v1

README.md

@@ -0,0 +1,14 @@
+---
+title: Video LLaMA
+emoji: 🚀
+colorFrom: purple
+colorTo: gray
+sdk: gradio
+sdk_version: 3.29.0
+app_file: app.py
+pinned: false
+license: other
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+

app.py

@@ -0,0 +1,192 @@
+"""
+Adapted from: https://github.com/Vision-CAIR/MiniGPT-4/blob/main/demo.py
+"""
+import argparse
+import os
+import random
+
+import numpy as np
+import torch
+import torch.backends.cudnn as cudnn
+import gradio as gr
+
+from video_llama.common.config import Config
+from video_llama.common.dist_utils import get_rank
+from video_llama.common.registry import registry
+from video_llama.conversation.conversation_video import Chat, Conversation, default_conversation,SeparatorStyle
+import decord
+decord.bridge.set_bridge('torch')
+
+#%%
+# imports modules for registration
+from video_llama.datasets.builders import *
+from video_llama.models import *
+from video_llama.processors import *
+from video_llama.runners import *
+from video_llama.tasks import *
+
+#%%
+def parse_args():
+    parser = argparse.ArgumentParser(description="Demo")
+    parser.add_argument("--cfg-path", default='eval_configs/video_llama_eval.yaml', help="path to configuration file.")
+    parser.add_argument("--gpu-id", type=int, default=0, help="specify the gpu to load the model.")
+    parser.add_argument(
+        "--options",
+        nargs="+",
+        help="override some settings in the used config, the key-value pair "
+        "in xxx=yyy format will be merged into config file (deprecate), "
+        "change to --cfg-options instead.",
+    )
+    args = parser.parse_args()
+    return args
+
+
+def setup_seeds(config):
+    seed = config.run_cfg.seed + get_rank()
+
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+
+    cudnn.benchmark = False
+    cudnn.deterministic = True
+
+
+# ========================================
+#             Model Initialization
+# ========================================
+
+print('Initializing Chat')
+args = parse_args()
+cfg = Config(args)
+
+model_config = cfg.model_cfg
+model_config.device_8bit = args.gpu_id
+model_cls = registry.get_model_class(model_config.arch)
+model = model_cls.from_config(model_config).to('cuda:{}'.format(args.gpu_id))
+
+vis_processor_cfg = cfg.datasets_cfg.webvid.vis_processor.train
+vis_processor = registry.get_processor_class(vis_processor_cfg.name).from_config(vis_processor_cfg)
+chat = Chat(model, vis_processor, device='cuda:{}'.format(args.gpu_id))
+print('Initialization Finished')
+
+# ========================================
+#             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_imgorvideo(gr_video, gr_img, text_input, chat_state):
+    if gr_img is None and gr_video is None:
+        return None, None, None, gr.update(interactive=True), chat_state, None
+    elif gr_img is not None and gr_video is None:
+        print(gr_img)
+        chat_state = Conversation(
+            system= "You are able to understand the visual content that the user provides."
+           "Follow the instructions carefully and explain your answers in detail.",
+            roles=("Human", "Assistant"),
+            messages=[],
+            offset=0,
+            sep_style=SeparatorStyle.SINGLE,
+            sep="###",
+        )
+        img_list = []
+        llm_message = chat.upload_img(gr_img, chat_state, img_list)
+        return gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=True, placeholder='Type and press Enter'), gr.update(value="Start Chatting", interactive=False), chat_state, img_list
+    elif gr_video is not None and gr_img is None:
+        print(gr_video)
+        chat_state = default_conversation.copy()
+        chat_state = Conversation(
+            system= "You are able to understand the visual content that the user provides."
+           "Follow the instructions carefully and explain your answers in detail.",
+            roles=("Human", "Assistant"),
+            messages=[],
+            offset=0,
+            sep_style=SeparatorStyle.SINGLE,
+            sep="###",
+        )
+        img_list = []
+        llm_message = chat.upload_video(gr_video, chat_state, img_list)
+        return gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=True, placeholder='Type and press Enter'), gr.update(value="Start Chatting", interactive=False), chat_state, img_list
+    else:
+        # img_list = []
+        return gr.update(interactive=False), gr.update(interactive=False, placeholder='Currently, only one input is supported'), gr.update(value="Currently, only one input is supported", interactive=False), chat_state, None
+
+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
+    chat.ask(user_message, chat_state)
+    chatbot = chatbot + [[user_message, None]]
+    return '', chatbot, chat_state
+
+
+def gradio_answer(chatbot, chat_state, img_list, num_beams, temperature):
+    llm_message = chat.answer(conv=chat_state,
+                              img_list=img_list,
+                              num_beams=num_beams,
+                              temperature=temperature,
+                              max_new_tokens=300,
+                              max_length=2000)[0]
+    chatbot[-1][1] = llm_message
+    print(chat_state.get_prompt())
+    print(chat_state)
+    return chatbot, chat_state, img_list
+
+title = """<h1 align="center">Demo of Video-LLaMA</h1>"""
+description = """<h3>This is the demo of Video-LLaMA. Upload your images/videos and start chatting!</h3>"""
+
+
+#TODO show examples below
+
+with gr.Blocks() as demo:
+    gr.Markdown(title)
+    gr.Markdown(description)
+
+    with gr.Row():
+        with gr.Column(scale=0.5):
+            video = gr.Video()
+            image = gr.Image(type="pil")
+
+            upload_button = gr.Button(value="Upload & Start Chat", interactive=True, variant="primary")
+            clear = gr.Button("Restart")
+            
+            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",
+            )
+
+        with gr.Column():
+            chat_state = gr.State()
+            img_list = gr.State()
+            chatbot = gr.Chatbot(label='Video-LLaMA')
+            text_input = gr.Textbox(label='User', placeholder='Please upload your image/video first', interactive=False)
+
+    
+    upload_button.click(upload_imgorvideo, [video, image, text_input, chat_state], [video, image, 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, [chatbot, chat_state, img_list, num_beams, temperature], [chatbot, chat_state, img_list]
+    )
+    clear.click(gradio_reset, [chat_state, img_list], [chatbot, video, image, text_input, upload_button, chat_state, img_list], queue=False)
+
+demo.launch(share=False, enable_queue=False)
+
+# %%

ckpt/blip2_pretrained_flant5xxl.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b3839ea6c617f315ead9bf4036bbb0f0cf6bf62695ecfc14968ea626af03a29
+size 433481467

ckpt/finetune-vicuna7b-v2-nofrozen_imageQ.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc4b32437c90df51bc3faa29deaa9b25ab77e1707ac79066f17ae3193ebe8bfc
+size 1527692539

ckpt/finetune-vicuna7b-v2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0680ad8eb14c2a3273b7be71309ab6b06c9f426e87ad4675a903371fe0fa8162
+size 265436777

ckpt/pretrain-billa7b-zh.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f50a51db3055e1be6461f6dec833fbbbba28650287d26c8787664c8ee31dcf0f
+size 265435689

eval_configs/video_llama_eval.yaml

@@ -0,0 +1,32 @@
+model:
+  arch: video_llama
+  model_type: pretrain_vicuna
+  freeze_vit: True
+  freeze_qformer: True
+  max_txt_len: 512
+  end_sym: "###"
+  low_resource: False
+
+
+  llama_model: "DAMO-NLP-SG/vicuna-7b"
+
+  fusion_head_layers: 2
+  max_frame_pos: 32
+  fusion_header_type: "seqTransf"
+
+  ckpt: 'ckpt/finetune-vicuna7b-v2-nofrozen_imageQ.pth'
+  q_former_model: 'ckpt/blip2_pretrained_flant5xxl.pth'
+
+datasets:
+  webvid:
+    vis_processor:
+      train:
+        name: "alpro_video_eval"
+        n_frms: 8
+        image_size: 224
+    text_processor:
+      train:
+        name: "blip_caption"
+
+run:
+  task: video_text_pretrain

requirements.txt

@@ -0,0 +1,11 @@
+transformers==4.28.0
+tqdm
+decord
+timm
+einops
+opencv_python
+torchvision
+
+salesforce-lavis
+bitsandbytes
+accelerate

video_llama/__init__.py

@@ -0,0 +1,31 @@
+"""
+ Copyright (c) 2022, 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 os
+import sys
+
+from omegaconf import OmegaConf
+
+from video_llama.common.registry import registry
+
+from video_llama.datasets.builders import *
+from video_llama.models import *
+from video_llama.processors import *
+from video_llama.tasks import *
+
+
+root_dir = os.path.dirname(os.path.abspath(__file__))
+default_cfg = OmegaConf.load(os.path.join(root_dir, "configs/default.yaml"))
+
+registry.register_path("library_root", root_dir)
+repo_root = os.path.join(root_dir, "..")
+registry.register_path("repo_root", repo_root)
+cache_root = os.path.join(repo_root, default_cfg.env.cache_root)
+registry.register_path("cache_root", cache_root)
+
+registry.register("MAX_INT", sys.maxsize)
+registry.register("SPLIT_NAMES", ["train", "val", "test"])

video_llama/app.py

@@ -0,0 +1,192 @@
+"""
+Adapted from: https://github.com/Vision-CAIR/MiniGPT-4/blob/main/demo.py
+"""
+import argparse
+import os
+import random
+
+import numpy as np
+import torch
+import torch.backends.cudnn as cudnn
+import gradio as gr
+
+from video_llama.common.config import Config
+from video_llama.common.dist_utils import get_rank
+from video_llama.common.registry import registry
+from video_llama.conversation.conversation_video import Chat, Conversation, default_conversation,SeparatorStyle
+import decord
+decord.bridge.set_bridge('torch')
+
+#%%
+# imports modules for registration
+from video_llama.datasets.builders import *
+from video_llama.models import *
+from video_llama.processors import *
+from video_llama.runners import *
+from video_llama.tasks import *
+
+#%%
+def parse_args():
+    parser = argparse.ArgumentParser(description="Demo")
+    parser.add_argument("--cfg-path", default='eval_configs/video_llama_eval.yaml', help="path to configuration file.")
+    parser.add_argument("--gpu-id", type=int, default=0, help="specify the gpu to load the model.")
+    parser.add_argument(
+        "--options",
+        nargs="+",
+        help="override some settings in the used config, the key-value pair "
+        "in xxx=yyy format will be merged into config file (deprecate), "
+        "change to --cfg-options instead.",
+    )
+    args = parser.parse_args()
+    return args
+
+
+def setup_seeds(config):
+    seed = config.run_cfg.seed + get_rank()
+
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+
+    cudnn.benchmark = False
+    cudnn.deterministic = True
+
+
+# ========================================
+#             Model Initialization
+# ========================================
+
+print('Initializing Chat')
+args = parse_args()
+cfg = Config(args)
+
+model_config = cfg.model_cfg
+model_config.device_8bit = args.gpu_id
+model_cls = registry.get_model_class(model_config.arch)
+model = model_cls.from_config(model_config).to('cuda:{}'.format(args.gpu_id))
+
+vis_processor_cfg = cfg.datasets_cfg.webvid.vis_processor.train
+vis_processor = registry.get_processor_class(vis_processor_cfg.name).from_config(vis_processor_cfg)
+chat = Chat(model, vis_processor, device='cuda:{}'.format(args.gpu_id))
+print('Initialization Finished')
+
+# ========================================
+#             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_imgorvideo(gr_video, gr_img, text_input, chat_state):
+    if gr_img is None and gr_video is None:
+        return None, None, None, gr.update(interactive=True), chat_state, None
+    elif gr_img is not None and gr_video is None:
+        print(gr_img)
+        chat_state = Conversation(
+            system= "You are able to understand the visual content that the user provides."
+           "Follow the instructions carefully and explain your answers in detail.",
+            roles=("Human", "Assistant"),
+            messages=[],
+            offset=0,
+            sep_style=SeparatorStyle.SINGLE,
+            sep="###",
+        )
+        img_list = []
+        llm_message = chat.upload_img(gr_img, chat_state, img_list)
+        return gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=True, placeholder='Type and press Enter'), gr.update(value="Start Chatting", interactive=False), chat_state, img_list
+    elif gr_video is not None and gr_img is None:
+        print(gr_video)
+        chat_state = default_conversation.copy()
+        chat_state = Conversation(
+            system= "You are able to understand the visual content that the user provides."
+           "Follow the instructions carefully and explain your answers in detail.",
+            roles=("Human", "Assistant"),
+            messages=[],
+            offset=0,
+            sep_style=SeparatorStyle.SINGLE,
+            sep="###",
+        )
+        img_list = []
+        llm_message = chat.upload_video(gr_video, chat_state, img_list)
+        return gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=True, placeholder='Type and press Enter'), gr.update(value="Start Chatting", interactive=False), chat_state, img_list
+    else:
+        # img_list = []
+        return gr.update(interactive=False), gr.update(interactive=False, placeholder='Currently, only one input is supported'), gr.update(value="Currently, only one input is supported", interactive=False), chat_state, None
+
+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
+    chat.ask(user_message, chat_state)
+    chatbot = chatbot + [[user_message, None]]
+    return '', chatbot, chat_state
+
+
+def gradio_answer(chatbot, chat_state, img_list, num_beams, temperature):
+    llm_message = chat.answer(conv=chat_state,
+                              img_list=img_list,
+                              num_beams=num_beams,
+                              temperature=temperature,
+                              max_new_tokens=300,
+                              max_length=2000)[0]
+    chatbot[-1][1] = llm_message
+    print(chat_state.get_prompt())
+    print(chat_state)
+    return chatbot, chat_state, img_list
+
+title = """<h1 align="center">Demo of Video-LLaMA</h1>"""
+description = """<h3>This is the demo of Video-LLaMA. Upload your images/videos and start chatting!</h3>"""
+
+
+#TODO show examples below
+
+with gr.Blocks() as demo:
+    gr.Markdown(title)
+    gr.Markdown(description)
+
+    with gr.Row():
+        with gr.Column(scale=0.5):
+            video = gr.Video()
+            image = gr.Image(type="pil")
+
+            upload_button = gr.Button(value="Upload & Start Chat", interactive=True, variant="primary")
+            clear = gr.Button("Restart")
+            
+            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",
+            )
+
+        with gr.Column():
+            chat_state = gr.State()
+            img_list = gr.State()
+            chatbot = gr.Chatbot(label='Video-LLaMA')
+            text_input = gr.Textbox(label='User', placeholder='Please upload your image/video first', interactive=False)
+
+    
+    upload_button.click(upload_imgorvideo, [video, image, text_input, chat_state], [video, image, 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, [chatbot, chat_state, img_list, num_beams, temperature], [chatbot, chat_state, img_list]
+    )
+    clear.click(gradio_reset, [chat_state, img_list], [chatbot, video, image, text_input, upload_button, chat_state, img_list], queue=False)
+
+demo.launch(share=False, enable_queue=False)
+
+# %%

video_llama/ckpt/blip2_pretrained_flant5xxl.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b3839ea6c617f315ead9bf4036bbb0f0cf6bf62695ecfc14968ea626af03a29
+size 433481467

video_llama/ckpt/finetune-vicuna7b-v2-nofrozen_imageQ.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46af76d307c14d28c56534e4bf8654343e5512aa1285fc1c1fdb5728c418e7ca
+size 623104000

video_llama/ckpt/pretrain-billa7b-zh.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f50a51db3055e1be6461f6dec833fbbbba28650287d26c8787664c8ee31dcf0f
+size 265435689

video_llama/common/config.py

@@ -0,0 +1,468 @@
+"""
+ Copyright (c) 2022, 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 logging
+import json
+from typing import Dict
+
+from omegaconf import OmegaConf
+from video_llama.common.registry import registry
+
+
+class Config:
+    def __init__(self, args):
+        self.config = {}
+
+        self.args = args
+
+        # Register the config and configuration for setup
+        registry.register("configuration", self)
+
+        user_config = self._build_opt_list(self.args.options)
+
+        config = OmegaConf.load(self.args.cfg_path)
+
+        runner_config = self.build_runner_config(config)
+        model_config = self.build_model_config(config, **user_config)
+        dataset_config = self.build_dataset_config(config)
+
+        # Validate the user-provided runner configuration
+        # model and dataset configuration are supposed to be validated by the respective classes
+        # [TODO] validate the model/dataset configuration
+        # self._validate_runner_config(runner_config)
+
+        # Override the default configuration with user options.
+        self.config = OmegaConf.merge(
+            runner_config, model_config, dataset_config, user_config
+        )
+
+    def _validate_runner_config(self, runner_config):
+        """
+        This method validates the configuration, such that
+            1) all the user specified options are valid;
+            2) no type mismatches between the user specified options and the config.
+        """
+        runner_config_validator = create_runner_config_validator()
+        runner_config_validator.validate(runner_config)
+
+    def _build_opt_list(self, opts):
+        opts_dot_list = self._convert_to_dot_list(opts)
+        return OmegaConf.from_dotlist(opts_dot_list)
+
+    @staticmethod
+    def build_model_config(config, **kwargs):
+        model = config.get("model", None)
+        assert model is not None, "Missing model configuration file."
+
+        model_cls = registry.get_model_class(model.arch)
+        assert model_cls is not None, f"Model '{model.arch}' has not been registered."
+
+        model_type = kwargs.get("model.model_type", None)
+        if not model_type:
+            model_type = model.get("model_type", None)
+        # else use the model type selected by user.
+
+        assert model_type is not None, "Missing model_type."
+
+        model_config_path = model_cls.default_config_path(model_type=model_type)
+
+        model_config = OmegaConf.create()
+        # hierarchy override, customized config > default config
+        model_config = OmegaConf.merge(
+            model_config,
+            OmegaConf.load(model_config_path),
+            {"model": config["model"]},
+        )
+
+        return model_config
+
+    @staticmethod
+    def build_runner_config(config):
+        return {"run": config.run}
+
+    @staticmethod
+    def build_dataset_config(config):
+        datasets = config.get("datasets", None)
+        if datasets is None:
+            raise KeyError(
+                "Expecting 'datasets' as the root key for dataset configuration."
+            )
+
+        dataset_config = OmegaConf.create()
+
+        for dataset_name in datasets:
+            builder_cls = registry.get_builder_class(dataset_name)
+
+            dataset_config_type = datasets[dataset_name].get("type", "default")
+            dataset_config_path = builder_cls.default_config_path(
+                type=dataset_config_type
+            )
+
+            # hierarchy override, customized config > default config
+            dataset_config = OmegaConf.merge(
+                dataset_config,
+                OmegaConf.load(dataset_config_path),
+                {"datasets": {dataset_name: config["datasets"][dataset_name]}},
+            )
+
+        return dataset_config
+
+    def _convert_to_dot_list(self, opts):
+        if opts is None:
+            opts = []
+
+        if len(opts) == 0:
+            return opts
+
+        has_equal = opts[0].find("=") != -1
+
+        if has_equal:
+            return opts
+
+        return [(opt + "=" + value) for opt, value in zip(opts[0::2], opts[1::2])]
+
+    def get_config(self):
+        return self.config
+
+    @property
+    def run_cfg(self):
+        return self.config.run
+
+    @property
+    def datasets_cfg(self):
+        return self.config.datasets
+
+    @property
+    def model_cfg(self):
+        return self.config.model
+
+    def pretty_print(self):
+        logging.info("\n=====  Running Parameters    =====")
+        logging.info(self._convert_node_to_json(self.config.run))
+
+        logging.info("\n======  Dataset Attributes  ======")
+        datasets = self.config.datasets
+
+        for dataset in datasets:
+            if dataset in self.config.datasets:
+                logging.info(f"\n======== {dataset} =======")
+                dataset_config = self.config.datasets[dataset]
+                logging.info(self._convert_node_to_json(dataset_config))
+            else:
+                logging.warning(f"No dataset named '{dataset}' in config. Skipping")
+
+        logging.info(f"\n======  Model Attributes  ======")
+        logging.info(self._convert_node_to_json(self.config.model))
+
+    def _convert_node_to_json(self, node):
+        container = OmegaConf.to_container(node, resolve=True)
+        return json.dumps(container, indent=4, sort_keys=True)
+
+    def to_dict(self):
+        return OmegaConf.to_container(self.config)
+
+
+def node_to_dict(node):
+    return OmegaConf.to_container(node)
+
+
+class ConfigValidator:
+    """
+    This is a preliminary implementation to centralize and validate the configuration.
+    May be altered in the future.
+
+    A helper class to validate configurations from yaml file.
+
+    This serves the following purposes:
+        1. Ensure all the options in the yaml are defined, raise error if not.
+        2. when type mismatches are found, the validator will raise an error.
+        3. a central place to store and display helpful messages for supported configurations.
+
+    """
+
+    class _Argument:
+        def __init__(self, name, choices=None, type=None, help=None):
+            self.name = name
+            self.val = None
+            self.choices = choices
+            self.type = type
+            self.help = help
+
+        def __str__(self):
+            s = f"{self.name}={self.val}"
+            if self.type is not None:
+                s += f", ({self.type})"
+            if self.choices is not None:
+                s += f", choices: {self.choices}"
+            if self.help is not None:
+                s += f", ({self.help})"
+            return s
+
+    def __init__(self, description):
+        self.description = description
+
+        self.arguments = dict()
+
+        self.parsed_args = None
+
+    def __getitem__(self, key):
+        assert self.parsed_args is not None, "No arguments parsed yet."
+
+        return self.parsed_args[key]
+
+    def __str__(self) -> str:
+        return self.format_help()
+
+    def add_argument(self, *args, **kwargs):
+        """
+        Assume the first argument is the name of the argument.
+        """
+        self.arguments[args[0]] = self._Argument(*args, **kwargs)
+
+    def validate(self, config=None):
+        """
+        Convert yaml config (dict-like) to list, required by argparse.
+        """
+        for k, v in config.items():
+            assert (
+                k in self.arguments
+            ), f"""{k} is not a valid argument. Support arguments are {self.format_arguments()}."""
+
+            if self.arguments[k].type is not None:
+                try:
+                    self.arguments[k].val = self.arguments[k].type(v)
+                except ValueError:
+                    raise ValueError(f"{k} is not a valid {self.arguments[k].type}.")
+
+            if self.arguments[k].choices is not None:
+                assert (
+                    v in self.arguments[k].choices
+                ), f"""{k} must be one of {self.arguments[k].choices}."""
+
+        return config
+
+    def format_arguments(self):
+        return str([f"{k}" for k in sorted(self.arguments.keys())])
+
+    def format_help(self):
+        # description + key-value pair string for each argument
+        help_msg = str(self.description)
+        return help_msg + ", available arguments: " + self.format_arguments()
+
+    def print_help(self):
+        # display help message
+        print(self.format_help())
+
+
+def create_runner_config_validator():
+    validator = ConfigValidator(description="Runner configurations")
+
+    validator.add_argument(
+        "runner",
+        type=str,
+        choices=["runner_base", "runner_iter"],
+        help="""Runner to use. The "runner_base" uses epoch-based training while iter-based
+            runner runs based on iters. Default: runner_base""",
+    )
+    # add argumetns for training dataset ratios
+    validator.add_argument(
+        "train_dataset_ratios",
+        type=Dict[str, float],
+        help="""Ratios of training dataset. This is used in iteration-based runner.
+        Do not support for epoch-based runner because how to define an epoch becomes tricky.
+        Default: None""",
+    )
+    validator.add_argument(
+        "max_iters",
+        type=float,
+        help="Maximum number of iterations to run.",
+    )
+    validator.add_argument(
+        "max_epoch",
+        type=int,
+        help="Maximum number of epochs to run.",
+    )
+    # add arguments for iters_per_inner_epoch
+    validator.add_argument(
+        "iters_per_inner_epoch",
+        type=float,
+        help="Number of iterations per inner epoch. This is required when runner is runner_iter.",
+    )
+    lr_scheds_choices = registry.list_lr_schedulers()
+    validator.add_argument(
+        "lr_sched",
+        type=str,
+        choices=lr_scheds_choices,
+        help="Learning rate scheduler to use, from {}".format(lr_scheds_choices),
+    )
+    task_choices = registry.list_tasks()
+    validator.add_argument(
+        "task",
+        type=str,
+        choices=task_choices,
+        help="Task to use, from {}".format(task_choices),
+    )
+    # add arguments for init_lr
+    validator.add_argument(
+        "init_lr",
+        type=float,
+        help="Initial learning rate. This will be the learning rate after warmup and before decay.",
+    )
+    # add arguments for min_lr
+    validator.add_argument(
+        "min_lr",
+        type=float,
+        help="Minimum learning rate (after decay).",
+    )
+    # add arguments for warmup_lr
+    validator.add_argument(
+        "warmup_lr",
+        type=float,
+        help="Starting learning rate for warmup.",
+    )
+    # add arguments for learning rate decay rate
+    validator.add_argument(
+        "lr_decay_rate",
+        type=float,
+        help="Learning rate decay rate. Required if using a decaying learning rate scheduler.",
+    )
+    # add arguments for weight decay
+    validator.add_argument(
+        "weight_decay",
+        type=float,
+        help="Weight decay rate.",
+    )
+    # add arguments for training batch size
+    validator.add_argument(
+        "batch_size_train",
+        type=int,
+        help="Training batch size.",
+    )
+    # add arguments for evaluation batch size
+    validator.add_argument(
+        "batch_size_eval",
+        type=int,
+        help="Evaluation batch size, including validation and testing.",
+    )
+    # add arguments for number of workers for data loading
+    validator.add_argument(
+        "num_workers",
+        help="Number of workers for data loading.",
+    )
+    # add arguments for warm up steps
+    validator.add_argument(
+        "warmup_steps",
+        type=int,
+        help="Number of warmup steps. Required if a warmup schedule is used.",
+    )
+    # add arguments for random seed
+    validator.add_argument(
+        "seed",
+        type=int,
+        help="Random seed.",
+    )
+    # add arguments for output directory
+    validator.add_argument(
+        "output_dir",
+        type=str,
+        help="Output directory to save checkpoints and logs.",
+    )
+    # add arguments for whether only use evaluation
+    validator.add_argument(
+        "evaluate",
+        help="Whether to only evaluate the model. If true, training will not be performed.",
+    )
+    # add arguments for splits used for training, e.g. ["train", "val"]
+    validator.add_argument(
+        "train_splits",
+        type=list,
+        help="Splits to use for training.",
+    )
+    # add arguments for splits used for validation, e.g. ["val"]
+    validator.add_argument(
+        "valid_splits",
+        type=list,
+        help="Splits to use for validation. If not provided, will skip the validation.",
+    )
+    # add arguments for splits used for testing, e.g. ["test"]
+    validator.add_argument(
+        "test_splits",
+        type=list,
+        help="Splits to use for testing. If not provided, will skip the testing.",
+    )
+    # add arguments for accumulating gradient for iterations
+    validator.add_argument(
+        "accum_grad_iters",
+        type=int,
+        help="Number of iterations to accumulate gradient for.",
+    )
+
+    # ====== distributed training ======
+    validator.add_argument(
+        "device",
+        type=str,
+        choices=["cpu", "cuda"],
+        help="Device to use. Support 'cuda' or 'cpu' as for now.",
+    )
+    validator.add_argument(
+        "world_size",
+        type=int,
+        help="Number of processes participating in the job.",
+    )
+    validator.add_argument("dist_url", type=str)
+    validator.add_argument("distributed", type=bool)
+    # add arguments to opt using distributed sampler during evaluation or not
+    validator.add_argument(
+        "use_dist_eval_sampler",
+        type=bool,
+        help="Whether to use distributed sampler during evaluation or not.",
+    )
+
+    # ====== task specific ======
+    # generation task specific arguments
+    # add arguments for maximal length of text output
+    validator.add_argument(
+        "max_len",
+        type=int,
+        help="Maximal length of text output.",
+    )
+    # add arguments for minimal length of text output
+    validator.add_argument(
+        "min_len",
+        type=int,
+        help="Minimal length of text output.",
+    )
+    # add arguments number of beams
+    validator.add_argument(
+        "num_beams",
+        type=int,
+        help="Number of beams used for beam search.",
+    )
+
+    # vqa task specific arguments
+    # add arguments for number of answer candidates
+    validator.add_argument(
+        "num_ans_candidates",
+        type=int,
+        help="""For ALBEF and BLIP, these models first rank answers according to likelihood to select answer candidates.""",
+    )
+    # add arguments for inference method
+    validator.add_argument(
+        "inference_method",
+        type=str,
+        choices=["genearte", "rank"],
+        help="""Inference method to use for question answering. If rank, requires a answer list.""",
+    )
+
+    # ====== model specific ======
+    validator.add_argument(
+        "k_test",
+        type=int,
+        help="Number of top k most similar samples from ITC/VTC selection to be tested.",
+    )
+
+    return validator

video_llama/common/dist_utils.py

@@ -0,0 +1,137 @@
+"""
+ Copyright (c) 2022, 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 datetime
+import functools
+import os
+
+import torch
+import torch.distributed as dist
+import timm.models.hub as timm_hub
+
+
+def setup_for_distributed(is_master):
+    """
+    This function disables printing when not in master process
+    """
+    import builtins as __builtin__
+
+    builtin_print = __builtin__.print
+
+    def print(*args, **kwargs):
+        force = kwargs.pop("force", False)
+        if is_master or force:
+            builtin_print(*args, **kwargs)
+
+    __builtin__.print = print
+
+
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+
+def get_world_size():
+    if not is_dist_avail_and_initialized():
+        return 1
+    return dist.get_world_size()
+
+
+def get_rank():
+    if not is_dist_avail_and_initialized():
+        return 0
+    return dist.get_rank()
+
+
+def is_main_process():
+    return get_rank() == 0
+
+
+def init_distributed_mode(args):
+    if "RANK" in os.environ and "WORLD_SIZE" in os.environ:
+        args.rank = int(os.environ["RANK"])
+        args.world_size = int(os.environ["WORLD_SIZE"])
+        args.gpu = int(os.environ["LOCAL_RANK"])
+    elif "SLURM_PROCID" in os.environ:
+        args.rank = int(os.environ["SLURM_PROCID"])
+        args.gpu = args.rank % torch.cuda.device_count()
+    else:
+        print("Not using distributed mode")
+        args.distributed = False
+        return
+
+    args.distributed = True
+
+    torch.cuda.set_device(args.gpu)
+    args.dist_backend = "nccl"
+    print(
+        "| distributed init (rank {}, world {}): {}".format(
+            args.rank, args.world_size, args.dist_url
+        ),
+        flush=True,
+    )
+    torch.distributed.init_process_group(
+        backend=args.dist_backend,
+        init_method=args.dist_url,
+        world_size=args.world_size,
+        rank=args.rank,
+        timeout=datetime.timedelta(
+            days=365
+        ),  # allow auto-downloading and de-compressing
+    )
+    torch.distributed.barrier()
+    setup_for_distributed(args.rank == 0)
+
+
+def get_dist_info():
+    if torch.__version__ < "1.0":
+        initialized = dist._initialized
+    else:
+        initialized = dist.is_initialized()
+    if initialized:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    else:  # non-distributed training
+        rank = 0
+        world_size = 1
+    return rank, world_size
+
+
+def main_process(func):
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs):
+        rank, _ = get_dist_info()
+        if rank == 0:
+            return func(*args, **kwargs)
+
+    return wrapper
+
+
+def download_cached_file(url, check_hash=True, progress=False):
+    """
+    Download a file from a URL and cache it locally. If the file already exists, it is not downloaded again.
+    If distributed, only the main process downloads the file, and the other processes wait for the file to be downloaded.
+    """
+
+    def get_cached_file_path():
+        # a hack to sync the file path across processes
+        parts = torch.hub.urlparse(url)
+        filename = os.path.basename(parts.path)
+        cached_file = os.path.join(timm_hub.get_cache_dir(), filename)
+
+        return cached_file
+
+    if is_main_process():
+        timm_hub.download_cached_file(url, check_hash, progress)
+
+    if is_dist_avail_and_initialized():
+        dist.barrier()
+
+    return get_cached_file_path()

video_llama/common/gradcam.py

@@ -0,0 +1,24 @@
+import numpy as np
+from matplotlib import pyplot as plt
+from scipy.ndimage import filters
+from skimage import transform as skimage_transform
+
+
+def getAttMap(img, attMap, blur=True, overlap=True):
+    attMap -= attMap.min()
+    if attMap.max() > 0:
+        attMap /= attMap.max()
+    attMap = skimage_transform.resize(attMap, (img.shape[:2]), order=3, mode="constant")
+    if blur:
+        attMap = filters.gaussian_filter(attMap, 0.02 * max(img.shape[:2]))
+        attMap -= attMap.min()
+        attMap /= attMap.max()
+    cmap = plt.get_cmap("jet")
+    attMapV = cmap(attMap)
+    attMapV = np.delete(attMapV, 3, 2)
+    if overlap:
+        attMap = (
+            1 * (1 - attMap**0.7).reshape(attMap.shape + (1,)) * img
+            + (attMap**0.7).reshape(attMap.shape + (1,)) * attMapV
+        )
+    return attMap

video_llama/common/logger.py

@@ -0,0 +1,195 @@
+"""
+ Copyright (c) 2022, 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 datetime
+import logging
+import time
+from collections import defaultdict, deque
+
+import torch
+import torch.distributed as dist
+
+from video_llama.common import dist_utils
+
+
+class SmoothedValue(object):
+    """Track a series of values and provide access to smoothed values over a
+    window or the global series average.
+    """
+
+    def __init__(self, window_size=20, fmt=None):
+        if fmt is None:
+            fmt = "{median:.4f} ({global_avg:.4f})"
+        self.deque = deque(maxlen=window_size)
+        self.total = 0.0
+        self.count = 0
+        self.fmt = fmt
+
+    def update(self, value, n=1):
+        self.deque.append(value)
+        self.count += n
+        self.total += value * n
+
+    def synchronize_between_processes(self):
+        """
+        Warning: does not synchronize the deque!
+        """
+        if not dist_utils.is_dist_avail_and_initialized():
+            return
+        t = torch.tensor([self.count, self.total], dtype=torch.float64, device="cuda")
+        dist.barrier()
+        dist.all_reduce(t)
+        t = t.tolist()
+        self.count = int(t[0])
+        self.total = t[1]
+
+    @property
+    def median(self):
+        d = torch.tensor(list(self.deque))
+        return d.median().item()
+
+    @property
+    def avg(self):
+        d = torch.tensor(list(self.deque), dtype=torch.float32)
+        return d.mean().item()
+
+    @property
+    def global_avg(self):
+        return self.total / self.count
+
+    @property
+    def max(self):
+        return max(self.deque)
+
+    @property
+    def value(self):
+        return self.deque[-1]
+
+    def __str__(self):
+        return self.fmt.format(
+            median=self.median,
+            avg=self.avg,
+            global_avg=self.global_avg,
+            max=self.max,
+            value=self.value,
+        )
+
+
+class MetricLogger(object):
+    def __init__(self, delimiter="\t"):
+        self.meters = defaultdict(SmoothedValue)
+        self.delimiter = delimiter
+
+    def update(self, **kwargs):
+        for k, v in kwargs.items():
+            if isinstance(v, torch.Tensor):
+                v = v.item()
+            assert isinstance(v, (float, int))
+            self.meters[k].update(v)
+
+    def __getattr__(self, attr):
+        if attr in self.meters:
+            return self.meters[attr]
+        if attr in self.__dict__:
+            return self.__dict__[attr]
+        raise AttributeError(
+            "'{}' object has no attribute '{}'".format(type(self).__name__, attr)
+        )
+
+    def __str__(self):
+        loss_str = []
+        for name, meter in self.meters.items():
+            loss_str.append("{}: {}".format(name, str(meter)))
+        return self.delimiter.join(loss_str)
+
+    def global_avg(self):
+        loss_str = []
+        for name, meter in self.meters.items():
+            loss_str.append("{}: {:.4f}".format(name, meter.global_avg))
+        return self.delimiter.join(loss_str)
+
+    def synchronize_between_processes(self):
+        for meter in self.meters.values():
+            meter.synchronize_between_processes()
+
+    def add_meter(self, name, meter):
+        self.meters[name] = meter
+
+    def log_every(self, iterable, print_freq, header=None):
+        i = 0
+        if not header:
+            header = ""
+        start_time = time.time()
+        end = time.time()
+        iter_time = SmoothedValue(fmt="{avg:.4f}")
+        data_time = SmoothedValue(fmt="{avg:.4f}")
+        space_fmt = ":" + str(len(str(len(iterable)))) + "d"
+        log_msg = [
+            header,
+            "[{0" + space_fmt + "}/{1}]",
+            "eta: {eta}",
+            "{meters}",
+            "time: {time}",
+            "data: {data}",
+        ]
+        if torch.cuda.is_available():
+            log_msg.append("max mem: {memory:.0f}")
+        log_msg = self.delimiter.join(log_msg)
+        MB = 1024.0 * 1024.0
+        for obj in iterable:
+            data_time.update(time.time() - end)
+            yield obj
+            iter_time.update(time.time() - end)
+            if i % print_freq == 0 or i == len(iterable) - 1:
+                eta_seconds = iter_time.global_avg * (len(iterable) - i)
+                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
+                if torch.cuda.is_available():
+                    print(
+                        log_msg.format(
+                            i,
+                            len(iterable),
+                            eta=eta_string,
+                            meters=str(self),
+                            time=str(iter_time),
+                            data=str(data_time),
+                            memory=torch.cuda.max_memory_allocated() / MB,
+                        )
+                    )
+                else:
+                    print(
+                        log_msg.format(
+                            i,
+                            len(iterable),
+                            eta=eta_string,
+                            meters=str(self),
+                            time=str(iter_time),
+                            data=str(data_time),
+                        )
+                    )
+            i += 1
+            end = time.time()
+        total_time = time.time() - start_time
+        total_time_str = str(datetime.timedelta(seconds=int(total_time)))
+        print(
+            "{} Total time: {} ({:.4f} s / it)".format(
+                header, total_time_str, total_time / len(iterable)
+            )
+        )
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def setup_logger():
+    logging.basicConfig(
+        level=logging.INFO if dist_utils.is_main_process() else logging.WARN,
+        format="%(asctime)s [%(levelname)s] %(message)s",
+        handlers=[logging.StreamHandler()],
+    )

video_llama/common/optims.py

@@ -0,0 +1,119 @@
+"""
+ Copyright (c) 2022, 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 math
+
+from video_llama.common.registry import registry
+
+
+@registry.register_lr_scheduler("linear_warmup_step_lr")
+class LinearWarmupStepLRScheduler:
+    def __init__(
+        self,
+        optimizer,
+        max_epoch,
+        min_lr,
+        init_lr,
+        decay_rate=1,
+        warmup_start_lr=-1,
+        warmup_steps=0,
+        **kwargs
+    ):
+        self.optimizer = optimizer
+
+        self.max_epoch = max_epoch
+        self.min_lr = min_lr
+
+        self.decay_rate = decay_rate
+
+        self.init_lr = init_lr
+        self.warmup_steps = warmup_steps
+        self.warmup_start_lr = warmup_start_lr if warmup_start_lr >= 0 else init_lr
+
+    def step(self, cur_epoch, cur_step):
+        if cur_epoch == 0:
+            warmup_lr_schedule(
+                step=cur_step,
+                optimizer=self.optimizer,
+                max_step=self.warmup_steps,
+                init_lr=self.warmup_start_lr,
+                max_lr=self.init_lr,
+            )
+        else:
+            step_lr_schedule(
+                epoch=cur_epoch,
+                optimizer=self.optimizer,
+                init_lr=self.init_lr,
+                min_lr=self.min_lr,
+                decay_rate=self.decay_rate,
+            )
+
+
+@registry.register_lr_scheduler("linear_warmup_cosine_lr")
+class LinearWarmupCosineLRScheduler:
+    def __init__(
+        self,
+        optimizer,
+        max_epoch,
+        iters_per_epoch,
+        min_lr,
+        init_lr,
+        warmup_steps=0,
+        warmup_start_lr=-1,
+        **kwargs
+    ):
+        self.optimizer = optimizer
+
+        self.max_epoch = max_epoch
+        self.iters_per_epoch = iters_per_epoch
+        self.min_lr = min_lr
+
+        self.init_lr = init_lr
+        self.warmup_steps = warmup_steps
+        self.warmup_start_lr = warmup_start_lr if warmup_start_lr >= 0 else init_lr
+
+    def step(self, cur_epoch, cur_step):
+        total_cur_step = cur_epoch * self.iters_per_epoch + cur_step
+        if total_cur_step < self.warmup_steps:
+            warmup_lr_schedule(
+                step=cur_step,
+                optimizer=self.optimizer,
+                max_step=self.warmup_steps,
+                init_lr=self.warmup_start_lr,
+                max_lr=self.init_lr,
+            )
+        else:
+            cosine_lr_schedule(
+                epoch=total_cur_step,
+                optimizer=self.optimizer,
+                max_epoch=self.max_epoch * self.iters_per_epoch,
+                init_lr=self.init_lr,
+                min_lr=self.min_lr,
+            )
+
+
+def cosine_lr_schedule(optimizer, epoch, max_epoch, init_lr, min_lr):
+    """Decay the learning rate"""
+    lr = (init_lr - min_lr) * 0.5 * (
+        1.0 + math.cos(math.pi * epoch / max_epoch)
+    ) + min_lr
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = lr
+
+
+def warmup_lr_schedule(optimizer, step, max_step, init_lr, max_lr):
+    """Warmup the learning rate"""
+    lr = min(max_lr, init_lr + (max_lr - init_lr) * step / max(max_step, 1))
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = lr
+
+
+def step_lr_schedule(optimizer, epoch, init_lr, min_lr, decay_rate):
+    """Decay the learning rate"""
+    lr = max(min_lr, init_lr * (decay_rate**epoch))
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = lr

video_llama/common/registry.py

@@ -0,0 +1,329 @@
+"""
+ Copyright (c) 2022, 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
+"""
+
+
+class Registry:
+    mapping = {
+        "builder_name_mapping": {},
+        "task_name_mapping": {},
+        "processor_name_mapping": {},
+        "model_name_mapping": {},
+        "lr_scheduler_name_mapping": {},
+        "runner_name_mapping": {},
+        "state": {},
+        "paths": {},
+    }
+
+    @classmethod
+    def register_builder(cls, name):
+        r"""Register a dataset builder to registry with key 'name'
+
+        Args:
+            name: Key with which the builder will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+            from video_llama.datasets.base_dataset_builder import BaseDatasetBuilder
+        """
+
+        def wrap(builder_cls):
+            from video_llama.datasets.builders.base_dataset_builder import BaseDatasetBuilder
+
+            assert issubclass(
+                builder_cls, BaseDatasetBuilder
+            ), "All builders must inherit BaseDatasetBuilder class, found {}".format(
+                builder_cls
+            )
+            if name in cls.mapping["builder_name_mapping"]:
+                raise KeyError(
+                    "Name '{}' already registered for {}.".format(
+                        name, cls.mapping["builder_name_mapping"][name]
+                    )
+                )
+            cls.mapping["builder_name_mapping"][name] = builder_cls
+            return builder_cls
+
+        return wrap
+
+    @classmethod
+    def register_task(cls, name):
+        r"""Register a task to registry with key 'name'
+
+        Args:
+            name: Key with which the task will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+        """
+
+        def wrap(task_cls):
+            from video_llama.tasks.base_task import BaseTask
+
+            assert issubclass(
+                task_cls, BaseTask
+            ), "All tasks must inherit BaseTask class"
+            if name in cls.mapping["task_name_mapping"]:
+                raise KeyError(
+                    "Name '{}' already registered for {}.".format(
+                        name, cls.mapping["task_name_mapping"][name]
+                    )
+                )
+            cls.mapping["task_name_mapping"][name] = task_cls
+            return task_cls
+
+        return wrap
+
+    @classmethod
+    def register_model(cls, name):
+        r"""Register a task to registry with key 'name'
+
+        Args:
+            name: Key with which the task will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+        """
+
+        def wrap(model_cls):
+            from video_llama.models import BaseModel
+
+            assert issubclass(
+                model_cls, BaseModel
+            ), "All models must inherit BaseModel class"
+            if name in cls.mapping["model_name_mapping"]:
+                raise KeyError(
+                    "Name '{}' already registered for {}.".format(
+                        name, cls.mapping["model_name_mapping"][name]
+                    )
+                )
+            cls.mapping["model_name_mapping"][name] = model_cls
+            return model_cls
+
+        return wrap
+
+    @classmethod
+    def register_processor(cls, name):
+        r"""Register a processor to registry with key 'name'
+
+        Args:
+            name: Key with which the task will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+        """
+
+        def wrap(processor_cls):
+            from video_llama.processors import BaseProcessor
+
+            assert issubclass(
+                processor_cls, BaseProcessor
+            ), "All processors must inherit BaseProcessor class"
+            if name in cls.mapping["processor_name_mapping"]:
+                raise KeyError(
+                    "Name '{}' already registered for {}.".format(
+                        name, cls.mapping["processor_name_mapping"][name]
+                    )
+                )
+            cls.mapping["processor_name_mapping"][name] = processor_cls
+            return processor_cls
+
+        return wrap
+
+    @classmethod
+    def register_lr_scheduler(cls, name):
+        r"""Register a model to registry with key 'name'
+
+        Args:
+            name: Key with which the task will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+        """
+
+        def wrap(lr_sched_cls):
+            if name in cls.mapping["lr_scheduler_name_mapping"]:
+                raise KeyError(
+                    "Name '{}' already registered for {}.".format(
+                        name, cls.mapping["lr_scheduler_name_mapping"][name]
+                    )
+                )
+            cls.mapping["lr_scheduler_name_mapping"][name] = lr_sched_cls
+            return lr_sched_cls
+
+        return wrap
+
+    @classmethod
+    def register_runner(cls, name):
+        r"""Register a model to registry with key 'name'
+
+        Args:
+            name: Key with which the task will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+        """
+
+        def wrap(runner_cls):
+            if name in cls.mapping["runner_name_mapping"]:
+                raise KeyError(
+                    "Name '{}' already registered for {}.".format(
+                        name, cls.mapping["runner_name_mapping"][name]
+                    )
+                )
+            cls.mapping["runner_name_mapping"][name] = runner_cls
+            return runner_cls
+
+        return wrap
+
+    @classmethod
+    def register_path(cls, name, path):
+        r"""Register a path to registry with key 'name'
+
+        Args:
+            name: Key with which the path will be registered.
+
+        Usage:
+
+            from video_llama.common.registry import registry
+        """
+        assert isinstance(path, str), "All path must be str."
+        if name in cls.mapping["paths"]:
+            raise KeyError("Name '{}' already registered.".format(name))
+        cls.mapping["paths"][name] = path
+
+    @classmethod
+    def register(cls, name, obj):
+        r"""Register an item to registry with key 'name'
+
+        Args:
+            name: Key with which the item will be registered.
+
+        Usage::
+
+            from video_llama.common.registry import registry
+
+            registry.register("config", {})
+        """
+        path = name.split(".")
+        current = cls.mapping["state"]
+
+        for part in path[:-1]:
+            if part not in current:
+                current[part] = {}
+            current = current[part]
+
+        current[path[-1]] = obj
+
+    # @classmethod
+    # def get_trainer_class(cls, name):
+    #     return cls.mapping["trainer_name_mapping"].get(name, None)
+
+    @classmethod
+    def get_builder_class(cls, name):
+        return cls.mapping["builder_name_mapping"].get(name, None)
+
+    @classmethod
+    def get_model_class(cls, name):
+        return cls.mapping["model_name_mapping"].get(name, None)
+
+    @classmethod
+    def get_task_class(cls, name):
+        return cls.mapping["task_name_mapping"].get(name, None)
+
+    @classmethod
+    def get_processor_class(cls, name):
+        return cls.mapping["processor_name_mapping"].get(name, None)
+
+    @classmethod
+    def get_lr_scheduler_class(cls, name):
+        return cls.mapping["lr_scheduler_name_mapping"].get(name, None)
+
+    @classmethod
+    def get_runner_class(cls, name):
+        return cls.mapping["runner_name_mapping"].get(name, None)
+
+    @classmethod
+    def list_runners(cls):
+        return sorted(cls.mapping["runner_name_mapping"].keys())
+
+    @classmethod
+    def list_models(cls):
+        return sorted(cls.mapping["model_name_mapping"].keys())
+
+    @classmethod
+    def list_tasks(cls):
+        return sorted(cls.mapping["task_name_mapping"].keys())
+
+    @classmethod
+    def list_processors(cls):
+        return sorted(cls.mapping["processor_name_mapping"].keys())
+
+    @classmethod
+    def list_lr_schedulers(cls):
+        return sorted(cls.mapping["lr_scheduler_name_mapping"].keys())
+
+    @classmethod
+    def list_datasets(cls):
+        return sorted(cls.mapping["builder_name_mapping"].keys())
+
+    @classmethod
+    def get_path(cls, name):
+        return cls.mapping["paths"].get(name, None)
+
+    @classmethod
+    def get(cls, name, default=None, no_warning=False):
+        r"""Get an item from registry with key 'name'
+
+        Args:
+            name (string): Key whose value needs to be retrieved.
+            default: If passed and key is not in registry, default value will
+                     be returned with a warning. Default: None
+            no_warning (bool): If passed as True, warning when key doesn't exist
+                               will not be generated. Useful for MMF's
+                               internal operations. Default: False
+        """
+        original_name = name
+        name = name.split(".")
+        value = cls.mapping["state"]
+        for subname in name:
+            value = value.get(subname, default)
+            if value is default:
+                break
+
+        if (
+            "writer" in cls.mapping["state"]
+            and value == default
+            and no_warning is False
+        ):
+            cls.mapping["state"]["writer"].warning(
+                "Key {} is not present in registry, returning default value "
+                "of {}".format(original_name, default)
+            )
+        return value
+
+    @classmethod
+    def unregister(cls, name):
+        r"""Remove an item from registry with key 'name'
+
+        Args:
+            name: Key which needs to be removed.
+        Usage::
+
+            from mmf.common.registry import registry
+
+            config = registry.unregister("config")
+        """
+        return cls.mapping["state"].pop(name, None)
+
+
+registry = Registry()

video_llama/common/utils.py

@@ -0,0 +1,424 @@
+"""
+ Copyright (c) 2022, 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 io
+import json
+import logging
+import os
+import pickle
+import re
+import shutil
+import urllib
+import urllib.error
+import urllib.request
+from typing import Optional
+from urllib.parse import urlparse
+
+import numpy as np
+import pandas as pd
+import yaml
+from iopath.common.download import download
+from iopath.common.file_io import file_lock, g_pathmgr
+from video_llama.common.registry import registry
+from torch.utils.model_zoo import tqdm
+from torchvision.datasets.utils import (
+    check_integrity,
+    download_file_from_google_drive,
+    extract_archive,
+)
+
+
+def now():
+    from datetime import datetime
+
+    return datetime.now().strftime("%Y%m%d%H%M")[:-1]
+
+
+def is_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ("http", "https")
+
+
+def get_cache_path(rel_path):
+    return os.path.expanduser(os.path.join(registry.get_path("cache_root"), rel_path))
+
+
+def get_abs_path(rel_path):
+    return os.path.join(registry.get_path("library_root"), rel_path)
+
+
+def load_json(filename):
+    with open(filename, "r") as f:
+        return json.load(f)
+
+
+# The following are adapted from torchvision and vissl
+# torchvision: https://github.com/pytorch/vision
+# vissl: https://github.com/facebookresearch/vissl/blob/main/vissl/utils/download.py
+
+
+def makedir(dir_path):
+    """
+    Create the directory if it does not exist.
+    """
+    is_success = False
+    try:
+        if not g_pathmgr.exists(dir_path):
+            g_pathmgr.mkdirs(dir_path)
+        is_success = True
+    except BaseException:
+        print(f"Error creating directory: {dir_path}")
+    return is_success
+
+
+def get_redirected_url(url: str):
+    """
+    Given a URL, returns the URL it redirects to or the
+    original URL in case of no indirection
+    """
+    import requests
+
+    with requests.Session() as session:
+        with session.get(url, stream=True, allow_redirects=True) as response:
+            if response.history:
+                return response.url
+            else:
+                return url
+
+
+def to_google_drive_download_url(view_url: str) -> str:
+    """
+    Utility function to transform a view URL of google drive
+    to a download URL for google drive
+    Example input:
+        https://drive.google.com/file/d/137RyRjvTBkBiIfeYBNZBtViDHQ6_Ewsp/view
+    Example output:
+        https://drive.google.com/uc?export=download&id=137RyRjvTBkBiIfeYBNZBtViDHQ6_Ewsp
+    """
+    splits = view_url.split("/")
+    assert splits[-1] == "view"
+    file_id = splits[-2]
+    return f"https://drive.google.com/uc?export=download&id={file_id}"
+
+
+def download_google_drive_url(url: str, output_path: str, output_file_name: str):
+    """
+    Download a file from google drive
+    Downloading an URL from google drive requires confirmation when
+    the file of the size is too big (google drive notifies that
+    anti-viral checks cannot be performed on such files)
+    """
+    import requests
+
+    with requests.Session() as session:
+
+        # First get the confirmation token and append it to the URL
+        with session.get(url, stream=True, allow_redirects=True) as response:
+            for k, v in response.cookies.items():
+                if k.startswith("download_warning"):
+                    url = url + "&confirm=" + v
+
+        # Then download the content of the file
+        with session.get(url, stream=True, verify=True) as response:
+            makedir(output_path)
+            path = os.path.join(output_path, output_file_name)
+            total_size = int(response.headers.get("Content-length", 0))
+            with open(path, "wb") as file:
+                from tqdm import tqdm
+
+                with tqdm(total=total_size) as progress_bar:
+                    for block in response.iter_content(
+                        chunk_size=io.DEFAULT_BUFFER_SIZE
+                    ):
+                        file.write(block)
+                        progress_bar.update(len(block))
+
+
+def _get_google_drive_file_id(url: str) -> Optional[str]:
+    parts = urlparse(url)
+
+    if re.match(r"(drive|docs)[.]google[.]com", parts.netloc) is None:
+        return None
+
+    match = re.match(r"/file/d/(?P<id>[^/]*)", parts.path)
+    if match is None:
+        return None
+
+    return match.group("id")
+
+
+def _urlretrieve(url: str, filename: str, chunk_size: int = 1024) -> None:
+    with open(filename, "wb") as fh:
+        with urllib.request.urlopen(
+            urllib.request.Request(url, headers={"User-Agent": "vissl"})
+        ) as response:
+            with tqdm(total=response.length) as pbar:
+                for chunk in iter(lambda: response.read(chunk_size), ""):
+                    if not chunk:
+                        break
+                    pbar.update(chunk_size)
+                    fh.write(chunk)
+
+
+def download_url(
+    url: str,
+    root: str,
+    filename: Optional[str] = None,
+    md5: Optional[str] = None,
+) -> None:
+    """Download a file from a url and place it in root.
+    Args:
+        url (str): URL to download file from
+        root (str): Directory to place downloaded file in
+        filename (str, optional): Name to save the file under.
+                                  If None, use the basename of the URL.
+        md5 (str, optional): MD5 checksum of the download. If None, do not check
+    """
+    root = os.path.expanduser(root)
+    if not filename:
+        filename = os.path.basename(url)
+    fpath = os.path.join(root, filename)
+
+    makedir(root)
+
+    # check if file is already present locally
+    if check_integrity(fpath, md5):
+        print("Using downloaded and verified file: " + fpath)
+        return
+
+    # expand redirect chain if needed
+    url = get_redirected_url(url)
+
+    # check if file is located on Google Drive
+    file_id = _get_google_drive_file_id(url)
+    if file_id is not None:
+        return download_file_from_google_drive(file_id, root, filename, md5)
+
+    # download the file
+    try:
+        print("Downloading " + url + " to " + fpath)
+        _urlretrieve(url, fpath)
+    except (urllib.error.URLError, IOError) as e:  # type: ignore[attr-defined]
+        if url[:5] == "https":
+            url = url.replace("https:", "http:")
+            print(
+                "Failed download. Trying https -> http instead."
+                " Downloading " + url + " to " + fpath
+            )
+            _urlretrieve(url, fpath)
+        else:
+            raise e
+
+    # check integrity of downloaded file
+    if not check_integrity(fpath, md5):
+        raise RuntimeError("File not found or corrupted.")
+
+
+def download_and_extract_archive(
+    url: str,
+    download_root: str,
+    extract_root: Optional[str] = None,
+    filename: Optional[str] = None,
+    md5: Optional[str] = None,
+    remove_finished: bool = False,
+) -> None:
+    download_root = os.path.expanduser(download_root)
+    if extract_root is None:
+        extract_root = download_root
+    if not filename:
+        filename = os.path.basename(url)
+
+    download_url(url, download_root, filename, md5)
+
+    archive = os.path.join(download_root, filename)
+    print("Extracting {} to {}".format(archive, extract_root))
+    extract_archive(archive, extract_root, remove_finished)
+
+
+def cache_url(url: str, cache_dir: str) -> str:
+    """
+    This implementation downloads the remote resource and caches it locally.
+    The resource will only be downloaded if not previously requested.
+    """
+    parsed_url = urlparse(url)
+    dirname = os.path.join(cache_dir, os.path.dirname(parsed_url.path.lstrip("/")))
+    makedir(dirname)
+    filename = url.split("/")[-1]
+    cached = os.path.join(dirname, filename)
+    with file_lock(cached):
+        if not os.path.isfile(cached):
+            logging.info(f"Downloading {url} to {cached} ...")
+            cached = download(url, dirname, filename=filename)
+    logging.info(f"URL {url} cached in {cached}")
+    return cached
+
+
+# TODO (prigoyal): convert this into RAII-style API
+def create_file_symlink(file1, file2):
+    """
+    Simply create the symlinks for a given file1 to file2.
+    Useful during model checkpointing to symlinks to the
+    latest successful checkpoint.
+    """
+    try:
+        if g_pathmgr.exists(file2):
+            g_pathmgr.rm(file2)
+        g_pathmgr.symlink(file1, file2)
+    except Exception as e:
+        logging.info(f"Could NOT create symlink. Error: {e}")
+
+
+def save_file(data, filename, append_to_json=True, verbose=True):
+    """
+    Common i/o utility to handle saving data to various file formats.
+    Supported:
+        .pkl, .pickle, .npy, .json
+    Specifically for .json, users have the option to either append (default)
+    or rewrite by passing in Boolean value to append_to_json.
+    """
+    if verbose:
+        logging.info(f"Saving data to file: {filename}")
+    file_ext = os.path.splitext(filename)[1]
+    if file_ext in [".pkl", ".pickle"]:
+        with g_pathmgr.open(filename, "wb") as fopen:
+            pickle.dump(data, fopen, pickle.HIGHEST_PROTOCOL)
+    elif file_ext == ".npy":
+        with g_pathmgr.open(filename, "wb") as fopen:
+            np.save(fopen, data)
+    elif file_ext == ".json":
+        if append_to_json:
+            with g_pathmgr.open(filename, "a") as fopen:
+                fopen.write(json.dumps(data, sort_keys=True) + "\n")
+                fopen.flush()
+        else:
+            with g_pathmgr.open(filename, "w") as fopen:
+                fopen.write(json.dumps(data, sort_keys=True) + "\n")
+                fopen.flush()
+    elif file_ext == ".yaml":
+        with g_pathmgr.open(filename, "w") as fopen:
+            dump = yaml.dump(data)
+            fopen.write(dump)
+            fopen.flush()
+    else:
+        raise Exception(f"Saving {file_ext} is not supported yet")
+
+    if verbose:
+        logging.info(f"Saved data to file: {filename}")
+
+
+def load_file(filename, mmap_mode=None, verbose=True, allow_pickle=False):
+    """
+    Common i/o utility to handle loading data from various file formats.
+    Supported:
+        .pkl, .pickle, .npy, .json
+    For the npy files, we support reading the files in mmap_mode.
+    If the mmap_mode of reading is not successful, we load data without the
+    mmap_mode.
+    """
+    if verbose:
+        logging.info(f"Loading data from file: {filename}")
+
+    file_ext = os.path.splitext(filename)[1]
+    if file_ext == ".txt":
+        with g_pathmgr.open(filename, "r") as fopen:
+            data = fopen.readlines()
+    elif file_ext in [".pkl", ".pickle"]:
+        with g_pathmgr.open(filename, "rb") as fopen:
+            data = pickle.load(fopen, encoding="latin1")
+    elif file_ext == ".npy":
+        if mmap_mode:
+            try:
+                with g_pathmgr.open(filename, "rb") as fopen:
+                    data = np.load(
+                        fopen,
+                        allow_pickle=allow_pickle,
+                        encoding="latin1",
+                        mmap_mode=mmap_mode,
+                    )
+            except ValueError as e:
+                logging.info(
+                    f"Could not mmap {filename}: {e}. Trying without g_pathmgr"
+                )
+                data = np.load(
+                    filename,
+                    allow_pickle=allow_pickle,
+                    encoding="latin1",
+                    mmap_mode=mmap_mode,
+                )
+                logging.info("Successfully loaded without g_pathmgr")
+            except Exception:
+                logging.info("Could not mmap without g_pathmgr. Trying without mmap")
+                with g_pathmgr.open(filename, "rb") as fopen:
+                    data = np.load(fopen, allow_pickle=allow_pickle, encoding="latin1")
+        else:
+            with g_pathmgr.open(filename, "rb") as fopen:
+                data = np.load(fopen, allow_pickle=allow_pickle, encoding="latin1")
+    elif file_ext == ".json":
+        with g_pathmgr.open(filename, "r") as fopen:
+            data = json.load(fopen)
+    elif file_ext == ".yaml":
+        with g_pathmgr.open(filename, "r") as fopen:
+            data = yaml.load(fopen, Loader=yaml.FullLoader)
+    elif file_ext == ".csv":
+        with g_pathmgr.open(filename, "r") as fopen:
+            data = pd.read_csv(fopen)
+    else:
+        raise Exception(f"Reading from {file_ext} is not supported yet")
+    return data
+
+
+def abspath(resource_path: str):
+    """
+    Make a path absolute, but take into account prefixes like
+    "http://" or "manifold://"
+    """
+    regex = re.compile(r"^\w+://")
+    if regex.match(resource_path) is None:
+        return os.path.abspath(resource_path)
+    else:
+        return resource_path
+
+
+def makedir(dir_path):
+    """
+    Create the directory if it does not exist.
+    """
+    is_success = False
+    try:
+        if not g_pathmgr.exists(dir_path):
+            g_pathmgr.mkdirs(dir_path)
+        is_success = True
+    except BaseException:
+        logging.info(f"Error creating directory: {dir_path}")
+    return is_success
+
+
+def is_url(input_url):
+    """
+    Check if an input string is a url. look for http(s):// and ignoring the case
+    """
+    is_url = re.match(r"^(?:http)s?://", input_url, re.IGNORECASE) is not None
+    return is_url
+
+
+def cleanup_dir(dir):
+    """
+    Utility for deleting a directory. Useful for cleaning the storage space
+    that contains various training artifacts like checkpoints, data etc.
+    """
+    if os.path.exists(dir):
+        logging.info(f"Deleting directory: {dir}")
+        shutil.rmtree(dir)
+    logging.info(f"Deleted contents of directory: {dir}")
+
+
+def get_file_size(filename):
+    """
+    Given a file, get the size of file in MB
+    """
+    size_in_mb = os.path.getsize(filename) / float(1024**2)
+    return size_in_mb

video_llama/configs/datasets/cc_sbu/align.yaml

@@ -0,0 +1,5 @@
+datasets:
+  cc_sbu_align:
+    data_type: images
+    build_info:
+      storage: /path/to/cc_sbu_align_dataset

video_llama/configs/datasets/cc_sbu/defaults.yaml

@@ -0,0 +1,5 @@
+datasets:
+  cc_sbu:
+    data_type: images
+    build_info:
+      storage: /path/to/cc_sbu_dataset/{00000..00001}.tar

video_llama/configs/datasets/instruct/llava_instruct.yaml

@@ -0,0 +1,6 @@
+datasets:
+  llava_instruct:
+    data_type: image
+    build_info:
+      anno_dir: /path/llava_instruct_150k.json
+      videos_dir: /path/train2014/train2014/

video_llama/configs/datasets/instruct/webvid_instruct.yaml

@@ -0,0 +1,6 @@
+datasets:
+  webvid_instruct:
+    data_type: image
+    build_info:
+      anno_dir: /path/webvid_align/videochat_instruct_11k.json
+      videos_dir: /path/webvid_align/videos/

video_llama/configs/datasets/laion/defaults.yaml

@@ -0,0 +1,5 @@
+datasets:
+  laion:
+    data_type: images
+    build_info:
+      storage: path/laion/laion_dataset/{00000..00001}.tar

video_llama/configs/datasets/webvid/defaults.yaml

@@ -0,0 +1,6 @@
+datasets:
+  webvid:
+    data_type: video
+    build_info:
+      anno_dir: path/webvid/webvid_tain_data/annotations/
+      videos_dir: path//webvid/webvid_tain_data/videos/

video_llama/configs/default.yaml

@@ -0,0 +1,5 @@
+env:
+  # For default users
+  # cache_root: "cache"
+  # For internal use with persistent storage
+  cache_root: "/export/home/.cache/minigpt4"

video_llama/configs/models/minigpt4.yaml

@@ -0,0 +1,33 @@
+model:
+  arch: mini_gpt4
+
+  # vit encoder
+  image_size: 224
+  drop_path_rate: 0
+  use_grad_checkpoint: False
+  vit_precision: "fp16"
+  freeze_vit: True
+  freeze_qformer: True
+
+  # Q-Former
+  num_query_token: 32
+
+  # Vicuna
+  llama_model: "ckpt/vicuna-13b/"
+
+  # generation configs
+  prompt: ""
+
+preprocess:
+    vis_processor:
+        train:
+          name: "blip2_image_train"
+          image_size: 224
+        eval:
+          name: "blip2_image_eval"
+          image_size: 224
+    text_processor:
+        train:
+          name: "blip_caption"
+        eval:
+          name: "blip_caption"

video_llama/configs/models/video_llama.yaml

@@ -0,0 +1,36 @@
+model:
+  arch: video_llama
+
+  # vit encoder
+  image_size: 224
+  drop_path_rate: 0
+  use_grad_checkpoint: False
+  vit_precision: "fp16"
+  freeze_vit: True
+  freeze_qformer: True
+
+  # Q-Former
+  num_query_token: 32
+
+  # Vicuna
+  llama_model: "ckpt/vicuna-7b/"
+
+  # generation configs
+  prompt: ""
+
+preprocess:
+    vis_processor:
+        train:
+          name: "alpro_video_train"
+          image_size: 224
+          n_frms: 8
+        eval:
+          name: "alpro_video_eval"
+          image_size: 224
+          n_frms: 8
+    text_processor:
+        train:
+          name: "blip_caption"
+        eval:
+          name: "blip_caption"
+    

video_llama/conversation/conversation_video.py

@@ -0,0 +1,248 @@
+"""
+Conversation prompt template of Video-LLaMA.
+Adapted from: https://github.com/Vision-CAIR/MiniGPT-4/blob/main/minigpt4/conversation/conversation.py 
+"""
+import argparse
+import time
+from PIL import Image
+
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaTokenizer
+from transformers import StoppingCriteria, StoppingCriteriaList
+
+import dataclasses
+from enum import auto, Enum
+from typing import List, Tuple, Any
+import os
+from video_llama.common.registry import registry
+from video_llama.processors.video_processor import ToTHWC,ToUint8,load_video
+from video_llama.processors import Blip2ImageEvalProcessor
+class SeparatorStyle(Enum):
+    """Different separator style."""
+    SINGLE = auto()
+    TWO = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that keeps all conversation history."""
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    # system_img: List[Image.Image] = []
+    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
+    sep: str = "###"
+    sep2: str = None
+
+    skip_next: bool = False
+    conv_id: Any = None
+
+    def get_prompt(self):
+        if self.sep_style == SeparatorStyle.SINGLE:
+            ret = self.system + self.sep
+            for role, message in self.messages:
+                if message:
+                    ret += role + ": " + message + self.sep
+                else:
+                    ret += role + ":"
+            return ret
+        elif self.sep_style == SeparatorStyle.TWO:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+            return ret
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+    def append_message(self, role, message):
+        self.messages.append([role, message])
+
+    def to_gradio_chatbot(self):
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def copy(self):
+        return Conversation(
+            system=self.system,
+            # system_img=self.system_img,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            conv_id=self.conv_id)
+
+    def dict(self):
+        return {
+            "system": self.system,
+            # "system_img": self.system_img,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+            "sep": self.sep,
+            "sep2": self.sep2,
+            "conv_id": self.conv_id,
+        }
+
+
+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
+
+
+CONV_VISION = Conversation(
+    system="Give the following image: <Img>ImageContent</Img>. "
+           "You will be able to see the image once I provide it to you. Please answer my questions.",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+default_conversation = Conversation(
+    system="",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+
+class Chat:
+    def __init__(self, model, vis_processor, device='cuda:0'):
+        self.device = device
+        self.model = model
+        self.vis_processor = vis_processor
+        self.image_vis_processor = Blip2ImageEvalProcessor()
+        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):
+        if len(conv.messages) > 0 and conv.messages[-1][0] == conv.roles[0] \
+                and ('</Video>' in conv.messages[-1][1] or '</Image>' in conv.messages[-1][1]):  # last message is image.
+            conv.messages[-1][1] = ' '.join([conv.messages[-1][1], text])
+        else:
+            conv.append_message(conv.roles[0], text)
+
+    def answer(self, conv, img_list, max_new_tokens=300, num_beams=1, min_length=1, top_p=0.9,
+               repetition_penalty=1.0, length_penalty=1, temperature=1.0, max_length=2000):
+        conv.append_message(conv.roles[1], None)
+        embs = self.get_context_emb(conv, img_list)
+
+        current_max_len = embs.shape[1] + max_new_tokens
+        if current_max_len - max_length > 0:
+            print('Warning: The number of tokens in current conversation exceeds the max length. '
+                  'The model will not see the contexts outside the range.')
+        begin_idx = max(0, current_max_len - max_length)
+
+        embs = embs[:, begin_idx:]
+        
+        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()
+    
+    def upload_video(self, video, conv, img_list):
+
+        msg = ""
+        if isinstance(video, str):  # is a video path
+            ext = os.path.splitext(video)[-1].lower()
+            print(video)
+            # image = self.vis_processor(image).unsqueeze(0).to(self.device)
+            video, msg = load_video(
+                video_path=video,
+                n_frms=8,
+                height=224,
+                width=224,
+                sampling ="uniform", return_msg = True
+            )
+            video = self.vis_processor.transform(video)
+            video = video.unsqueeze(0).to(self.device)
+            # print(image)
+        else:
+            raise NotImplementedError
+
+        image_emb, _ = self.model.encode_img(video)
+        img_list.append(image_emb)
+        conv.append_message(conv.roles[0], "<Video><ImageHere></Video> "+ msg)
+        return "Received."
+
+    def upload_img(self, image, conv, img_list):
+
+        msg = ""
+        if isinstance(image, str):  # is a image path
+            raw_image = Image.open(image).convert('RGB') # 增加一个时间维度
+            image = self.image_vis_processor(raw_image).unsqueeze(0).unsqueeze(2).to(self.device)
+        elif isinstance(image, Image.Image):
+            raw_image = image
+            image = self.image_vis_processor(raw_image).unsqueeze(0).unsqueeze(2).to(self.device)
+        elif isinstance(image, torch.Tensor):
+            if len(image.shape) == 3:
+                image = image.unsqueeze(0)
+            image = image.to(self.device)
+        else:
+            raise NotImplementedError
+
+        image_emb, _ = self.model.encode_img(image)
+        img_list.append(image_emb)
+        # Todo msg=""
+        conv.append_message(conv.roles[0], "<Image><ImageHere></Image> "+ msg)
+
+        return "Received."
+
+    def get_context_emb(self, conv, img_list):
+        prompt = conv.get_prompt()
+        prompt_segs = prompt.split('<ImageHere>')
+        assert len(prompt_segs) == len(img_list) + 1, "Unmatched numbers of image placeholders and images."
+        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
+
+

video_llama/datasets/builders/__init__.py

@@ -0,0 +1,77 @@
+"""
+ Copyright (c) 2022, 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
+"""
+
+from video_llama.datasets.builders.base_dataset_builder import load_dataset_config
+from video_llama.datasets.builders.image_text_pair_builder import (
+    CCSBUBuilder,
+    LaionBuilder,
+    CCSBUAlignBuilder
+)
+from video_llama.datasets.builders.video_caption_builder import WebvidBuilder
+from video_llama.common.registry import registry
+from video_llama.datasets.builders.instruct_builder import WebvidInstruct_Builder,LlavaInstruct_Builder
+__all__ = [
+    "CCSBUBuilder",
+    "LaionBuilder",
+    "CCSBUAlignBuilder",
+    "WebvidBuilder",
+    "LlavaInstruct_Builder",
+    "WebvidInstruct_Builder"
+    
+]
+
+
+def load_dataset(name, cfg_path=None, vis_path=None, data_type=None):
+    """
+    Example
+
+    >>> dataset = load_dataset("coco_caption", cfg=None)
+    >>> splits = dataset.keys()
+    >>> print([len(dataset[split]) for split in splits])
+
+    """
+    if cfg_path is None:
+        cfg = None
+    else:
+        cfg = load_dataset_config(cfg_path)
+
+    try:
+        builder = registry.get_builder_class(name)(cfg)
+    except TypeError:
+        print(
+            f"Dataset {name} not found. Available datasets:\n"
+            + ", ".join([str(k) for k in dataset_zoo.get_names()])
+        )
+        exit(1)
+
+    if vis_path is not None:
+        if data_type is None:
+            # use default data type in the config
+            data_type = builder.config.data_type
+
+        assert (
+            data_type in builder.config.build_info
+        ), f"Invalid data_type {data_type} for {name}."
+
+        builder.config.build_info.get(data_type).storage = vis_path
+
+    dataset = builder.build_datasets()
+    return dataset
+
+
+class DatasetZoo:
+    def __init__(self) -> None:
+        self.dataset_zoo = {
+            k: list(v.DATASET_CONFIG_DICT.keys())
+            for k, v in sorted(registry.mapping["builder_name_mapping"].items())
+        }
+
+    def get_names(self):
+        return list(self.dataset_zoo.keys())
+
+
+dataset_zoo = DatasetZoo()

video_llama/datasets/builders/base_dataset_builder.py

@@ -0,0 +1,236 @@
+"""
+ This file is from
+ Copyright (c) 2022, 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 logging
+import os
+import shutil
+import warnings
+
+from omegaconf import OmegaConf
+import torch.distributed as dist
+from torchvision.datasets.utils import download_url
+
+import video_llama.common.utils as utils
+from video_llama.common.dist_utils import is_dist_avail_and_initialized, is_main_process
+from video_llama.common.registry import registry
+from video_llama.processors.base_processor import BaseProcessor
+
+
+
+class BaseDatasetBuilder:
+    train_dataset_cls, eval_dataset_cls = None, None
+
+    def __init__(self, cfg=None):
+        super().__init__()
+
+        if cfg is None:
+            # help to create datasets from default config.
+            self.config = load_dataset_config(self.default_config_path())
+        elif isinstance(cfg, str):
+            self.config = load_dataset_config(cfg)
+        else:
+            # when called from task.build_dataset()
+            self.config = cfg
+
+        self.data_type = self.config.data_type
+
+        self.vis_processors = {"train": BaseProcessor(), "eval": BaseProcessor()}
+        self.text_processors = {"train": BaseProcessor(), "eval": BaseProcessor()}
+
+    def build_datasets(self):
+        # download, split, etc...
+        # only called on 1 GPU/TPU in distributed
+
+        if is_main_process():
+            self._download_data()
+
+        if is_dist_avail_and_initialized():
+            dist.barrier()
+
+        # at this point, all the annotations and image/videos should be all downloaded to the specified locations.
+        logging.info("Building datasets...")
+        datasets = self.build()  # dataset['train'/'val'/'test']
+
+        return datasets
+
+    def build_processors(self):
+        vis_proc_cfg = self.config.get("vis_processor")
+        txt_proc_cfg = self.config.get("text_processor")
+
+        if vis_proc_cfg is not None:
+            vis_train_cfg = vis_proc_cfg.get("train")
+            vis_eval_cfg = vis_proc_cfg.get("eval")
+
+            self.vis_processors["train"] = self._build_proc_from_cfg(vis_train_cfg)
+            self.vis_processors["eval"] = self._build_proc_from_cfg(vis_eval_cfg)
+
+        if txt_proc_cfg is not None:
+            txt_train_cfg = txt_proc_cfg.get("train")
+            txt_eval_cfg = txt_proc_cfg.get("eval")
+
+            self.text_processors["train"] = self._build_proc_from_cfg(txt_train_cfg)
+            self.text_processors["eval"] = self._build_proc_from_cfg(txt_eval_cfg)
+
+    @staticmethod
+    def _build_proc_from_cfg(cfg):
+        return (
+            registry.get_processor_class(cfg.name).from_config(cfg)
+            if cfg is not None
+            else None
+        )
+
+    @classmethod
+    def default_config_path(cls, type="default"):
+        return utils.get_abs_path(cls.DATASET_CONFIG_DICT[type])
+
+    def _download_data(self):
+        self._download_ann()
+        self._download_vis()
+
+    def _download_ann(self):
+        """
+        Download annotation files if necessary.
+        All the vision-language datasets should have annotations of unified format.
+
+        storage_path can be:
+          (1) relative/absolute: will be prefixed with env.cache_root to make full path if relative.
+          (2) basename/dirname: will be suffixed with base name of URL if dirname is provided.
+
+        Local annotation paths should be relative.
+        """
+        anns = self.config.build_info.annotations
+
+        splits = anns.keys()
+
+        cache_root = registry.get_path("cache_root")
+
+        for split in splits:
+            info = anns[split]
+
+            urls, storage_paths = info.get("url", None), info.storage
+
+            if isinstance(urls, str):
+                urls = [urls]
+            if isinstance(storage_paths, str):
+                storage_paths = [storage_paths]
+
+            assert len(urls) == len(storage_paths)
+
+            for url_or_filename, storage_path in zip(urls, storage_paths):
+                # if storage_path is relative, make it full by prefixing with cache_root.
+                if not os.path.isabs(storage_path):
+                    storage_path = os.path.join(cache_root, storage_path)
+
+                dirname = os.path.dirname(storage_path)
+                if not os.path.exists(dirname):
+                    os.makedirs(dirname)
+
+                if os.path.isfile(url_or_filename):
+                    src, dst = url_or_filename, storage_path
+                    if not os.path.exists(dst):
+                        shutil.copyfile(src=src, dst=dst)
+                    else:
+                        logging.info("Using existing file {}.".format(dst))
+                else:
+                    if os.path.isdir(storage_path):
+                        # if only dirname is provided, suffix with basename of URL.
+                        raise ValueError(
+                            "Expecting storage_path to be a file path, got directory {}".format(
+                                storage_path
+                            )
+                        )
+                    else:
+                        filename = os.path.basename(storage_path)
+
+                    download_url(url=url_or_filename, root=dirname, filename=filename)
+
+    def _download_vis(self):
+
+        storage_path = self.config.build_info.get(self.data_type).storage
+        storage_path = utils.get_cache_path(storage_path)
+
+        if not os.path.exists(storage_path):
+            warnings.warn(
+                f"""
+                The specified path {storage_path} for visual inputs does not exist.
+                Please provide a correct path to the visual inputs or
+                refer to datasets/download_scripts/README.md for downloading instructions.
+                """
+            )
+
+    def build(self):
+        """
+        Create by split datasets inheriting torch.utils.data.Datasets.
+
+        # build() can be dataset-specific. Overwrite to customize.
+        """
+        self.build_processors()
+
+        build_info = self.config.build_info
+
+        ann_info = build_info.annotations
+        vis_info = build_info.get(self.data_type)
+
+        datasets = dict()
+        for split in ann_info.keys():
+            if split not in ["train", "val", "test"]:
+                continue
+
+            is_train = split == "train"
+
+            # processors
+            vis_processor = (
+                self.vis_processors["train"]
+                if is_train
+                else self.vis_processors["eval"]
+            )
+            text_processor = (
+                self.text_processors["train"]
+                if is_train
+                else self.text_processors["eval"]
+            )
+
+            # annotation path
+            ann_paths = ann_info.get(split).storage
+            if isinstance(ann_paths, str):
+                ann_paths = [ann_paths]
+
+            abs_ann_paths = []
+            for ann_path in ann_paths:
+                if not os.path.isabs(ann_path):
+                    ann_path = utils.get_cache_path(ann_path)
+                abs_ann_paths.append(ann_path)
+            ann_paths = abs_ann_paths
+
+            # visual data storage path
+            vis_path = os.path.join(vis_info.storage, split)
+
+            if not os.path.isabs(vis_path):
+                # vis_path = os.path.join(utils.get_cache_path(), vis_path)
+                vis_path = utils.get_cache_path(vis_path)
+
+            if not os.path.exists(vis_path):
+                warnings.warn("storage path {} does not exist.".format(vis_path))
+
+            # create datasets
+            dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls
+            datasets[split] = dataset_cls(
+                vis_processor=vis_processor,
+                text_processor=text_processor,
+                ann_paths=ann_paths,
+                vis_root=vis_path,
+            )
+
+        return datasets
+
+
+def load_dataset_config(cfg_path):
+    cfg = OmegaConf.load(cfg_path).datasets
+    cfg = cfg[list(cfg.keys())[0]]
+
+    return cfg

video_llama/datasets/builders/image_text_pair_builder.py

@@ -0,0 +1,106 @@
+import os
+import logging
+import warnings
+
+from video_llama.common.registry import registry
+from video_llama.datasets.builders.base_dataset_builder import BaseDatasetBuilder
+from video_llama.datasets.datasets.laion_dataset import LaionDataset
+from video_llama.datasets.datasets.cc_sbu_dataset import CCSBUDataset, CCSBUAlignDataset
+
+
+@registry.register_builder("cc_sbu")
+class CCSBUBuilder(BaseDatasetBuilder):
+    train_dataset_cls = CCSBUDataset
+
+    DATASET_CONFIG_DICT = {"default": "configs/datasets/cc_sbu/defaults.yaml"}
+
+    def _download_ann(self):
+        pass
+
+    def _download_vis(self):
+        pass
+
+    def build(self):
+        self.build_processors()
+
+        build_info = self.config.build_info
+
+        datasets = dict()
+        split = "train"
+
+        # create datasets
+        # [NOTE] return inner_datasets (wds.DataPipeline)
+        dataset_cls = self.train_dataset_cls
+        datasets[split] = dataset_cls(
+            vis_processor=self.vis_processors[split],
+            text_processor=self.text_processors[split],
+            location=build_info.storage,
+        ).inner_dataset
+
+        return datasets
+
+
+@registry.register_builder("laion")
+class LaionBuilder(BaseDatasetBuilder):
+    train_dataset_cls = LaionDataset
+
+    DATASET_CONFIG_DICT = {"default": "configs/datasets/laion/defaults.yaml"}
+
+    def _download_ann(self):
+        pass
+
+    def _download_vis(self):
+        pass
+
+    def build(self):
+        self.build_processors()
+
+        build_info = self.config.build_info
+
+        datasets = dict()
+        split = "train"
+
+        # create datasets
+        # [NOTE] return inner_datasets (wds.DataPipeline)
+        dataset_cls = self.train_dataset_cls
+        datasets[split] = dataset_cls(
+            vis_processor=self.vis_processors[split],
+            text_processor=self.text_processors[split],
+            location=build_info.storage,
+        ).inner_dataset
+
+        return datasets
+
+
+@registry.register_builder("cc_sbu_align")
+class CCSBUAlignBuilder(BaseDatasetBuilder):
+    train_dataset_cls = CCSBUAlignDataset
+
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/cc_sbu/align.yaml",
+    }
+
+    def build_datasets(self):
+        # at this point, all the annotations and image/videos should be all downloaded to the specified locations.
+        logging.info("Building datasets...")
+        self.build_processors()
+
+        build_info = self.config.build_info
+        storage_path = build_info.storage
+
+        datasets = dict()
+
+        if not os.path.exists(storage_path):
+            warnings.warn("storage path {} does not exist.".format(storage_path))
+
+        # create datasets
+        dataset_cls = self.train_dataset_cls
+        datasets['train'] = dataset_cls(
+            vis_processor=self.vis_processors["train"],
+            text_processor=self.text_processors["train"],
+            ann_paths=[os.path.join(storage_path, 'filter_cap.json')],
+            vis_root=os.path.join(storage_path, 'image'),
+        )
+
+        return datasets
+

video_llama/datasets/builders/instruct_builder.py

@@ -0,0 +1,78 @@
+import os
+import logging
+import warnings
+
+from video_llama.common.registry import registry
+from video_llama.datasets.builders.base_dataset_builder import BaseDatasetBuilder
+from video_llama.datasets.datasets.laion_dataset import LaionDataset
+from video_llama.datasets.datasets.llava_instruct_dataset import Instruct_Dataset
+from video_llama.datasets.datasets.video_instruct_dataset import Video_Instruct_Dataset
+
+@registry.register_builder("instruct")
+class Instruct_Builder(BaseDatasetBuilder):
+    train_dataset_cls = Instruct_Dataset
+
+    DATASET_CONFIG_DICT = {"default": "configs/datasets/instruct/defaults.yaml"}
+
+    def _download_ann(self):
+        pass
+
+    def _download_vis(self):
+        pass
+
+    def build(self):
+        self.build_processors()
+        datasets = dict()
+        split = "train"
+
+        build_info = self.config.build_info
+        dataset_cls = self.train_dataset_cls
+        if self.config.num_video_query_token:
+            num_video_query_token = self.config.num_video_query_token 
+        else:
+            num_video_query_token = 32
+
+        if self.config.tokenizer_name:
+            tokenizer_name = self.config.tokenizer_name 
+        else:
+            tokenizer_name = '/mnt/workspace/ckpt/vicuna-13b/'
+
+
+        datasets[split] = dataset_cls(
+            vis_processor=self.vis_processors[split],
+            text_processor=self.text_processors[split],
+            vis_root=build_info.videos_dir,
+            ann_root=build_info.anno_dir,
+            num_video_query_token = num_video_query_token,
+            tokenizer_name = tokenizer_name,
+            data_type = self.config.data_type
+        )
+
+        return datasets
+
+@registry.register_builder("webvid_instruct")
+class WebvidInstruct_Builder(Instruct_Builder):
+    train_dataset_cls = Video_Instruct_Dataset
+
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/instruct/webvid_instruct.yaml",
+    }
+
+@registry.register_builder("webvid_instruct_zh")
+class WebvidInstruct_zh_Builder(Instruct_Builder):
+    train_dataset_cls = Video_Instruct_Dataset
+
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/instruct/webvid_instruct.yaml",
+    }
+
+
+
+@registry.register_builder("llava_instruct")
+class LlavaInstruct_Builder(Instruct_Builder):
+    train_dataset_cls = Instruct_Dataset
+
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/instruct/llava_instruct.yaml",
+    }
+

video_llama/datasets/builders/video_caption_builder.py

@@ -0,0 +1,34 @@
+import os
+import logging
+import warnings
+
+from video_llama.common.registry import registry
+from video_llama.datasets.builders.base_dataset_builder import BaseDatasetBuilder
+from video_llama.datasets.datasets.webvid_datasets import WebvidDataset
+
+@registry.register_builder("webvid")
+class WebvidBuilder(BaseDatasetBuilder):
+    train_dataset_cls = WebvidDataset
+    DATASET_CONFIG_DICT = {"default": "configs/datasets/webvid/defaults.yaml"}
+    
+    def _download_ann(self):
+        pass
+
+    def _download_vis(self):
+        pass
+
+    def build(self):
+        self.build_processors()
+        datasets = dict()
+        split = "train"
+
+        build_info = self.config.build_info
+        dataset_cls = self.train_dataset_cls
+        datasets[split] = dataset_cls(
+            vis_processor=self.vis_processors[split],
+            text_processor=self.text_processors[split],
+            vis_root=build_info.videos_dir,
+            ann_root=build_info.anno_dir
+        )
+
+        return datasets

video_llama/datasets/data_utils.py

@@ -0,0 +1,196 @@
+"""
+ Copyright (c) 2022, 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 gzip
+import logging
+import os
+import random as rnd
+import tarfile
+import zipfile
+import random
+from typing import List
+from tqdm import tqdm
+
+import decord
+from decord import VideoReader
+import webdataset as wds
+import numpy as np
+import torch
+from torch.utils.data.dataset import IterableDataset
+
+from video_llama.common.registry import registry
+from video_llama.datasets.datasets.base_dataset import ConcatDataset
+
+
+decord.bridge.set_bridge("torch")
+MAX_INT = registry.get("MAX_INT")
+
+
+class ChainDataset(wds.DataPipeline):
+    r"""Dataset for chaining multiple :class:`DataPipeline` s.
+
+    This class is useful to assemble different existing dataset streams. The
+    chaining operation is done on-the-fly, so concatenating large-scale
+    datasets with this class will be efficient.
+
+    Args:
+        datasets (iterable of IterableDataset): datasets to be chained together
+    """
+    def __init__(self, datasets: List[wds.DataPipeline]) -> None:
+        super().__init__()
+        self.datasets = datasets
+        self.prob = []
+        self.names = []
+        for dataset in self.datasets:
+            if hasattr(dataset, 'name'):
+                self.names.append(dataset.name)
+            else:
+                self.names.append('Unknown')
+            if hasattr(dataset, 'sample_ratio'):
+                self.prob.append(dataset.sample_ratio)
+            else:
+                self.prob.append(1)
+                logging.info("One of the datapipeline doesn't define ratio and set to 1 automatically.")
+
+    def __iter__(self):
+        datastreams = [iter(dataset) for dataset in self.datasets]
+        while True:
+            select_datastream = random.choices(datastreams, weights=self.prob, k=1)[0]
+            yield next(select_datastream)
+
+
+def apply_to_sample(f, sample):
+    if len(sample) == 0:
+        return {}
+
+    def _apply(x):
+        if torch.is_tensor(x):
+            return f(x)
+        elif isinstance(x, dict):
+            return {key: _apply(value) for key, value in x.items()}
+        elif isinstance(x, list):
+            return [_apply(x) for x in x]
+        else:
+            return x
+
+    return _apply(sample)
+
+
+def move_to_cuda(sample):
+    def _move_to_cuda(tensor):
+        return tensor.cuda()
+
+    return apply_to_sample(_move_to_cuda, sample)
+
+
+def prepare_sample(samples, cuda_enabled=True):
+    if cuda_enabled:
+        samples = move_to_cuda(samples)
+
+    # TODO fp16 support
+
+    return samples
+
+
+def reorg_datasets_by_split(datasets):
+    """
+    Organizes datasets by split.
+
+    Args:
+        datasets: dict of torch.utils.data.Dataset objects by name.
+
+    Returns:
+        Dict of datasets by split {split_name: List[Datasets]}.
+    """
+    # if len(datasets) == 1:
+    #     return datasets[list(datasets.keys())[0]]
+    # else:
+    reorg_datasets = dict()
+
+    # reorganize by split
+    for _, dataset in datasets.items():
+        for split_name, dataset_split in dataset.items():
+            if split_name not in reorg_datasets:
+                reorg_datasets[split_name] = [dataset_split]
+            else:
+                reorg_datasets[split_name].append(dataset_split)
+
+    return reorg_datasets
+
+
+def concat_datasets(datasets):
+    """
+    Concatenates multiple datasets into a single dataset.
+
+    It supports may-style datasets and DataPipeline from WebDataset. Currently, does not support
+    generic IterableDataset because it requires creating separate samplers.
+
+    Now only supports conctenating training datasets and assuming validation and testing
+    have only a single dataset. This is because metrics should not be computed on the concatenated
+    datasets.
+
+    Args:
+        datasets: dict of torch.utils.data.Dataset objects by split.
+
+    Returns:
+        Dict of concatenated datasets by split, "train" is the concatenation of multiple datasets,
+        "val" and "test" remain the same.
+
+        If the input training datasets contain both map-style and DataPipeline datasets, returns
+        a tuple, where the first element is a concatenated map-style dataset and the second
+        element is a chained DataPipeline dataset.
+
+    """
+    # concatenate datasets in the same split
+    for split_name in datasets:
+        if split_name != "train":
+            assert (
+                len(datasets[split_name]) == 1
+            ), "Do not support multiple {} datasets.".format(split_name)
+            datasets[split_name] = datasets[split_name][0]
+        else:
+            iterable_datasets, map_datasets = [], []
+            for dataset in datasets[split_name]:
+                if isinstance(dataset, wds.DataPipeline):
+                    logging.info(
+                        "Dataset {} is IterableDataset, can't be concatenated.".format(
+                            dataset
+                        )
+                    )
+                    iterable_datasets.append(dataset)
+                elif isinstance(dataset, IterableDataset):
+                    raise NotImplementedError(
+                        "Do not support concatenation of generic IterableDataset."
+                    )
+                else:
+                    map_datasets.append(dataset)
+
+            # if len(iterable_datasets) > 0:
+            # concatenate map-style datasets and iterable-style datasets separately
+            if len(iterable_datasets) > 1:
+                chained_datasets = (
+                    ChainDataset(iterable_datasets)
+                )
+            elif len(iterable_datasets) == 1:
+                chained_datasets = iterable_datasets[0]
+            else:
+                chained_datasets = None
+
+            concat_datasets = (
+                ConcatDataset(map_datasets) if len(map_datasets) > 0 else None
+            )
+
+            train_datasets = concat_datasets, chained_datasets
+            train_datasets = tuple([x for x in train_datasets if x is not None])
+            train_datasets = (
+                train_datasets[0] if len(train_datasets) == 1 else train_datasets
+            )
+
+            datasets[split_name] = train_datasets
+
+    return datasets
+

video_llama/datasets/datasets/base_dataset.py

@@ -0,0 +1,68 @@
+"""
+ Copyright (c) 2022, 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 json
+from typing import Iterable
+
+from torch.utils.data import Dataset, ConcatDataset
+from torch.utils.data.dataloader import default_collate
+
+
+class BaseDataset(Dataset):
+    def __init__(
+        self, vis_processor=None, text_processor=None, vis_root=None, ann_paths=[]
+    ):
+        """
+        vis_root (string): Root directory of images (e.g. coco/images/)
+        ann_root (string): directory to store the annotation file
+        """
+        self.vis_root = vis_root
+
+        self.annotation = []
+        for ann_path in ann_paths:
+            self.annotation.extend(json.load(open(ann_path, "r"))['annotations'])
+
+        self.vis_processor = vis_processor
+        self.text_processor = text_processor
+
+        self._add_instance_ids()
+
+    def __len__(self):
+        return len(self.annotation)
+
+    def collater(self, samples):
+        return default_collate(samples)
+
+    def set_processors(self, vis_processor, text_processor):
+        self.vis_processor = vis_processor
+        self.text_processor = text_processor
+
+    def _add_instance_ids(self, key="instance_id"):
+        for idx, ann in enumerate(self.annotation):
+            ann[key] = str(idx)
+
+
+class ConcatDataset(ConcatDataset):
+    def __init__(self, datasets: Iterable[Dataset]) -> None:
+        super().__init__(datasets)
+
+    def collater(self, samples):
+        # TODO For now only supports datasets with same underlying collater implementations
+
+        all_keys = set()
+        for s in samples:
+            all_keys.update(s)
+
+        shared_keys = all_keys
+        for s in samples:
+            shared_keys = shared_keys & set(s.keys())
+
+        samples_shared_keys = []
+        for s in samples:
+            samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})
+
+        return self.datasets[0].collater(samples_shared_keys)

video_llama/datasets/datasets/caption_datasets.py

@@ -0,0 +1,85 @@
+"""
+ Copyright (c) 2022, 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 os
+from collections import OrderedDict
+
+from video_llama.datasets.datasets.base_dataset import BaseDataset
+from PIL import Image
+
+
+class __DisplMixin:
+    def displ_item(self, index):
+        sample, ann = self.__getitem__(index), self.annotation[index]
+
+        return OrderedDict(
+            {
+                "file": ann["image"],
+                "caption": ann["caption"],
+                "image": sample["image"],
+            }
+        )
+
+
+class CaptionDataset(BaseDataset, __DisplMixin):
+    def __init__(self, vis_processor, text_processor, vis_root, ann_paths):
+        """
+        vis_root (string): Root directory of images (e.g. coco/images/)
+        ann_root (string): directory to store the annotation file
+        """
+        super().__init__(vis_processor, text_processor, vis_root, ann_paths)
+
+        self.img_ids = {}
+        n = 0
+        for ann in self.annotation:
+            img_id = ann["image_id"]
+            if img_id not in self.img_ids.keys():
+                self.img_ids[img_id] = n
+                n += 1
+
+    def __getitem__(self, index):
+
+        # TODO this assumes image input, not general enough
+        ann = self.annotation[index]
+
+        img_file = '{:0>12}.jpg'.format(ann["image_id"])
+        image_path = os.path.join(self.vis_root, img_file)
+        image = Image.open(image_path).convert("RGB")
+
+        image = self.vis_processor(image)
+        caption = self.text_processor(ann["caption"])
+
+        return {
+            "image": image,
+            "text_input": caption,
+            "image_id": self.img_ids[ann["image_id"]],
+        }
+
+
+class CaptionEvalDataset(BaseDataset, __DisplMixin):
+    def __init__(self, vis_processor, text_processor, vis_root, ann_paths):
+        """
+        vis_root (string): Root directory of images (e.g. coco/images/)
+        ann_root (string): directory to store the annotation file
+        split (string): val or test
+        """
+        super().__init__(vis_processor, text_processor, vis_root, ann_paths)
+
+    def __getitem__(self, index):
+
+        ann = self.annotation[index]
+
+        image_path = os.path.join(self.vis_root, ann["image"])
+        image = Image.open(image_path).convert("RGB")
+
+        image = self.vis_processor(image)
+
+        return {
+            "image": image,
+            "image_id": ann["image_id"],
+            "instance_id": ann["instance_id"],
+        }

video_llama/datasets/datasets/cc_sbu_dataset.py

@@ -0,0 +1,49 @@
+import os
+from PIL import Image
+import webdataset as wds
+from video_llama.datasets.datasets.base_dataset import BaseDataset
+from video_llama.datasets.datasets.caption_datasets import CaptionDataset
+
+
+class CCSBUDataset(BaseDataset):
+    def __init__(self, vis_processor, text_processor, location):
+        super().__init__(vis_processor=vis_processor, text_processor=text_processor)
+
+        self.inner_dataset = wds.DataPipeline(
+            wds.ResampledShards(location),
+            wds.tarfile_to_samples(handler=wds.warn_and_continue),
+            wds.shuffle(1000, handler=wds.warn_and_continue),
+            wds.decode("pilrgb", handler=wds.warn_and_continue),
+            wds.to_tuple("jpg", "json", handler=wds.warn_and_continue),
+            wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),
+            wds.map(self.to_dict, handler=wds.warn_and_continue),
+        )
+
+    def to_dict(self, sample):
+        return {
+            "image": sample[0],
+            "text_input": self.text_processor(sample[1]["caption"]),
+            "type":'image',
+        }
+
+
+class CCSBUAlignDataset(CaptionDataset):
+
+    def __getitem__(self, index):
+
+        # TODO this assumes image input, not general enough
+        ann = self.annotation[index]
+
+        img_file = '{}.jpg'.format(ann["image_id"])
+        image_path = os.path.join(self.vis_root, img_file)
+        image = Image.open(image_path).convert("RGB")
+
+        image = self.vis_processor(image)
+        caption = ann["caption"]
+
+        return {
+            "image": image,
+            "text_input": caption,
+            "image_id": self.img_ids[ann["image_id"]],
+            "type":'image',
+        }

video_llama/datasets/datasets/dataloader_utils.py

@@ -0,0 +1,162 @@
+"""
+ Copyright (c) 2022, 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 time
+import random
+import torch
+from video_llama.datasets.data_utils import move_to_cuda
+from torch.utils.data import DataLoader
+
+
+class MultiIterLoader:
+    """
+    A simple wrapper for iterating over multiple iterators.
+
+    Args:
+        loaders (List[Loader]): List of Iterator loaders.
+        ratios (List[float]): List of ratios to sample from each loader. If None, all loaders are sampled uniformly.
+    """
+
+    def __init__(self, loaders, ratios=None):
+        # assert all loaders has __next__ method
+        for loader in loaders:
+            assert hasattr(
+                loader, "__next__"
+            ), "Loader {} has no __next__ method.".format(loader)
+
+        if ratios is None:
+            ratios = [1.0] * len(loaders)
+        else:
+            assert len(ratios) == len(loaders)
+            ratios = [float(ratio) / sum(ratios) for ratio in ratios]
+
+        self.loaders = loaders
+        self.ratios = ratios
+
+    def __next__(self):
+        # random sample from each loader by ratio
+        loader_idx = random.choices(range(len(self.loaders)), self.ratios, k=1)[0]
+        return next(self.loaders[loader_idx])
+
+
+class PrefetchLoader(object):
+    """
+    Modified from https://github.com/ChenRocks/UNITER.
+
+    overlap compute and cuda data transfer
+    (copied and then modified from nvidia apex)
+    """
+
+    def __init__(self, loader):
+        self.loader = loader
+        self.stream = torch.cuda.Stream()
+
+    def __iter__(self):
+        loader_it = iter(self.loader)
+        self.preload(loader_it)
+        batch = self.next(loader_it)
+        while batch is not None:
+            is_tuple = isinstance(batch, tuple)
+            if is_tuple:
+                task, batch = batch
+
+            if is_tuple:
+                yield task, batch
+            else:
+                yield batch
+            batch = self.next(loader_it)
+
+    def __len__(self):
+        return len(self.loader)
+
+    def preload(self, it):
+        try:
+            self.batch = next(it)
+        except StopIteration:
+            self.batch = None
+            return
+        # if record_stream() doesn't work, another option is to make sure
+        # device inputs are created on the main stream.
+        # self.next_input_gpu = torch.empty_like(self.next_input,
+        #                                        device='cuda')
+        # self.next_target_gpu = torch.empty_like(self.next_target,
+        #                                         device='cuda')
+        # Need to make sure the memory allocated for next_* is not still in use
+        # by the main stream at the time we start copying to next_*:
+        # self.stream.wait_stream(torch.cuda.current_stream())
+        with torch.cuda.stream(self.stream):
+            self.batch = move_to_cuda(self.batch)
+            # more code for the alternative if record_stream() doesn't work:
+            # copy_ will record the use of the pinned source tensor in this
+            # side stream.
+            # self.next_input_gpu.copy_(self.next_input, non_blocking=True)
+            # self.next_target_gpu.copy_(self.next_target, non_blocking=True)
+            # self.next_input = self.next_input_gpu
+            # self.next_target = self.next_target_gpu
+
+    def next(self, it):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        if batch is not None:
+            record_cuda_stream(batch)
+        self.preload(it)
+        return batch
+
+    def __getattr__(self, name):
+        method = self.loader.__getattribute__(name)
+        return method
+
+
+def record_cuda_stream(batch):
+    if isinstance(batch, torch.Tensor):
+        batch.record_stream(torch.cuda.current_stream())
+    elif isinstance(batch, list) or isinstance(batch, tuple):
+        for t in batch:
+            record_cuda_stream(t)
+    elif isinstance(batch, dict):
+        for t in batch.values():
+            record_cuda_stream(t)
+    else:
+        pass
+
+
+class IterLoader:
+    """
+    A wrapper to convert DataLoader as an infinite iterator.
+
+    Modified from:
+        https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py
+    """
+
+    def __init__(self, dataloader: DataLoader, use_distributed: bool = False):
+        self._dataloader = dataloader
+        self.iter_loader = iter(self._dataloader)
+        self._use_distributed = use_distributed
+        self._epoch = 0
+
+    @property
+    def epoch(self) -> int:
+        return self._epoch
+
+    def __next__(self):
+        try:
+            data = next(self.iter_loader)
+        except StopIteration:
+            self._epoch += 1
+            if hasattr(self._dataloader.sampler, "set_epoch") and self._use_distributed:
+                self._dataloader.sampler.set_epoch(self._epoch)
+            time.sleep(2)  # Prevent possible deadlock during epoch transition
+            self.iter_loader = iter(self._dataloader)
+            data = next(self.iter_loader)
+
+        return data
+
+    def __iter__(self):
+        return self
+
+    def __len__(self):
+        return len(self._dataloader)

video_llama/datasets/datasets/laion_dataset.py

@@ -0,0 +1,31 @@
+"""
+ Copyright (c) 2022, 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 webdataset as wds
+from video_llama.datasets.datasets.base_dataset import BaseDataset
+
+
+class LaionDataset(BaseDataset):
+    def __init__(self, vis_processor, text_processor, location):
+        super().__init__(vis_processor=vis_processor, text_processor=text_processor)
+
+        self.inner_dataset = wds.DataPipeline(
+            wds.ResampledShards(location),
+            wds.tarfile_to_samples(handler=wds.warn_and_continue),
+            wds.shuffle(1000, handler=wds.warn_and_continue),
+            wds.decode("pilrgb", handler=wds.warn_and_continue),
+            wds.to_tuple("jpg", "json", handler=wds.warn_and_continue),
+            wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),
+            wds.map(self.to_dict, handler=wds.warn_and_continue),
+        )
+
+    def to_dict(self, sample):
+        return {
+            "image": sample[0],
+            "text_input": self.text_processor(sample[1]["caption"]),
+        }
+

video_llama/datasets/datasets/llava_instruct_dataset.py

@@ -0,0 +1,228 @@
+import os
+from video_llama.datasets.datasets.base_dataset import BaseDataset
+from video_llama.datasets.datasets.caption_datasets import CaptionDataset
+import pandas as pd
+import decord
+from decord import VideoReader
+import random
+import torch
+from torch.utils.data.dataloader import default_collate
+from PIL import Image
+from typing import Dict, Optional, Sequence
+import transformers
+import pathlib
+import json
+from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaTokenizer
+from video_llama.conversation.conversation_video import Conversation,SeparatorStyle
+DEFAULT_IMAGE_PATCH_TOKEN = '<ImageHere>'
+DEFAULT_IMAGE_TOKEN = "<image>"
+import copy
+IGNORE_INDEX = -100
+image_conversation = Conversation(
+    system="",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+IGNORE_INDEX = -100
+
+class Instruct_Dataset(BaseDataset):
+    def __init__(self, vis_processor, text_processor, vis_root, ann_root,num_video_query_token=32,tokenizer_name = '/mnt/workspace/ckpt/vicuna-13b/',data_type = 'image'):
+        """
+        vis_root (string): Root directory of Llava images (e.g. webvid_eval/video/)
+        ann_root (string): Root directory of video (e.g. webvid_eval/annotations/)
+        split (string): val or test
+        """
+        super().__init__(vis_processor=vis_processor, text_processor=text_processor)
+
+        data_path = pathlib.Path(ann_root)
+        with data_path.open(encoding='utf-8') as f:
+            self.annotation = json.load(f)
+
+        self.vis_root = vis_root
+        self.resize_size = 224
+        self.num_frm = 8
+        self.tokenizer = LlamaTokenizer.from_pretrained(tokenizer_name, use_fast=False)
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        self.num_video_query_token = num_video_query_token
+        self.IMAGE_PATCH_TOKEN_ID = self.tokenizer.get_vocab()[DEFAULT_IMAGE_PATCH_TOKEN]
+
+        self.transform = AlproVideoTrainProcessor(
+            image_size=self.resize_size, n_frms = self.num_frm
+        ).transform
+        self.data_type = data_type
+
+    def _get_image_path(self, sample):
+        rel_video_fp ='COCO_train2014_' + sample['image']
+        full_video_fp = os.path.join(self.vis_root,  rel_video_fp)
+        return full_video_fp
+
+    def __getitem__(self, index):
+        num_retries = 10  # skip error videos
+        for _ in range(num_retries):
+            try:
+                sample = self.annotation[index]
+
+                image_path = self._get_image_path(sample)
+                conversation_list = sample['conversations']
+                image = Image.open(image_path).convert("RGB")
+
+                image = self.vis_processor(image)
+                # text = self.text_processor(text)
+                sources = preprocess_multimodal(copy.deepcopy(conversation_list), None, cur_token_len=self.num_video_query_token)
+                data_dict = preprocess(
+                    sources,
+                    self.tokenizer)
+                data_dict = dict(input_ids=data_dict["input_ids"][0],
+                                labels=data_dict["labels"][0])
+
+                # image exist in the data
+                data_dict['image'] = image
+            except:
+                print(f"Failed to load examples with image: {image_path}. "
+                            f"Will randomly sample an example as a replacement.")
+                index = random.randint(0, len(self) - 1)
+                continue
+            break
+        else:  
+            raise RuntimeError(f"Failed to fetch image after {num_retries} retries.")
+        # "image_id" is kept to stay compatible with the COCO evaluation format
+        return {
+            "image": image,
+            "text_input": data_dict["input_ids"],
+            "labels": data_dict["labels"],
+            "type":'image',
+        }
+
+    def __len__(self):
+        return len(self.annotation)
+
+    def collater(self, instances):
+        input_ids, labels = tuple([instance[key] for instance in instances]
+                                  for key in ("text_input", "labels"))
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id)
+        labels = torch.nn.utils.rnn.pad_sequence(labels,
+                                                 batch_first=True,
+                                                 padding_value=IGNORE_INDEX)
+        batch = dict(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
+        )
+
+        if 'image' in instances[0]:
+            images = [instance['image'] for instance in instances]
+            if all(x is not None and x.shape == images[0].shape for x in images):
+                batch['images'] = torch.stack(images)
+            else:
+                batch['images'] = images
+        batch['conv_type'] = 'multi'
+        return batch
+
+
+def preprocess_multimodal(
+    conversation_list: Sequence[str],
+    multimodal_cfg: dict,
+    cur_token_len: int,
+) -> Dict:
+    # 将conversational list中
+    is_multimodal = True
+    # image_token_len = multimodal_cfg['image_token_len']
+    image_token_len = cur_token_len
+
+    for sentence in conversation_list:
+        replace_token = '<Image>'+DEFAULT_IMAGE_PATCH_TOKEN * image_token_len+'/<Image>'
+        sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token)
+
+    return [conversation_list]
+
+def _add_speaker_and_signal(header, source, get_conversation=True):
+    """Add speaker and start/end signal on each round."""
+    BEGIN_SIGNAL = "###"
+    END_SIGNAL = "\n"
+    conversation = header
+    for sentence in source:
+        from_str = sentence["from"]
+        if from_str.lower() == "human":
+            from_str = image_conversation.roles[0]
+        elif from_str.lower() == "gpt":
+            from_str = image_conversation.roles[1]
+        else:
+            from_str = 'unknown'
+        sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
+                             sentence["value"] + END_SIGNAL)
+        if get_conversation:
+            conversation += sentence["value"]
+    conversation += BEGIN_SIGNAL
+    return conversation
+        
+def _tokenize_fn(strings: Sequence[str],
+                tokenizer: transformers.PreTrainedTokenizer) -> Dict:
+    """Tokenize a list of strings."""
+    tokenized_list = [
+        tokenizer(
+            text,
+            return_tensors="pt",
+            padding="longest",
+            max_length=512,
+            truncation=True,
+        ) for text in strings
+    ]
+    input_ids = labels = [
+        tokenized.input_ids[0] for tokenized in tokenized_list
+    ]
+    input_ids_lens = labels_lens = [
+        tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
+        for tokenized in tokenized_list
+    ]
+    return dict(
+        input_ids=input_ids,
+        labels=labels,
+        input_ids_lens=input_ids_lens,
+        labels_lens=labels_lens,
+    )
+
+def preprocess(
+    sources: Sequence[str],
+    tokenizer: transformers.PreTrainedTokenizer,
+) -> Dict:
+    """
+    Given a list of sources, each is a conversation list. This transform:
+    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
+    2. Concatenate conversations together;
+    3. Tokenize the concatenated conversation;
+    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
+    """
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        header = f"{image_conversation.system}\n\n"
+        conversation = _add_speaker_and_signal(header, source)
+        conversations.append(conversation)
+    # tokenize conversations
+    conversations_tokenized = _tokenize_fn(conversations, tokenizer)
+    input_ids = conversations_tokenized["input_ids"]
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source],
+                                      tokenizer)["input_ids_lens"]
+        speakers = [sentence["from"] for sentence in source]
+        _mask_targets(target, tokenized_lens, speakers)
+
+    return dict(input_ids=input_ids, labels=targets)
+
+def _mask_targets(target, tokenized_lens, speakers):
+    # cur_idx = 0
+    cur_idx = tokenized_lens[0]
+    tokenized_lens = tokenized_lens[1:]
+    target[:cur_idx] = IGNORE_INDEX
+    for tokenized_len, speaker in zip(tokenized_lens, speakers):
+        if speaker == "human":
+            target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
+        cur_idx += tokenized_len

video_llama/datasets/datasets/video_instruct_dataset.py

@@ -0,0 +1,253 @@
+import os
+from video_llama.datasets.datasets.base_dataset import BaseDataset
+from video_llama.datasets.datasets.caption_datasets import CaptionDataset
+import pandas as pd
+import decord
+from decord import VideoReader
+import random
+import torch
+from torch.utils.data.dataloader import default_collate
+from PIL import Image
+from typing import Dict, Optional, Sequence
+import transformers
+import pathlib
+import json
+from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaTokenizer
+import copy
+from video_llama.processors import transforms_video,AlproVideoTrainProcessor
+from torchvision import transforms
+from video_llama.processors.video_processor import ToTHWC,ToUint8,load_video
+from video_llama.conversation.conversation_video import Conversation,SeparatorStyle
+
+DEFAULT_IMAGE_PATCH_TOKEN = '<ImageHere>'
+video_conversation = Conversation(
+    system="",
+    roles=("Human", "Assistant"),
+    messages=[],
+    offset=0,
+    sep_style=SeparatorStyle.SINGLE,
+    sep="###",
+)
+IGNORE_INDEX = -100
+
+class Video_Instruct_Dataset(BaseDataset):
+    def __init__(self, vis_processor, text_processor, vis_root, ann_root,num_video_query_token=32,tokenizer_name = '/mnt/workspace/ckpt/vicuna-13b/',data_type = 'video'):
+        """
+        vis_root (string): Root directory of Llava images (e.g. webvid_eval/video/)
+        ann_root (string): Root directory of video (e.g. webvid_eval/annotations/)
+        split (string): val or test
+        """
+        super().__init__(vis_processor=vis_processor, text_processor=text_processor)
+
+        data_path = pathlib.Path(ann_root)
+        with data_path.open(encoding='utf-8') as f:
+            self.annotation = json.load(f)
+
+        self.num_video_query_token = num_video_query_token
+        self.vis_root = vis_root
+        self.resize_size = 224
+        self.num_frm = 8
+        self.tokenizer = LlamaTokenizer.from_pretrained(tokenizer_name, use_fast=False)
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        self.IMAGE_PATCH_TOKEN_ID = self.tokenizer.get_vocab()[DEFAULT_IMAGE_PATCH_TOKEN]
+
+        self.transform = AlproVideoTrainProcessor(
+            image_size=self.resize_size, n_frms = self.num_frm
+        ).transform
+        self.data_type = data_type
+          
+    def _get_video_path(self, sample):
+        rel_video_fp = sample['video']
+        full_video_fp = os.path.join(self.vis_root,  rel_video_fp)
+        return full_video_fp
+
+    def __getitem__(self, index):
+        num_retries = 10  # skip error videos
+        for _ in range(num_retries):
+            try:
+                sample = self.annotation[index]
+
+                video_path = self._get_video_path(sample)
+                conversation_list = sample['QA']
+
+                video, msg = load_video(
+                    video_path=video_path,
+                    n_frms=self.num_frm,
+                    height=self.resize_size,
+                    width=self.resize_size,
+                    sampling ="uniform", return_msg = True
+                )
+                video = self.transform(video)
+                if 'cn' in self.data_type:
+                    msg = ""
+                # 添加视频<DEFAULT_IMAGE_PATCH_TOKEN>,以及msg到convsation list 0
+                sources = preprocess_multimodal(copy.deepcopy(conversation_list), None, cur_token_len=self.num_video_query_token,msg = msg)
+                new_sources = convert_source_vicuna_format(sources)
+
+                data_dict = preprocess(
+                    new_sources,
+                    self.tokenizer)
+                data_dict = dict(input_ids=data_dict["input_ids"][0],
+                                labels=data_dict["labels"][0])
+                # image exist in the data
+                data_dict['image'] = video
+            except:
+                print(f"Failed to load examples with video: {video_path}. "
+                            f"Will randomly sample an example as a replacement.")
+                index = random.randint(0, len(self) - 1)
+                continue
+            break
+        else:  
+            raise RuntimeError(f"Failed to fetch video after {num_retries} retries.")
+        # "image_id" is kept to stay compatible with the COCO evaluation format
+        return {
+            "image": video,
+            "text_input": data_dict["input_ids"],
+            "labels": data_dict["labels"],
+            "type":'video',
+        }
+
+    def __len__(self):
+        return len(self.annotation)
+
+    def collater(self, instances):
+        input_ids, labels = tuple([instance[key] for instance in instances]
+                                  for key in ("text_input", "labels"))
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id)
+        labels = torch.nn.utils.rnn.pad_sequence(labels,
+                                                 batch_first=True,
+                                                 padding_value=IGNORE_INDEX)
+        batch = dict(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
+        )
+
+        if 'image' in instances[0]:
+            images = [instance['image'] for instance in instances]
+            if all(x is not None and x.shape == images[0].shape for x in images):
+                batch['images'] = torch.stack(images)
+            else:
+                batch['images'] = images
+        batch['conv_type'] = 'multi'
+        return batch
+
+def convert_source_vicuna_format(sources):
+    new_sources = []
+    for source in sources:
+        new_source = []
+        for i, sentence in enumerate(source):
+            role_0_msg = sentence['q']
+            role_1_msg = sentence['a']
+            new_source.append({
+                'from':'human',
+                'value': role_0_msg,
+            })
+            new_source.append({
+                'from':'gpt',
+                'value': role_1_msg,
+            })
+        new_sources.append(new_source)
+    return new_sources
+
+def preprocess_multimodal(
+    conversation_list: Sequence[str],
+    multimodal_cfg: dict,
+    cur_token_len: int,
+    msg=''
+) -> Dict:
+    # 将conversational list中
+    is_multimodal = True
+    # image_token_len = multimodal_cfg['image_token_len']
+    image_token_len = cur_token_len
+    conversation_list[0]["q"] = "<Video>"+DEFAULT_IMAGE_PATCH_TOKEN * image_token_len +"</Video> " + msg + conversation_list[0]["q"]
+    return [conversation_list]
+
+def _add_speaker_and_signal(header, source, get_conversation=True):
+    """Add speaker and start/end signal on each round."""
+    BEGIN_SIGNAL = "###"
+    END_SIGNAL = "\n"
+    conversation = header
+    for sentence in source:
+        from_str = sentence["from"]
+        if from_str.lower() == "human":
+            from_str = video_conversation.roles[0]
+        elif from_str.lower() == "gpt":
+            from_str = video_conversation.roles[1]
+        else:
+            from_str = 'unknown'
+        sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
+                             sentence["value"] + END_SIGNAL)
+        if get_conversation:
+            conversation += sentence["value"]
+    conversation += BEGIN_SIGNAL
+    return conversation
+        
+def _tokenize_fn(strings: Sequence[str],
+                tokenizer: transformers.PreTrainedTokenizer) -> Dict:
+    """Tokenize a list of strings."""
+    tokenized_list = [
+        tokenizer(
+            text,
+            return_tensors="pt",
+            padding="longest",
+            max_length=512,
+            truncation=True,
+        ) for text in strings
+    ]
+    input_ids = labels = [
+        tokenized.input_ids[0] for tokenized in tokenized_list
+    ]
+    input_ids_lens = labels_lens = [
+        tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
+        for tokenized in tokenized_list
+    ]
+    return dict(
+        input_ids=input_ids,
+        labels=labels,
+        input_ids_lens=input_ids_lens,
+        labels_lens=labels_lens,
+    )
+
+def preprocess(
+    sources: Sequence[str],
+    tokenizer: transformers.PreTrainedTokenizer,
+) -> Dict:
+    """
+    Given a list of sources, each is a conversation list. This transform:
+    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
+    2. Concatenate conversations together;
+    3. Tokenize the concatenated conversation;
+    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
+    """
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        header = f"{video_conversation.system}\n\n"
+        conversation = _add_speaker_and_signal(header, source)
+        conversations.append(conversation)
+    # tokenize conversations
+    conversations_tokenized = _tokenize_fn(conversations, tokenizer)
+    input_ids = conversations_tokenized["input_ids"]
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source],
+                                      tokenizer)["input_ids_lens"]
+        speakers = [sentence["from"] for sentence in source]
+        _mask_targets(target, tokenized_lens, speakers)
+
+    return dict(input_ids=input_ids, labels=targets)
+
+def _mask_targets(target, tokenized_lens, speakers):
+    # cur_idx = 0
+    cur_idx = tokenized_lens[0]
+    tokenized_lens = tokenized_lens[1:]
+    target[:cur_idx] = IGNORE_INDEX
+    for tokenized_len, speaker in zip(tokenized_lens, speakers):
+        if speaker == "human":
+            target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
+        cur_idx += tokenized_len

video_llama/datasets/datasets/webvid_datasets.py

@@ -0,0 +1,122 @@
+"""
+ Copyright (c) 2022, salesforce.com, inc.
+ All rights reserved.
+ SPDX-License-Identifier: BSD-3-Clause
+ For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+import os
+from video_llama.datasets.datasets.base_dataset import BaseDataset
+from video_llama.datasets.datasets.caption_datasets import CaptionDataset
+import pandas as pd
+import decord
+from decord import VideoReader
+import random
+import torch
+from torch.utils.data.dataloader import default_collate
+class WebvidDataset(BaseDataset):
+    def __init__(self, vis_processor, text_processor, vis_root, ann_root):
+        """
+        vis_root (string): Root directory of video (e.g. webvid_eval/video/)
+        ann_root (string): Root directory of video (e.g. webvid_eval/annotations/)
+        split (string): val or test
+        """
+        super().__init__(vis_processor=vis_processor, text_processor=text_processor)
+
+
+        # 读取一个路径下所有的
+
+        ts_df = []
+        for file_name in os.listdir(ann_root):
+            if file_name.endswith('.csv'):
+                df = pd.read_csv(os.path.join(ann_root, file_name))
+                ts_df.append(df)
+
+        merged_df = pd.concat(ts_df)
+        self.annotation = merged_df
+        self.vis_root = vis_root
+        self.resize_size = 224
+        self.num_frm = 8
+        self.frm_sampling_strategy = 'headtail'
+
+    def _get_video_path(self, sample):
+        rel_video_fp = os.path.join(sample['page_dir'], str(sample['videoid']) + '.mp4')
+        full_video_fp = os.path.join(self.vis_root,  rel_video_fp)
+        return full_video_fp
+
+    def __getitem__(self, index):
+        num_retries = 10  # skip error videos
+        for _ in range(num_retries):
+            sample = self.annotation.iloc[index]
+            sample_dict = sample.to_dict()
+            video_id = sample_dict['videoid']
+
+            if 'name' in sample_dict.keys():
+                text = sample_dict['name'].strip()
+            else:
+                raise NotImplementedError("Un-supported text annotation format.")
+
+            # fetch video
+            video_path = self._get_video_path(sample_dict) 
+            # if os.path.exists(video_path):
+            try:
+                video = self.vis_processor(video_path)
+            except:
+                print(f"Failed to load examples with video: {video_path}. "
+                            f"Will randomly sample an example as a replacement.")
+                index = random.randint(0, len(self) - 1)
+                continue
+            caption = self.text_processor(text)
+
+            # print(video.size())
+            if video is None or caption is None \
+                    or video.size()!=torch.Size([3,self.vis_processor.n_frms,224,224]):
+                print(f"Failed to load examples with video: {video_path}. "
+                            f"Will randomly sample an example as a replacement.")
+                index = random.randint(0, len(self) - 1)
+                continue
+            else:
+                break
+        else:  
+            raise RuntimeError(f"Failed to fetch video after {num_retries} retries.")
+        # "image_id" is kept to stay compatible with the COCO evaluation format
+        return {
+            "image": video,
+            "text_input": caption,
+            "type":'video',
+        }
+
+    def __len__(self):
+        return len(self.annotation)
+
+    # def collater(self, samples):
+    #     new_result = {}
+    #     new_result['image'] = default_collate( [sample["image"] for sample in samples])
+    #     new_result['text_input'] = default_collate( [sample["text_input"] for sample in samples])
+    #     return new_result
+        
+class WebvidDatasetEvalDataset(BaseDataset):
+    def __init__(self, vis_processor, text_processor, vis_root, ann_paths):
+        """
+        vis_root (string): Root directory of images (e.g. coco/images/)
+        ann_root (string): directory to store the annotation file
+        split (string): val or test
+        """
+        super().__init__(vis_processor, text_processor, vis_root, ann_paths)
+
+    def __getitem__(self, index):
+
+        ann = self.annotation[index]
+
+        vname = ann["video"]
+        video_path = os.path.join(self.vis_root, vname)
+
+        video = self.vis_processor(video_path)
+
+        return {
+            "video": video,
+            "image_id": ann["image_id"],
+            "instance_id": ann["instance_id"],
+        }
+
+

video_llama/models/Qformer.py

@@ -0,0 +1,1217 @@
+"""
+Adapted from salesforce@LAVIS. Below is the original copyright:
+ * 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 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 BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        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.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.self = BertSelfAttention(config, is_cross_attention)
+        self.output = BertSelfOutput(config)
+        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):
+        super().__init__()
+        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.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
+        self.attention = BertAttention(config)
+        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
+            )
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+        self.intermediate_query = BertIntermediate(config)
+        self.output_query = BertOutput(config)
+
+    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,
+        )

video_llama/models/__init__.py

@@ -0,0 +1,201 @@
+"""
+Adapted from salesforce@LAVIS Vision-CAIR@MiniGPT-4. Below is the original copyright:
+ Copyright (c) 2022, 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 logging
+import torch
+from omegaconf import OmegaConf
+
+from video_llama.common.registry import registry
+from video_llama.models.base_model import BaseModel
+from video_llama.models.blip2 import Blip2Base
+from video_llama.models.video_llama import VideoLLAMA
+from video_llama.processors.base_processor import BaseProcessor
+
+
+__all__ = [
+    "load_model",
+    "BaseModel",
+    "Blip2Base",
+    "VideoLLAMA"
+]
+
+
+def load_model(name, model_type, is_eval=False, device="cpu", checkpoint=None):
+    """
+    Load supported models.
+
+    To list all available models and types in registry:
+    >>> from video_llama.models import model_zoo
+    >>> print(model_zoo)
+
+    Args:
+        name (str): name of the model.
+        model_type (str): type of the model.
+        is_eval (bool): whether the model is in eval mode. Default: False.
+        device (str): device to use. Default: "cpu".
+        checkpoint (str): path or to checkpoint. Default: None.
+            Note that expecting the checkpoint to have the same keys in state_dict as the model.
+
+    Returns:
+        model (torch.nn.Module): model.
+    """
+
+    model = registry.get_model_class(name).from_pretrained(model_type=model_type)
+
+    if checkpoint is not None:
+        model.load_checkpoint(checkpoint)
+
+    if is_eval:
+        model.eval()
+
+    if device == "cpu":
+        model = model.float()
+
+    return model.to(device)
+
+
+def load_preprocess(config):
+    """
+    Load preprocessor configs and construct preprocessors.
+
+    If no preprocessor is specified, return BaseProcessor, which does not do any preprocessing.
+
+    Args:
+        config (dict): preprocessor configs.
+
+    Returns:
+        vis_processors (dict): preprocessors for visual inputs.
+        txt_processors (dict): preprocessors for text inputs.
+
+        Key is "train" or "eval" for processors used in training and evaluation respectively.
+    """
+
+    def _build_proc_from_cfg(cfg):
+        return (
+            registry.get_processor_class(cfg.name).from_config(cfg)
+            if cfg is not None
+            else BaseProcessor()
+        )
+
+    vis_processors = dict()
+    txt_processors = dict()
+
+    vis_proc_cfg = config.get("vis_processor")
+    txt_proc_cfg = config.get("text_processor")
+
+    if vis_proc_cfg is not None:
+        vis_train_cfg = vis_proc_cfg.get("train")
+        vis_eval_cfg = vis_proc_cfg.get("eval")
+    else:
+        vis_train_cfg = None
+        vis_eval_cfg = None
+
+    vis_processors["train"] = _build_proc_from_cfg(vis_train_cfg)
+    vis_processors["eval"] = _build_proc_from_cfg(vis_eval_cfg)
+
+    if txt_proc_cfg is not None:
+        txt_train_cfg = txt_proc_cfg.get("train")
+        txt_eval_cfg = txt_proc_cfg.get("eval")
+    else:
+        txt_train_cfg = None
+        txt_eval_cfg = None
+
+    txt_processors["train"] = _build_proc_from_cfg(txt_train_cfg)
+    txt_processors["eval"] = _build_proc_from_cfg(txt_eval_cfg)
+
+    return vis_processors, txt_processors
+
+
+def load_model_and_preprocess(name, model_type, is_eval=False, device="cpu"):
+    """
+    Load model and its related preprocessors.
+
+    List all available models and types in registry:
+    >>> from video_llama.models import model_zoo
+    >>> print(model_zoo)
+
+    Args:
+        name (str): name of the model.
+        model_type (str): type of the model.
+        is_eval (bool): whether the model is in eval mode. Default: False.
+        device (str): device to use. Default: "cpu".
+
+    Returns:
+        model (torch.nn.Module): model.
+        vis_processors (dict): preprocessors for visual inputs.
+        txt_processors (dict): preprocessors for text inputs.
+    """
+    model_cls = registry.get_model_class(name)
+
+    # load model
+    model = model_cls.from_pretrained(model_type=model_type)
+
+    if is_eval:
+        model.eval()
+
+    # load preprocess
+    cfg = OmegaConf.load(model_cls.default_config_path(model_type))
+    if cfg is not None:
+        preprocess_cfg = cfg.preprocess
+
+        vis_processors, txt_processors = load_preprocess(preprocess_cfg)
+    else:
+        vis_processors, txt_processors = None, None
+        logging.info(
+            f"""No default preprocess for model {name} ({model_type}).
+                This can happen if the model is not finetuned on downstream datasets,
+                or it is not intended for direct use without finetuning.
+            """
+        )
+
+    if device == "cpu" or device == torch.device("cpu"):
+        model = model.float()
+
+    return model.to(device), vis_processors, txt_processors
+
+
+class ModelZoo:
+    """
+    A utility class to create string representation of available model architectures and types.
+
+    >>> from video_llama.models import model_zoo
+    >>> # list all available models
+    >>> print(model_zoo)
+    >>> # show total number of models
+    >>> print(len(model_zoo))
+    """
+
+    def __init__(self) -> None:
+        self.model_zoo = {
+            k: list(v.PRETRAINED_MODEL_CONFIG_DICT.keys())
+            for k, v in registry.mapping["model_name_mapping"].items()
+        }
+
+    def __str__(self) -> str:
+        return (
+            "=" * 50
+            + "\n"
+            + f"{'Architectures':<30} {'Types'}\n"
+            + "=" * 50
+            + "\n"
+            + "\n".join(
+                [
+                    f"{name:<30} {', '.join(types)}"
+                    for name, types in self.model_zoo.items()
+                ]
+            )
+        )
+
+    def __iter__(self):
+        return iter(self.model_zoo.items())
+
+    def __len__(self):
+        return sum([len(v) for v in self.model_zoo.values()])
+
+
+model_zoo = ModelZoo()