Upload 119 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+unimernet/processors/formula_processor_helper/frost/frost1.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,12 +1,11 @@
----
-title: UniMERNet Demo
-emoji: 🌖
-colorFrom: purple
-colorTo: blue
-sdk: gradio
-sdk_version: 4.43.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: UniMERNet
+emoji: 👁
+colorFrom: purple
+colorTo: blue
+sdk: gradio
+sdk_version: 4.42.0
+app_file: app.py
+pinned: false
+license: apache-2.0
+---

app.py

@@ -0,0 +1,106 @@
+import os
+os.system('pip install -U transformers==4.44.2')
+import sys
+import shutil
+import torch
+import argparse
+import gradio as gr
+import numpy as np
+from PIL import Image
+from huggingface_hub import snapshot_download
+import spaces
+
+# == download weights ==
+tiny_model_dir = snapshot_download('wanderkid/unimernet_tiny', local_dir='./models/unimernet_tiny')
+small_model_dir = snapshot_download('wanderkid/unimernet_small', local_dir='./models/unimernet_small')
+base_model_dir = snapshot_download('wanderkid/unimernet_base', local_dir='./models/unimernet_base')
+os.system("ls -l models/unimernet_tiny")
+# os.system(f"sed -i 's/MODEL_DIR/{tiny_model_dir}/g' cfg_tiny.yaml")
+# os.system(f"sed -i 's/MODEL_DIR/{small_model_dir}/g' cfg_small.yaml")
+# os.system(f"sed -i 's/MODEL_DIR/{base_model_dir}/g' cfg_base.yaml")
+# root_path = os.path.abspath(os.getcwd())
+# os.makedirs(os.path.join(root_path, "models"), exist_ok=True)
+# shutil.move(tiny_model_dir, os.path.join(root_path, "models", "unimernet_tiny"))
+# shutil.move(small_model_dir, os.path.join(root_path, "models", "unimernet_small"))
+# shutil.move(base_model_dir, os.path.join(root_path, "models", "unimernet_base"))
+# == download weights ==
+
+sys.path.insert(0, os.path.join(os.getcwd(), ".."))
+from unimernet.common.config import Config
+import unimernet.tasks as tasks
+from unimernet.processors import load_processor
+
+
+def load_model_and_processor(cfg_path):
+    args = argparse.Namespace(cfg_path=cfg_path, options=None)
+    cfg = Config(args)
+    task = tasks.setup_task(cfg)
+    model = task.build_model(cfg)
+    vis_processor = load_processor('formula_image_eval', cfg.config.datasets.formula_rec_eval.vis_processor.eval)
+    return model, vis_processor
+
+@spaces.GPU
+def recognize_image(input_img, model_type):
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    if model_type == "base":
+        model = model_base.to(device)
+    elif model_type == "small":
+        model = model_small.to(device)
+    else:
+        model = model_tiny.to(device)
+        
+    if len(input_img.shape) == 3:
+            input_img = input_img[:, :, ::-1].copy()
+
+    img = Image.fromarray(input_img)
+    image = vis_processor(img).unsqueeze(0).to(device)
+    output = model.generate({"image": image})
+    latex_code = output["pred_str"][0]
+    return latex_code
+    
+def gradio_reset():
+    return gr.update(value=None), gr.update(value=None)
+
+    
+if __name__ == "__main__":
+    root_path = os.path.abspath(os.getcwd())
+    # == load model ==
+    model_tiny, vis_processor = load_model_and_processor(os.path.join(root_path, "cfg_tiny.yaml"))
+    model_small, vis_processor = load_model_and_processor(os.path.join(root_path, "cfg_small.yaml"))
+    model_base, vis_processor = load_model_and_processor(os.path.join(root_path, "cfg_base.yaml"))
+    print("== load all models ==")
+    # == load model ==
+    
+    with open("header.html", "r") as file:
+        header = file.read()
+    with gr.Blocks() as demo:
+        gr.HTML(header)
+        
+        with gr.Row():
+            with gr.Column():
+                model_type = gr.Radio(
+                        choices=["tiny", "small", "base"],
+                        value="tiny",
+                        label="Model Type",
+                        interactive=True,
+                    )
+                input_img = gr.Image(label=" ", interactive=True)
+                with gr.Row():
+                    clear = gr.Button("Clear")
+                    predict = gr.Button(value="Recognize", interactive=True, variant="primary")
+                    
+                with gr.Accordion("Examples:"):
+                    example_root = os.path.join(os.path.dirname(__file__), "examples")
+                    gr.Examples(
+                        examples=[os.path.join(example_root, _) for _ in os.listdir(example_root) if
+                                    _.endswith("png")],
+                        inputs=input_img,
+                    )
+            with gr.Column():
+                gr.Button(value="Predict Latex:", interactive=False)
+                pred_latex = gr.Textbox(label='Latex', interactive=False)
+    
+        clear.click(gradio_reset, inputs=None, outputs=[input_img, pred_latex])
+        predict.click(recognize_image, inputs=[input_img, model_type], outputs=[pred_latex])
+    
+    demo.launch(server_name="0.0.0.0", server_port=7860, debug=True)

cfg_base.yaml

@@ -0,0 +1,46 @@
+model:
+  arch: unimernet
+  model_type: unimernet
+  model_config:
+    model_name: ./models/unimernet_base
+    max_seq_len: 1536
+
+  load_pretrained: True
+  pretrained: './models/unimernet_base/unimernet_base.pth'
+  tokenizer_config:
+    path: ./models/unimernet_base
+
+datasets:
+  formula_rec_eval:
+    vis_processor:
+      eval:
+        name: "formula_image_eval"
+        image_size:
+          - 192
+          - 672
+   
+run:
+  runner: runner_iter
+  task: unimernet_train
+
+  batch_size_train: 64
+  batch_size_eval: 64
+  num_workers: 1
+
+  iters_per_inner_epoch: 2000
+  max_iters: 60000
+
+  seed: 42
+  output_dir: "../output/demo"
+
+  evaluate: True
+  test_splits: [ "eval" ]
+
+  device: "cuda"
+  world_size: 1
+  dist_url: "env://"
+  distributed: True
+  distributed_type: ddp  # or fsdp when train llm
+
+  generate_cfg:
+    temperature: 0.0

cfg_small.yaml

@@ -0,0 +1,46 @@
+model:
+  arch: unimernet
+  model_type: unimernet
+  model_config:
+    model_name: ./models/unimernet_small
+    max_seq_len: 1536
+
+  load_pretrained: True
+  pretrained: './models/unimernet_small/unimernet_small.pth'
+  tokenizer_config:
+    path: ./models/unimernet_small
+
+datasets:
+  formula_rec_eval:
+    vis_processor:
+      eval:
+        name: "formula_image_eval"
+        image_size:
+          - 192
+          - 672
+   
+run:
+  runner: runner_iter
+  task: unimernet_train
+
+  batch_size_train: 64
+  batch_size_eval: 64
+  num_workers: 1
+
+  iters_per_inner_epoch: 2000
+  max_iters: 60000
+
+  seed: 42
+  output_dir: "../output/demo"
+
+  evaluate: True
+  test_splits: [ "eval" ]
+
+  device: "cuda"
+  world_size: 1
+  dist_url: "env://"
+  distributed: True
+  distributed_type: ddp  # or fsdp when train llm
+
+  generate_cfg:
+    temperature: 0.0

cfg_tiny.yaml

@@ -0,0 +1,46 @@
+model:
+  arch: unimernet
+  model_type: unimernet
+  model_config:
+    model_name: ./models/unimernet_tiny
+    max_seq_len: 1536
+
+  load_pretrained: True
+  pretrained: './models/unimernet_tiny/unimernet_tiny.pth'
+  tokenizer_config:
+    path: ./models/unimernet_tiny
+
+datasets:
+  formula_rec_eval:
+    vis_processor:
+      eval:
+        name: "formula_image_eval"
+        image_size:
+          - 192
+          - 672
+   
+run:
+  runner: runner_iter
+  task: unimernet_train
+
+  batch_size_train: 64
+  batch_size_eval: 64
+  num_workers: 1
+
+  iters_per_inner_epoch: 2000
+  max_iters: 60000
+
+  seed: 42
+  output_dir: "../output/demo"
+
+  evaluate: True
+  test_splits: [ "eval" ]
+
+  device: "cuda"
+  world_size: 1
+  dist_url: "env://"
+  distributed: True
+  distributed_type: ddp  # or fsdp when train llm
+
+  generate_cfg:
+    temperature: 0.0

gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+unimernet/processors/formula_processor_helper/frost/frost1.png filter=lfs diff=lfs merge=lfs -text

header.html

@@ -0,0 +1,109 @@
+<html><head>
+    <!-- <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bulma@0.9.3/css/bulma.min.css"> -->
+    <link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.4/css/all.css">
+  <style>
+    .link-block {
+      border: 1px solid transparent;
+      border-radius: 24px;
+      background-color: rgba(54, 54, 54, 1);
+      cursor: pointer !important;
+    }
+    .link-block:hover {
+      background-color: rgba(54, 54, 54, 0.75) !important;
+      cursor: pointer !important;
+    }
+    .external-link {
+      display: inline-flex;
+      align-items: center;
+      height: 36px;
+      line-height: 36px;
+      padding: 0 16px;
+      cursor: pointer !important;
+    }
+    .external-link,
+    .external-link:hover {
+      cursor: pointer !important;
+    }
+    a {
+      text-decoration: none;
+    }
+  </style></head>
+  
+  <body>
+    <div style="
+        display: flex;
+        flex-direction: column;
+        justify-content: center;
+        align-items: center;
+        text-align: center;
+        background: linear-gradient(45deg, #007bff 0%, #0056b3 100%);
+        padding: 24px;
+        gap: 24px;
+        border-radius: 8px;
+      ">
+      <div style="
+          display: flex;
+          flex-direction: column;
+          align-items: center;
+          gap: 16px;
+        ">
+        <div style="display: flex; flex-direction: column; gap: 8px">
+          <h1 style="
+              font-size: 48px;
+              color: #fafafa;
+              margin: 0;
+              font-family: 'Trebuchet MS', 'Lucida Sans Unicode',
+                'Lucida Grande', 'Lucida Sans', Arial, sans-serif;
+            ">
+            UniMERNet
+          </h1>
+        </div>
+      </div>
+  
+      <p style="
+          margin: 0;
+          line-height: 1.6rem;
+          font-size: 16px;
+          color: #fafafa;
+          opacity: 0.8;
+        ">
+        A Universal Network for Real-World Mathematical Expression Recognition.<br>
+      </p>
+      <style>
+        .link-block {
+          display: inline-block;
+        }
+        .link-block + .link-block {
+          margin-left: 20px;
+        }
+      </style>
+  
+      <div class="column has-text-centered">
+        <div class="publication-links">
+          <!-- Code Link. -->
+          <span class="link-block">
+            <a href="https://github.com/opendatalab/UniMERNet" class="external-link button is-normal is-rounded is-dark" style="text-decoration: none; cursor: pointer">
+              <span class="icon" style="margin-right: 4px">
+                <i class="fab fa-github" style="color: white; margin-right: 4px"></i>
+              </span>
+              <span style="color: white">Code</span>
+            </a>
+          </span>
+  
+          <!-- Paper Link. -->
+          <span class="link-block">
+            <a href="https://arxiv.org/pdf/2404.15254" class="external-link button is-normal is-rounded is-dark" style="text-decoration: none; cursor: pointer">
+              <span class="icon" style="margin-right: 8px">
+                <i class="fas fa-globe" style="color: white"></i>
+              </span>
+              <span style="color: white">Paper</span>
+            </a>
+          </span>
+        </div>
+      </div>
+  
+      <!-- New Demo Links -->
+    </div>
+  
+  
+  </body></html>

requirements.txt

@@ -0,0 +1,3 @@
+unimernet==0.1.6
+gradio
+huggingface_hub

unimernet/__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 file in the repo root or https://opensource.org/licenses/BSD-3-Clause
+"""
+
+import os
+import sys
+
+from omegaconf import OmegaConf
+
+from unimernet.common.registry import registry
+
+from unimernet.datasets.builders import *
+from unimernet.models import *
+from unimernet.processors import *
+from unimernet.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"])

unimernet/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 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 unimernet.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()
+        # hiararchy 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
+            )
+
+            # hiararchy 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

unimernet/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 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()

unimernet/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

unimernet/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 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 unimernet.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()],
+    )

unimernet/common/optims.py

@@ -0,0 +1,120 @@
+"""
+ 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 math
+
+from unimernet.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,
+        min_lr,
+        init_lr,
+        iters_per_epoch,
+        warmup_steps=0,
+        warmup_start_lr=-1,
+        **kwargs
+    ):
+        self.optimizer = optimizer
+
+        self.max_epoch = max_epoch
+        self.min_lr = min_lr
+
+        self.init_lr = init_lr
+        self.warmup_steps = warmup_steps
+        self.iters_per_epoch = iters_per_epoch
+        self.warmup_start_lr = warmup_start_lr if warmup_start_lr >= 0 else init_lr
+
+    def step(self, cur_epoch, cur_step):
+        # assuming the warmup iters less than one epoch
+        total_steps = cur_epoch * self.iters_per_epoch + cur_step
+        if total_steps < 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_steps,
+                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

unimernet/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 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 unimernet.common.registry import registry
+            from unimernet.datasets.base_dataset_builder import BaseDatasetBuilder
+        """
+
+        def wrap(builder_cls):
+            from unimernet.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 unimernet.common.registry import registry
+        """
+
+        def wrap(task_cls):
+            from unimernet.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 unimernet.common.registry import registry
+        """
+
+        def wrap(model_cls):
+            from unimernet.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 unimernet.common.registry import registry
+        """
+
+        def wrap(processor_cls):
+            from unimernet.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 unimernet.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 unimernet.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 unimernet.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 unimernet.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()

unimernet/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 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 unimernet.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

unimernet/configs/datasets/formula/formula_eval.yaml

@@ -0,0 +1,6 @@
+datasets:
+  formula_rec_eval:
+    data_type: images
+    build_info:
+      images: /mnt/petrelfs/share_data/hanxiao/latex-ocr/pdf/val
+      annotation: /mnt/petrelfs/share_data/hanxiao/latex-ocr/pdf/pdfmath.txt

unimernet/configs/datasets/formula/formula_train.yaml

@@ -0,0 +1,6 @@
+datasets:
+  formula_rec_train:
+    data_type: images
+    build_info:
+      images: /mnt/petrelfs/share_data/hanxiao/latex-ocr/pdf/train
+      annotation: /mnt/petrelfs/share_data/hanxiao/latex-ocr/pdf/pdfmath.txt

unimernet/configs/datasets/formula/multi_scale_formula_train.yaml

@@ -0,0 +1,21 @@
+datasets:
+  multi_scale_formula_rec_train:
+    data_type: images
+    build_info:
+      images: /mnt/petrelfs/share_data/hanxiao/latex-ocr/pdf/train
+      annotation: /mnt/petrelfs/share_data/hanxiao/latex-ocr/pdf/pdfmath.txt
+
+    vis_processor:
+      train:
+        name: "formula_image_multi_scale_train"
+        all_scales:
+          - [ 96, 336 ]
+          - [ 128, 448 ]
+          - [ 192, 672 ]
+          - [ 288, 1008 ]
+          - [ 384, 1344 ]
+
+    text_processor:
+      train:
+        name: "blip_caption"
+        max_words: 256

unimernet/configs/default.yaml

@@ -0,0 +1,10 @@
+ # 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
+
+env:
+  # For default users
+  # cache_root: "cache"
+  # For internal use with persistent storage
+  cache_root: "/export/home/.cache/vigc"

unimernet/configs/models/unimernet_base.yaml

@@ -0,0 +1,31 @@
+model:
+  arch: unimernet
+  load_finetuned: False
+  load_pretrained: False
+  pretrained: "path/to/pretrained/weight"
+  finetuned: ""
+  tokenizer_name: nougat
+  tokenizer_config:
+    path: ./models/unimernet
+  model_name: unimernet
+  model_config:
+    max_seq_len: 384
+
+
+preprocess:
+  vis_processor:
+    train:
+      name: "formula_image_train"
+      image_size:
+        - 192
+        - 672
+    eval:
+      name: "formula_image_eval"
+      image_size:
+        - 192
+        - 672
+  text_processor:
+    train:
+      name: "blip_caption"
+    eval:
+      name: "blip_caption"

unimernet/datasets/builders/__init__.py

@@ -0,0 +1,69 @@
+"""
+ 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
+"""
+
+from unimernet.datasets.builders.base_dataset_builder import load_dataset_config
+from unimernet.common.registry import registry
+from unimernet.datasets.builders.formula import FormulaRecTrainBuilder, FormulaRecEvalBuilder, \
+    MultiScaleFormulaRecTrainBuilder
+
+__all__ = [
+    "FormulaRecTrainBuilder",
+    "FormulaRecEvalBuilder",
+    "MultiScaleFormulaRecTrainBuilder",
+]
+
+
+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()

unimernet/datasets/builders/base_dataset_builder.py

@@ -0,0 +1,233 @@
+"""
+ 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 logging
+import os
+import shutil
+import warnings
+
+import unimernet.common.utils as utils
+import torch.distributed as dist
+from unimernet.common.dist_utils import is_dist_avail_and_initialized, is_main_process
+from unimernet.common.registry import registry
+from unimernet.processors.base_processor import BaseProcessor
+from omegaconf import OmegaConf
+from torchvision.datasets.utils import download_url
+
+
+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 = vis_info.storage
+
+            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

unimernet/datasets/builders/formula.py

@@ -0,0 +1,105 @@
+import logging
+from unimernet.common.registry import registry
+from unimernet.datasets.builders.base_dataset_builder import BaseDatasetBuilder
+from unimernet.datasets.datasets.formula import Im2LatexDataset
+from unimernet.datasets.datasets.formula_multi_scale import MultiScaleIm2LatexDataset
+
+
+@registry.register_builder("formula_rec_train")
+class FormulaRecTrainBuilder(BaseDatasetBuilder):
+    train_dataset_cls = Im2LatexDataset
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/formula/formula_train.yaml"
+    }
+    LOG_INFO = "Formula Recgnition Train"
+
+    def build_datasets(self):
+        logging.info(f"Building {self.LOG_INFO} datasets ...")
+        self.build_processors()
+
+        build_info = self.config.build_info
+        anno_path = build_info.annotation,
+        vis_root = build_info.images
+        anno_path = [anno_path] if isinstance(anno_path, str) else anno_path
+        vis_root = [vis_root] if isinstance(vis_root, str) else vis_root
+        datasets = dict()
+
+        # create datasets
+        dataset_cls = self.train_dataset_cls
+        datasets['train'] = dataset_cls(
+            vis_processor=self.vis_processors["train"],
+            text_processor=self.text_processors["train"],
+            vis_root=vis_root,
+            anno_path=anno_path,
+        )
+        print(datasets['train'][0])
+
+        return datasets
+
+
+@registry.register_builder("multi_scale_formula_rec_train")
+class MultiScaleFormulaRecTrainBuilder(BaseDatasetBuilder):
+    train_dataset_cls = MultiScaleIm2LatexDataset
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/formula/multi_scale_formula_train.yaml"
+    }
+    LOG_INFO = "Multi Scale Formula Recgnition Train"
+
+    def build_datasets(self):
+        logging.info(f"Building {self.LOG_INFO} datasets ...")
+        self.build_processors()
+
+        build_info = self.config.build_info
+        anno_path = build_info.annotation,
+        vis_root = build_info.images
+
+        anno_path = [anno_path] if isinstance(anno_path, str) else anno_path
+        vis_root = [vis_root] if isinstance(vis_root, str) else vis_root
+
+        datasets = dict()
+
+        # create datasets
+        dataset_cls = self.train_dataset_cls
+        datasets['train'] = dataset_cls(
+            vis_processor=self.vis_processors["train"],
+            text_processor=self.text_processors["train"],
+            vis_root=vis_root,
+            anno_path=anno_path,
+        )
+        print(datasets['train'][0])
+
+        return datasets
+
+
+@registry.register_builder("formula_rec_eval")
+class FormulaRecEvalBuilder(BaseDatasetBuilder):
+    eval_dataset_cls = Im2LatexDataset
+    DATASET_CONFIG_DICT = {
+        "default": "configs/datasets/formula/formula_eval.yaml"
+    }
+    LOG_INFO = "Formula Recgnition Eval"
+
+    def build_datasets(self):
+        logging.info(f"Building {self.LOG_INFO} datasets ...")
+        self.build_processors()
+
+        build_info = self.config.build_info
+        anno_path = build_info.annotation,
+        vis_root = build_info.images
+
+        anno_path = [anno_path] if isinstance(anno_path, str) else anno_path
+        vis_root = [vis_root] if isinstance(vis_root, str) else vis_root
+
+        datasets = dict()
+
+        # create datasets
+        dataset_cls = self.eval_dataset_cls
+        datasets['eval'] = dataset_cls(
+            vis_processor=self.vis_processors["eval"],
+            text_processor=self.text_processors["eval"],
+            vis_root=vis_root,
+            anno_path=anno_path,
+        )
+        print(datasets['eval'][0])
+
+        return datasets

unimernet/datasets/data_utils.py

@@ -0,0 +1,284 @@
+"""
+ 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 gzip
+import logging
+import os
+import random as rnd
+import tarfile
+import zipfile
+
+import decord
+import webdataset as wds
+import numpy as np
+import torch
+from torch.utils.data.dataset import IterableDataset, ChainDataset
+from decord import VideoReader
+from unimernet.common.registry import registry
+from unimernet.datasets.datasets.base_dataset import ConcatDataset
+from tqdm import tqdm
+
+decord.bridge.set_bridge("torch")
+MAX_INT = registry.get("MAX_INT")
+
+
+def load_video(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling="uniform"):
+    vr = VideoReader(uri=video_path, height=height, width=width)
+
+    vlen = len(vr)
+    start, end = 0, vlen
+
+    n_frms = min(n_frms, vlen)
+
+    if sampling == "uniform":
+        indices = np.arange(start, end, vlen / n_frms).astype(int)
+    elif sampling == "headtail":
+        indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))
+        indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))
+        indices = indices_h + indices_t
+    else:
+        raise NotImplementedError
+
+    # get_batch -> T, H, W, C
+    frms = vr.get_batch(indices).permute(3, 0, 1, 2).float()  # (C, T, H, W)
+
+    return frms
+
+
+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
+            chained_datasets = (
+                ChainDataset(iterable_datasets) if len(iterable_datasets) > 0 else 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
+
+
+def extract_archive(from_path, to_path=None, overwrite=False):
+    """Extract archive.
+
+    Args:
+        from_path: the path of the archive.
+        to_path: the root path of the extracted files (directory of from_path)
+        overwrite: overwrite existing files (False)
+
+    Returns:
+        List of paths to extracted files even if not overwritten.
+
+    Examples:
+        >>> url = 'http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/validation.tar.gz'
+        >>> from_path = './validation.tar.gz'
+        >>> to_path = './'
+        >>> torchtext.utils.download_from_url(url, from_path)
+        >>> torchtext.utils.extract_archive(from_path, to_path)
+        >>> ['.data/val.de', '.data/val.en']
+        >>> torchtext.utils.download_from_url(url, from_path)
+        >>> torchtext.utils.extract_archive(from_path, to_path)
+        >>> ['.data/val.de', '.data/val.en']
+
+    """
+
+    if to_path is None:
+        to_path = os.path.dirname(from_path)
+
+    if from_path.endswith((".tar.gz", ".tgz")):
+        logging.info("Opening tar file {} to {}.".format(from_path, to_path))
+        with tarfile.open(from_path, "r") as tar:
+            files = []
+            for file_ in tqdm(tar):
+                file_path = os.path.join(to_path, file_.name)
+                if file_.isfile():
+                    files.append(file_path)
+                    if os.path.exists(file_path):
+                        logging.info("{} already extracted.".format(file_path))
+                        if not overwrite:
+                            continue
+                tar.extract(file_, to_path)
+            logging.info("Finished extracting tar file {}.".format(from_path))
+            return files
+
+    elif from_path.endswith(".zip"):
+        assert zipfile.is_zipfile(from_path), from_path
+        logging.info("Opening zip file {} to {}.".format(from_path, to_path))
+        with zipfile.ZipFile(from_path, "r") as zfile:
+            files = []
+            for file_ in tqdm(zfile.namelist()):
+                file_path = os.path.join(to_path, file_)
+                files.append(file_path)
+                if os.path.exists(file_path):
+                    logging.info("{} already extracted.".format(file_path))
+                    if not overwrite:
+                        continue
+                zfile.extract(file_, to_path)
+        files = [f for f in files if os.path.isfile(f)]
+        logging.info("Finished extracting zip file {}.".format(from_path))
+        return files
+
+    elif from_path.endswith(".gz"):
+        logging.info("Opening gz file {} to {}.".format(from_path, to_path))
+        default_block_size = 65536
+        filename = from_path[:-3]
+        files = [filename]
+        with gzip.open(from_path, "rb") as gzfile, open(filename, "wb") as d_file:
+            while True:
+                block = gzfile.read(default_block_size)
+                if not block:
+                    break
+                else:
+                    d_file.write(block)
+            d_file.write(block)
+        logging.info("Finished extracting gz file {}.".format(from_path))
+        return files
+
+    else:
+        raise NotImplementedError(
+            "We currently only support tar.gz, .tgz, .gz and zip achives."
+        )
+
+
+def save_frames_grid(img_array, out_path):
+    import torch
+    from PIL import Image
+    from torchvision.utils import make_grid
+
+    if len(img_array.shape) == 3:
+        img_array = img_array.unsqueeze(0)
+    elif len(img_array.shape) == 5:
+        b, t, c, h, w = img_array.shape
+        img_array = img_array.view(-1, c, h, w)
+    elif len(img_array.shape) == 4:
+        pass
+    else:
+        raise NotImplementedError(
+            "Supports only (b,t,c,h,w)-shaped inputs. First two dimensions can be ignored."
+        )
+
+    assert img_array.shape[1] == 3, "Exepcting input shape of (H, W, 3), i.e. RGB-only."
+
+    grid = make_grid(img_array)
+    ndarr = grid.permute(1, 2, 0).to("cpu", torch.uint8).numpy()
+
+    img = Image.fromarray(ndarr)
+
+    img.save(out_path)

unimernet/datasets/datasets/base_dataset.py

@@ -0,0 +1,103 @@
+import json
+from PIL import Image, ImageFile
+import os.path as osp
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+
+from io import BytesIO
+from typing import Iterable
+from torch.utils.data import Dataset, ConcatDataset
+import torch
+
+
+class BaseDataset(Dataset):
+
+    def __init__(self, vis_processor, text_processor, vis_root, anno_path):
+
+        self.vis_root = vis_root
+        # if isinstance(anno_path, tuple) or isinstance(anno_path, list):
+        #     anno_path = anno_path[0]
+        self.anno_path = anno_path
+
+        self.vis_processor = vis_processor
+        self.text_processor = text_processor
+
+        self.samples = self.init_samples()
+        self.reader = self.init_reader()
+
+        print('total {} {} samples'.format(self.__len__(), self.__class__.__name__))
+
+        for idx in range(10):
+            self.__getitem__(idx)
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, index):
+        raise NotImplementedError
+
+    def init_samples(self):
+        # read annotation from ceph
+        if self.anno_path.startswith('cluster'):
+            from petrel_client.client import Client
+            client = Client("~/petreloss.conf")
+            samples = json.loads(client.get(self.anno_path))
+        else:
+            samples = json.load(open(self.anno_path, 'r'))
+        return samples
+
+    def init_reader(self):
+        if self.vis_root.startswith('cluster'):
+            from petrel_client.client import Client
+            client = Client("~/petreloss.conf")
+            reader = {'type': 'PetrelReader', 'body': client.get}
+        else:
+            reader = {'type': 'LocalReader', 'body': Image.open}
+        return reader
+
+    def _read_image(self, sample, image_key="image"):
+        img_file = sample[image_key]
+        image_path = osp.join(self.vis_root, img_file)
+        image = self.reader['body'](image_path)
+        if isinstance(image, bytes):
+            bytes_stream = BytesIO(image)
+            image = Image.open(bytes_stream)
+        image = image.convert("RGB")
+        return image
+
+    def collater(self, samples):
+        image_list, question_list, answer_list = [], [], []
+
+        for sample in samples:
+            image_list.append(sample["image"])
+            question_list.append(sample["text_input"])
+            answer_list.append(sample["text_output"])
+
+        return {
+            "image": torch.stack(image_list, dim=0),
+            "text_input": question_list,
+            "text_output": answer_list,
+            "data_type": "vqa",
+        }
+
+
+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)

unimernet/datasets/datasets/dataloader_utils.py

@@ -0,0 +1,200 @@
+"""
+ 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 time
+import random
+import torch
+from unimernet.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])
+
+    def __len__(self):
+        return sum([len(_) for _ in self.loaders if hasattr(_, "__len__")])
+
+
+class ConcatLoader:
+    """
+    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):
+        # assert all loaders has __next__ method
+        for loader in loaders:
+            assert hasattr(
+                loader, "__len__"
+            ), "Loader {} has no __len__ method.".format(loader)
+
+        self._epoch = 0
+        self._loader_lens = [len(_) for _ in loaders]
+        self._rest_lens = self._loader_lens.copy()
+
+        self.loaders = loaders
+
+    def __next__(self):
+        # random sample from each loader by ratio
+        loader_idx = random.choices(range(len(self.loaders)), self._rest_lens, k=1)[0]
+        self._rest_lens[loader_idx] -= 1
+        if sum(self._rest_lens) == 0:
+            self._epoch += 1
+            self._rest_lens = self._loader_lens.copy()
+        return next(self.loaders[loader_idx])
+
+    def __len__(self):
+        return sum([len(_) for _ in self.loaders if hasattr(_, "__len__")])
+
+
+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)