Commit a61abe60 authored by narugo1992's avatar narugo1992
Browse files

dev(narugo): update, ci skip

parent 839d669f

docs/source/api_doc/metrics/ccip.benchmark.py

+4 −4
Original line number Diff line number Diff line 
import random



from benchmark import BaseBenchmark, create_plot_cli

from imgutils.metrics.ccip import get_ccip_features, get_ccip_similarity, _VALID_MODEL_NAMES

from imgutils.metrics.ccip import batch_ccip_features, get_ccip_difference, _VALID_MODEL_NAMES





class CCIPFeatureBenchmark(BaseBenchmark):
@@ -19,7 +19,7 @@ class CCIPFeatureBenchmark(BaseBenchmark): 


    def run(self):

        image_file = random.choice(self.all_images)

        _ = get_ccip_features([image_file], model_name=self.model_name)

        _ = batch_ccip_features([image_file], model_name=self.model_name)





class CCIPDiffBenchmark(BaseBenchmark):
@@ -28,7 +28,7 @@ class CCIPDiffBenchmark(BaseBenchmark): 
        self.model_name = model_name



    def prepare(self):

        self.feats = list(get_ccip_features(random.sample(self.all_images, k=30), model_name=self.model_name))

        self.feats = list(batch_ccip_features(random.sample(self.all_images, k=30), model_name=self.model_name))



    def load(self):

        from imgutils.metrics.ccip import _open_metric_model
@@ -41,7 +41,7 @@ class CCIPDiffBenchmark(BaseBenchmark): 
    def run(self):

        feat1 = random.choice(self.feats)

        feat2 = random.choice(self.feats)

        _ = get_ccip_similarity(feat1, feat2, model_name=self.model_name)

        _ = get_ccip_difference(feat1, feat2, model_name=self.model_name)





if __name__ == '__main__':

imgutils/metrics/__init__.py

+1 −0
Original line number Diff line number Diff line
@@ -3,5 +3,6 @@ Overview: 
    Tools for computing visual metrics.

"""

from .aesthetic import *

from .ccip import *

from .lpips import *

from .psnr_ import *

imgutils/metrics/ccip.py

+34 −45
Original line number Diff line number Diff line 
import json

from functools import lru_cache

from typing import Union, List



import numpy as np

from PIL import Image

from huggingface_hub import hf_hub_download

from sklearn.cluster import DBSCAN

from tqdm.auto import tqdm



from ..data import MultiImagesTyping, load_images, ImageTyping

from ..utils import open_onnx_model



__all__ = [

    'get_ccip_features',

    'get_ccip_similarity',

    'batch_ccip_similarity',

    'ccip_clustering',

    'get_ccip_feature',

    'batch_ccip_features',

    'get_ccip_difference',

    'batch_ccip_differences',

]
@@ -35,29 +34,39 @@ def _preprocess_image(image: Image.Image, size: int = 384): 
@lru_cache()

def _open_feat_model(model_name):

    return open_onnx_model(hf_hub_download(

        f'deepghs/imgutils-models',

        f'ccip/{model_name}_feat.onnx',

        f'deepghs/ccip_onnx',

        f'{model_name}/model_feat.onnx',

    ))





@lru_cache()

def _open_metric_model(model_name, safe: bool = False):

def _open_metric_model(model_name):

    return open_onnx_model(hf_hub_download(

        f'deepghs/imgutils-models',

        f'ccip/{model_name}_{"safe_" if safe else ""}metrics.onnx',

        f'deepghs/ccip_onnx',

        f'{model_name}/model_metrics.onnx',

    ))





_VALID_MODEL_NAMES = [

    # 'ccip-caformer-23_randaug_fp32',

    'ccip-caformer-5_fp32',

    'ccip-caformer-4_fp32',

    'ccip-caformer-2_fp32',

]

_DEFAULT_MODEL_NAMES = _VALID_MODEL_NAMES[0]

@lru_cache()

def _open_metrics(model_name):

    with open(hf_hub_download(f'deepghs/ccip_onnx', f'{model_name}/metrics.json'), 'r') as f:

        return json.load(f)





@lru_cache()

def _open_cluster_metrics(model_name):

    with open(hf_hub_download(f'deepghs/ccip_onnx', f'{model_name}/cluster.json'), 'r') as f:

        return json.load(f)





_DEFAULT_MODEL_NAMES = 'ccip-caformer-24-randaug-pruned'





def get_ccip_feature(image: ImageTyping, size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):

    return batch_ccip_features([image], size, model_name)[0]



def get_ccip_features(images: MultiImagesTyping, size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):



def batch_ccip_features(images: MultiImagesTyping, size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):

    images = load_images(images, mode='RGB')

    data = np.stack([_preprocess_image(item, size=size) for item in images]).astype(np.float32)

    output, = _open_feat_model(model_name).run(['output'], {'input': data})
@@ -73,7 +82,7 @@ def _preprocess_feats(x, size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES 
            if isinstance(item, np.ndarray):

                feats.append(item)

            else:

                feats.append(get_ccip_features(load_images([item]), size, model_name)[0])

                feats.append(batch_ccip_features(load_images([item]), size, model_name)[0])



        return np.stack(feats)

    else:
@@ -83,33 +92,13 @@ def _preprocess_feats(x, size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES 
_FeatureOrImage = Union[ImageTyping, np.ndarray]





def get_ccip_similarity(x: _FeatureOrImage, y: _FeatureOrImage, safe: bool = False,

def get_ccip_difference(x: _FeatureOrImage, y: _FeatureOrImage,

                        size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES) -> float:

    return batch_ccip_similarity([x, y], safe, size, model_name)[0, 1].item()

    return batch_ccip_differences([x, y], size, model_name)[0, 1].item()





def batch_ccip_similarity(images: Union[np.ndarray, List[_FeatureOrImage]], safe: bool = False,

def batch_ccip_differences(images: Union[np.ndarray, List[_FeatureOrImage]],

                           size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):

    input_ = _preprocess_feats(images, size, model_name).astype(np.float32)

    output, = _open_metric_model(model_name, safe=safe).run(['output'], {'input': input_})

    output, = _open_metric_model(model_name).run(['output'], {'input': input_})

    return output
 




def ccip_clustering(images: MultiImagesTyping, threshold: float = 0.6, min_samples: int = 2, safe: bool = True,

                    size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):

    images = load_images(images, mode='RGB')

    features = []

    for image in tqdm(images, desc='Feature Extract'):

        features.append(get_ccip_features([image], size, model_name)[0])



    if not features:

        return []

    feats = np.stack(features)

    differences = 1 - batch_ccip_similarity(feats, safe, size, model_name)



    def _metric(x, y):

        return differences[int(x), int(y)]



    samples = np.array(range(len(images))).reshape(-1, 1)

    clustering = DBSCAN(eps=1 - threshold, min_samples=min_samples, metric=_metric).fit(samples)

    return clustering.labels_.tolist()

requirements-zoo.txt

+2 −1
Original line number Diff line number Diff line
@@ -21,3 +21,4 @@ ultralytics 
controlnet_aux

lighttuner

natsort

tabulate
 
 No newline at end of file

zoo/ccip/plot.py

+1 −1
Original line number Diff line number Diff line
@@ -29,7 +29,7 @@ def plt_confusion_matrix(ax, y_true, y_pred, title: str = 'Confusion Matrix', 
    cm = confusion_matrix(y_true, y_pred, normalize=normalize)

    disp = ConfusionMatrixDisplay(

        confusion_matrix=cm,

        display_labels=['Similar', 'Diff'],

        display_labels=['Diff', 'Similar'],

    )

    disp.plot(ax=ax, cmap=cmap or plt.cm.Blues)

    ax.set_yticklabels(ax.get_yticklabels(), rotation=90)