Feat/content analyser pro (#859)

* Update to Yolo powered content analyser

* Update to Yolo powered content analyser

* Fix typing

* Drop bounding boxes and NMS check

* Drop bounding boxes and NMS check

* Fix CI

facefusion/content_analyser.py

@@ -1,6 +1,6 @@
 from functools import lru_cache
+from typing import List
 
-import cv2
 import numpy
 from tqdm import tqdm
 
@@ -8,11 +8,9 @@ from facefusion import inference_manager, state_manager, wording
 from facefusion.download import conditional_download_hashes, conditional_download_sources, resolve_download_url
 from facefusion.filesystem import resolve_relative_path
 from facefusion.thread_helper import conditional_thread_semaphore
-from facefusion.typing import DownloadScope, Fps, InferencePool, ModelOptions, ModelSet, VisionFrame
-from facefusion.vision import detect_video_fps, get_video_frame, read_image
+from facefusion.typing import Detection, DownloadScope, Fps, InferencePool, ModelOptions, ModelSet, Score, VisionFrame
+from facefusion.vision import detect_video_fps, get_video_frame, read_image, resize_frame_resolution
 
-PROBABILITY_LIMIT = 0.80
-RATE_LIMIT = 10
 STREAM_COUNTER = 0
 
 
@@ -20,26 +18,25 @@ STREAM_COUNTER = 0
 def create_static_model_set(download_scope : DownloadScope) -> ModelSet:
 	return\
 	{
-		'open_nsfw':
+		'yolo_nsfw':
 		{
 			'hashes':
 			{
 				'content_analyser':
 				{
-					'url': resolve_download_url('models-3.0.0', 'open_nsfw.hash'),
-					'path': resolve_relative_path('../.assets/models/open_nsfw.hash')
+					'url': resolve_download_url('models-3.2.0', 'yolo_11m_nsfw.hash'),
+					'path': resolve_relative_path('../.assets/models/yolo_11m_nsfw.hash')
 				}
 			},
 			'sources':
 			{
 				'content_analyser':
 				{
-					'url': resolve_download_url('models-3.0.0', 'open_nsfw.onnx'),
-					'path': resolve_relative_path('../.assets/models/open_nsfw.onnx')
+					'url': resolve_download_url('models-3.2.0', 'yolo_11m_nsfw.onnx'),
+					'path': resolve_relative_path('../.assets/models/yolo_11m_nsfw.onnx')
 				}
 			},
-			'size': (224, 224),
-			'mean': [ 104, 117, 123 ]
+			'size': (640, 640)
 		}
 	}
 
@@ -54,7 +51,7 @@ def clear_inference_pool() -> None:
 
 
 def get_model_options() -> ModelOptions:
-	return create_static_model_set('full').get('open_nsfw')
+	return create_static_model_set('full').get('yolo_nsfw')
 
 
 def pre_check() -> bool:
@@ -74,31 +71,9 @@ def analyse_stream(vision_frame : VisionFrame, video_fps : Fps) -> bool:
 
 
 def analyse_frame(vision_frame : VisionFrame) -> bool:
-	vision_frame = prepare_frame(vision_frame)
-	probability = forward(vision_frame)
+	nsfw_scores = detect_nsfw(vision_frame)
 
-	return probability > PROBABILITY_LIMIT
-
-
-def forward(vision_frame : VisionFrame) -> float:
-	content_analyser = get_inference_pool().get('content_analyser')
-
-	with conditional_thread_semaphore():
-		probability = content_analyser.run(None,
-		{
-			'input': vision_frame
-		})[0][0][1]
-
-	return probability
-
-
-def prepare_frame(vision_frame : VisionFrame) -> VisionFrame:
-	model_size = get_model_options().get('size')
-	model_mean = get_model_options().get('mean')
-	vision_frame = cv2.resize(vision_frame, model_size).astype(numpy.float32)
-	vision_frame -= numpy.array(model_mean).astype(numpy.float32)
-	vision_frame = numpy.expand_dims(vision_frame, axis = 0)
-	return vision_frame
+	return len(nsfw_scores) > 0
 
 
 @lru_cache(maxsize = None)
@@ -115,12 +90,52 @@ def analyse_video(video_path : str, trim_frame_start : int, trim_frame_end : int
 	counter = 0
 
 	with tqdm(total = len(frame_range), desc = wording.get('analysing'), unit = 'frame', ascii = ' =', disable = state_manager.get_item('log_level') in [ 'warn', 'error' ]) as progress:
+
 		for frame_number in frame_range:
 			if frame_number % int(video_fps) == 0:
 				vision_frame = get_video_frame(video_path, frame_number)
 				if analyse_frame(vision_frame):
 					counter += 1
 			rate = counter * int(video_fps) / len(frame_range) * 100
-			progress.update()
 			progress.set_postfix(rate = rate)
-	return rate > RATE_LIMIT
+			progress.update()
+
+	return rate > 10.0
+
+
+def detect_nsfw(vision_frame : VisionFrame) -> List[Score]:
+	nsfw_scores = []
+	model_size = get_model_options().get('size')
+	temp_vision_frame = resize_frame_resolution(vision_frame, model_size)
+	detect_vision_frame = prepare_detect_frame(temp_vision_frame)
+	detection = forward(detect_vision_frame)
+	detection = numpy.squeeze(detection).T
+	nsfw_scores_raw = numpy.amax(detection[:, 4:], axis = 1)
+	keep_indices = numpy.where(nsfw_scores_raw > 0.2)[0]
+
+	if numpy.any(keep_indices):
+		nsfw_scores_raw = nsfw_scores_raw[keep_indices]
+		nsfw_scores = nsfw_scores_raw.ravel().tolist()
+
+	return nsfw_scores
+
+
+def forward(vision_frame : VisionFrame) -> Detection:
+	content_analyser = get_inference_pool().get('content_analyser')
+
+	with conditional_thread_semaphore():
+		detection = content_analyser.run(None,
+		{
+			'input': vision_frame
+		})
+
+	return detection
+
+
+def prepare_detect_frame(temp_vision_frame : VisionFrame) -> VisionFrame:
+	model_size = get_model_options().get('size')
+	detect_vision_frame = numpy.zeros((model_size[0], model_size[1], 3))
+	detect_vision_frame[:temp_vision_frame.shape[0], :temp_vision_frame.shape[1], :] = temp_vision_frame
+	detect_vision_frame = detect_vision_frame / 255.0
+	detect_vision_frame = numpy.expand_dims(detect_vision_frame.transpose(2, 0, 1), axis = 0).astype(numpy.float32)
+	return detect_vision_frame

facefusion/face_helper.py

@@ -208,9 +208,9 @@ def estimate_face_angle(face_landmark_68 : FaceLandmark68) -> Angle:
 	return face_angle
 
 
-def apply_nms(bounding_boxes : List[BoundingBox], face_scores : List[Score], score_threshold : float, nms_threshold : float) -> Sequence[int]:
+def apply_nms(bounding_boxes : List[BoundingBox], scores : List[Score], score_threshold : float, nms_threshold : float) -> Sequence[int]:
 	normed_bounding_boxes = [ (x1, y1, x2 - x1, y2 - y1) for (x1, y1, x2, y2) in bounding_boxes ]
-	keep_indices = cv2.dnn.NMSBoxes(normed_bounding_boxes, face_scores, score_threshold = score_threshold, nms_threshold = nms_threshold)
+	keep_indices = cv2.dnn.NMSBoxes(normed_bounding_boxes, scores, score_threshold = score_threshold, nms_threshold = nms_threshold)
 	return keep_indices