[Python] how to covert a frame (np array) to a pytorch tensor

i need to convert a frame of the video (which is a nparray) to a pytorch tensor, do some particular actions with it and convert it back but i'm struggling

so, i have a frame returned from video_capture.read() and, as i understood, it's a np array. firstly, i convert it to a tensor and checks if looks correctly (sorry i can't add photos for some reason) then i analyze it (no mistakes), try to rotate it and here's a problem.

1. frame.show() shows the tensor in different colours, looks super wrong
2. self.check_tensor(rotated_tensor) after rotation shows just a black screen

can somebody please help me to fix this, i'm so exhausted, chatgpt confuses me even more and don't understand anything... i guess the problem with colours is related to how i convert tensor to pil image, but i tried several changes (commented lines) and nothing hepled. also is there a way to avoid converting tensor to pil image before the rotation? can't i just rotate a tensor?

def tensor_to_image(tensor):
tensor = (tensor * 255).byte()
tensor = tensor.squeeze(0)
tensor = tensor.permute(1, 2, 0)
image = Image.fromarray(np.array(tensor).astype(np.uint8))
image = cv2.cvtColor(np.asarray(image), cv2.COLOR_BGR2RGB)
image = Image.fromarray(np.asarray(image))
return image

def rotate_tensor(frame_tensor, landmarks):
roll = calc_face_angle(landmarks)
frame = tf.to_pil_image(frame_tensor.squeeze(0))
#frame = tensor_to_image(frame_tensor)
frame.show()

if not np.isnan(roll):
rotated_frame = frame.rotate(roll, resample=Image.BICUBIC, expand=True)
else:
print("Failed to calculate face angle for rotation")
return frame_tensor

#rotated_tensor = tf.to_tensor(rotated_frame).unsqueeze(0)

transform = transforms.ToTensor() # Используем torchvision для преобразования в тензор
rotated_tensor = transform(rotated_frame).unsqueeze(0)
return rotated_tensor


def check_tensor(self, frame_tensor):
frame_numpy = frame_tensor.squeeze(0).permute(1, 2, 0).byte().numpy()
#frame_numpy = cv2.cvtColor(frame_numpy, cv2.COLOR_RGB2BGR)

cv2.imshow("Frame", frame_numpy)
cv2.waitKey(0)
cv2.destroyAllWindows()


def analyze_video(self, video_path):
video_capture = cv2.VideoCapture(video_path)

for i in range(1):
ret, frame = video_capture.read()
if not ret:
break

# преобразуем фрейм в тензор
frame_tensor = torch.from_numpy(frame).float()
frame_tensor = frame_tensor.permute(2, 0, 1).unsqueeze(0)
#frame_tensor = frame_tensor[:, [2, 1, 0], :, :]

self.check_tensor(frame_tensor)


orig_prediction = self.analyze_frame(frame_tensor)

rotated_tensor = im.rotate_tensor(frame_tensor, orig_prediction.head())
self.check_tensor(rotated_tensor)

Continue reading...