app.py

import spaces
import gradio as gr
import copy
import random
import torch
import PIL
from PIL import Image, ImageDraw, ImageFont
import torchvision.transforms.functional as F
import numpy as np
import argparse
import json
import plotly.express as px
import pandas as pd
from util.slconfig import SLConfig, DictAction
from util.misc import nested_tensor_from_tensor_list
import datasets.transforms as T
import scipy.ndimage as ndimage
import matplotlib.pyplot as plt
# https://github.com/PhyscalX/gradio-image-prompter/tree/main/backend/gradio_image_prompter/templates/component
import io
from enum import Enum
import os
import subprocess
from subprocess import call
import shlex
os.environ["GRADIO_TEMP_DIR"] = os.path.join(os.getcwd(), "tmp")
cwd = os.getcwd()
print("Current working directory:", cwd)

# Installing dependencies not in requirements.txt
subprocess.run(
    shlex.split(
        "pip install gradio_image_prompter-0.1.0-py3-none-any.whl"
    )
)
from gradio_image_prompter import ImagePrompter

subprocess.run(
    shlex.split(
        "pip install MultiScaleDeformableAttention-1.0-cp310-cp310-linux_x86_64.whl"
    )
)

class AppSteps(Enum):
    JUST_TEXT = 1
    TEXT_AND_EXEMPLARS = 2
    JUST_EXEMPLARS = 3
    FULL_APP = 4

CONF_THRESH = 0.23

# MODEL:
def get_args_parser():
    """
    Example eval command:

    >> python main.py --output_dir ./gdino_test -c config/cfg_fsc147_vit_b_test.py --eval --datasets config/datasets_fsc147.json --pretrain_model_path ../checkpoints_and_logs/gdino_train/checkpoint_best_regular.pth --options text_encoder_type=checkpoints/bert-base-uncased --sam_tt_norm --crop
    """
    parser = argparse.ArgumentParser("Set transformer detector", add_help=False)
    parser.add_argument(
        "--device", default="cuda", help="device to use for inference"
    )
    parser.add_argument(
        "--options",
        nargs="+",
        action=DictAction,
        help="override some settings in the used config, the key-value pair "
        "in xxx=yyy format will be merged into config file.",
    )

    # dataset parameters
    parser.add_argument("--remove_difficult", action="store_true")
    parser.add_argument("--fix_size", action="store_true")

    # training parameters
    parser.add_argument("--note", default="", help="add some notes to the experiment")
    parser.add_argument("--resume", default="", help="resume from checkpoint")
    parser.add_argument(
        "--pretrain_model_path",
        help="load from other checkpoint",
        default="checkpoint_best_regular.pth",
    )
    parser.add_argument("--finetune_ignore", type=str, nargs="+")
    parser.add_argument(
        "--start_epoch", default=0, type=int, metavar="N", help="start epoch"
    )
    parser.add_argument("--eval", action="store_false")
    parser.add_argument("--num_workers", default=8, type=int)
    parser.add_argument("--test", action="store_true")
    parser.add_argument("--debug", action="store_true")
    parser.add_argument("--find_unused_params", action="store_true")
    parser.add_argument("--save_results", action="store_true")
    parser.add_argument("--save_log", action="store_true")

    # distributed training parameters
    parser.add_argument(
        "--world_size", default=1, type=int, help="number of distributed processes"
    )
    parser.add_argument(
        "--dist_url", default="env://", help="url used to set up distributed training"
    )
    parser.add_argument(
        "--rank", default=0, type=int, help="number of distributed processes"
    )
    parser.add_argument(
        "--local_rank", type=int, help="local rank for DistributedDataParallel"
    )
    parser.add_argument(
        "--local-rank", type=int, help="local rank for DistributedDataParallel"
    )
    parser.add_argument("--amp", action="store_true", help="Train with mixed precision")
    return parser

def get_device():
    if torch.cuda.is_available():
        print("USING GPU")
        return torch.device('cuda')
    else:
        print("USING CPU")
        return torch.device('cpu')

# Get counting model.
def build_model_and_transforms(args):
    normalize = T.Compose(
        [T.ToTensor(), T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]
    )
    data_transform = T.Compose(
        [
            T.RandomResize([800], max_size=1333),
            normalize,
        ]
    )
    cfg = SLConfig.fromfile("cfg_app.py")
    cfg.merge_from_dict({"text_encoder_type": "checkpoints/bert-base-uncased"})
    cfg_dict = cfg._cfg_dict.to_dict()
    args_vars = vars(args)
    for k, v in cfg_dict.items():
        if k not in args_vars:
            setattr(args, k, v)
        else:
            raise ValueError("Key {} can used by args only".format(k))

    # fix the seed for reproducibility
    seed = 42
    torch.manual_seed(seed)
    np.random.seed(seed)
    random.seed(seed)

    # we use register to maintain models from catdet6 on.
    from models.registry import MODULE_BUILD_FUNCS

    assert args.modelname in MODULE_BUILD_FUNCS._module_dict

    build_func = MODULE_BUILD_FUNCS.get(args.modelname)
    model, _, _ = build_func(args)

    checkpoint = torch.load(args.pretrain_model_path, map_location="cpu")["model"]
    model.load_state_dict(checkpoint, strict=False)

    model.eval()

    return model, data_transform


parser = argparse.ArgumentParser("Counting Application", parents=[get_args_parser()])
args = parser.parse_args()

device = get_device()
model, transform = build_model_and_transforms(args)

examples = [
    ["strawberry.jpg", "strawberry", {"image": "strawberry.jpg"}],
    ["strawberry.jpg", "blueberry", {"image": "strawberry.jpg"}],
    ["bird-1.JPG", "bird", {"image": "bird-2.JPG"}],
    ["fish.jpg", "fish", {"image": "fish.jpg"}],
    ["women.jpg", "girl", {"image": "women.jpg"}],
    ["women.jpg", "boy", {"image": "women.jpg"}],
    ["balloon.jpg", "hot air balloon", {"image": "balloon.jpg"}],
    ["deer.jpg", "deer", {"image": "deer.jpg"}],
    ["apple.jpg", "apple", {"image": "apple.jpg"}],
    ["egg.jpg", "egg", {"image": "egg.jpg"}],
    ["stamp.jpg", "stamp", {"image": "stamp.jpg"}],
    ["green-pea.jpg", "green pea", {"image": "green-pea.jpg"}],
    ["lego.jpg", "lego", {"image": "lego.jpg"}]
]

# APP:
def get_box_inputs(prompts):
    box_inputs = []
    for prompt in prompts:
        if prompt[2] == 2.0 and prompt[5] == 3.0:
            box_inputs.append([prompt[0], prompt[1], prompt[3], prompt[4]])

    return box_inputs

def get_ind_to_filter(text, word_ids, keywords):
    if len(keywords) <= 0:
        return list(range(len(word_ids)))
    input_words = text.split()
    keywords = keywords.split(",")
    keywords = [keyword.strip() for keyword in keywords]

    word_inds = []
    for keyword in keywords:
        if keyword in input_words:
            if len(word_inds) <= 0:
                ind = input_words.index(keyword)
                word_inds.append(ind)
            else:
                ind = input_words.index(keyword, word_inds[-1])
                word_inds.append(ind)
        else:
            raise Exception("Only specify keywords in the input text!")

    inds_to_filter = []
    for ind in range(len(word_ids)):
        word_id = word_ids[ind]
        if word_id in word_inds:
            inds_to_filter.append(ind)

    return inds_to_filter

@spaces.GPU
def count(image, text, prompts, state, device):

    model.to(device)
    
    keywords = "" # do not handle this for now
    
    # Handle no prompt case.
    if prompts is None:
        prompts = {"image": image, "points": []}
    input_image, _ = transform(image, {"exemplars": torch.tensor([])})
    input_image = input_image.unsqueeze(0).to(device)
    exemplars = get_box_inputs(prompts["points"])
    
    input_image_exemplars, exemplars = transform(prompts["image"], {"exemplars": torch.tensor(exemplars)})
    input_image_exemplars = input_image_exemplars.unsqueeze(0).to(device)
    exemplars = [exemplars["exemplars"].to(device)]

    print("model device: " + str(next(model.parameters()).device))
    print("input image device: " + str(input_image.device))
    print("input image exemplars device: " + str(input_image_exemplars.device))
    print("exemplars device: " + str(exemplars[0].device))
    
    with torch.no_grad():
        model_output = model(
                nested_tensor_from_tensor_list(input_image),
                nested_tensor_from_tensor_list(input_image_exemplars),
                exemplars,
                [torch.tensor([0]).to(device) for _ in range(len(input_image))],
                captions=[text + " ."] * len(input_image),
            )
    
    ind_to_filter = get_ind_to_filter(text, model_output["token"][0].word_ids, keywords)
    logits = model_output["pred_logits"].sigmoid()[0][:, ind_to_filter]
    boxes = model_output["pred_boxes"][0]
    if len(keywords.strip()) > 0:
        box_mask = (logits > CONF_THRESH).sum(dim=-1) == len(ind_to_filter)
    else:
        box_mask = logits.max(dim=-1).values > CONF_THRESH
    logits = logits[box_mask, :].cpu().numpy()
    boxes = boxes[box_mask, :].cpu().numpy()
    
    # Plot results.
    (w, h) = image.size
    det_map = np.zeros((h, w))
    det_map[(h * boxes[:, 1]).astype(int), (w * boxes[:, 0]).astype(int)] = 1
    det_map = ndimage.gaussian_filter(
        det_map, sigma=(w // 200, w // 200), order=0
    )
    plt.imshow(image)
    plt.imshow(det_map[None, :].transpose(1, 2, 0), 'jet', interpolation='none', alpha=0.7)
    plt.axis('off')
    img_buf = io.BytesIO()
    plt.savefig(img_buf, format='png', bbox_inches='tight')

    output_img = Image.open(img_buf)

    if AppSteps.TEXT_AND_EXEMPLARS not in state:
        exemplar_image = ImagePrompter(type='pil', label='Visual Exemplar Image', value=prompts, interactive=True, visible=True)
        new_submit_btn = gr.Button("Count", variant="primary", interactive=False)
        state = [AppSteps.JUST_TEXT, AppSteps.TEXT_AND_EXEMPLARS]
        main_instructions_comp = gr.Markdown(visible=False)
        step_3 = gr.Tab(visible=False)
    elif AppSteps.FULL_APP not in state:
        exemplar_image = ImagePrompter(type='pil', label='Visual Exemplar Image', value=prompts, interactive=True, visible=True)
        new_submit_btn = submit_btn
        state = [AppSteps.JUST_TEXT, AppSteps.TEXT_AND_EXEMPLARS, AppSteps.FULL_APP]
        main_instructions_comp = gr.Markdown(visible=True)
        step_3 = gr.Tab(visible=True)
    else:
        exemplar_image = ImagePrompter(type='pil', label='Visual Exemplar Image', value=prompts, interactive=True, visible=True)
        new_submit_btn = submit_btn
        main_instructions_comp = gr.Markdown(visible=True)
        step_3 = gr.Tab(visible=True)

    out_label = "Detected instances predicted with"
    if len(text.strip()) > 0:
        out_label += " text"
        if exemplars[0].size()[0] == 1:
            out_label += " and " + str(exemplars[0].size()[0]) + " visual exemplar."        
        elif exemplars[0].size()[0] > 1:
            out_label += " and " + str(exemplars[0].size()[0]) + " visual exemplars."
        else:
            out_label += "."
    elif exemplars[0].size()[0] > 0:
        if exemplars[0].size()[0] == 1:
            out_label += " " + str(exemplars[0].size()[0]) + " visual exemplar."
        else:
            out_label += " " + str(exemplars[0].size()[0]) + " visual exemplars."
    else:
        out_label = "Nothing specified to detect."
 
    return (gr.Image(output_img, visible=True, label=out_label, show_label=True), gr.Number(label="Predicted Count", visible=True, value=boxes.shape[0]), new_submit_btn, gr.Tab(visible=True), step_3, state)

@spaces.GPU
def count_main(image, text, prompts, device):
    model.to(device)
    
    keywords = "" # do not handle this for now
    # Handle no prompt case.
    if prompts is None:
        prompts = {"image": image, "points": []}
    input_image, _ = transform(image, {"exemplars": torch.tensor([])})
    input_image = input_image.unsqueeze(0).to(device)
    exemplars = get_box_inputs(prompts["points"])
    
    input_image_exemplars, exemplars = transform(prompts["image"], {"exemplars": torch.tensor(exemplars)})
    input_image_exemplars = input_image_exemplars.unsqueeze(0).to(device)
    exemplars = [exemplars["exemplars"].to(device)]
    
    with torch.no_grad():
        model_output = model(
                nested_tensor_from_tensor_list(input_image),
                nested_tensor_from_tensor_list(input_image_exemplars),
                exemplars,
                [torch.tensor([0]).to(device) for _ in range(len(input_image))],
                captions=[text + " ."] * len(input_image),
            )
    
    ind_to_filter = get_ind_to_filter(text, model_output["token"][0].word_ids, keywords)
    logits = model_output["pred_logits"].sigmoid()[0][:, ind_to_filter]
    boxes = model_output["pred_boxes"][0]
    if len(keywords.strip()) > 0:
        box_mask = (logits > CONF_THRESH).sum(dim=-1) == len(ind_to_filter)
    else:
        box_mask = logits.max(dim=-1).values > CONF_THRESH
    logits = logits[box_mask, :].cpu().numpy()
    boxes = boxes[box_mask, :].cpu().numpy()
    
    # Plot results.
    (w, h) = image.size
    det_map = np.zeros((h, w))
    det_map[(h * boxes[:, 1]).astype(int), (w * boxes[:, 0]).astype(int)] = 1
    det_map = ndimage.gaussian_filter(
        det_map, sigma=(w // 200, w // 200), order=0
    )
    plt.imshow(image)
    plt.imshow(det_map[None, :].transpose(1, 2, 0), 'jet', interpolation='none', alpha=0.7)
    plt.axis('off')
    img_buf = io.BytesIO()
    plt.savefig(img_buf, format='png', bbox_inches='tight')

    output_img = Image.open(img_buf)

    out_label = "Detected instances predicted with"
    if len(text.strip()) > 0:
        out_label += " text"
        if exemplars[0].size()[0] == 1:
            out_label += " and " + str(exemplars[0].size()[0]) + " visual exemplar."        
        elif exemplars[0].size()[0] > 1:
            out_label += " and " + str(exemplars[0].size()[0]) + " visual exemplars."
        else:
            out_label += "."
    elif exemplars[0].size()[0] > 0:
        if exemplars[0].size()[0] == 1:
            out_label += " " + str(exemplars[0].size()[0]) + " visual exemplar."
        else:
            out_label += " " + str(exemplars[0].size()[0]) + " visual exemplars."
    else:
        out_label = "Nothing specified to detect."
    
    return (gr.Image(output_img, visible=True, label=out_label, show_label=True), gr.Number(label="Predicted Count", visible=True, value=boxes.shape[0]))

def remove_label(image):
    return gr.Image(show_label=False)

def check_submit_btn(exemplar_image_prompts, state):
    if AppSteps.TEXT_AND_EXEMPLARS not in state or len(state) == 3:
        return gr.Button("Count", variant="primary", interactive=True)
    elif exemplar_image_prompts is None:
        return gr.Button("Count", variant="primary", interactive=False)
    elif len(get_box_inputs(exemplar_image_prompts["points"])) > 0:
        return gr.Button("Count", variant="primary", interactive=True)
    else:
        return gr.Button("Count", variant="primary", interactive=False)

exemplar_img_drawing_instructions_part_1 = '<p><strong>Congrats, you have counted the strawberries!</strong> You can also draw a box around the object you want to count. <strong>Click and drag the mouse on the image below to draw a box around one of the strawberries.</strong> You can click the back button in the top right of the image to delete the box and try again.<img src="file/button-legend.jpg" width="750"></p>'
exemplar_img_drawing_instructions_part_2 = '<p>The boxes you draw are called \"visual exemplars,\" image examples of what you want the model to count. You can add more boxes around more examples of strawberries in the image above to increase the accuracy of the predicted count. You can also use strawberries from a different image to specify the object to count by uploading or pasting a new image above and drawing boxes around strawberries in it.</p>'
instructions_main = """
# How to Use the App
As shown earlier, there are 3 ways to specify the object to count: (1) with text only, (2) with text and any number of boxes (i.e., "visual exemplars") around example objects, and (3) with visual exemplars only. What is being used is indicated in the top left of the output image. How to try each case is detailed below.

<ol>
  <li><strong>Text Only: </strong> Only provide text describing the object to count in the textbox titled "What would you like to count?" Delete all boxes drawn on the visual exemplar image.</li>
  <li><strong>Text + Visual Exemplars: </strong> Provide text describing the object to count in the textbox titled "What would you like to count?" and draw at least one box around an example object in the visual exemplar image.</li>
  <li><strong>Visual Exemplars Only: </strong> Remove all text in the textbox titled "What would you like to count?" and draw at least one box around an example object in the visual exemplar image.</li>
</ol>

## Click on the "App" tab at the top of the screen to exit the tutorial and start using the main app!
"""

with gr.Blocks(title="CountGD: Multi-Modal Open-World Counting", theme="soft", head="""<meta name="viewport" content="width=device-width, initial-scale=1, user-scalable=1">""") as demo:
    state = gr.State(value=[AppSteps.JUST_TEXT])
    device = gr.State(device)
    with gr.Tab("Tutorial"):
        with gr.Row():
            with gr.Column():
                with gr.Tab("Step 3", visible=False) as step_3:
                    main_instructions = gr.Markdown(instructions_main)
                with gr.Tab("Step 2", visible=False) as step_2:
                    gr.Markdown(exemplar_img_drawing_instructions_part_1)
                    exemplar_image = ImagePrompter(type='pil', label='Visual Exemplar Image', show_label=True, value={"image": "strawberry.jpg", "points": []}, interactive=True)
                    with gr.Accordion("Open for Further Information", open=False):
                        gr.Markdown(exemplar_img_drawing_instructions_part_2)
                with gr.Tab("Step 1", visible=True) as step_1:
                    input_image = gr.Image(type='pil', label='Input Image', show_label='True', value="strawberry.jpg", interactive=False, width="30vw")
                    gr.Markdown('# Click "Count" to count the strawberries.')
                
            with gr.Column():
                with gr.Tab("Output Image"):
                    detected_instances = gr.Image(label="Detected Instances", show_label='True', interactive=False, visible=True, width="40vw")
                    
        with gr.Row():
            input_text = gr.Textbox(label="What would you like to count?", value="strawberry", interactive=True)
            pred_count = gr.Number(label="Predicted Count", visible=False)
        submit_btn = gr.Button("Count", variant="primary", interactive=True)
                
        submit_btn.click(fn=remove_label, inputs=[detected_instances], outputs=[detected_instances]).then(fn=count, inputs=[input_image, input_text, exemplar_image, state, device], outputs=[detected_instances, pred_count, submit_btn, step_2, step_3, state])
        exemplar_image.change(check_submit_btn, inputs=[exemplar_image, state], outputs=[submit_btn])
    with gr.Tab("App", visible=True) as main_app:
        
        gr.Markdown(
              """
              # <center>CountGD: Multi-Modal Open-World Counting
              <center><h3>Count objects with text, visual exemplars, or both together.</h3>
              <h3>Scroll down to try more examples</h3>
              <h3><a href='https://arxiv.org/abs/2407.04619' target='_blank' rel='noopener'>[paper]</a>
                <a href='https://github.com/niki-amini-naieni/CountGD/' target='_blank' rel='noopener'>[code]</a></h3>
              Limitation: this app does not support fine-grained counting based on attributes or visual grounding inputs yet.</center>
              """
            )
            
        with gr.Row():
            with gr.Column():
              input_image_main = gr.Image(type='pil', label='Input Image', show_label='True', value="strawberry.jpg", interactive=True)
              input_text_main = gr.Textbox(label="What would you like to count?", placeholder="", value="strawberry")
              exemplar_image_main = ImagePrompter(type='pil', label='Visual Exemplar Image', show_label=True, value={"image": "strawberry.jpg", "points": []}, interactive=True)
            with gr.Column():
              detected_instances_main = gr.Image(label="Detected Instances", show_label='True', interactive=False)
              pred_count_main = gr.Number(label="Predicted Count")
              submit_btn_main = gr.Button("Count", variant="primary")
              clear_btn_main = gr.ClearButton(variant="secondary")
        gr.Examples(label="Examples: click on a row to load the example. Add visual exemplars by drawing boxes on the loaded \"Visual Exemplar Image.\"", examples=examples, inputs=[input_image_main, input_text_main, exemplar_image_main])
        submit_btn_main.click(fn=remove_label, inputs=[detected_instances_main], outputs=[detected_instances_main]).then(fn=count_main, inputs=[input_image_main, input_text_main, exemplar_image_main, device], outputs=[detected_instances_main, pred_count_main])
        clear_btn_main.add([input_image_main, input_text_main, exemplar_image_main, detected_instances_main, pred_count_main])
        

demo.launch(allowed_paths=['back-icon.jpg', 'paste-icon.jpg', 'upload-icon.jpg', 'button-legend.jpg'])