app.py

import json

import gradio as gr
from PIL import Image
import safetensors.torch
import spaces
import timm
from timm.models import VisionTransformer
import torch
from torchvision.transforms import transforms
from torchvision.transforms import InterpolationMode
import torchvision.transforms.functional as TF
from huggingface_hub import hf_hub_download
import numpy as np
import matplotlib.cm as cm

class Fit(torch.nn.Module):
    def __init__(
        self,
        bounds: tuple[int, int] | int,
        interpolation = InterpolationMode.LANCZOS,
        grow: bool = True,
        pad: float | None = None
    ):
        super().__init__()

        self.bounds = (bounds, bounds) if isinstance(bounds, int) else bounds
        self.interpolation = interpolation
        self.grow = grow
        self.pad = pad

    def forward(self, img: Image) -> Image:
        wimg, himg = img.size
        hbound, wbound = self.bounds

        hscale = hbound / himg
        wscale = wbound / wimg

        if not self.grow:
            hscale = min(hscale, 1.0)
            wscale = min(wscale, 1.0)

        scale = min(hscale, wscale)
        if scale == 1.0:
            return img

        hnew = min(round(himg * scale), hbound)
        wnew = min(round(wimg * scale), wbound)

        img = TF.resize(img, (hnew, wnew), self.interpolation)

        if self.pad is None:
            return img

        hpad = hbound - hnew
        wpad = wbound - wnew

        tpad = hpad // 2
        bpad = hpad - tpad

        lpad = wpad // 2
        rpad = wpad - lpad

        return TF.pad(img, (lpad, tpad, rpad, bpad), self.pad)

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}(" +
            f"bounds={self.bounds}, " +
            f"interpolation={self.interpolation.value}, " +
            f"grow={self.grow}, " +
            f"pad={self.pad})"
        )

class CompositeAlpha(torch.nn.Module):
    def __init__(
        self,
        background: tuple[float, float, float] | float,
    ):
        super().__init__()

        self.background = (background, background, background) if isinstance(background, float) else background
        self.background = torch.tensor(self.background).unsqueeze(1).unsqueeze(2)

    def forward(self, img: torch.Tensor) -> torch.Tensor:
        if img.shape[-3] == 3:
            return img

        alpha = img[..., 3, None, :, :]

        img[..., :3, :, :] *= alpha

        background = self.background.expand(-1, img.shape[-2], img.shape[-1])
        if background.ndim == 1:
            background = background[:, None, None]
        elif background.ndim == 2:
            background = background[None, :, :]

        img[..., :3, :, :] += (1.0 - alpha) * background
        return img[..., :3, :, :]

    def __repr__(self) -> str:
        return (
            f"{self.__class__.__name__}(" +
            f"background={self.background})"
        )

transform = transforms.Compose([
    Fit((384, 384)),
    transforms.ToTensor(),
    CompositeAlpha(0.5),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
    transforms.CenterCrop((384, 384)),
])

model = timm.create_model(
    "vit_so400m_patch14_siglip_384.webli",
    pretrained=False,
    num_classes=9083,
) # type: VisionTransformer

class GatedHead(torch.nn.Module):
    def __init__(self,
        num_features: int,
        num_classes: int
    ):
        super().__init__()
        self.num_classes = num_classes
        self.linear = torch.nn.Linear(num_features, num_classes * 2)

        self.act = torch.nn.Sigmoid()
        self.gate = torch.nn.Sigmoid()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.linear(x)
        x = self.act(x[:, :self.num_classes]) * self.gate(x[:, self.num_classes:])
        return x

model.head = GatedHead(min(model.head.weight.shape), 9083)

cached_model = hf_hub_download(
    repo_id="RedRocket/JointTaggerProject",
    subfolder="JTP_PILOT2",
    filename="JTP_PILOT2-e3-vit_so400m_patch14_siglip_384.safetensors"
)

safetensors.torch.load_model(model, cached_model)
model.eval()

with open("tagger_tags.json", "r") as file:
    tags = json.load(file) # type: dict
allowed_tags = list(tags.keys())

for idx, tag in enumerate(allowed_tags):
    allowed_tags[idx] = tag.replace("_", " ")

sorted_tag_score = {}
input_image = None


@spaces.GPU(duration=5)
def run_classifier(image, threshold):
    global sorted_tag_score, input_image
    input_image = image.convert('RGBA')
    img = input_image
    tensor = transform(img).unsqueeze(0)

    with torch.no_grad():
        probits = model(tensor)[0]
        values, indices = probits.topk(250)

    tag_score = dict()
    for i in range(indices.size(0)):
        tag_score[allowed_tags[indices[i]]] = values[i].item()
    sorted_tag_score = dict(sorted(tag_score.items(), key=lambda item: item[1], reverse=True))

    return create_tags(threshold)

def create_tags(threshold):
    global sorted_tag_score
    filtered_tag_score = {key: value for key, value in sorted_tag_score.items() if value > threshold}
    text_no_impl = ", ".join(filtered_tag_score.keys())
    return text_no_impl, filtered_tag_score

def clear_image():
    global sorted_tag_score, input_image
    input_image = None
    sorted_tag_score = {}
    return "", {}

target_tag_index = None

# Store hooks and intermediate values
gradients = {}
activations = {}

def hook_forward(module, input, output):
    activations['value'] = output

def hook_backward(module, grad_in, grad_out):
    gradients['value'] = grad_out[0]

def cam_inference(threshold, evt: gr.SelectData):
    target_tag = evt.value
    print(f"target_tag: {target_tag}")
    global input_image, sorted_tag_score, target_tag_index, gradients, activations
    img = input_image
    tensor = transform(img).unsqueeze(0)

    gradients = {}
    activations = {}
    cam = None
    target_tag_index = None

    if target_tag:
        if target_tag not in allowed_tags:
            print(f"Warning: Target tag '{target_tag}' not found in allowed tags.")
            target_tag = None
        else:
            target_tag_index = allowed_tags.index(target_tag)
            handle_forward = model.norm.register_forward_hook(hook_forward)
            handle_backward = model.norm.register_full_backward_hook(hook_backward)

    probits = model(tensor)[0].cpu()
    
    
    if target_tag is not None and target_tag_index is not None:
        model.zero_grad()
        target_score = probits[target_tag_index]
        target_score.backward(retain_graph=True)

        grads = gradients.get('value')
        acts = activations.get('value')

        if grads is not None and acts is not None:
            patch_grads = grads
            patch_acts = acts

            weights = torch.mean(patch_grads, dim=1).squeeze(0)

            cam_1d = torch.einsum('pe,e->p', patch_acts.squeeze(0), weights)
            cam_1d = torch.relu(cam_1d)

            cam = cam_1d.reshape(27, 27).detach().cpu().numpy()


        handle_forward.remove()
        handle_backward.remove()
        gradients = {}
        activations = {}

    return create_cam_visualization_pil(cam, vis_threshold=threshold)

def create_cam_visualization_pil(cam, alpha=0.6, vis_threshold=0.2):
    """
    Overlays CAM on image and returns a PIL image.

    Args:
        image_pil: PIL Image (RGB)
        cam: 2D numpy array (activation map)
        alpha: float, blending factor
        vis_threshold: float, minimum normalized CAM value to show color

    Returns:
        PIL.Image.Image with overlay
    """

    global input_image
    # Convert to RGB (in case RGBA or others)
    image_pil = input_image
    w, h = image_pil.size

    # Resize CAM to match image
    cam_resized = np.array(Image.fromarray(cam).resize((w, h), resample=Image.Resampling.BILINEAR))

    # Normalize CAM to [0, 1]
    cam_norm = (cam_resized - cam_resized.min()) / (np.ptp(cam_resized) + 1e-8)

    # Create heatmap using matplotlib colormap
    colormap = cm.get_cmap('jet')
    cam_colored = colormap(cam_norm)[:, :, :3]  # RGB
    cam_alpha = (cam_norm >= vis_threshold).astype(np.float32) * alpha  # Alpha mask

    cam_rgba = np.dstack((cam_colored, cam_alpha))  # Shape: (H, W, 4)
    cam_image = Image.fromarray((cam_rgba * 255).astype(np.uint8), mode="RGBA")

    # Composite over original
    composite = Image.alpha_composite(image_pil, cam_image)

    return composite


with gr.Blocks(css=".output-class { display: none; }") as demo:
    gr.Markdown("""
    ## Joint Tagger Project: JTP-PILOT² Demo **BETA**
    This tagger is designed for use on furry images (though may very well work on out-of-distribution images, potentially with funny results).  A threshold of 0.2 is recommended.  Lower thresholds often turn up more valid tags, but can also result in some amount of hallucinated tags.

    This tagger is the result of joint efforts between members of the RedRocket team, with distinctions given to Thessalo for creating the foundation for this project with his efforts, RedHotTensors for redesigning the process into a second-order method that models information expectation, and drhead for dataset prep, creation of training code and supervision of training runs.

    Special thanks to Minotoro at frosting.ai for providing the compute power for this project.
    """)
    with gr.Row():
        with gr.Column():
            image_input = gr.Image(label="Source", sources=['upload'], type='pil', height=512, show_label=False)
            threshold_slider = gr.Slider(minimum=0.00, maximum=1.00, step=0.01, value=0.20, label="Threshold")
        with gr.Column():
            tag_string = gr.Textbox(label="Tag String")
            label_box = gr.Label(label="Tag Predictions", num_top_classes=250, show_label=False)

    image_input.upload(
        fn=run_classifier,
        inputs=[image_input, threshold_slider],
        outputs=[tag_string, label_box]
    )

    image_input.clear(
        fn=clear_image,
        inputs=[],
        outputs=[tag_string, label_box]
    )

    threshold_slider.input(
        fn=create_tags,
        inputs=[threshold_slider],
        outputs=[tag_string, label_box]
    )

    label_box.select(
        fn=cam_inference,
        inputs=[threshold_slider],
        outputs=[image_input]
    )

if __name__ == "__main__":
    demo.launch()