simba-uw-tf-dev 4.6.2__py3-none-any.whl → 4.7.1__py3-none-any.whl

simba/sandbox/cuda/egocentric_rotator.py

@@ -1,374 +1,374 @@
-import functools
-import multiprocessing
-import os
-from typing import Optional, Tuple, Union
-
-import cv2
-import numpy as np
-
-try:
-    import cupy as cp
-    from cupyx.scipy.ndimage import affine_transform
-    CUPY_AVAILABLE = True
-except ImportError:
-    import numpy as cp
-    from scipy.ndimage import affine_transform
-    CUPY_AVAILABLE = False
-
-from simba.utils.checks import (check_file_exist_and_readable,
-                                check_if_dir_exists, check_if_valid_rgb_tuple,
-                                check_int, check_valid_array,
-                                check_valid_boolean, check_valid_tuple)
-from simba.utils.data import (align_target_warpaffine_vectors,
-                              center_rotation_warpaffine_vectors,
-                              egocentrically_align_pose)
-from simba.utils.enums import Defaults, Formats
-from simba.utils.printing import SimbaTimer, stdout_success
-from simba.utils.read_write import (concatenate_videos_in_folder,
-                                    create_directory, find_core_cnt,
-                                    get_fn_ext, get_video_meta_data, read_df,
-                                    read_frm_of_video,
-                                    read_img_batch_from_video_gpu,
-                                    remove_a_folder,
-                                    _read_img_batch_from_video_helper)
-
-
-def egocentric_video_aligner(frm_range: np.ndarray,
-                             video_path: Union[str, os.PathLike],
-                             temp_dir: Union[str, os.PathLike],
-                             video_name: str,
-                             centers: np.ndarray,
-                             rotation_vectors: np.ndarray,
-                             target: Tuple[int, int],
-                             fill_clr: Tuple[int, int, int] = (255, 255, 255),
-                             verbose: bool = False,
-                             gpu: bool = True):
-
-    video_meta = get_video_meta_data(video_path=video_path)
-
-    batch, frm_range = frm_range[0], frm_range[1]
-    save_path = os.path.join(temp_dir, f'{batch}.mp4')
-    fourcc = cv2.VideoWriter_fourcc(*f'{Formats.MP4_CODEC.value}')
-    writer = cv2.VideoWriter(save_path, fourcc, video_meta['fps'], (video_meta['width'], video_meta['height']))
-    batch_rotation_vectors = rotation_vectors[frm_range[0]: frm_range[-1]+1]
-    batch_centers = centers[frm_range[0]: frm_range[-1]+1]
-    m_rotates = center_rotation_warpaffine_vectors(rotation_vectors=batch_rotation_vectors, centers=batch_centers)
-    m_translations = align_target_warpaffine_vectors(centers=batch_centers, target=np.array(target))
-
-    if gpu:
-        # Combine rotation and translation matrices into single transform
-        # This reduces two sequential operations to one
-        batch_size = len(frm_range)
-        m_combined = np.zeros((batch_size, 2, 3), dtype=np.float32)
-        
-        for i in range(batch_size):
-            # Convert rotation matrix (2x3) to 3x3 homogeneous
-            m_rot_3x3 = np.eye(3, dtype=np.float32)
-            m_rot_3x3[:2, :] = m_rotates[i].astype(np.float32)
-            
-            # Convert translation matrix (2x3) to 3x3 homogeneous
-            m_trans_3x3 = np.eye(3, dtype=np.float32)
-            m_trans_3x3[:2, :] = m_translations[i].astype(np.float32)
-            
-            # Combine: translation after rotation (matches sequential cv2.warpAffine order)
-            m_combined_3x3 = m_trans_3x3 @ m_rot_3x3
-            
-            # Convert back to 2x3 for warpAffine compatibility
-            m_combined[i] = m_combined_3x3[:2, :]
-        
-        # Process frames in batches using GPU reading
-        # Use same batch size as original (30) for optimal I/O overlap
-        # Main optimization: combined matrix (one warpAffine instead of two)
-        img_counter = 0
-        frm_batches = np.array_split(frm_range, (len(frm_range) + 30 - 1) // 30)
-        for frm_batch_cnt, frm_ids in enumerate(frm_batches):
-            frms = read_img_batch_from_video_gpu(video_path=video_path, start_frm=frm_ids[0], end_frm=frm_ids[-1], verbose=False)
-            frms = np.stack(list(frms.values()), axis=0)
-            for img_cnt, img in enumerate(frms):
-                # Use combined matrix for single warpAffine (faster than two separate calls)
-                m = m_combined[img_counter].astype(np.float32)
-                final_frame = cv2.warpAffine(img, m, (video_meta['width'], video_meta['height']), borderValue=fill_clr)
-                writer.write(final_frame)
-                if verbose:
-                    frame_id = frm_ids[img_cnt]
-                    print(f'Creating frame {frame_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
-                img_counter += 1
-        
-        # Legacy CuPy code (commented out - CPU is faster for this use case)
-        if False and CUPY_AVAILABLE:
-            # Pre-compute all inverse matrices upfront (much faster than per-frame)
-            # For CuPy affine_transform, we need inverse matrices
-            m_inv_matrices = []
-            m_offsets = []
-            for i in range(batch_size):
-                m = m_combined[i]
-                matrix_2x2 = m[:2, :2].astype(np.float32)
-                offset = m[:2, 2].astype(np.float32)
-                m_inv_matrices.append(cp.asarray(matrix_2x2))
-                m_offsets.append(cp.asarray(offset))
-            # Batch invert all matrices at once
-            m_inv_matrices_gpu = cp.stack(m_inv_matrices)
-            m_inv_matrices_gpu = cp.linalg.inv(m_inv_matrices_gpu)
-            m_offsets_gpu = cp.stack(m_offsets)
-            
-            # Create async reader for GPU
-            async_reader = AsyncVideoFrameReader(
-                video_path=video_path,
-                batch_size=batch_size_gpu,
-                max_que_size=3,
-                start_idx=frm_range[0],
-                end_idx=frm_range[-1] + 1,
-                gpu=True,  # Use GPU reading
-                verbose=False
-            )
-            async_reader.start()
-            
-            # Process batches as they become available from async reader
-            # Batch process and transfer to minimize GPU->CPU overhead
-            processed_frames_batch = []
-            frame_ids_batch = []
-            
-            while True:
-                batch_result = get_async_frame_batch(batch_reader=async_reader, timeout=10)
-                if batch_result is None:
-                    # Write any remaining frames
-                    if processed_frames_batch:
-                        for frame in processed_frames_batch:
-                            writer.write(frame)
-                    break
-                
-                start_idx, end_idx, frms = batch_result
-                batch_len = end_idx - start_idx + 1
-                frms_gpu = cp.asarray(frms)
-                
-                # Process all frames in batch on GPU first
-                batch_transformed = []
-                batch_frame_indices = []
-                
-                for i in range(batch_len):
-                    # Map frame index from video to frm_range index
-                    frame_id = start_idx + i
-                    frame_idx_in_range = np.where(frm_range == frame_id)[0]
-                    if len(frame_idx_in_range) == 0:
-                        continue
-                    frame_idx_in_range = frame_idx_in_range[0]
-                    batch_frame_indices.append((i, frame_idx_in_range))
-                
-                # Process all frames in this batch on GPU
-                for i, frame_idx_in_range in batch_frame_indices:
-                    img_gpu = frms_gpu[i]
-                    matrix_inv = m_inv_matrices_gpu[frame_idx_in_range]
-                    offset = m_offsets_gpu[frame_idx_in_range]
-                    
-                    if len(img_gpu.shape) == 3:  # Multi-channel
-                        transformed_channels = []
-                        for c in range(img_gpu.shape[2]):
-                            transformed_ch = affine_transform(
-                                img_gpu[:, :, c],
-                                matrix=matrix_inv,
-                                offset=offset,
-                                output_shape=(video_meta['height'], video_meta['width']),
-                                order=1,
-                                mode='constant',
-                                cval=fill_clr[c] if c < len(fill_clr) else fill_clr[0],
-                                prefilter=False
-                            )
-                            transformed_channels.append(transformed_ch)
-                        transformed = cp.stack(transformed_channels, axis=2)
-                    else:  # Single channel
-                        transformed = affine_transform(
-                            img_gpu,
-                            matrix=matrix_inv,
-                            offset=offset,
-                            output_shape=(video_meta['height'], video_meta['width']),
-                            order=1,
-                            mode='constant',
-                            cval=fill_clr[0] if len(fill_clr) > 0 else 0,
-                            prefilter=False
-                        )
-                    batch_transformed.append(transformed)
-                
-                # Batch transfer all frames from GPU to CPU at once
-                if batch_transformed:
-                    # Stack all transformed frames and transfer in one go
-                    batch_transformed_stack = cp.stack(batch_transformed)
-                    batch_cpu = cp.asnumpy(batch_transformed_stack).astype(np.uint8)
-                    
-                    # Write all frames from this batch
-                    for frame_idx, (i, frame_idx_in_range) in enumerate(batch_frame_indices):
-                        final_frame = batch_cpu[frame_idx]
-                        writer.write(final_frame)
-                        
-                        if verbose:
-                            frame_id = start_idx + i
-                            print(f'Creating frame {frame_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
-            
-            async_reader.kill()
-        
-        else:
-            # Fallback to CPU with combined matrix and batch reading
-            # Process frames in batches
-            # Use helper function directly to avoid nested multiprocessing (we're already in a worker process)
-            # Larger batch size reduces overhead
-            batch_size_gpu = 500
-            frm_batches = np.array_split(frm_range, (len(frm_range) + batch_size_gpu - 1) // batch_size_gpu)
-            
-            # Create a mapping from frame_id to index in frm_range for fast lookup
-            frm_id_to_idx = {frame_id: idx for idx, frame_id in enumerate(frm_range)}
-            
-            for frm_batch_cnt, frm_ids in enumerate(frm_batches):
-                # Read batch of frames directly using helper (no multiprocessing)
-                frm_idx_array = np.array(frm_ids)
-                frms_dict = _read_img_batch_from_video_helper(
-                    frm_idx=frm_idx_array,
-                    video_path=video_path,
-                    greyscale=False,
-                    verbose=False,
-                    black_and_white=False,
-                    clahe=False
-                )
-                frms = np.stack([frms_dict[f] for f in frm_ids], axis=0)
-                
-                # Process all frames in batch using optimized CPU cv2.warpAffine with combined matrices
-                for i, frame_id in enumerate(frm_ids):
-                    # Fast dictionary lookup instead of np.where
-                    frame_idx_in_range = frm_id_to_idx.get(frame_id)
-                    if frame_idx_in_range is None:
-                        continue
-                    
-                    img = frms[i]
-                    m = m_combined[frame_idx_in_range].astype(np.float32)
-                    final_frame = cv2.warpAffine(img, m, (video_meta['width'], video_meta['height']), borderValue=fill_clr)
-                    writer.write(final_frame)
-                    
-                    if verbose:
-                        print(f'Creating frame {frame_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
-    else:
-        cap = cv2.VideoCapture(video_path)
-        for frm_idx, frm_id in enumerate(frm_range):
-            img = read_frm_of_video(video_path=cap, frame_index=frm_id)
-            rotated_frame = cv2.warpAffine(img, m_rotates[frm_idx], (video_meta['width'], video_meta['height']), borderValue=fill_clr)
-            final_frame = cv2.warpAffine(rotated_frame, m_translations[frm_idx], (video_meta['width'], video_meta['height']), borderValue=fill_clr)
-            writer.write(final_frame)
-            if verbose:
-                print(f'Creating frame {frm_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
-    writer.release()
-    return batch + 1
-
-class EgocentricVideoRotator():
-    """
-    Perform egocentric rotation of a video using CPU multiprocessing.
-
-    .. video:: _static/img/EgocentricalAligner_2.webm
-       :width: 800
-       :autoplay:
-       :loop:
-
-    .. seealso::
-       To perform joint egocentric alignment of both pose and video, or pose only, use :func:`~simba.data_processors.egocentric_aligner.EgocentricalAligner`.
-       To produce rotation vectors, use :func:`~simba.utils.data.egocentrically_align_pose_numba` or :func:`~simba.utils.data.egocentrically_align_pose`.
-
-    :param Union[str, os.PathLike] video_path: Path to a video file.
-    :param np.ndarray centers: A 2D array of shape `(num_frames, 2)` containing the original locations of `anchor_1_idx` in each frame before alignment. Returned by :func:`~simba.utils.data.egocentrically_align_pose_numba` or :func:`~simba.utils.data.egocentrically_align_pose`.
-    :param np.ndarray rotation_vectors: A 3D array of shape `(num_frames, 2, 2)` containing the rotation matrices applied to each frame. Returned by :func:`~simba.utils.data.egocentrically_align_pose_numba` or :func:`~simba.utils.data.egocentrically_align_pose`.
-    :param bool verbose: If True, prints progress. Deafult True.
-    :param Tuple[int, int, int] fill_clr: The color of the additional pixels. Deafult black. (0, 0, 0).
-    :param int core_cnt: Number of CPU cores to use for video rotation; `-1` uses all available cores.
-    :param Optional[Union[str, os.PathLike]] save_path: The location where to store the rotated video. If None, saves the video as the same dir as the input video with the `_rotated` suffix.
-
-    :example:
-    >>> DATA_PATH = "C:\501_MA142_Gi_Saline_0513.csv"
-    >>> VIDEO_PATH = "C:\501_MA142_Gi_Saline_0513.mp4"
-    >>> SAVE_PATH = "C:\501_MA142_Gi_Saline_0513_rotated.mp4"
-    >>> ANCHOR_LOC = np.array([250, 250])
-
-    >>> df = read_df(file_path=DATA_PATH, file_type='csv')
-    >>> bp_cols = [x for x in df.columns if not x.endswith('_p')]
-    >>> data = df[bp_cols].values.reshape(len(df), int(len(bp_cols)/2), 2).astype(np.int32)
-    >>> _, centers, rotation_vectors = egocentrically_align_pose(data=data, anchor_1_idx=6, anchor_2_idx=2, anchor_location=ANCHOR_LOC, direction=0)
-    >>> rotater = EgocentricVideoRotator(video_path=VIDEO_PATH, centers=centers, rotation_vectors=rotation_vectors, anchor_location=ANCHOR_LOC, save_path=SAVE_PATH)
-    >>> rotater.run()
-    """
-
-    def __init__(self,
-                 video_path: Union[str, os.PathLike],
-                 centers: np.ndarray,
-                 rotation_vectors: np.ndarray,
-                 anchor_location: Tuple[int, int],
-                 verbose: bool = True,
-                 fill_clr: Tuple[int, int, int] = (0, 0, 0),
-                 core_cnt: int = -1,
-                 save_path: Optional[Union[str, os.PathLike]] = None,
-                 gpu: Optional[bool] = True):
-
-        check_file_exist_and_readable(file_path=video_path)
-        self.video_meta_data = get_video_meta_data(video_path=video_path)
-        check_valid_array(data=centers, source=f'{self.__class__.__name__} centers', accepted_ndims=(2,), accepted_axis_1_shape=[2, ], accepted_axis_0_shape=[self.video_meta_data['frame_count']], accepted_dtypes=Formats.NUMERIC_DTYPES.value)
-        check_valid_array(data=rotation_vectors, source=f'{self.__class__.__name__} rotation_vectors', accepted_ndims=(3,), accepted_axis_0_shape=[self.video_meta_data['frame_count']], accepted_dtypes=Formats.NUMERIC_DTYPES.value)
-        check_valid_tuple(x=anchor_location, source=f'{self.__class__.__name__} anchor_location', accepted_lengths=(2,), valid_dtypes=(int,))
-        for i in anchor_location: check_int(name=f'{self.__class__.__name__} anchor_location', value=i, min_value=1)
-        check_valid_boolean(value=[verbose], source=f'{self.__class__.__name__} verbose')
-        check_if_valid_rgb_tuple(data=fill_clr)
-        check_int(name=f'{self.__class__.__name__} core_cnt', value=core_cnt, min_value=-1, unaccepted_vals=[0])
-        if core_cnt > find_core_cnt()[0] or core_cnt == -1:
-            self.core_cnt = find_core_cnt()[0]
-        else:
-            self.core_cnt = core_cnt
-        video_dir, self.video_name, _ = get_fn_ext(filepath=video_path)
-        if save_path is not None:
-            self.save_dir = os.path.dirname(save_path)
-            check_if_dir_exists(in_dir=self.save_dir, source=f'{self.__class__.__name__} save_path')
-        else:
-            self.save_dir = video_dir
-            save_path = os.path.join(video_dir, f'{self.video_name}_rotated.mp4')
-        self.video_path, self.save_path = video_path, save_path
-        self.centers, self.rotation_vectors, self.gpu = centers, rotation_vectors, gpu
-        self.verbose, self.fill_clr, self.anchor_loc = verbose, fill_clr, anchor_location
-
-    def run(self):
-        video_timer = SimbaTimer(start=True)
-        temp_dir = os.path.join(self.save_dir, 'temp')
-        if not os.path.isdir(temp_dir):
-            create_directory(paths=temp_dir)
-        else:
-            remove_a_folder(folder_dir=temp_dir)
-            create_directory(paths=temp_dir)
-        frm_list = np.arange(0, self.video_meta_data['frame_count'])
-        frm_list = np.array_split(frm_list, self.core_cnt)
-        frm_list = [(cnt, x) for cnt, x in enumerate(frm_list)]
-        if self.verbose:
-            print(f"Creating rotated video {self.video_name}, multiprocessing (chunksize: {1}, cores: {self.core_cnt})...")
-        with multiprocessing.Pool(self.core_cnt, maxtasksperchild=Defaults.LARGE_MAX_TASK_PER_CHILD.value) as pool:
-            constants = functools.partial(egocentric_video_aligner,
-                                          temp_dir=temp_dir,
-                                          video_name=self.video_name,
-                                          video_path=self.video_path,
-                                          centers=self.centers,
-                                          rotation_vectors=self.rotation_vectors,
-                                          target=self.anchor_loc,
-                                          verbose=self.verbose,
-                                          fill_clr=self.fill_clr,
-                                          gpu=self.gpu)
-            for cnt, result in enumerate(pool.imap(constants, frm_list, chunksize=1)):
-                if self.verbose:
-                    print(f"Rotate batch {result}/{self.core_cnt} complete...")
-            pool.terminate()
-            pool.join()
-
-        concatenate_videos_in_folder(in_folder=temp_dir, save_path=self.save_path, remove_splits=True, gpu=self.gpu, verbose=self.verbose)
-        video_timer.stop_timer()
-        stdout_success(msg=f"Egocentric rotation video {self.save_path} complete", elapsed_time=video_timer.elapsed_time_str, source=self.__class__.__name__)
-
-if __name__ == "__main__":
-    DATA_PATH = r"C:\Users\sroni\OneDrive\Desktop\desktop\rotate_ex\data\501_MA142_Gi_Saline_0513.csv"
-    VIDEO_PATH = r"C:\Users\sroni\OneDrive\Desktop\desktop\rotate_ex\videos\501_MA142_Gi_Saline_0513.mp4"
-    SAVE_PATH = r"C:\Users\sroni\OneDrive\Desktop\desktop\rotate_ex\videos\501_MA142_Gi_Saline_0513_rotated.mp4"
-    ANCHOR_LOC = np.array([250, 250])
-
-    df = read_df(file_path=DATA_PATH, file_type='csv')
-    bp_cols = [x for x in df.columns if not x.endswith('_p')]
-    data = df[bp_cols].values.reshape(len(df), int(len(bp_cols)/2), 2).astype(np.int32)
-
-    _, centers, rotation_vectors = egocentrically_align_pose(data=data, anchor_1_idx=5, anchor_2_idx=2, anchor_location=ANCHOR_LOC, direction=0)
-    rotater = EgocentricVideoRotator(video_path=VIDEO_PATH, centers=centers, rotation_vectors=rotation_vectors, anchor_location=(400, 100), save_path=SAVE_PATH, verbose=True, core_cnt=16, gpu=True)
-    rotater.run()
+import functools
+import multiprocessing
+import os
+from typing import Optional, Tuple, Union
+
+import cv2
+import numpy as np
+
+try:
+    import cupy as cp
+    from cupyx.scipy.ndimage import affine_transform
+    CUPY_AVAILABLE = True
+except ImportError:
+    import numpy as cp
+    from scipy.ndimage import affine_transform
+    CUPY_AVAILABLE = False
+
+from simba.utils.checks import (check_file_exist_and_readable,
+                                check_if_dir_exists, check_if_valid_rgb_tuple,
+                                check_int, check_valid_array,
+                                check_valid_boolean, check_valid_tuple)
+from simba.utils.data import (align_target_warpaffine_vectors,
+                              center_rotation_warpaffine_vectors,
+                              egocentrically_align_pose)
+from simba.utils.enums import Defaults, Formats
+from simba.utils.printing import SimbaTimer, stdout_success
+from simba.utils.read_write import (concatenate_videos_in_folder,
+                                    create_directory, find_core_cnt,
+                                    get_fn_ext, get_video_meta_data, read_df,
+                                    read_frm_of_video,
+                                    read_img_batch_from_video_gpu,
+                                    remove_a_folder,
+                                    _read_img_batch_from_video_helper)
+
+
+def egocentric_video_aligner(frm_range: np.ndarray,
+                             video_path: Union[str, os.PathLike],
+                             temp_dir: Union[str, os.PathLike],
+                             video_name: str,
+                             centers: np.ndarray,
+                             rotation_vectors: np.ndarray,
+                             target: Tuple[int, int],
+                             fill_clr: Tuple[int, int, int] = (255, 255, 255),
+                             verbose: bool = False,
+                             gpu: bool = True):
+
+    video_meta = get_video_meta_data(video_path=video_path)
+
+    batch, frm_range = frm_range[0], frm_range[1]
+    save_path = os.path.join(temp_dir, f'{batch}.mp4')
+    fourcc = cv2.VideoWriter_fourcc(*f'{Formats.MP4_CODEC.value}')
+    writer = cv2.VideoWriter(save_path, fourcc, video_meta['fps'], (video_meta['width'], video_meta['height']))
+    batch_rotation_vectors = rotation_vectors[frm_range[0]: frm_range[-1]+1]
+    batch_centers = centers[frm_range[0]: frm_range[-1]+1]
+    m_rotates = center_rotation_warpaffine_vectors(rotation_vectors=batch_rotation_vectors, centers=batch_centers)
+    m_translations = align_target_warpaffine_vectors(centers=batch_centers, target=np.array(target))
+
+    if gpu:
+        # Combine rotation and translation matrices into single transform
+        # This reduces two sequential operations to one
+        batch_size = len(frm_range)
+        m_combined = np.zeros((batch_size, 2, 3), dtype=np.float32)
+        
+        for i in range(batch_size):
+            # Convert rotation matrix (2x3) to 3x3 homogeneous
+            m_rot_3x3 = np.eye(3, dtype=np.float32)
+            m_rot_3x3[:2, :] = m_rotates[i].astype(np.float32)
+            
+            # Convert translation matrix (2x3) to 3x3 homogeneous
+            m_trans_3x3 = np.eye(3, dtype=np.float32)
+            m_trans_3x3[:2, :] = m_translations[i].astype(np.float32)
+            
+            # Combine: translation after rotation (matches sequential cv2.warpAffine order)
+            m_combined_3x3 = m_trans_3x3 @ m_rot_3x3
+            
+            # Convert back to 2x3 for warpAffine compatibility
+            m_combined[i] = m_combined_3x3[:2, :]
+        
+        # Process frames in batches using GPU reading
+        # Use same batch size as original (30) for optimal I/O overlap
+        # Main optimization: combined matrix (one warpAffine instead of two)
+        img_counter = 0
+        frm_batches = np.array_split(frm_range, (len(frm_range) + 30 - 1) // 30)
+        for frm_batch_cnt, frm_ids in enumerate(frm_batches):
+            frms = read_img_batch_from_video_gpu(video_path=video_path, start_frm=frm_ids[0], end_frm=frm_ids[-1], verbose=False)
+            frms = np.stack(list(frms.values()), axis=0)
+            for img_cnt, img in enumerate(frms):
+                # Use combined matrix for single warpAffine (faster than two separate calls)
+                m = m_combined[img_counter].astype(np.float32)
+                final_frame = cv2.warpAffine(img, m, (video_meta['width'], video_meta['height']), borderValue=fill_clr)
+                writer.write(final_frame)
+                if verbose:
+                    frame_id = frm_ids[img_cnt]
+                    print(f'Creating frame {frame_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
+                img_counter += 1
+        
+        # Legacy CuPy code (commented out - CPU is faster for this use case)
+        if False and CUPY_AVAILABLE:
+            # Pre-compute all inverse matrices upfront (much faster than per-frame)
+            # For CuPy affine_transform, we need inverse matrices
+            m_inv_matrices = []
+            m_offsets = []
+            for i in range(batch_size):
+                m = m_combined[i]
+                matrix_2x2 = m[:2, :2].astype(np.float32)
+                offset = m[:2, 2].astype(np.float32)
+                m_inv_matrices.append(cp.asarray(matrix_2x2))
+                m_offsets.append(cp.asarray(offset))
+            # Batch invert all matrices at once
+            m_inv_matrices_gpu = cp.stack(m_inv_matrices)
+            m_inv_matrices_gpu = cp.linalg.inv(m_inv_matrices_gpu)
+            m_offsets_gpu = cp.stack(m_offsets)
+            
+            # Create async reader for GPU
+            async_reader = AsyncVideoFrameReader(
+                video_path=video_path,
+                batch_size=batch_size_gpu,
+                max_que_size=3,
+                start_idx=frm_range[0],
+                end_idx=frm_range[-1] + 1,
+                gpu=True,  # Use GPU reading
+                verbose=False
+            )
+            async_reader.start()
+            
+            # Process batches as they become available from async reader
+            # Batch process and transfer to minimize GPU->CPU overhead
+            processed_frames_batch = []
+            frame_ids_batch = []
+            
+            while True:
+                batch_result = get_async_frame_batch(batch_reader=async_reader, timeout=10)
+                if batch_result is None:
+                    # Write any remaining frames
+                    if processed_frames_batch:
+                        for frame in processed_frames_batch:
+                            writer.write(frame)
+                    break
+                
+                start_idx, end_idx, frms = batch_result
+                batch_len = end_idx - start_idx + 1
+                frms_gpu = cp.asarray(frms)
+                
+                # Process all frames in batch on GPU first
+                batch_transformed = []
+                batch_frame_indices = []
+                
+                for i in range(batch_len):
+                    # Map frame index from video to frm_range index
+                    frame_id = start_idx + i
+                    frame_idx_in_range = np.where(frm_range == frame_id)[0]
+                    if len(frame_idx_in_range) == 0:
+                        continue
+                    frame_idx_in_range = frame_idx_in_range[0]
+                    batch_frame_indices.append((i, frame_idx_in_range))
+                
+                # Process all frames in this batch on GPU
+                for i, frame_idx_in_range in batch_frame_indices:
+                    img_gpu = frms_gpu[i]
+                    matrix_inv = m_inv_matrices_gpu[frame_idx_in_range]
+                    offset = m_offsets_gpu[frame_idx_in_range]
+                    
+                    if len(img_gpu.shape) == 3:  # Multi-channel
+                        transformed_channels = []
+                        for c in range(img_gpu.shape[2]):
+                            transformed_ch = affine_transform(
+                                img_gpu[:, :, c],
+                                matrix=matrix_inv,
+                                offset=offset,
+                                output_shape=(video_meta['height'], video_meta['width']),
+                                order=1,
+                                mode='constant',
+                                cval=fill_clr[c] if c < len(fill_clr) else fill_clr[0],
+                                prefilter=False
+                            )
+                            transformed_channels.append(transformed_ch)
+                        transformed = cp.stack(transformed_channels, axis=2)
+                    else:  # Single channel
+                        transformed = affine_transform(
+                            img_gpu,
+                            matrix=matrix_inv,
+                            offset=offset,
+                            output_shape=(video_meta['height'], video_meta['width']),
+                            order=1,
+                            mode='constant',
+                            cval=fill_clr[0] if len(fill_clr) > 0 else 0,
+                            prefilter=False
+                        )
+                    batch_transformed.append(transformed)
+                
+                # Batch transfer all frames from GPU to CPU at once
+                if batch_transformed:
+                    # Stack all transformed frames and transfer in one go
+                    batch_transformed_stack = cp.stack(batch_transformed)
+                    batch_cpu = cp.asnumpy(batch_transformed_stack).astype(np.uint8)
+                    
+                    # Write all frames from this batch
+                    for frame_idx, (i, frame_idx_in_range) in enumerate(batch_frame_indices):
+                        final_frame = batch_cpu[frame_idx]
+                        writer.write(final_frame)
+                        
+                        if verbose:
+                            frame_id = start_idx + i
+                            print(f'Creating frame {frame_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
+            
+            async_reader.kill()
+        
+        else:
+            # Fallback to CPU with combined matrix and batch reading
+            # Process frames in batches
+            # Use helper function directly to avoid nested multiprocessing (we're already in a worker process)
+            # Larger batch size reduces overhead
+            batch_size_gpu = 500
+            frm_batches = np.array_split(frm_range, (len(frm_range) + batch_size_gpu - 1) // batch_size_gpu)
+            
+            # Create a mapping from frame_id to index in frm_range for fast lookup
+            frm_id_to_idx = {frame_id: idx for idx, frame_id in enumerate(frm_range)}
+            
+            for frm_batch_cnt, frm_ids in enumerate(frm_batches):
+                # Read batch of frames directly using helper (no multiprocessing)
+                frm_idx_array = np.array(frm_ids)
+                frms_dict = _read_img_batch_from_video_helper(
+                    frm_idx=frm_idx_array,
+                    video_path=video_path,
+                    greyscale=False,
+                    verbose=False,
+                    black_and_white=False,
+                    clahe=False
+                )
+                frms = np.stack([frms_dict[f] for f in frm_ids], axis=0)
+                
+                # Process all frames in batch using optimized CPU cv2.warpAffine with combined matrices
+                for i, frame_id in enumerate(frm_ids):
+                    # Fast dictionary lookup instead of np.where
+                    frame_idx_in_range = frm_id_to_idx.get(frame_id)
+                    if frame_idx_in_range is None:
+                        continue
+                    
+                    img = frms[i]
+                    m = m_combined[frame_idx_in_range].astype(np.float32)
+                    final_frame = cv2.warpAffine(img, m, (video_meta['width'], video_meta['height']), borderValue=fill_clr)
+                    writer.write(final_frame)
+                    
+                    if verbose:
+                        print(f'Creating frame {frame_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
+    else:
+        cap = cv2.VideoCapture(video_path)
+        for frm_idx, frm_id in enumerate(frm_range):
+            img = read_frm_of_video(video_path=cap, frame_index=frm_id)
+            rotated_frame = cv2.warpAffine(img, m_rotates[frm_idx], (video_meta['width'], video_meta['height']), borderValue=fill_clr)
+            final_frame = cv2.warpAffine(rotated_frame, m_translations[frm_idx], (video_meta['width'], video_meta['height']), borderValue=fill_clr)
+            writer.write(final_frame)
+            if verbose:
+                print(f'Creating frame {frm_id}/{video_meta["frame_count"]} ({video_name}, CPU core: {batch + 1}).')
+    writer.release()
+    return batch + 1
+
+class EgocentricVideoRotator():
+    """
+    Perform egocentric rotation of a video using CPU multiprocessing.
+
+    .. video:: _static/img/EgocentricalAligner_2.webm
+       :width: 800
+       :autoplay:
+       :loop:
+
+    .. seealso::
+       To perform joint egocentric alignment of both pose and video, or pose only, use :func:`~simba.data_processors.egocentric_aligner.EgocentricalAligner`.
+       To produce rotation vectors, use :func:`~simba.utils.data.egocentrically_align_pose_numba` or :func:`~simba.utils.data.egocentrically_align_pose`.
+
+    :param Union[str, os.PathLike] video_path: Path to a video file.
+    :param np.ndarray centers: A 2D array of shape `(num_frames, 2)` containing the original locations of `anchor_1_idx` in each frame before alignment. Returned by :func:`~simba.utils.data.egocentrically_align_pose_numba` or :func:`~simba.utils.data.egocentrically_align_pose`.
+    :param np.ndarray rotation_vectors: A 3D array of shape `(num_frames, 2, 2)` containing the rotation matrices applied to each frame. Returned by :func:`~simba.utils.data.egocentrically_align_pose_numba` or :func:`~simba.utils.data.egocentrically_align_pose`.
+    :param bool verbose: If True, prints progress. Deafult True.
+    :param Tuple[int, int, int] fill_clr: The color of the additional pixels. Deafult black. (0, 0, 0).
+    :param int core_cnt: Number of CPU cores to use for video rotation; `-1` uses all available cores.
+    :param Optional[Union[str, os.PathLike]] save_path: The location where to store the rotated video. If None, saves the video as the same dir as the input video with the `_rotated` suffix.
+
+    :example:
+    >>> DATA_PATH = "C:\501_MA142_Gi_Saline_0513.csv"
+    >>> VIDEO_PATH = "C:\501_MA142_Gi_Saline_0513.mp4"
+    >>> SAVE_PATH = "C:\501_MA142_Gi_Saline_0513_rotated.mp4"
+    >>> ANCHOR_LOC = np.array([250, 250])
+
+    >>> df = read_df(file_path=DATA_PATH, file_type='csv')
+    >>> bp_cols = [x for x in df.columns if not x.endswith('_p')]
+    >>> data = df[bp_cols].values.reshape(len(df), int(len(bp_cols)/2), 2).astype(np.int32)
+    >>> _, centers, rotation_vectors = egocentrically_align_pose(data=data, anchor_1_idx=6, anchor_2_idx=2, anchor_location=ANCHOR_LOC, direction=0)
+    >>> rotater = EgocentricVideoRotator(video_path=VIDEO_PATH, centers=centers, rotation_vectors=rotation_vectors, anchor_location=ANCHOR_LOC, save_path=SAVE_PATH)
+    >>> rotater.run()
+    """
+
+    def __init__(self,
+                 video_path: Union[str, os.PathLike],
+                 centers: np.ndarray,
+                 rotation_vectors: np.ndarray,
+                 anchor_location: Tuple[int, int],
+                 verbose: bool = True,
+                 fill_clr: Tuple[int, int, int] = (0, 0, 0),
+                 core_cnt: int = -1,
+                 save_path: Optional[Union[str, os.PathLike]] = None,
+                 gpu: Optional[bool] = True):
+
+        check_file_exist_and_readable(file_path=video_path)
+        self.video_meta_data = get_video_meta_data(video_path=video_path)
+        check_valid_array(data=centers, source=f'{self.__class__.__name__} centers', accepted_ndims=(2,), accepted_axis_1_shape=[2, ], accepted_axis_0_shape=[self.video_meta_data['frame_count']], accepted_dtypes=Formats.NUMERIC_DTYPES.value)
+        check_valid_array(data=rotation_vectors, source=f'{self.__class__.__name__} rotation_vectors', accepted_ndims=(3,), accepted_axis_0_shape=[self.video_meta_data['frame_count']], accepted_dtypes=Formats.NUMERIC_DTYPES.value)
+        check_valid_tuple(x=anchor_location, source=f'{self.__class__.__name__} anchor_location', accepted_lengths=(2,), valid_dtypes=(int,))
+        for i in anchor_location: check_int(name=f'{self.__class__.__name__} anchor_location', value=i, min_value=1)
+        check_valid_boolean(value=[verbose], source=f'{self.__class__.__name__} verbose')
+        check_if_valid_rgb_tuple(data=fill_clr)
+        check_int(name=f'{self.__class__.__name__} core_cnt', value=core_cnt, min_value=-1, unaccepted_vals=[0])
+        if core_cnt > find_core_cnt()[0] or core_cnt == -1:
+            self.core_cnt = find_core_cnt()[0]
+        else:
+            self.core_cnt = core_cnt
+        video_dir, self.video_name, _ = get_fn_ext(filepath=video_path)
+        if save_path is not None:
+            self.save_dir = os.path.dirname(save_path)
+            check_if_dir_exists(in_dir=self.save_dir, source=f'{self.__class__.__name__} save_path')
+        else:
+            self.save_dir = video_dir
+            save_path = os.path.join(video_dir, f'{self.video_name}_rotated.mp4')
+        self.video_path, self.save_path = video_path, save_path
+        self.centers, self.rotation_vectors, self.gpu = centers, rotation_vectors, gpu
+        self.verbose, self.fill_clr, self.anchor_loc = verbose, fill_clr, anchor_location
+
+    def run(self):
+        video_timer = SimbaTimer(start=True)
+        temp_dir = os.path.join(self.save_dir, 'temp')
+        if not os.path.isdir(temp_dir):
+            create_directory(paths=temp_dir)
+        else:
+            remove_a_folder(folder_dir=temp_dir)
+            create_directory(paths=temp_dir)
+        frm_list = np.arange(0, self.video_meta_data['frame_count'])
+        frm_list = np.array_split(frm_list, self.core_cnt)
+        frm_list = [(cnt, x) for cnt, x in enumerate(frm_list)]
+        if self.verbose:
+            print(f"Creating rotated video {self.video_name}, multiprocessing (chunksize: {1}, cores: {self.core_cnt})...")
+        with multiprocessing.Pool(self.core_cnt, maxtasksperchild=Defaults.LARGE_MAX_TASK_PER_CHILD.value) as pool:
+            constants = functools.partial(egocentric_video_aligner,
+                                          temp_dir=temp_dir,
+                                          video_name=self.video_name,
+                                          video_path=self.video_path,
+                                          centers=self.centers,
+                                          rotation_vectors=self.rotation_vectors,
+                                          target=self.anchor_loc,
+                                          verbose=self.verbose,
+                                          fill_clr=self.fill_clr,
+                                          gpu=self.gpu)
+            for cnt, result in enumerate(pool.imap(constants, frm_list, chunksize=1)):
+                if self.verbose:
+                    print(f"Rotate batch {result}/{self.core_cnt} complete...")
+            pool.terminate()
+            pool.join()
+
+        concatenate_videos_in_folder(in_folder=temp_dir, save_path=self.save_path, remove_splits=True, gpu=self.gpu, verbose=self.verbose)
+        video_timer.stop_timer()
+        stdout_success(msg=f"Egocentric rotation video {self.save_path} complete", elapsed_time=video_timer.elapsed_time_str, source=self.__class__.__name__)
+
+if __name__ == "__main__":
+    DATA_PATH = r"C:\Users\sroni\OneDrive\Desktop\desktop\rotate_ex\data\501_MA142_Gi_Saline_0513.csv"
+    VIDEO_PATH = r"C:\Users\sroni\OneDrive\Desktop\desktop\rotate_ex\videos\501_MA142_Gi_Saline_0513.mp4"
+    SAVE_PATH = r"C:\Users\sroni\OneDrive\Desktop\desktop\rotate_ex\videos\501_MA142_Gi_Saline_0513_rotated.mp4"
+    ANCHOR_LOC = np.array([250, 250])
+
+    df = read_df(file_path=DATA_PATH, file_type='csv')
+    bp_cols = [x for x in df.columns if not x.endswith('_p')]
+    data = df[bp_cols].values.reshape(len(df), int(len(bp_cols)/2), 2).astype(np.int32)
+
+    _, centers, rotation_vectors = egocentrically_align_pose(data=data, anchor_1_idx=5, anchor_2_idx=2, anchor_location=ANCHOR_LOC, direction=0)
+    rotater = EgocentricVideoRotator(video_path=VIDEO_PATH, centers=centers, rotation_vectors=rotation_vectors, anchor_location=(400, 100), save_path=SAVE_PATH, verbose=True, core_cnt=16, gpu=True)
+    rotater.run()