+        self.device = device
+        self.model = model
+        stop_words_ids = [torch.tensor([835]).to(self.device),
+                          torch.tensor([2277, 29937]).to(self.device)]  # '###' can be encoded in two different ways.
+        self.stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
+
+    def ask(self,text,conv):
+        conv.messages.append([conv.roles[0], text + '\n'])
+        return conv
+
+    def answer(self, conv,  img_list, max_new_tokens=200, num_beams=1, min_length=1, top_p=0.9,
+               repetition_penalty=1.0, length_penalty=1, temperature=1.0):
+        conv.messages.append([conv.roles[1], None])
+        embs = self.get_context_emb(conv, img_list)
+        outputs = self.model.llama_model.generate(
+            inputs_embeds=embs,
+            max_new_tokens=max_new_tokens,
+            stopping_criteria=self.stopping_criteria,
+            num_beams=num_beams,
+            do_sample=True,
+            min_length=min_length,
+            top_p=top_p,
+            repetition_penalty=repetition_penalty,
+            length_penalty=length_penalty,
+            temperature=temperature,
+        )
+        output_token = outputs[0]
+        if output_token[0] == 0:  # the model might output a unknow token <unk> at the beginning. remove it
+                output_token = output_token[1:]
+        if output_token[0] == 1:  # some users find that there is a start token <s> at the beginning. remove it
+                output_token = output_token[1:]
+        output_text = self.model.llama_tokenizer.decode(output_token, add_special_tokens=False)
+        output_text = output_text.split('###')[0]  # remove the stop sign '###'
+        output_text = output_text.split('Assistant:')[-1].strip()
+        conv.messages[-1][1] = output_text
+        return output_text, output_token.cpu().numpy(), conv
+        
+    def get_index(self, num_frames, num_segments):
+        seg_size = float(num_frames - 1) / num_segments
+        start = int(seg_size / 2)
+        offsets = np.array([
+            start + int(np.round(seg_size * idx)) for idx in range(num_segments)
+        ])
+        return offsets
+
+    def load_video(self, video_path, num_segments=8, return_msg=False):
+        vr = VideoReader(video_path, ctx=cpu(0))
+        num_frames = len(vr)
+        frame_indices = self.get_index(num_frames, num_segments)
+        
+        duration = len(vr) // vr.get_avg_fps()
+        index = np.linspace(0, len(vr)-1, num=int(duration))
+        buffer = vr.get_batch(index).asnumpy()
+        # transform
+        input_mean = [0.48145466, 0.4578275, 0.40821073]
+        input_std = [0.26862954, 0.26130258, 0.27577711]
+        
+        transform = T.Compose([
+            GroupScale(int(224), interpolation=InterpolationMode.BICUBIC),
+            GroupCenterCrop(224),
+            Stack(),
+            ToTorchFormatTensor(),
+            GroupNormalize(input_mean, input_std) 
+        ])
+
+        images_group = list()
+        for frame in buffer:
+            img = Image.fromarray(frame)
+            images_group.append(img)
+        images_group = list()
+        for frame_index in frame_indices:
+            img = Image.fromarray(vr[frame_index].asnumpy())
+            images_group.append(img)
+        torch_imgs_224 = transform(images_group)
+        if return_msg:
+            fps = float(vr.get_avg_fps())
+            sec = ", ".join([str(round(f / fps, 1)) for f in frame_indices])
+            # " " should be added in the start and end
+            msg = f"The video contains {len(frame_indices)} frames sampled at {sec} seconds."
+            return torch_imgs_224, msg
+        else:
+            return torch_imgs_224
+
+    def upload_video(self, image, conv, img_list, num_segments):
+        if isinstance(image, str):  # is a image path
+            vid_chat, msg = self.load_video(image, num_segments=num_segments, return_msg=True)
+            TC, H, W = vid_chat.shape
+            image = vid_chat.reshape(1, TC//3, 3, H, W).to(self.device)
+
+        else:
+            raise NotImplementedError
+        print("Input video shape:", vid_chat.shape)
+        image_emb, _ = self.model.encode_img(image)
+        img_list.append(image_emb)
+        conv.messages.append([
+            conv.roles[0], 
+            f"<Video><VideoHere></Video> {msg}\n"
+        ])
+        msg = "Received."
+        # self.conv.append_message(self.conv.roles[1], msg)
+        return msg, img_list, conv
+    
+    def upload_img(self, image, conv, img_list):
+        img = image#Image.open(image)#.convert('RGB')
+        transform = T.Compose(
+            [
+                T.Resize(
+                    (224, 224), interpolation=InterpolationMode.BICUBIC
+                ),
+                T.ToTensor(),
+                T.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+            ]
+        )
+
+        img = transform(img).unsqueeze(0).unsqueeze(0).cuda()
+        image_emb, _ = self.model.encode_img(img)
+        img_list.append(image_emb)
+        conv.messages.append([
+            conv.roles[0],
+            f"<Image><ImageHere></Image>\n"
+        ])
+        msg = "Received."
+        # self.conv.append_message(self.conv.roles[1], msg)
+        return msg,img_list, conv
+
+    def get_context_emb(self, conv, img_list):
+        prompt = get_prompt(conv)
+        #print(prompt)
+        if '<VideoHere>' in prompt:
+            prompt_segs = prompt.split('<VideoHere>')
+        else:
+            prompt_segs = prompt.split('<ImageHere>')
+        assert len(prompt_segs) == len(img_list) + 1, "Unmatched numbers of visual placeholders and videos."
+        seg_tokens = [
+            self.model.llama_tokenizer(
+                seg, return_tensors="pt", add_special_tokens=i == 0).to(self.device).input_ids
+            # only add bos to the first seg
+            for i, seg in enumerate(prompt_segs)
+        ]
+        seg_embs = [self.model.llama_model.model.embed_tokens(seg_t) for seg_t in seg_tokens]
+        mixed_embs = [emb for pair in zip(seg_embs[:-1], img_list) for emb in pair] + [seg_embs[-1]]
+        mixed_embs = torch.cat(mixed_embs, dim=1)
+        return mixed_embs

demo.py

@@ -0,0 +1,203 @@
+import torch
+import gradio as gr
+from gradio.themes.utils import colors, fonts, sizes
+
+from conversation import Chat
+
+# videochat
+from utils.config import Config
+from utils.easydict import EasyDict
+from models.videochat import VideoChat
+
+
+# ========================================
+#             Model Initialization
+# ========================================
+def init_model():
+    print('Initializing VideoChat')
+    config_file = "configs/config.json"
+    cfg = Config.from_file(config_file)
+    model = VideoChat(config=cfg.model)
+    model = model.to(torch.device(cfg.device))
+    model = model.eval()
+    chat = Chat(model)
+    print('Initialization Finished')
+    return chat
+
+
+# ========================================
+#             Gradio Setting
+# ========================================
+def gradio_reset(chat_state, img_list):
+    if chat_state is not None:
+        chat_state.messages = []
+    if img_list is not None:
+        img_list = []
+    return None, gr.update(value=None, interactive=True), gr.update(value=None, interactive=True), gr.update(placeholder='Please upload your video first', interactive=False),gr.update(value="Upload & Start Chat", interactive=True), chat_state, img_list
+
+
+def upload_img(gr_img, gr_video, chat_state, num_segments):
+    # print(gr_img, gr_video)
+    chat_state = EasyDict({
+        "system": "",
+        "roles": ("Human", "Assistant"),
+        "messages": [],
+        "sep": "###"
+    })
+    img_list = []
+    if gr_img is None and gr_video is None:
+        return None, None, gr.update(interactive=True), chat_state, None
+    if gr_video: 
+        llm_message, img_list, chat_state = chat.upload_video(gr_video, chat_state, img_list, num_segments)
+        return gr.update(interactive=True), gr.update(interactive=True), gr.update(interactive=True, placeholder='Type and press Enter'), gr.update(value="Start Chatting", interactive=False), chat_state, img_list
+    if gr_img:
+        llm_message, img_list,chat_state = chat.upload_img(gr_img, chat_state, img_list)
+        return gr.update(interactive=True), gr.update(interactive=True), gr.update(interactive=True, placeholder='Type and press Enter'), gr.update(value="Start Chatting", interactive=False), chat_state, img_list
+
+
+def gradio_ask(user_message, chatbot, chat_state):
+    if len(user_message) == 0:
+        return gr.update(interactive=True, placeholder='Input should not be empty!'), chatbot, chat_state
+    #print(chat_state)
+    chat_state =  chat.ask(user_message, chat_state)
+    chatbot = chatbot + [[user_message, None]]
+    return '', chatbot, chat_state
+
+
+def gradio_answer(gr_img, gr_video,chatbot, chat_state, img_list, num_beams, temperature):
+    llm_message,llm_message_token, chat_state = chat.answer(conv=chat_state, img_list=img_list, max_new_tokens=1000, num_beams=num_beams, temperature=temperature)
+    llm_message = llm_message.replace("<s>", "") # handle <s>
+    chatbot[-1][1] = llm_message
+    print(f"========{gr_img}##<BOS>##{gr_video}========")
+    print(chat_state,flush=True)
+    print(f"========{gr_img}##<END>##{gr_video}========")
+    # print(f"Answer: {llm_message}")
+    return chatbot, chat_state, img_list
+
+
+class OpenGVLab(gr.themes.base.Base):
+    def __init__(
+        self,
+        *,
+        primary_hue=colors.blue,
+        secondary_hue=colors.sky,
+        neutral_hue=colors.gray,
+        spacing_size=sizes.spacing_md,
+        radius_size=sizes.radius_sm,
+        text_size=sizes.text_md,
+        font=(
+            fonts.GoogleFont("Noto Sans"),
+            "ui-sans-serif",
+            "sans-serif",
+        ),
+        font_mono=(
+            fonts.GoogleFont("IBM Plex Mono"),
+            "ui-monospace",
+            "monospace",
+        ),
+    ):
+        super().__init__(
+            primary_hue=primary_hue,
+            secondary_hue=secondary_hue,
+            neutral_hue=neutral_hue,
+            spacing_size=spacing_size,
+            radius_size=radius_size,
+            text_size=text_size,
+            font=font,
+            font_mono=font_mono,
+        )
+        super().set(
+            body_background_fill="*neutral_50",
+        )
+
+
+gvlabtheme = OpenGVLab(primary_hue=colors.blue,
+        secondary_hue=colors.sky,
+        neutral_hue=colors.gray,
+        spacing_size=sizes.spacing_md,
+        radius_size=sizes.radius_sm,
+        text_size=sizes.text_md,
+        )
+
+title = """<h1 align="center"><a href="https://github.com/OpenGVLab/Ask-Anything"><img src="https://i.328888.xyz/2023/05/11/iqrAkZ.md.png" alt="Ask-Anything" border="0" style="margin: 0 auto; height: 100px;" /></a> </h1>"""
+description ="""
+        <p> VideoChat, an end-to-end chat-centric video understanding system powered by <a href='https://github.com/OpenGVLab/InternVideo'>InternVideo</a>. It integrates video foundation models and large language models via a learnable neural interface, excelling in spatiotemporal reasoning, event localization, and causal relationship inference.</p>
+        <div style='display:flex; gap: 0.25rem; '>
+        <a src="https://img.shields.io/badge/Github-Code-blue?logo=github" href="https://github.com/OpenGVLab/Ask-Anything"> <img src="https://img.shields.io/badge/Github-Code-blue?logo=github">
+        <a src="https://img.shields.io/badge/cs.CV-2305.06355-b31b1b?logo=arxiv&logoColor=red" href="https://arxiv.org/abs/2305.06355"> <img src="https://img.shields.io/badge/cs.CV-2305.06355-b31b1b?logo=arxiv&logoColor=red">
+        <a src="https://img.shields.io/badge/WeChat-Group-green?logo=wechat" href="https://pjlab-gvm-data.oss-cn-shanghai.aliyuncs.com/papers/media/wechat_group.jpg"> <img src="https://img.shields.io/badge/WeChat-Group-green?logo=wechat">
+        <a src="https://img.shields.io/discord/1099920215724277770?label=Discord&logo=discord" href="https://discord.gg/A2Ex6Pph6A"> <img src="https://img.shields.io/discord/1099920215724277770?label=Discord&logo=discord"> </div>
+        """
+
+
+with gr.Blocks(title="InternVideo-VideoChat!",theme=gvlabtheme,css="#chatbot {overflow:auto; height:500px;} #InputVideo {overflow:visible; height:320px;} footer {visibility: none}") as demo:
+    gr.Markdown(title)
+    gr.Markdown(description)
+
+    with gr.Row():
+        with gr.Column(scale=0.5, visible=True) as video_upload:
+            with gr.Column(elem_id="image") as img_part:
+                with gr.Tab("Video", elem_id='video_tab'):
+                    up_video = gr.Video(interactive=True, include_audio=True, elem_id="video_upload")#.style(height=320)
+                with gr.Tab("Image", elem_id='image_tab'):
+                    up_image = gr.Image(type="pil", interactive=True, elem_id="image_upload")#.style(height=320)
+            upload_button = gr.Button(value="Upload & Start Chat", interactive=True, variant="primary")
+            
+            num_beams = gr.Slider(
+                minimum=1,
+                maximum=10,
+                value=1,
+                step=1,
+                interactive=True,
+                label="beam search numbers",
+            )
+            
+            temperature = gr.Slider(
+                minimum=0.1,
+                maximum=2.0,
+                value=1.0,
+                step=0.1,
+                interactive=True,
+                label="Temperature",
+            )
+            
+            num_segments = gr.Slider(
+                minimum=8,
+                maximum=64,
+                value=8,
+                step=1,
+                interactive=True,
+                label="Video Segments",
+            )
+        
+        
+        with gr.Column(visible=True)  as input_raws:
+            chat_state = gr.State(EasyDict({
+                "system": "",
+                "roles": ("Human", "Assistant"),
+                "messages": [],
+                "sep": "###"
+            }))
+            img_list = gr.State()
+            chatbot = gr.Chatbot(elem_id="chatbot",label='VideoChat')
+            with gr.Row():
+                with gr.Column(scale=0.7):
+                    text_input = gr.Textbox(show_label=False, placeholder='Please upload your video first', interactive=False).style(container=False)
+                with gr.Column(scale=0.15, min_width=0):
+                    run = gr.Button("💭Send")
+                with gr.Column(scale=0.15, min_width=0):
+                    clear = gr.Button("🔄Clear️")     
+    
+    chat = init_model()
+    upload_button.click(upload_img, [up_image, up_video, chat_state, num_segments], [up_image, up_video, text_input, upload_button, chat_state, img_list])
+    
+    text_input.submit(gradio_ask, [text_input, chatbot, chat_state], [text_input, chatbot, chat_state]).then(
+        gradio_answer, [up_image, up_video, chatbot, chat_state, img_list, num_beams, temperature], [chatbot, chat_state, img_list]
+    )
+    run.click(gradio_ask, [text_input, chatbot, chat_state], [text_input, chatbot, chat_state]).then(
+        gradio_answer, [up_image, up_video,chatbot, chat_state, img_list, num_beams, temperature], [chatbot, chat_state, img_list]
+    )
+    run.click(lambda: "", None, text_input)  
+    clear.click(gradio_reset, [chat_state, img_list], [chatbot, up_image, up_video, text_input, upload_button, chat_state, img_list], queue=False)
+
+demo.launch(server_name="0.0.0.0", favicon_path='bot_avatar.jpg', enable_queue=True,ssl_keyfile="vchat_cert/privkey1.pem",ssl_certfile="vchat_cert/cert1.pem",ssl_verify=False)

models/Qformer.py

@@ -0,0 +1,1237 @@
+"""
+ * Copyright (c) 2023, salesforce.com, inc.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+ * By Junnan Li
+ * Based on huggingface code base
+ * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert
+"""
+
+import math
+import os
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple, Dict, Any
+
+import torch
+from torch import Tensor, device, dtype, nn
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from timm.models.layers import drop_path
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    NextSentencePredictorOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+from transformers.models.bert.configuration_bert import BertConfig
+
+logger = logging.get_logger(__name__)
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id
+        )
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer(
+            "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))
+        )
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+
+        self.config = config
+
+    def forward(
+        self,
+        input_ids=None,
+        position_ids=None,
+        query_embeds=None,
+        past_key_values_length=0,
+    ):
+        if input_ids is not None:
+            seq_length = input_ids.size()[1]
+        else:
+            seq_length = 0
+
+        if position_ids is None:
+            position_ids = self.position_ids[
+                :, past_key_values_length : seq_length + past_key_values_length
+            ].clone()
+
+        if input_ids is not None:
+            embeddings = self.word_embeddings(input_ids)
+            if self.position_embedding_type == "absolute":
+                position_embeddings = self.position_embeddings(position_ids)
+                embeddings = embeddings + position_embeddings
+
+            if query_embeds is not None:
+                embeddings = torch.cat((query_embeds, embeddings), dim=1)
+        else:
+            embeddings = query_embeds
+
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config, is_cross_attention):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(
+            config, "embedding_size"
+        ):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_width, self.all_head_size)
+            self.value = nn.Linear(config.encoder_width, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(
+                2 * config.max_position_embeddings - 1, self.attention_head_size
+            )
+        self.save_attention = False
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (
+            self.num_attention_heads,
+            self.attention_head_size,
+        )
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        mixed_query_layer = self.query(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if (
+            self.position_embedding_type == "relative_key"
+            or self.position_embedding_type == "relative_key_query"
+        ):
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(-1, 1)
+            position_ids_r = torch.arange(
+                seq_length, dtype=torch.long, device=hidden_states.device
+            ).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(
+                distance + self.max_position_embeddings - 1
+            )
+            positional_embedding = positional_embedding.to(
+                dtype=query_layer.dtype
+            )  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum(
+                    "bhld,lrd->bhlr", query_layer, positional_embedding
+                )
+                relative_position_scores_key = torch.einsum(
+                    "bhrd,lrd->bhlr", key_layer, positional_embedding
+                )
+                attention_scores = (
+                    attention_scores
+                    + relative_position_scores_query
+                    + relative_position_scores_key
+                )
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (
+            (context_layer, attention_probs) if output_attentions else (context_layer,)
+        )
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+    
+    
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self) -> str:
+        return 'p={}'.format(self.drop_prob)
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config, drop_path=0.):
+        super().__init__()
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.drop_path(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False, drop_path=0.,):
+        super().__init__()
+        self.self = BertSelfAttention(config, is_cross_attention)
+        self.output = BertSelfOutput(config, drop_path=drop_path)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads,
+            self.self.num_attention_heads,
+            self.self.attention_head_size,
+            self.pruned_heads,
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = (
+            self.self.attention_head_size * self.self.num_attention_heads
+        )
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[
+            1:
+        ]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config, drop_path=0.):
+        super().__init__()
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.drop_path(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        drop_path = config.drop_path_list[layer_num]
+        self.attention = BertAttention(config, drop_path=drop_path)
+        self.layer_num = layer_num
+        if (
+            self.config.add_cross_attention
+            and layer_num % self.config.cross_attention_freq == 0
+        ):
+            self.crossattention = BertAttention(
+                config, is_cross_attention=self.config.add_cross_attention,
+                drop_path=drop_path
+            )
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config, drop_path=drop_path)
+
+        self.intermediate_query = BertIntermediate(config)
+        self.output_query = BertOutput(config, drop_path=drop_path)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        query_length=0,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = (
+            past_key_value[:2] if past_key_value is not None else None
+        )
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:-1]
+
+        present_key_value = self_attention_outputs[-1]
+
+        if query_length > 0:
+            query_attention_output = attention_output[:, :query_length, :]
+
+            if self.has_cross_attention:
+                assert (
+                    encoder_hidden_states is not None
+                ), "encoder_hidden_states must be given for cross-attention layers"
+                cross_attention_outputs = self.crossattention(
+                    query_attention_output,
+                    attention_mask,
+                    head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+                query_attention_output = cross_attention_outputs[0]
+                outputs = (
+                    outputs + cross_attention_outputs[1:-1]
+                )  # add cross attentions if we output attention weights
+
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk_query,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                query_attention_output,
+            )
+            if attention_output.shape[1] > query_length:
+                layer_output_text = apply_chunking_to_forward(
+                    self.feed_forward_chunk,
+                    self.chunk_size_feed_forward,
+                    self.seq_len_dim,
+                    attention_output[:, query_length:, :],
+                )
+                layer_output = torch.cat([layer_output, layer_output_text], dim=1)
+        else:
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                attention_output,
+            )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+    def feed_forward_chunk_query(self, attention_output):
+        intermediate_output = self.intermediate_query(attention_output)
+        layer_output = self.output_query(intermediate_output, attention_output)
+        return layer_output
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [BertLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        query_length=0,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = (
+            () if output_attentions and self.config.add_cross_attention else None
+        )
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(
+                            *inputs, past_key_value, output_attentions, query_length
+                        )
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+class BertModel(BertPreTrainedModel):
+    """
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in `Attention is
+    all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+
+    def __init__(self, config, add_pooling_layer=False):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self,
+        attention_mask: Tensor,
+        input_shape: Tuple[int],
+        device: device,
+        is_decoder: bool,
+        has_query: bool = False,
+    ) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (:obj:`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (:obj:`Tuple[int]`):
+                The shape of the input to the model.
+            device: (:obj:`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if is_decoder:
+                batch_size, seq_length = input_shape
+
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = (
+                    seq_ids[None, None, :].repeat(batch_size, seq_length, 1)
+                    <= seq_ids[None, :, None]
+                )
+
+                # add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    if has_query:  # UniLM style attention mask
+                        causal_mask = torch.cat(
+                            [
+                                torch.zeros(
+                                    (batch_size, prefix_seq_len, seq_length),
+                                    device=device,
+                                    dtype=causal_mask.dtype,
+                                ),
+                                causal_mask,
+                            ],
+                            axis=1,
+                        )
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones(
+                                (batch_size, causal_mask.shape[1], prefix_seq_len),
+                                device=device,
+                                dtype=causal_mask.dtype,
+                            ),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )
+                extended_attention_mask = (
+                    causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+                )
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(
+            dtype=self.dtype
+        )  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        is_decoder=False,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if input_ids is None:
+            assert (
+                query_embeds is not None
+            ), "You have to specify query_embeds when input_ids is None"
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] - self.config.query_length
+            if past_key_values is not None
+            else 0
+        )
+
+        query_length = query_embeds.shape[1] if query_embeds is not None else 0
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            query_embeds=query_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+
+        input_shape = embedding_output.size()[:-1]
+        batch_size, seq_length = input_shape
+        device = embedding_output.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                ((batch_size, seq_length + past_key_values_length)), device=device
+            )
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if is_decoder:
+            extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask,
+                input_ids.shape,
+                device,
+                is_decoder,
+                has_query=(query_embeds is not None),
+            )
+        else:
+            extended_attention_mask = self.get_extended_attention_mask(
+                attention_mask, input_shape, device, is_decoder
+            )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if type(encoder_hidden_states) == list:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[
+                    0
+                ].size()
+            else:
+                (
+                    encoder_batch_size,
+                    encoder_sequence_length,
+                    _,
+                ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+
+            if type(encoder_attention_mask) == list:
+                encoder_extended_attention_mask = [
+                    self.invert_attention_mask(mask) for mask in encoder_attention_mask
+                ]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask
+                )
+            else:
+                encoder_extended_attention_mask = self.invert_attention_mask(
+                    encoder_attention_mask
+                )
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            query_length=query_length,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = (
+            self.pooler(sequence_output) if self.pooler is not None else None
+        )
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=True,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=True,
+        reduction="mean",
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
+            ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        Returns:
+        Example::
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+            >>> prediction_logits = outputs.logits
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+        if labels is not None:
+            use_cache = False
+        if past_key_values is not None:
+            query_embeds = None
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        sequence_output = outputs[0]
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1] :, :]
+
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores[:, :-1, :].contiguous()
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)
+            lm_loss = loss_fct(
+                shifted_prediction_scores.view(-1, self.config.vocab_size),
+                labels.view(-1),
+            )
+            if reduction == "none":
+                lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, query_embeds, past=None, attention_mask=None, **model_kwargs
+    ):
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_ids.shape)
+        query_mask = input_ids.new_ones(query_embeds.shape[:-1])
+        attention_mask = torch.cat([query_mask, attention_mask], dim=-1)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "input_ids": input_ids,
+            "query_embeds": query_embeds,
+            "attention_mask": attention_mask,
+            "past_key_values": past,
+            "encoder_hidden_states": model_kwargs.get("encoder_hidden_states", None),
+            "encoder_attention_mask": model_kwargs.get("encoder_attention_mask", None),
+            "is_decoder": True,
+        }
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (
+                tuple(
+                    past_state.index_select(0, beam_idx) for past_state in layer_past
+                ),
+            )
+        return reordered_past
+
+
+class BertForMaskedLM(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        query_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=False,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
+            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
+        """
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            query_embeds=query_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        if query_embeds is not None:
+            sequence_output = outputs[0][:, query_embeds.shape[1] :, :]
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

models/blip2.py

@@ -0,0 +1,126 @@
+"""
+ Copyright (c) 2023, salesforce.com, inc.
+ All rights reserved.
+ SPDX-License-Identifier: BSD-3-Clause
+ For full license text, see the LICENSE_Lavis file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+import contextlib
+import os
+import logging
+
+import torch
+import torch.nn as nn
+
+from .Qformer import BertConfig, BertLMHeadModel
+from .eva_vit import create_eva_vit_g
+from transformers import BertTokenizer
+
+
+class Blip2Base(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @classmethod
+    def init_tokenizer(cls):
+        tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+        tokenizer.add_special_tokens({"bos_token": "[DEC]"})
+        return tokenizer
+    
+    @property
+    def device(self):
+        return list(self.parameters())[0].device
+
+    def maybe_autocast(self, dtype=torch.float16):
+        # if on cpu, don't use autocast
+        # if on gpu, use autocast with dtype if provided, otherwise use torch.float16
+        enable_autocast = self.device != torch.device("cpu")
+
+        if enable_autocast:
+            return torch.cuda.amp.autocast(dtype=dtype)
+        else:
+            return contextlib.nullcontext()
+
+    @classmethod
+    def init_Qformer(
+        cls, 
+        num_query_token, vision_width, 
+        qformer_hidden_dropout_prob=0.,
+        qformer_attention_probs_dropout_prob=0.,
+        qformer_drop_path_rate=0.,
+    ):
+        encoder_config = BertConfig.from_pretrained("bert-base-uncased")
+        encoder_config.encoder_width = vision_width
+        # insert cross-attention layer every other block
+        encoder_config.add_cross_attention = True
+        encoder_config.cross_attention_freq = 2
+        encoder_config.query_length = num_query_token
+        encoder_config.hidden_dropout_prob = qformer_hidden_dropout_prob
+        encoder_config.attention_probs_dropout_prob = qformer_attention_probs_dropout_prob
+        encoder_config.drop_path_list = [x.item() for x in torch.linspace(0, qformer_drop_path_rate, encoder_config.num_hidden_layers)]
+        print(f"Drop_path:{encoder_config.drop_path_list}")
+        print(encoder_config)
+        Qformer = BertLMHeadModel(config=encoder_config)
+        query_tokens = nn.Parameter(
+            torch.zeros(1, num_query_token, encoder_config.hidden_size)
+        )
+        query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)
+        return Qformer, query_tokens
+
+    @classmethod
+    def init_vision_encoder(
+        cls, 
+        model_name, img_size, drop_path_rate, 
+        use_grad_checkpoint, precision, vit_model_path,
+        temporal_downsample=True,
+        no_lmhra=False, 
+        double_lmhra=False,
+        lmhra_reduction=2.0, 
+        gmhra_layers=8, 
+        gmhra_drop_path_rate=0.,
+        gmhra_dropout=0.5, 
+    ):
+        assert model_name == "eva_clip_g", "vit model must be eva_clip_g for current version of VideoChat"
+        visual_encoder = create_eva_vit_g(
+            img_size, drop_path_rate, 
+            use_grad_checkpoint, precision, vit_model_path,
+            temporal_downsample=temporal_downsample,
+            no_lmhra=no_lmhra, 
+            double_lmhra=double_lmhra,
+            lmhra_reduction=lmhra_reduction, 
+            gmhra_layers=gmhra_layers, 
+            gmhra_drop_path_rate=gmhra_drop_path_rate,
+            gmhra_dropout=gmhra_dropout, 
+        )
+
+        ln_vision = LayerNorm(visual_encoder.num_features)
+        return visual_encoder, ln_vision
+
+    def load_from_pretrained(self, model_path):
+        if model_path is not None and os.path.isfile(model_path):
+            checkpoint = torch.load(model_path, map_location="cpu")
+        else:
+            raise RuntimeError("checkpoint url or path is invalid")
+
+        state_dict = checkpoint["model"]
+
+        msg = self.load_state_dict(state_dict, strict=False)
+
+        print(f"Load QFormer from {model_path}")
+        print(msg)
+
+        return msg
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)

models/eva_vit.py

@@ -0,0 +1,631 @@
+# Based on EVA, BEIT, timm and DeiT code bases
+# https://github.com/baaivision/EVA
+# https://github.com/rwightman/pytorch-image-models/tree/master/timm
+# https://github.com/microsoft/unilm/tree/master/beit
+# https://github.com/facebookresearch/deit/
+# https://github.com/facebookresearch/dino
+# --------------------------------------------------------'
+import os
+import math
+import logging
+from functools import partial
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+
+
+def _cfg(url='', **kwargs):
+    return {
+        'url': url,
+        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
+        'crop_pct': .9, 'interpolation': 'bicubic',
+        'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),
+        **kwargs
+    }
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self):
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        # x = self.drop(x)
+        # commit this for the orignal BERT implement 
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Local_MHRA(nn.Module):
+    def __init__(self, d_model, dw_reduction=1.5, pos_kernel_size=3):
+        super().__init__() 
+
+        padding = pos_kernel_size // 2
+        re_d_model = int(d_model // dw_reduction)
+        self.pos_embed = nn.Sequential(
+            nn.BatchNorm3d(d_model),
+            nn.Conv3d(d_model, re_d_model, kernel_size=1, stride=1, padding=0),
+            nn.Conv3d(re_d_model, re_d_model, kernel_size=(pos_kernel_size, 1, 1), stride=(1, 1, 1), padding=(padding, 0, 0), groups=re_d_model),
+            nn.Conv3d(re_d_model, d_model, kernel_size=1, stride=1, padding=0),
+        )
+
+        # init zero
+        # print('Init zero for Conv in pos_emb')
+        nn.init.constant_(self.pos_embed[3].weight, 0)
+        nn.init.constant_(self.pos_embed[3].bias, 0)
+
+    def forward(self, x):
+        out = self.pos_embed(x)
+        return out
+
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., window_size=None, attn_head_dim=None):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+
+        if window_size:
+            self.window_size = window_size
+            self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+            # cls to token & token 2 cls & cls to cls
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(window_size[0])
+            coords_w = torch.arange(window_size[1])
+            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+            relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+            relative_coords[:, :, 1] += window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+            relative_position_index = \
+                torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)
+            relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+            relative_position_index[0, 0:] = self.num_relative_distance - 3
+            relative_position_index[0:, 0] = self.num_relative_distance - 2
+            relative_position_index[0, 0] = self.num_relative_distance - 1
+
+            self.register_buffer("relative_position_index", relative_position_index)
+        else:
+            self.window_size = None
+            self.relative_position_bias_table = None
+            self.relative_position_index = None
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, rel_pos_bias=None):
+        B, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+        # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        if self.relative_position_bias_table is not None:
+            relative_position_bias = \
+                self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                    self.window_size[0] * self.window_size[1] + 1,
+                    self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+        if rel_pos_bias is not None:
+            attn = attn + rel_pos_bias
+        
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,
+                 window_size=None, attn_head_dim=None,
+                 no_lmhra=False, double_lmhra=True, lmhra_reduction=2.0, 
+                 ):
+        super().__init__()
+        self.no_lmhra = no_lmhra
+        self.double_lmhra = double_lmhra
+        if not no_lmhra:
+            self.lmhra1 = Local_MHRA(dim, dw_reduction=lmhra_reduction)
+            if double_lmhra:
+                self.lmhra2 = Local_MHRA(dim, dw_reduction=lmhra_reduction)
+
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if init_values is not None and init_values > 0:
+            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
+        else:
+            self.gamma_1, self.gamma_2 = None, None
+
+    def forward(self, x, rel_pos_bias=None, T=8):
+        # Local MHRA
+        if not self.no_lmhra:
+            # x: BT, HW+1, C
+            tmp_x = x[:, 1:, :]
+            BT, N, C = tmp_x.shape
+            B = BT // T
+            H = W = int(N ** 0.5)
+            tmp_x = tmp_x.view(B, T, H, W, C).permute(0, 4, 1, 2, 3).contiguous()
+            tmp_x = tmp_x + self.drop_path(self.lmhra1(tmp_x))
+            tmp_x = tmp_x.view(B, C, T, N).permute(0, 2, 3, 1).contiguous().view(BT, N, C)
+            x = torch.cat([x[:, :1, :], tmp_x], dim=1)
+
+        # MHSA
+        if self.gamma_1 is None:
+            x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))
+        else:
+            x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))
+
+        # Local MHRA
+        if not self.no_lmhra and self.double_lmhra:
+            tmp_x = x[:, 1:, :]
+            tmp_x = tmp_x.view(B, T, H, W, C).permute(0, 4, 1, 2, 3).contiguous()
+            tmp_x = tmp_x + self.drop_path(self.lmhra2(tmp_x))
+            tmp_x = tmp_x.view(B, C, T, N).permute(0, 2, 3, 1).contiguous().view(BT, N, C)
+            x = torch.cat([x[:, :1, :], tmp_x], dim=1)
+
+        # MLP
+        if self.gamma_1 is None:
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+        else:
+            x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, temporal_downsample=False):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+        if temporal_downsample:
+            self.proj = nn.Conv3d(
+                in_chans, embed_dim, kernel_size=(3, patch_size[0], patch_size[1]), 
+                stride=(2, patch_size[0], patch_size[1]), padding=(1, 0, 0)
+            )
+        else:
+            self.proj = nn.Conv3d(
+                in_chans, embed_dim, kernel_size=(1, patch_size[0], patch_size[1]), 
+                stride=(1, patch_size[0], patch_size[1]), padding=(0, 0, 0)
+            )
+
+    def forward(self, x, **kwargs):
+        B, C, T, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        assert H == self.img_size[0] and W == self.img_size[1], \
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x)
+        return x
+
+
+class RelativePositionBias(nn.Module):
+    def __init__(self, window_size, num_heads):
+        super().__init__()
+        self.window_size = window_size
+        self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+        # cls to token & token 2 cls & cls to cls
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(window_size[0])
+        coords_w = torch.arange(window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+        relative_position_index = \
+            torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)
+        relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        relative_position_index[0, 0:] = self.num_relative_distance - 3
+        relative_position_index[0:, 0] = self.num_relative_distance - 2
+        relative_position_index[0, 0] = self.num_relative_distance - 1
+
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        # trunc_normal_(self.relative_position_bias_table, std=.02)
+
+    def forward(self):
+        relative_position_bias = \
+            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1] + 1,
+                self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
+        return relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+
+
+class Global_MHRA(nn.Module):
+    def __init__(
+            self, d_model, n_head, attn_mask=None,
+            mlp_factor=4.0, drop_path=0., dropout=0.,
+        ):
+        super().__init__()
+        print(f'Drop path rate: {drop_path}')
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+        self.dpe = nn.Conv3d(d_model, d_model, kernel_size=3, stride=1, padding=1, bias=True, groups=d_model)
+        nn.init.constant_(self.dpe.bias, 0.)
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = nn.LayerNorm(d_model)
+        d_mlp = round(mlp_factor * d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_mlp)),
+            ("gelu", nn.GELU()),
+            ("dropout", nn.Dropout(dropout)),
+            ("c_proj", nn.Linear(d_mlp, d_model))
+        ]))
+        self.ln_2 = nn.LayerNorm(d_model)
+        self.ln_3 = nn.LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+        # zero init
+        nn.init.xavier_uniform_(self.attn.in_proj_weight)
+        nn.init.constant_(self.attn.out_proj.weight, 0.)
+        nn.init.constant_(self.attn.out_proj.bias, 0.)
+        nn.init.xavier_uniform_(self.mlp[0].weight)
+        nn.init.constant_(self.mlp[-1].weight, 0.)
+        nn.init.constant_(self.mlp[-1].bias, 0.)
+
+    def attention(self, x, y, T):
+        # x: 1, B, C
+        # y: BT, HW+1, C
+        BT, N, C = y.shape
+        B = BT // T
+        H = W = int(N ** 0.5)
+        y = y.view(B, T, N, C)
+        _, tmp_feats = y[:, :, :1], y[:, :, 1:]
+        tmp_feats = tmp_feats.view(B, T, H, W, C).permute(0, 4, 1, 2, 3).contiguous()
+        tmp_feats = self.dpe(tmp_feats.clone()).view(B, C, T, N - 1).permute(0, 2, 3, 1).contiguous()
+        y[:, :, 1:] = y[:, :, 1:] + tmp_feats
+        y = y.permute(1, 2, 0, 3).flatten(0, 1) # T(HW+1), B, C
+
+        d_model = self.ln_1.weight.size(0)
+        q = (x @ self.attn.in_proj_weight[:d_model].T) + self.attn.in_proj_bias[:d_model]
+
+        k = (y @ self.attn.in_proj_weight[d_model:-d_model].T) + self.attn.in_proj_bias[d_model:-d_model]
+        v = (y @ self.attn.in_proj_weight[-d_model:].T) + self.attn.in_proj_bias[-d_model:]
+        Tx, Ty, N = q.size(0), k.size(0), q.size(1)
+        q = q.view(Tx, N, self.attn.num_heads, self.attn.head_dim).permute(1, 2, 0, 3)
+        k = k.view(Ty, N, self.attn.num_heads, self.attn.head_dim).permute(1, 2, 0, 3)
+        v = v.view(Ty, N, self.attn.num_heads, self.attn.head_dim).permute(1, 2, 0, 3)
+        aff = (q @ k.transpose(-2, -1) / (self.attn.head_dim ** 0.5))
+
+        aff = aff.softmax(dim=-1)
+        out = aff @ v
+        out = out.permute(2, 0, 1, 3).flatten(2)
+        out = self.attn.out_proj(out)
+        return out
+
+    def forward(self, x, y, T):
+        x = x + self.drop_path(self.attention(self.ln_1(x), self.ln_3(y), T=T))
+        x = x + self.drop_path(self.mlp(self.ln_2(x)))
+        return x
+
+
+class VisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None,
+                 use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False,
+                 use_mean_pooling=True, init_scale=0.001, use_checkpoint=False,
+                 temporal_downsample=True,
+                 no_lmhra=False, double_lmhra=True, lmhra_reduction=1.5, 
+                 gmhra_layers=4, gmhra_drop_path_rate=0., gmhra_dropout=0.5, 
+                 ):
+        super().__init__()
+        self.image_size = img_size
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+
+        print(f"Temporal downsample: {temporal_downsample}")
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            temporal_downsample=temporal_downsample,
+        )
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        if use_abs_pos_emb:
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        else:
+            self.pos_embed = None
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        if use_shared_rel_pos_bias:
+            self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)
+        else:
+            self.rel_pos_bias = None
+        self.use_checkpoint = use_checkpoint
+        
+        print(f'No L_MHRA: {no_lmhra}')
+        print(f'Double L_MHRA: {double_lmhra}')
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+        self.use_rel_pos_bias = use_rel_pos_bias
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None,
+                no_lmhra=no_lmhra, double_lmhra=double_lmhra, lmhra_reduction=lmhra_reduction, 
+            )
+            for i in range(depth)])
+
+        # global MHRA
+        self.gmhra_layers = gmhra_layers
+        self.gmhra_layer_idx = [(depth - 1 - idx) for idx in range(gmhra_layers)]
+        print(f"GMHRA index: {self.gmhra_layer_idx}")
+        print(f"GMHRA dropout: {gmhra_dropout}")
+        if gmhra_layers > 0:
+            self.gmhra_cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        gmhra_dpr = [x.item() for x in torch.linspace(0, gmhra_drop_path_rate, gmhra_layers)]
+        self.gmhra = nn.ModuleList([
+            Global_MHRA(
+                embed_dim, num_heads, mlp_factor=mlp_ratio, 
+                drop_path=gmhra_dpr[i], dropout=gmhra_dropout,
+            ) for i in range(gmhra_layers)
+        ])
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+        trunc_normal_(self.cls_token, std=.02)
+        self.fix_init_weight()
+
+    def fix_init_weight(self):
+        def rescale(param, layer_id):
+            param.div_(math.sqrt(2.0 * layer_id))
+
+        for layer_id, layer in enumerate(self.blocks):
+            rescale(layer.attn.proj.weight.data, layer_id + 1)
+            rescale(layer.mlp.fc2.weight.data, layer_id + 1)
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        B, C, T, H, W = x.shape
+        x = x.permute(0, 2, 3, 4, 1).reshape(B * T, H * W, C)
+
+        cls_tokens = self.cls_token.expand(B * T, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        # the input of global MHRA should be (THW+1)xBx1
+        if self.gmhra_layers > 0:
+            gmhra_cls_token = self.gmhra_cls_token.repeat(1, B, 1)
+
+        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+        j = -1
+        for idx, blk in enumerate(self.blocks):
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, rel_pos_bias, T=T)
+            else:
+                x = blk(x, rel_pos_bias, T=T)
+            if idx in self.gmhra_layer_idx:
+                j += 1
+                tmp_x = x.clone()
+                gmhra_cls_token = self.gmhra[j](gmhra_cls_token, tmp_x, T=T)
+        z = torch.cat([x.view(B, -1, C), gmhra_cls_token.permute(1, 0, 2)], dim=1)
+        return z
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        return x
+    
+    
+def interpolate_pos_embed(model, checkpoint_model):
+    if 'pos_embed' in checkpoint_model:
+        pos_embed_checkpoint = checkpoint_model['pos_embed'].float()
+        embedding_size = pos_embed_checkpoint.shape[-1]
+        num_patches = model.patch_embed.num_patches
+        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
+        # height (== width) for the checkpoint position embedding
+        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
+        # height (== width) for the new position embedding
+        new_size = int(num_patches ** 0.5)
+        # class_token and dist_token are kept unchanged
+        if orig_size != new_size:
+            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
+            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
+            # only the position tokens are interpolated
+            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
+            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
+            pos_tokens = torch.nn.functional.interpolate(
+                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
+            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
+            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
+            checkpoint_model['pos_embed'] = new_pos_embed
+            
+            
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+        if isinstance(l, (nn.MultiheadAttention, Attention)):
+            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+    model.apply(_convert_weights_to_fp16)
+    
+
+def inflate_weight(weight_2d, time_dim, center=True):
+    print(f'Init center: {center}')
+    if center:
+        weight_3d = torch.zeros(*weight_2d.shape)
+        weight_3d = weight_3d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
+        middle_idx = time_dim // 2
+        weight_3d[:, :, middle_idx, :, :] = weight_2d
+    else:
+        weight_3d = weight_2d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
+        weight_3d = weight_3d / time_dim
+    return weight_3d
+
+
+def load_state_dict(model, state_dict, strict=True):
+    state_dict_3d = model.state_dict()
+    for k in state_dict.keys():
+        if k in state_dict_3d.keys() and state_dict[k].shape != state_dict_3d[k].shape:
+            if len(state_dict_3d[k].shape) <= 2:
+                print(f'Ignore: {k}')
+                continue
+            print(f'Inflate: {k}, {state_dict[k].shape} => {state_dict_3d[k].shape}')
+            time_dim = state_dict_3d[k].shape[2]
+            state_dict[k] = inflate_weight(state_dict[k], time_dim)
+    msg = model.load_state_dict(state_dict, strict=strict)
+    return msg
+    
+
+def create_eva_vit_g(
+        img_size=224, drop_path_rate=0.4, use_checkpoint=False,
+        precision="fp16", vit_model_path=None,
+        # UniFormerV2
+        temporal_downsample=True,
+        no_lmhra=False, 
+        double_lmhra=False,
+        lmhra_reduction=2.0, 
+        gmhra_layers=8, 
+        gmhra_drop_path_rate=0.,
+        gmhra_dropout=0.5, 
+    ):
+    model = VisionTransformer(
+        img_size=img_size,
+        patch_size=14,
+        use_mean_pooling=False,
+        embed_dim=1408,
+        depth=39,
+        num_heads=1408//88,
+        mlp_ratio=4.3637,
+        qkv_bias=True,
+        drop_path_rate=drop_path_rate,
+        norm_layer=partial(nn.LayerNorm, eps=1e-6),
+        use_checkpoint=use_checkpoint,
+        temporal_downsample=temporal_downsample,
+        no_lmhra=no_lmhra, 
+        double_lmhra=double_lmhra,
+        lmhra_reduction=lmhra_reduction, 
+        gmhra_layers=gmhra_layers, 
+        gmhra_drop_path_rate=gmhra_drop_path_rate,
+        gmhra_dropout=gmhra_dropout, 
+    )  
+    if vit_model_path is not None and os.path.isfile(vit_model_path):
+        state_dict = torch.load(vit_model_path, map_location="cpu")    
+        print(f"Load ViT model from: {vit_model_path}")
+        interpolate_pos_embed(model, state_dict)
+        msg = load_state_dict(model, state_dict, strict=False)
+        print(msg)
+    
+    if precision == "fp16":
+#         model.to("cuda") 
+        convert_weights_to_fp16(model)
+    return model
+
+
+if __name__ == '__main__':
+    import time
+    from fvcore.nn import FlopCountAnalysis
+    from fvcore.nn import flop_count_table
+    import numpy as np
+
+    seed = 4217
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    num_frames = 8
+
+    model = create_eva_vit_g(
+        img_size=224, drop_path_rate=0.4, use_checkpoint=False,
+        precision="fp16", vit_model_path=None,
+        temporal_downsample=True,
+        no_lmhra=False, 
+        double_lmhra=False,
+        lmhra_reduction=2.0, 
+        gmhra_layers=12,
+        gmhra_drop_path_rate=0.,
+        gmhra_dropout=0.5, 
+    )
+    video = torch.rand(1, 3, num_frames, 224, 224)
+    flops = FlopCountAnalysis(model, video)
+    s = time.time()
+    print(flop_count_table(flops, max_depth=1))
+    print(time.time()-s)

models/modeling_llama.py

@@ -0,0 +1,755 @@
+# This script is based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+
+""" PyTorch LLaMA model."""
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+class LlamaRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        LlamaRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        # convert into half-precision if necessary
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+class LlamaRotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+
+        # Build here to make `torch.jit.trace` work.
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
+        if seq_len > self.max_seq_len_cached:
+            self.max_seq_len_cached = seq_len
+            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+            self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
+    gather_indices = position_ids[:, None, :, None]  # [bs, 1, seq_len, 1]
+    gather_indices = gather_indices.repeat(1, cos.shape[1], 1, cos.shape[3])
+    cos = torch.gather(cos.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
+    sin = torch.gather(sin.repeat(gather_indices.shape[0], 1, 1, 1), 2, gather_indices)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class LlamaMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+        # [bsz, nh, t, hd]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class LlamaDecoderLayer(nn.Module):
+    def __init__(self, config: LlamaConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LlamaAttention(config=config)
+        self.mlp = LlamaMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+LLAMA_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`LlamaConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaPreTrainedModel(PreTrainedModel):
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["LlamaDecoderLayer"]
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    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_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, LlamaModel):
+            module.gradient_checkpointing = value
+
+
+LLAMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class LlamaModel(LlamaPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: LlamaConfig
+    """
+
+    def __init__(self, config: LlamaConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        query_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        if query_embeds is not None:
+            inputs_embeds = torch.cat([query_embeds, inputs_embeds], dim=1)
+            batch_size, seq_length, _ = inputs_embeds.shape
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class LlamaForCausalLM(LlamaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        query_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you consciours? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            query_embeds=query_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, query_embeds=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+                query_embeds = None
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "query_embeds": query_embeds,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+

models/video_transformers.py

@@ -0,0 +1,443 @@
+import torchvision
+import random
+from PIL import Image, ImageOps
+import numpy as np
+import numbers
+import math
+import torch
+
+
+class GroupRandomCrop(object):
+    def __init__(self, size):
+        if isinstance(size, numbers.Number):
+            self.size = (int(size), int(size))
+        else:
+            self.size = size
+
+    def __call__(self, img_group):
+
+        w, h = img_group[0].size
+        th, tw = self.size
+
+        out_images = list()
+
+        x1 = random.randint(0, w - tw)
+        y1 = random.randint(0, h - th)
+
+        for img in img_group:
+            assert(img.size[0] == w and img.size[1] == h)
+            if w == tw and h == th:
+                out_images.append(img)
+            else:
+                out_images.append(img.crop((x1, y1, x1 + tw, y1 + th)))
+
+        return out_images
+
+
+class MultiGroupRandomCrop(object):
+    def __init__(self, size, groups=1):
+        if isinstance(size, numbers.Number):
+            self.size = (int(size), int(size))
+        else:
+            self.size = size
+        self.groups = groups
+
+    def __call__(self, img_group):
+
+        w, h = img_group[0].size
+        th, tw = self.size
+
+        out_images = list()
+
+        for i in range(self.groups):
+            x1 = random.randint(0, w - tw)
+            y1 = random.randint(0, h - th)
+
+            for img in img_group:
+                assert(img.size[0] == w and img.size[1] == h)
+                if w == tw and h == th:
+                    out_images.append(img)
+                else:
+                    out_images.append(img.crop((x1, y1, x1 + tw, y1 + th)))
+
+        return out_images
+
+
+class GroupCenterCrop(object):
+    def __init__(self, size):
+        self.worker = torchvision.transforms.CenterCrop(size)
+
+    def __call__(self, img_group):
+        return [self.worker(img) for img in img_group]
+
+
+class GroupRandomHorizontalFlip(object):
+    """Randomly horizontally flips the given PIL.Image with a probability of 0.5
+    """
+
+    def __init__(self, is_flow=False):
+        self.is_flow = is_flow
+
+    def __call__(self, img_group, is_flow=False):
+        v = random.random()
+        if v < 0.5:
+            ret = [img.transpose(Image.FLIP_LEFT_RIGHT) for img in img_group]
+            if self.is_flow:
+                for i in range(0, len(ret), 2):
+                    # invert flow pixel values when flipping
+                    ret[i] = ImageOps.invert(ret[i])
+            return ret
+        else:
+            return img_group
+
+
+class GroupNormalize(object):
+    def __init__(self, mean, std):
+        self.mean = mean
+        self.std = std
+
+    def __call__(self, tensor):
+        rep_mean = self.mean * (tensor.size()[0] // len(self.mean))
+        rep_std = self.std * (tensor.size()[0] // len(self.std))
+
+        # TODO: make efficient
+        for t, m, s in zip(tensor, rep_mean, rep_std):
+            t.sub_(m).div_(s)
+
+        return tensor
+
+
+class GroupScale(object):
+    """ Rescales the input PIL.Image to the given 'size'.
+    'size' will be the size of the smaller edge.
+    For example, if height > width, then image will be
+    rescaled to (size * height / width, size)
+    size: size of the smaller edge
+    interpolation: Default: PIL.Image.BILINEAR
+    """
+
+    def __init__(self, size, interpolation=Image.BILINEAR):
+        self.worker = torchvision.transforms.Resize(size, interpolation)
+
+    def __call__(self, img_group):
+        return [self.worker(img) for img in img_group]
+
+
+class GroupOverSample(object):
+    def __init__(self, crop_size, scale_size=None, flip=True):
+        self.crop_size = crop_size if not isinstance(
+            crop_size, int) else (crop_size, crop_size)
+
+        if scale_size is not None:
+            self.scale_worker = GroupScale(scale_size)
+        else:
+            self.scale_worker = None
+        self.flip = flip
+
+    def __call__(self, img_group):
+
+        if self.scale_worker is not None:
+            img_group = self.scale_worker(img_group)
+
+        image_w, image_h = img_group[0].size
+        crop_w, crop_h = self.crop_size
+
+        offsets = GroupMultiScaleCrop.fill_fix_offset(
+            False, image_w, image_h, crop_w, crop_h)
+        oversample_group = list()
+        for o_w, o_h in offsets:
+            normal_group = list()
+            flip_group = list()
+            for i, img in enumerate(img_group):
+                crop = img.crop((o_w, o_h, o_w + crop_w, o_h + crop_h))
+                normal_group.append(crop)
+                flip_crop = crop.copy().transpose(Image.FLIP_LEFT_RIGHT)
+
+                if img.mode == 'L' and i % 2 == 0:
+                    flip_group.append(ImageOps.invert(flip_crop))
+                else:
+                    flip_group.append(flip_crop)
+
+            oversample_group.extend(normal_group)
+            if self.flip:
+                oversample_group.extend(flip_group)
+        return oversample_group
+
+
+class GroupFullResSample(object):
+    def __init__(self, crop_size, scale_size=None, flip=True):
+        self.crop_size = crop_size if not isinstance(
+            crop_size, int) else (crop_size, crop_size)
+
+        if scale_size is not None:
+            self.scale_worker = GroupScale(scale_size)
+        else:
+            self.scale_worker = None
+        self.flip = flip
+
+    def __call__(self, img_group):
+
+        if self.scale_worker is not None:
+            img_group = self.scale_worker(img_group)
+
+        image_w, image_h = img_group[0].size
+        crop_w, crop_h = self.crop_size
+
+        w_step = (image_w - crop_w) // 4
+        h_step = (image_h - crop_h) // 4
+
+        offsets = list()
+        offsets.append((0 * w_step, 2 * h_step))  # left
+        offsets.append((4 * w_step, 2 * h_step))  # right
+        offsets.append((2 * w_step, 2 * h_step))  # center
+
+        oversample_group = list()
+        for o_w, o_h in offsets:
+            normal_group = list()
+            flip_group = list()
+            for i, img in enumerate(img_group):
+                crop = img.crop((o_w, o_h, o_w + crop_w, o_h + crop_h))
+                normal_group.append(crop)
+                if self.flip:
+                    flip_crop = crop.copy().transpose(Image.FLIP_LEFT_RIGHT)
+
+                    if img.mode == 'L' and i % 2 == 0:
+                        flip_group.append(ImageOps.invert(flip_crop))
+                    else:
+                        flip_group.append(flip_crop)
+
+            oversample_group.extend(normal_group)
+            oversample_group.extend(flip_group)
+        return oversample_group
+
+
+class GroupMultiScaleCrop(object):
+
+    def __init__(self, input_size, scales=None, max_distort=1,
+                 fix_crop=True, more_fix_crop=True):
+        self.scales = scales if scales is not None else [1, .875, .75, .66]
+        self.max_distort = max_distort
+        self.fix_crop = fix_crop
+        self.more_fix_crop = more_fix_crop
+        self.input_size = input_size if not isinstance(input_size, int) else [
+            input_size, input_size]
+        self.interpolation = Image.BILINEAR
+
+    def __call__(self, img_group):
+
+        im_size = img_group[0].size
+
+        crop_w, crop_h, offset_w, offset_h = self._sample_crop_size(im_size)
+        crop_img_group = [
+            img.crop(
+                (offset_w,
+                 offset_h,
+                 offset_w +
+                 crop_w,
+                 offset_h +
+                 crop_h)) for img in img_group]
+        ret_img_group = [img.resize((self.input_size[0], self.input_size[1]), self.interpolation)
+                         for img in crop_img_group]
+        return ret_img_group
+
+    def _sample_crop_size(self, im_size):
+        image_w, image_h = im_size[0], im_size[1]
+
+        # find a crop size
+        base_size = min(image_w, image_h)
+        crop_sizes = [int(base_size * x) for x in self.scales]
+        crop_h = [
+            self.input_size[1] if abs(
+                x - self.input_size[1]) < 3 else x for x in crop_sizes]
+        crop_w = [
+            self.input_size[0] if abs(
+                x - self.input_size[0]) < 3 else x for x in crop_sizes]
+
+        pairs = []
+        for i, h in enumerate(crop_h):
+            for j, w in enumerate(crop_w):
+                if abs(i - j) <= self.max_distort:
+                    pairs.append((w, h))
+
+        crop_pair = random.choice(pairs)
+        if not self.fix_crop:
+            w_offset = random.randint(0, image_w - crop_pair[0])
+            h_offset = random.randint(0, image_h - crop_pair[1])
+        else:
+            w_offset, h_offset = self._sample_fix_offset(
+                image_w, image_h, crop_pair[0], crop_pair[1])
+
+        return crop_pair[0], crop_pair[1], w_offset, h_offset
+
+    def _sample_fix_offset(self, image_w, image_h, crop_w, crop_h):
+        offsets = self.fill_fix_offset(
+            self.more_fix_crop, image_w, image_h, crop_w, crop_h)
+        return random.choice(offsets)
+
+    @staticmethod
+    def fill_fix_offset(more_fix_crop, image_w, image_h, crop_w, crop_h):
+        w_step = (image_w - crop_w) // 4
+        h_step = (image_h - crop_h) // 4
+
+        ret = list()
+        ret.append((0, 0))  # upper left
+        ret.append((4 * w_step, 0))  # upper right
+        ret.append((0, 4 * h_step))  # lower left
+        ret.append((4 * w_step, 4 * h_step))  # lower right
+        ret.append((2 * w_step, 2 * h_step))  # center
+
+        if more_fix_crop:
+            ret.append((0, 2 * h_step))  # center left
+            ret.append((4 * w_step, 2 * h_step))  # center right
+            ret.append((2 * w_step, 4 * h_step))  # lower center
+            ret.append((2 * w_step, 0 * h_step))  # upper center
+
+            ret.append((1 * w_step, 1 * h_step))  # upper left quarter
+            ret.append((3 * w_step, 1 * h_step))  # upper right quarter
+            ret.append((1 * w_step, 3 * h_step))  # lower left quarter
+            ret.append((3 * w_step, 3 * h_step))  # lower righ quarter
+
+        return ret
+
+
+class GroupRandomSizedCrop(object):
+    """Random crop the given PIL.Image to a random size of (0.08 to 1.0) of the original size
+    and and a random aspect ratio of 3/4 to 4/3 of the original aspect ratio
+    This is popularly used to train the Inception networks
+    size: size of the smaller edge
+    interpolation: Default: PIL.Image.BILINEAR
+    """
+
+    def __init__(self, size, interpolation=Image.BILINEAR):
+        self.size = size
+        self.interpolation = interpolation
+
+    def __call__(self, img_group):
+        for attempt in range(10):
+            area = img_group[0].size[0] * img_group[0].size[1]
+            target_area = random.uniform(0.08, 1.0) * area
+            aspect_ratio = random.uniform(3. / 4, 4. / 3)
+
+            w = int(round(math.sqrt(target_area * aspect_ratio)))
+            h = int(round(math.sqrt(target_area / aspect_ratio)))
+
+            if random.random() < 0.5:
+                w, h = h, w
+
+            if w <= img_group[0].size[0] and h <= img_group[0].size[1]:
+                x1 = random.randint(0, img_group[0].size[0] - w)
+                y1 = random.randint(0, img_group[0].size[1] - h)
+                found = True
+                break
+        else:
+            found = False
+            x1 = 0
+            y1 = 0
+
+        if found:
+            out_group = list()
+            for img in img_group:
+                img = img.crop((x1, y1, x1 + w, y1 + h))
+                assert(img.size == (w, h))
+                out_group.append(
+                    img.resize(
+                        (self.size, self.size), self.interpolation))
+            return out_group
+        else:
+            # Fallback
+            scale = GroupScale(self.size, interpolation=self.interpolation)
+            crop = GroupRandomCrop(self.size)
+            return crop(scale(img_group))
+
+
+class ConvertDataFormat(object):
+    def __init__(self, model_type):
+        self.model_type = model_type
+
+    def __call__(self, images):
+        if self.model_type == '2D':
+            return images
+        tc, h, w = images.size()
+        t = tc // 3
+        images = images.view(t, 3, h, w)
+        images = images.permute(1, 0, 2, 3)
+        return images
+
+
+class Stack(object):
+
+    def __init__(self, roll=False):
+        self.roll = roll
+
+    def __call__(self, img_group):
+        if img_group[0].mode == 'L':
+            return np.concatenate([np.expand_dims(x, 2)
+                                   for x in img_group], axis=2)
+        elif img_group[0].mode == 'RGB':
+            if self.roll:
+                return np.concatenate([np.array(x)[:, :, ::-1]
+                                       for x in img_group], axis=2)
+            else:
+                #print(np.concatenate(img_group, axis=2).shape)
+                # print(img_group[0].shape)
+                return np.concatenate(img_group, axis=2)
+
+
+class ToTorchFormatTensor(object):
+    """ Converts a PIL.Image (RGB) or numpy.ndarray (H x W x C) in the range [0, 255]
+    to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0] """
+
+    def __init__(self, div=True):
+        self.div = div
+
+    def __call__(self, pic):
+        if isinstance(pic, np.ndarray):
+            # handle numpy array
+            img = torch.from_numpy(pic).permute(2, 0, 1).contiguous()
+        else:
+            # handle PIL Image
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(
+                    pic.tobytes()))
+            img = img.view(pic.size[1], pic.size[0], len(pic.mode))
+            # put it from HWC to CHW format
+            # yikes, this transpose takes 80% of the loading time/CPU
+            img = img.transpose(0, 1).transpose(0, 2).contiguous()
+        return img.float().div(255) if self.div else img.float()
+
+
+class IdentityTransform(object):
+
+    def __call__(self, data):
+        return data
+
+
+if __name__ == "__main__":
+    trans = torchvision.transforms.Compose([
+        GroupScale(256),
+        GroupRandomCrop(224),
+        Stack(),
+        ToTorchFormatTensor(),
+        GroupNormalize(
+            mean=[.485, .456, .406],
+            std=[.229, .224, .225]
+        )]
+    )
+
+    im = Image.open('../tensorflow-model-zoo.torch/lena_299.png')
+
+    color_group = [im] * 3
+    rst = trans(color_group)
+
+    gray_group = [im.convert('L')] * 9
+    gray_rst = trans(gray_group)
+
+    trans2 = torchvision.transforms.Compose([
+        GroupRandomSizedCrop(256),
+        Stack(),
+        ToTorchFormatTensor(),
+        GroupNormalize(
+            mean=[.485, .456, .406],
+            std=[.229, .224, .225])
+    ])
+    print(trans2(color_group))

models/videochat.py

@@ -0,0 +1,176 @@
+import random
+import logging
+
+import torch
+from torch.cuda.amp import autocast as autocast
+import torch.nn as nn
+
+from .blip2 import Blip2Base, disabled_train
+from .modeling_llama import LlamaForCausalLM
+from transformers import LlamaTokenizer, LlamaConfig
+
+
+class VideoChat(Blip2Base):
+    """
+    VideoChat model.
+    """
+    def __init__(self, config):
+        super().__init__()
+
+        vit_model = config.get("vit_model", "eva_clip_g")
+        vit_model_path = config.get("vit_model_path", None)
+        q_former_model_path = config.get("q_former_model_path", None)
+        llama_model_path = config.get("llama_model_path")
+        videochat_model_path = config.get("videochat_model_path", "")  
+        img_size = config.get("img_size")
+
+        drop_path_rate = config.get("drop_path_rate", 0)
+        use_grad_checkpoint = config.get("use_grad_checkpoint", False)
+        vit_precision = config.get("vit_precision", "fp16")
+        freeze_vit = config.get("freeze_vit", True)
+        freeze_qformer = config.get("freeze_qformer", True)
+        low_resource = config.get("low_resource", False) # use 8 bit and put vit in cpu
+        max_txt_len = config.get("max_txt_len", 32)
+
+        # uniformerv2
+        freeze_mhra = config.get("freeze_mhra", False)
+        temporal_downsample = config.get("temporal_downsample", True)
+        no_lmhra = config.get("no_lmhra", False)
+        double_lmhra = config.get("double_lmhra", False)
+        lmhra_reduction = config.get("lmhra_reduction", 2.0)
+        gmhra_layers = config.get("gmhra_layers", 8)
+        gmhra_drop_path_rate = config.get("gmhra_drop_path_rate", 0.)
+        gmhra_dropout = config.get("gmhra_dropout", 0.5)
+        # qformer
+        num_query_token = config.get("num_query_token")
+        extra_num_query_token = config.get("extra_num_query_token", 64)
+
+        self.tokenizer = self.init_tokenizer()
+        self.low_resource = low_resource
+
+        self.vit_precision = vit_precision
+        print(f'Loading VIT. Use fp16: {vit_precision}')
+        self.visual_encoder, self.ln_vision = self.init_vision_encoder(
+            vit_model, img_size, drop_path_rate, 
+            use_grad_checkpoint, vit_precision, vit_model_path,
+            temporal_downsample=temporal_downsample,
+            no_lmhra=no_lmhra, 
+            double_lmhra=double_lmhra,
+            lmhra_reduction=lmhra_reduction, 
+            gmhra_layers=gmhra_layers, 
+            gmhra_drop_path_rate=gmhra_drop_path_rate,
+            gmhra_dropout=gmhra_dropout, 
+        )
+        if freeze_vit:
+            print("freeze vision encoder")
+            if not freeze_mhra:
+                open_list = []
+                for name, param in self.visual_encoder.named_parameters():
+                    if 'mhra' not in name:
+                        param.requires_grad = False
+                    else:
+                        open_list.append(name)
+                print(f"open module: {open_list}")
+                print("open ln_vision")
+            else:
+                for name, param in self.visual_encoder.named_parameters():
+                    param.requires_grad = False
+                self.visual_encoder = self.visual_encoder.eval()
+                self.visual_encoder.train = disabled_train
+                for name, param in self.ln_vision.named_parameters():
+                    param.requires_grad = False
+                self.ln_vision = self.ln_vision.eval()
+                self.ln_vision.train = disabled_train
+        print('Loading VIT Done')
+
+        print('Loading Q-Former')
+        self.Qformer, self.query_tokens = self.init_Qformer(
+            num_query_token, self.visual_encoder.num_features,
+        )
+        self.Qformer.cls = None
+        self.Qformer.bert.embeddings.word_embeddings = None
+        self.Qformer.bert.embeddings.position_embeddings = None
+        for layer in self.Qformer.bert.encoder.layer:
+            layer.output = None
+            layer.intermediate = None
+        self.load_from_pretrained(model_path=q_former_model_path)
+        print(f"Add extra {extra_num_query_token} tokens in QFormer")
+        self.extra_query_tokens = nn.Parameter(
+            torch.zeros(1, extra_num_query_token, self.query_tokens.shape[-1])
+        )
+
+        if freeze_qformer:
+            print("freeze Qformer")
+            for name, param in self.Qformer.named_parameters():
+                param.requires_grad = False
+            self.Qformer = self.Qformer.eval()
+            self.Qformer.train = disabled_train
+            self.query_tokens.requires_grad = False
+        print('Loading Q-Former Done')
+
+        print('Loading LLAMA')
+        self.llama_tokenizer = LlamaTokenizer.from_pretrained(llama_model_path, use_fast=False)
+        self.llama_tokenizer.pad_token = self.llama_tokenizer.eos_token
+
+        if self.low_resource:
+                self.llama_model = LlamaForCausalLM.from_pretrained(
+                    llama_model_path,
+                    torch_dtype=torch.float16,
+                    load_in_8bit=True,
+                    device_map="auto"
+                )
+        else:
+                self.llama_model = LlamaForCausalLM.from_pretrained(
+                    llama_model_path,
+                    torch_dtype=torch.float16,
+                )
+
+        print("freeze LLAMA")
+        for name, param in self.llama_model.named_parameters():
+            param.requires_grad = False
+        print('Loading LLAMA Done')
+
+        self.llama_proj = nn.Linear(
+            self.Qformer.config.hidden_size, self.llama_model.config.hidden_size
+        )
+        self.max_txt_len = max_txt_len
+
+        # load weights of VideoChat
+        if videochat_model_path:
+            print(f"Load VideoChat from: {videochat_model_path}")
+            ckpt = torch.load(videochat_model_path, map_location="cpu")
+            msg = self.load_state_dict(ckpt['model'], strict=False)
+            print(msg)
+
+    def vit_to_cpu(self):
+        self.ln_vision.to("cpu")
+        self.ln_vision.float()
+        self.visual_encoder.to("cpu")
+        self.visual_encoder.float()
+
+    def encode_img(self, image):
+        device = image.device
+        if self.low_resource:
+            self.vit_to_cpu()
+            image = image.to("cpu")
+
+        with self.maybe_autocast():
+            T = image.shape[1]
+            # use_image = True if T == 1 else False
+            image = image.permute(0, 2, 1, 3, 4) # [B,T,C,H,W] -> [B,C,T,H,W]
+
+            image_embeds = self.ln_vision(self.visual_encoder(image)).to(device)
+            image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(device)
+
+            query_tokens = torch.cat([self.query_tokens, self.extra_query_tokens], dim=1)
+            query_tokens = query_tokens.expand(image_embeds.shape[0], -1, -1)
+            query_output = self.Qformer.bert(
+                query_embeds=query_tokens,
+                encoder_hidden_states=image_embeds,
+                encoder_attention_mask=image_atts,
+                return_dict=True,
+            )
+
+            inputs_llama = self.llama_proj(query_output.last_hidden_state)
+            atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
+        return inputs_llama, atts_llama

requirements.txt

@@ -0,0 +1,11 @@
+decord==0.6.0
+fvcore==0.1.5.post20221221
+gradio==3.29.0
+numpy==1.24.3
+Pillow==9.5.0
+PyYAML==6.0
+timm==0.6.13
+torch==1.12.1
+torchvision==0.13.1
+transformers==4.28.1
+sentencepiece==0.1.99

utils/config.py

@@ -0,0 +1,281 @@
+from __future__ import annotations
+
+import argparse
+import ast
+import json
+import os
+import os.path as osp
+import re
+import shutil
+import sys
+import tempfile
+from copy import deepcopy
+from importlib import import_module
+
+import yaml
+
+from .easydict import EasyDict
+
+__all__ = ["Config", "pretty_text"]
+
+
+BASE_KEY = "_base_"
+# BASE_CONFIG = {"OUTPUT_DIR": "./workspace", "SESSION": "base", "LOG_FILE": "log.txt"}
+BASE_CONFIG = {}
+
+cfg = None
+
+
+class Config(object):
+    """config"""
+
+    @classmethod
+    def pretty_text(cls, cfg: dict, indent=2) -> str:
+        """format dict to a string
+
+        Args:
+            cfg (EasyDict): the params.
+
+        Returns: The string to display.
+
+        """
+        msg = "{\n"
+        for i, (k, v) in enumerate(cfg.items()):
+            if isinstance(v, dict):
+                v = cls.pretty_text(v, indent + 4)
+            spaces = " " * indent
+            msg += spaces + "{}: {}".format(k, v)
+            if i == len(cfg) - 1:
+                msg += " }"
+            else:
+                msg += "\n"
+        return msg
+
+    @classmethod
+    def dump(cls, cfg, savepath=None):
+        """dump cfg to `json` file.
+
+        Args:
+            cfg (dict): The dict to dump.
+            savepath (str): The filepath to save the dumped dict.
+
+        Returns: TODO
+
+        """
+        if savepath is None:
+            savepath = osp.join(cfg.WORKSPACE, "config.json")
+        json.dump(cfg, open(savepath, "w"), indent=2)
+
+    @classmethod
+    def get_config(cls, default_config: dict = None):
+        """get a `Config` instance.
+
+        Args:
+            default_config (dict): The default config. `default_config` will be overrided
+                by config file `--cfg`, `--cfg` will be overrided by commandline args.
+
+        Returns: an EasyDict.
+        """
+        global cfg
+        if cfg is not None:
+            return cfg
+
+        # define arg parser.
+        parser = argparse.ArgumentParser()
+        # parser.add_argument("--cfg", help="load configs from yaml file", default="", type=str)
+        parser.add_argument(
+            "config_file", help="the configuration file to load. support: .yaml, .json, .py"
+        )
+        parser.add_argument(
+            "opts",
+            default=None,
+            nargs="*",
+            help="overrided configs. List. Format: 'key1 name1 key2 name2'",
+        )
+        args = parser.parse_args()
+
+        cfg = EasyDict(BASE_CONFIG)
+        if osp.isfile(args.config_file):
+            cfg_from_file = cls.from_file(args.config_file)
+            cfg = merge_a_into_b(cfg_from_file, cfg)
+        cfg = cls.merge_list(cfg, args.opts)
+        cfg = eval_dict_leaf(cfg)
+
+        # update some keys to make them show at the last
+        for k in BASE_CONFIG:
+            cfg[k] = cfg.pop(k)
+        return cfg
+
+    @classmethod
+    def from_file(cls, filepath: str) -> EasyDict:
+        """Build config from file. Supported filetypes: `.py`,`.yaml`,`.json`.
+
+        Args:
+            filepath (str): The config file path.
+
+        Returns: TODO
+
+        """
+        filepath = osp.abspath(osp.expanduser(filepath))
+        if not osp.isfile(filepath):
+            raise IOError(f"File does not exist: {filepath}")
+        if filepath.endswith(".py"):
+            with tempfile.TemporaryDirectory() as temp_config_dir:
+
+                shutil.copytree(osp.dirname(filepath), osp.join(temp_config_dir, "tmp_config"))
+                sys.path.insert(0, temp_config_dir)
+                mod = import_module("tmp_config." + osp.splitext(osp.basename(filepath))[0])
+                # mod = import_module(temp_module_name)
+                sys.path.pop(0)
+                cfg_dict = {
+                    name: value
+                    for name, value in mod.__dict__.items()
+                    if not name.startswith("__")
+                }
+                for k in list(sys.modules.keys()):
+                    if "tmp_config" in k:
+                        del sys.modules[k]
+        elif filepath.endswith((".yml", ".yaml")):
+            cfg_dict = yaml.load(open(filepath, "r"), Loader=yaml.Loader)
+        elif filepath.endswith(".json"):
+            cfg_dict = json.load(open(filepath, "r"))
+        else:
+            raise IOError("Only py/yml/yaml/json type are supported now!")
+
+        cfg_text = filepath + "\n"
+        with open(filepath, "r") as f:
+            cfg_text += f.read()
+
+        if BASE_KEY in cfg_dict:  # load configs in `BASE_KEY`
+            cfg_dir = osp.dirname(filepath)
+            base_filename = cfg_dict.pop(BASE_KEY)
+            base_filename = (
+                base_filename if isinstance(base_filename, list) else [base_filename]
+            )
+
+            cfg_dict_list = list()
+            for f in base_filename:
+                _cfg_dict = Config.from_file(osp.join(cfg_dir, f))
+                cfg_dict_list.append(_cfg_dict)
+
+            base_cfg_dict = dict()
+            for c in cfg_dict_list:
+                if len(base_cfg_dict.keys() & c.keys()) > 0:
+                    raise KeyError("Duplicate key is not allowed among bases")
+                base_cfg_dict.update(c)
+
+            cfg_dict = merge_a_into_b(cfg_dict, base_cfg_dict)
+
+        return EasyDict(cfg_dict)
+
+    @classmethod
+    def merge_list(cls, cfg, opts: list):
+        """merge commandline opts.
+
+        Args:
+            cfg: (dict): The config to be merged.
+            opts (list): The list to merge. Format: [key1, name1, key2, name2,...].
+                The keys can be nested. For example, ["a.b", v] will be considered
+                as `dict(a=dict(b=v))`.
+
+        Returns: dict.
+
+        """
+        assert len(opts) % 2 == 0, f"length of opts must be even. Got: {opts}"
+        for i in range(0, len(opts), 2):
+            full_k, v = opts[i], opts[i + 1]
+            keys = full_k.split(".")
+            sub_d = cfg
+            for i, k in enumerate(keys):
+                if not hasattr(sub_d, k):
+                    raise ValueError(f"The key {k} not exist in the config. Full key:{full_k}")
+                if i != len(keys) - 1:
+                    sub_d = sub_d[k]
+                else:
+                    sub_d[k] = v
+        return cfg
+
+
+def merge_a_into_b(a, b, inplace=False):
+    """The values in a will override values in b.
+
+    Args:
+        a (dict): source dict.
+        b (dict): target dict.
+
+    Returns: dict. recursively merge dict a into dict b.
+
+    """
+    if not inplace:
+        b = deepcopy(b)
+    for key in a:
+        if key in b:
+            if isinstance(a[key], dict) and isinstance(b[key], dict):
+                b[key] = merge_a_into_b(a[key], b[key], inplace=True)
+            else:
+                b[key] = a[key]
+        else:
+            b[key] = a[key]
+    return b
+
+
+def eval_dict_leaf(d, orig_dict=None):
+    """eval values of dict leaf.
+
+    Args:
+        d (dict): The dict to eval.
+
+    Returns: dict.
+
+    """
+    if orig_dict is None:
+        orig_dict = d
+    for k, v in d.items():
+        if not isinstance(v, dict):
+            d[k] = eval_string(v, orig_dict)
+        else:
+            eval_dict_leaf(v, orig_dict)
+    return d
+
+
+def eval_string(string, d):
+    """automatically evaluate string to corresponding types.
+
+    For example:
+        not a string  -> return the original input
+        '0'  -> 0
+        '0.2' -> 0.2
+        '[0, 1, 2]' -> [0,1,2]
+        'eval(1+2)' -> 3
+        'eval(range(5))' -> [0,1,2,3,4]
+        '${a}' -> d.a
+
+
+
+    Args:
+        string (str): The value to evaluate.
+        d (dict): The
+
+    Returns: the corresponding type
+
+    """
+    if not isinstance(string, str):
+        return string
+    # if len(string) > 1 and string[0] == "[" and string[-1] == "]":
+    #     return eval(string)
+    if string[0:5] == "eval(":
+        return eval(string[5:-1])
+
+    s0 = string
+    s1 = re.sub(r"\${(.*)}", r"d.\1", s0)
+    if s1 != s0:
+        while s1 != s0:
+            s0 = s1
+            s1 = re.sub(r"\${(.*)}", r"d.\1", s0)
+        return eval(s1)
+
+    try:
+        v = ast.literal_eval(string)
+    except:
+        v = string
+    return v

utils/easydict.py

@@ -0,0 +1,149 @@
+class EasyDict(dict):
+    """
+    Get attributes
+
+    >>> d = EasyDict({'foo':3})
+    >>> d['foo']
+    3
+    >>> d.foo
+    3
+    >>> d.bar
+    Traceback (most recent call last):
+    ...
+    AttributeError: 'EasyDict' object has no attribute 'bar'
+
+    Works recursively
+
+    >>> d = EasyDict({'foo':3, 'bar':{'x':1, 'y':2}})
+    >>> isinstance(d.bar, dict)
+    True
+    >>> d.bar.x
+    1
+
+    Bullet-proof
+
+    >>> EasyDict({})
+    {}
+    >>> EasyDict(d={})
+    {}
+    >>> EasyDict(None)
+    {}
+    >>> d = {'a': 1}
+    >>> EasyDict(**d)
+    {'a': 1}
+
+    Set attributes
+
+    >>> d = EasyDict()
+    >>> d.foo = 3
+    >>> d.foo
+    3
+    >>> d.bar = {'prop': 'value'}
+    >>> d.bar.prop
+    'value'
+    >>> d
+    {'foo': 3, 'bar': {'prop': 'value'}}
+    >>> d.bar.prop = 'newer'
+    >>> d.bar.prop
+    'newer'
+
+
+    Values extraction
+
+    >>> d = EasyDict({'foo':0, 'bar':[{'x':1, 'y':2}, {'x':3, 'y':4}]})
+    >>> isinstance(d.bar, list)
+    True
+    >>> from operator import attrgetter
+    >>> map(attrgetter('x'), d.bar)
+    [1, 3]
+    >>> map(attrgetter('y'), d.bar)
+    [2, 4]
+    >>> d = EasyDict()
+    >>> d.keys()
+    []
+    >>> d = EasyDict(foo=3, bar=dict(x=1, y=2))
+    >>> d.foo
+    3
+    >>> d.bar.x
+    1
+
+    Still like a dict though
+
+    >>> o = EasyDict({'clean':True})
+    >>> o.items()
+    [('clean', True)]
+
+    And like a class
+
+    >>> class Flower(EasyDict):
+    ...     power = 1
+    ...
+    >>> f = Flower()
+    >>> f.power
+    1
+    >>> f = Flower({'height': 12})
+    >>> f.height
+    12
+    >>> f['power']
+    1
+    >>> sorted(f.keys())
+    ['height', 'power']
+
+    update and pop items
+    >>> d = EasyDict(a=1, b='2')
+    >>> e = EasyDict(c=3.0, a=9.0)
+    >>> d.update(e)
+    >>> d.c
+    3.0
+    >>> d['c']
+    3.0
+    >>> d.get('c')
+    3.0
+    >>> d.update(a=4, b=4)
+    >>> d.b
+    4
+    >>> d.pop('a')
+    4
+    >>> d.a
+    Traceback (most recent call last):
+    ...
+    AttributeError: 'EasyDict' object has no attribute 'a'
+    """
+
+    def __init__(self, d=None, **kwargs):
+        if d is None:
+            d = {}
+        if kwargs:
+            d.update(**kwargs)
+        for k, v in d.items():
+            setattr(self, k, v)
+        # Class attributes
+        for k in self.__class__.__dict__.keys():
+            if not (k.startswith("__") and k.endswith("__")) and not k in ("update", "pop"):
+                setattr(self, k, getattr(self, k))
+
+    def __setattr__(self, name, value):
+        if isinstance(value, (list, tuple)):
+            value = [self.__class__(x) if isinstance(x, dict) else x for x in value]
+        elif isinstance(value, dict) and not isinstance(value, self.__class__):
+            value = self.__class__(value)
+        super(EasyDict, self).__setattr__(name, value)
+        super(EasyDict, self).__setitem__(name, value)
+
+    __setitem__ = __setattr__
+
+    def update(self, e=None, **f):
+        d = e or dict()
+        d.update(f)
+        for k in d:
+            setattr(self, k, d[k])
+
+    def pop(self, k, d=None):
+        if hasattr(self, k):
+            delattr(self, k)
+        return super(EasyDict, self).pop(k, d)
+
+
+if __name__ == "__main__":
+    import doctest
+