mirror of
https://github.com/iperov/DeepFaceLive
synced 2025-07-06 13:02:16 -07:00
removing CuPy. FaceMerger now works with any OpenCL1.2-compatible GPU.
This commit is contained in:
iperov
parent
c2ba2bab9d
commit
4fe2da23c0
6 changed files with 134 additions and 292 deletions
|
|
@ -3,7 +3,7 @@ from enum import IntEnum
|
|||
|
||||
import numexpr as ne
|
||||
import numpy as np
|
||||
from xlib import cupy as lib_cp
|
||||
from xlib import avecl as lib_cl
|
||||
from xlib import os as lib_os
|
||||
from xlib.image import ImageProcessor
|
||||
from xlib.mp import csw as lib_csw
|
||||
|
|
@ -57,12 +57,13 @@ class FaceMergerWorker(BackendWorker):
|
|||
cs.face_opacity.call_on_number(self.on_cs_face_opacity)
|
||||
|
||||
cs.device.enable()
|
||||
cs.device.set_choices( ['CPU'] + lib_cp.get_available_devices(), none_choice_name='@misc.menu_select')
|
||||
|
||||
cs.device.set_choices( ['CPU'] + lib_cl.get_available_devices_info(), none_choice_name='@misc.menu_select')
|
||||
|
||||
cs.device.select(state.device if state.device is not None else 'CPU')
|
||||
|
||||
|
||||
def on_cs_device(self, idxs, device : lib_cp.CuPyDeviceInfo):
|
||||
def on_cs_device(self, idxs, device : lib_cl.DeviceInfo):
|
||||
state, cs = self.get_state(), self.get_control_sheet()
|
||||
if device is not None and state.device == device:
|
||||
cs.face_x_offset.enable()
|
||||
|
|
@ -94,17 +95,6 @@ class FaceMergerWorker(BackendWorker):
|
|||
cs.face_opacity.set_config(lib_csw.Number.Config(min=0.0, max=1.0, step=0.01, decimals=2, allow_instant_update=True))
|
||||
cs.face_opacity.set_number(state.face_opacity if state.face_opacity is not None else 1.0)
|
||||
|
||||
if device != 'CPU':
|
||||
self.is_gpu = True
|
||||
|
||||
global cp
|
||||
import cupy as cp # BUG eats 1.8Gb paging file per process, so import on demand
|
||||
cp.cuda.Device( device.get_index() ).use()
|
||||
|
||||
self.cp_mask_clip_kernel = cp.ElementwiseKernel('T x', 'T z', 'z = x < 0.004 ? 0 : x > 1.0 ? 1.0 : x', 'mask_clip_kernel')
|
||||
self.cp_merge_kernel = cp.ElementwiseKernel('T bg, T face, T mask', 'T z', 'z = bg*(1.0-mask) + face*mask', 'merge_kernel')
|
||||
self.cp_merge_kernel_opacity = cp.ElementwiseKernel('T bg, T face, T mask, T opacity', 'T z', 'z = bg*(1.0-mask) + bg*mask*(1.0-opacity) + face*mask*opacity', 'merge_kernel_opacity')
|
||||
|
||||
else:
|
||||
state.device = device
|
||||
self.save_state()
|
||||
|
|
@ -165,6 +155,74 @@ class FaceMergerWorker(BackendWorker):
|
|||
self.save_state()
|
||||
self.reemit_frame_signal.send()
|
||||
|
||||
def _merge_on_cpu(self, frame_image, face_align_mask_img, face_swap_img, face_swap_mask_img, aligned_to_source_uni_mat, frame_width, frame_height ):
|
||||
state = self.get_state()
|
||||
|
||||
frame_image = ImageProcessor(frame_image).to_ufloat32().get_image('HWC')
|
||||
face_align_mask_img = ImageProcessor(face_align_mask_img).to_ufloat32().get_image('HW')
|
||||
face_swap_mask_img = ImageProcessor(face_swap_mask_img).to_ufloat32().get_image('HW')
|
||||
|
||||
|
||||
if state.face_mask_type == FaceMaskType.SRC:
|
||||
face_mask = face_align_mask_img
|
||||
elif state.face_mask_type == FaceMaskType.CELEB:
|
||||
face_mask = face_swap_mask_img
|
||||
elif state.face_mask_type == FaceMaskType.SRC_M_CELEB:
|
||||
face_mask = face_align_mask_img*face_swap_mask_img
|
||||
|
||||
# Combine face mask
|
||||
face_mask_ip = ImageProcessor(face_mask).erode_blur(state.face_mask_erode, state.face_mask_blur, fade_to_border=True) \
|
||||
.warpAffine(aligned_to_source_uni_mat, frame_width, frame_height)
|
||||
face_mask_ip.clip2( (1.0/255.0), 0.0, 1.0, 1.0)
|
||||
frame_face_mask = face_mask_ip.get_image('HWC')
|
||||
|
||||
frame_face_swap_img = ImageProcessor(face_swap_img) \
|
||||
.to_ufloat32().warpAffine(aligned_to_source_uni_mat, frame_width, frame_height).get_image('HWC')
|
||||
|
||||
# Combine final frame
|
||||
opacity = state.face_opacity
|
||||
if opacity == 1.0:
|
||||
frame_final = ne.evaluate('frame_image*(1.0-frame_face_mask) + frame_face_swap_img*frame_face_mask')
|
||||
else:
|
||||
frame_final = ne.evaluate('frame_image*(1.0-frame_face_mask) + frame_image*frame_face_mask*(1.0-opacity) + frame_face_swap_img*frame_face_mask*opacity')
|
||||
|
||||
return frame_final
|
||||
|
||||
def _merge_on_gpu(self, frame_image, face_align_mask_img, face_swap_img, face_swap_mask_img, aligned_to_source_uni_mat, frame_width, frame_height ):
|
||||
state = self.get_state()
|
||||
|
||||
if state.face_mask_type == FaceMaskType.SRC:
|
||||
face_mask_t = lib_cl.Tensor.from_value(face_align_mask_img, device=state.device)
|
||||
face_mask_t = face_mask_t.transpose( (2,0,1), op_text='O = (I <= 128 ? 0 : 1);', dtype=np.uint8)
|
||||
elif state.face_mask_type == FaceMaskType.CELEB:
|
||||
face_mask_t = lib_cl.Tensor.from_value(face_swap_mask_img, device=state.device)
|
||||
face_mask_t = face_mask_t.transpose( (2,0,1), op_text='O = (I <= 128 ? 0 : 1);', dtype=np.uint8)
|
||||
|
||||
elif state.face_mask_type == FaceMaskType.SRC_M_CELEB:
|
||||
face_mask_t = lib_cl.any_wise('float X = (((float)I0) / 255.0) * (((float)I1) / 255.0); O = (X <= 0.5 ? 0 : 1);',
|
||||
lib_cl.Tensor.from_value(face_align_mask_img, device=state.device),
|
||||
lib_cl.Tensor.from_value(face_swap_mask_img, device=state.device),
|
||||
dtype=np.uint8).transpose( (2,0,1) )
|
||||
|
||||
face_mask_t = lib_cl.binary_morph(face_mask_t, state.face_mask_erode, state.face_mask_blur, fade_to_border=True, dtype=np.float32)
|
||||
|
||||
face_swap_img_t = lib_cl.Tensor.from_value(face_swap_img, device=state.device)
|
||||
face_swap_img_t = face_swap_img_t.transpose( (2,0,1), op_text='O = ((O_TYPE)I) / 255.0', dtype=np.float32)
|
||||
|
||||
frame_face_mask_t = lib_cl.remap_np_affine(face_mask_t, aligned_to_source_uni_mat, output_size=(frame_height, frame_width) )
|
||||
frame_face_swap_img_t = lib_cl.remap_np_affine(face_swap_img_t, aligned_to_source_uni_mat, output_size=(frame_height, frame_width) )
|
||||
|
||||
frame_image_t = lib_cl.Tensor.from_value(frame_image, device=state.device).transpose( (2,0,1) )
|
||||
|
||||
opacity = state.face_opacity
|
||||
if opacity == 1.0:
|
||||
frame_final_t = lib_cl.any_wise('float I0f = (((float)I0) / 255.0); I1 = (I1 <= (1.0/255.0) ? 0.0 : I1 > 1.0 ? 1.0 : I1); O = I0f*(1.0-I1) + I2*I1', frame_image_t, frame_face_mask_t, frame_face_swap_img_t, dtype=np.float32)
|
||||
else:
|
||||
frame_final_t = lib_cl.any_wise('float I0f = (((float)I0) / 255.0); I1 = (I1 <= (1.0/255.0) ? 0.0 : I1 > 1.0 ? 1.0 : I1); O = I0f*(1.0-I1) + I0f*I1*(1.0-I3) + I2*I1*I3', frame_image_t, frame_face_mask_t, frame_face_swap_img_t, opacity, dtype=np.float32)
|
||||
|
||||
return frame_final_t.transpose( (1,2,0) ).np()
|
||||
|
||||
|
||||
def on_tick(self):
|
||||
state, cs = self.get_state(), self.get_control_sheet()
|
||||
|
||||
|
|
@ -189,72 +247,29 @@ class FaceMergerWorker(BackendWorker):
|
|||
face_swap_mask = face_swap.get_face_mask()
|
||||
if face_swap_mask is not None:
|
||||
|
||||
face_align_img = bcd.get_image(face_align.get_image_name())
|
||||
face_swap_img = bcd.get_image(face_swap.get_image_name())
|
||||
|
||||
face_align_img_shape, _ = bcd.get_image_shape_dtype(face_align.get_image_name())
|
||||
face_align_mask_img = bcd.get_image(face_align_mask.get_image_name())
|
||||
face_swap_img = bcd.get_image(face_swap.get_image_name())
|
||||
face_swap_mask_img = bcd.get_image(face_swap_mask.get_image_name())
|
||||
source_to_aligned_uni_mat = face_align.get_source_to_aligned_uni_mat()
|
||||
|
||||
face_mask_type = state.face_mask_type
|
||||
|
||||
if all_is_not_None(face_align_img, face_align_mask_img, face_swap_img, face_swap_mask_img, face_mask_type):
|
||||
face_height, face_width = face_align_img.shape[:2]
|
||||
|
||||
if self.is_gpu:
|
||||
frame_image = cp.asarray(frame_image)
|
||||
face_align_mask_img = cp.asarray(face_align_mask_img)
|
||||
face_swap_mask_img = cp.asarray(face_swap_mask_img)
|
||||
face_swap_img = cp.asarray(face_swap_img)
|
||||
|
||||
frame_image_ip = ImageProcessor(frame_image).to_ufloat32()
|
||||
frame_image, (_, frame_height, frame_width, _) = frame_image_ip.get_image('HWC'), frame_image_ip.get_dims()
|
||||
face_align_mask_img = ImageProcessor(face_align_mask_img).to_ufloat32().get_image('HW')
|
||||
face_swap_mask_img = ImageProcessor(face_swap_mask_img).to_ufloat32().get_image('HW')
|
||||
|
||||
if all_is_not_None(face_align_img_shape, face_align_mask_img, face_swap_img, face_swap_mask_img):
|
||||
face_height, face_width = face_align_img_shape[:2]
|
||||
frame_height, frame_width = frame_image.shape[:2]
|
||||
aligned_to_source_uni_mat = source_to_aligned_uni_mat.invert()
|
||||
aligned_to_source_uni_mat = aligned_to_source_uni_mat.source_translated(-state.face_x_offset, -state.face_y_offset)
|
||||
aligned_to_source_uni_mat = aligned_to_source_uni_mat.source_scaled_around_center(state.face_scale,state.face_scale)
|
||||
aligned_to_source_uni_mat = aligned_to_source_uni_mat.to_exact_mat (face_width, face_height, frame_width, frame_height)
|
||||
|
||||
if face_mask_type == FaceMaskType.SRC:
|
||||
face_mask = face_align_mask_img
|
||||
elif face_mask_type == FaceMaskType.CELEB:
|
||||
face_mask = face_swap_mask_img
|
||||
elif face_mask_type == FaceMaskType.SRC_M_CELEB:
|
||||
face_mask = face_align_mask_img*face_swap_mask_img
|
||||
|
||||
# Combine face mask
|
||||
face_mask_ip = ImageProcessor(face_mask).erode_blur(state.face_mask_erode, state.face_mask_blur, fade_to_border=True) \
|
||||
.warpAffine(aligned_to_source_uni_mat, frame_width, frame_height)
|
||||
if self.is_gpu:
|
||||
face_mask_ip.apply( lambda img: self.cp_mask_clip_kernel(img) )
|
||||
if state.device == 'CPU':
|
||||
merged_frame = self._merge_on_cpu(frame_image, face_align_mask_img, face_swap_img, face_swap_mask_img, aligned_to_source_uni_mat, frame_width, frame_height )
|
||||
else:
|
||||
face_mask_ip.clip2( (1.0/255.0), 0.0, 1.0, 1.0)
|
||||
frame_face_mask = face_mask_ip.get_image('HWC')
|
||||
|
||||
frame_face_swap_img = ImageProcessor(face_swap_img) \
|
||||
.to_ufloat32().warpAffine(aligned_to_source_uni_mat, frame_width, frame_height).get_image('HWC')
|
||||
|
||||
# Combine final frame
|
||||
opacity = state.face_opacity
|
||||
if self.is_gpu:
|
||||
if opacity == 1.0:
|
||||
frame_final = self.cp_merge_kernel(frame_image, frame_face_swap_img, frame_face_mask)
|
||||
else:
|
||||
frame_final = self.cp_merge_kernel_opacity(frame_image, frame_face_swap_img, frame_face_mask, opacity)
|
||||
frame_final = cp.asnumpy(frame_final)
|
||||
else:
|
||||
if opacity == 1.0:
|
||||
frame_final = ne.evaluate('frame_image*(1.0-frame_face_mask) + frame_face_swap_img*frame_face_mask')
|
||||
else:
|
||||
frame_final = ne.evaluate('frame_image*(1.0-frame_face_mask) + frame_image*frame_face_mask*(1.0-opacity) + frame_face_swap_img*frame_face_mask*opacity')
|
||||
|
||||
merged_frame = self._merge_on_gpu(frame_image, face_align_mask_img, face_swap_img, face_swap_mask_img, aligned_to_source_uni_mat, frame_width, frame_height )
|
||||
# keep image in float32 in order not to extra load FaceMerger
|
||||
|
||||
merged_frame_name = f'{frame_name}_merged'
|
||||
bcd.set_merged_frame_name(merged_frame_name)
|
||||
bcd.set_image(merged_frame_name, frame_final)
|
||||
bcd.set_image(merged_frame_name, merged_frame)
|
||||
break
|
||||
|
||||
self.stop_profile_timing()
|
||||
|
|
@ -297,7 +312,7 @@ class Sheet:
|
|||
self.face_opacity = lib_csw.Number.Host()
|
||||
|
||||
class WorkerState(BackendWorkerState):
|
||||
device : lib_cp.CuPyDeviceInfo = None
|
||||
device : lib_cl.DeviceInfo = None
|
||||
face_x_offset : float = None
|
||||
face_y_offset : float = None
|
||||
face_scale : float = None
|
||||
|
|
@ -305,3 +320,4 @@ class WorkerState(BackendWorkerState):
|
|||
face_mask_erode : int = None
|
||||
face_mask_blur : int = None
|
||||
face_opacity : float = None
|
||||
|
||||
|
|
|
|||
|
|
@ -9,7 +9,7 @@ RUN ln -s /usr/bin/python3 /usr/bin/python
|
|||
RUN git clone https://github.com/iperov/DeepFaceLive.git
|
||||
|
||||
RUN python -m pip install --upgrade pip
|
||||
RUN python -m pip install onnxruntime-gpu==1.8.1 cupy-cuda113 numpy==1.21.2 scipy==1.5.4 numexpr opencv-python==4.5.3.56 opencv-contrib-python==4.5.3.56 pyqt6==6.1.1 onnx==1.10.1 torch==1.8.1 torchvision==0.9.1
|
||||
RUN python -m pip install onnxruntime-gpu==1.8.1 numpy==1.21.2 numexpr opencv-python==4.5.3.56 opencv-contrib-python==4.5.3.56 pyqt6==6.1.1 onnx==1.10.1 torch==1.8.1 torchvision==0.9.1
|
||||
|
||||
WORKDIR /app/DeepFaceLive
|
||||
COPY example.sh example.sh
|
||||
|
|
|
|||
|
|
@ -472,7 +472,6 @@ def build_deepfacelive_windows(release_dir, cache_dir, python_ver='3.7.9', backe
|
|||
# PIP INSTALLATIONS
|
||||
|
||||
builder.install_pip_package('numpy==1.21.2')
|
||||
builder.install_pip_package('scipy==1.5.4')
|
||||
builder.install_pip_package('numexpr')
|
||||
builder.install_pip_package('opencv-python==4.5.3.56')
|
||||
builder.install_pip_package('opencv-contrib-python==4.5.3.56')
|
||||
|
|
@ -482,7 +481,6 @@ def build_deepfacelive_windows(release_dir, cache_dir, python_ver='3.7.9', backe
|
|||
if backend == 'cuda':
|
||||
builder.install_pip_package('torch==1.8.1+cu111 torchvision==0.9.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html')
|
||||
builder.install_pip_package('onnxruntime-gpu==1.9.0')
|
||||
builder.install_pip_package('cupy-cuda111===9.0.0')
|
||||
elif backend == 'directml':
|
||||
if python_ver[:3] == '3.7':
|
||||
builder.install_pip_package('https://github.com/iperov/DeepFaceLive/releases/download/ort-dml/onnxruntime_directml-1.8.2-cp37-cp37m-win_amd64.whl')
|
||||
|
|
|
|||
|
|
@ -1 +0,0 @@
|
|||
from .device import get_available_devices, CuPyDeviceInfo
|
||||
|
|
@ -1,71 +0,0 @@
|
|||
from typing import List
|
||||
|
||||
from .. import appargs as lib_appargs
|
||||
|
||||
|
||||
class CuPyDeviceInfo:
|
||||
"""
|
||||
Represents picklable CuPy device info
|
||||
"""
|
||||
def __init__(self, index=None, name=None, total_memory=None):
|
||||
self._index : int = index
|
||||
self._name : str = name
|
||||
self._total_memory : int = total_memory
|
||||
|
||||
def __getstate__(self):
|
||||
return self.__dict__.copy()
|
||||
|
||||
def __setstate__(self, d):
|
||||
self.__init__()
|
||||
self.__dict__.update(d)
|
||||
|
||||
def is_cpu(self) -> bool: return self._index == -1
|
||||
|
||||
def get_index(self) -> int:
|
||||
return self._index
|
||||
|
||||
def get_name(self) -> str:
|
||||
return self._name
|
||||
|
||||
def get_total_memory(self) -> int:
|
||||
return self._total_memory
|
||||
|
||||
def __eq__(self, other):
|
||||
if self is not None and other is not None and isinstance(self, CuPyDeviceInfo) and isinstance(other, CuPyDeviceInfo):
|
||||
return self._index == other._index
|
||||
return False
|
||||
|
||||
def __hash__(self):
|
||||
return self._index
|
||||
|
||||
def __str__(self):
|
||||
if self.is_cpu():
|
||||
return "CPU"
|
||||
else:
|
||||
return f"[{self._index}] {self._name} [{(self._total_memory / 1024**3) :.3}Gb]"
|
||||
|
||||
def __repr__(self):
|
||||
return f'{self.__class__.__name__} object: ' + self.__str__()
|
||||
|
||||
|
||||
_cupy_devices = None
|
||||
def get_available_devices() -> List[CuPyDeviceInfo]:
|
||||
"""
|
||||
returns a list of available CuPyDeviceInfo
|
||||
"""
|
||||
if lib_appargs.get_arg_bool('NO_CUDA'):
|
||||
return []
|
||||
|
||||
global _cupy_devices
|
||||
|
||||
if _cupy_devices is None:
|
||||
import cupy as cp # BUG eats 1.8Gb paging file per process, so import on demand
|
||||
devices = []
|
||||
|
||||
for i in range (cp.cuda.runtime.getDeviceCount()):
|
||||
device_props = cp.cuda.runtime.getDeviceProperties(i)
|
||||
devices.append ( CuPyDeviceInfo(index=i, name=device_props['name'].decode('utf-8'), total_memory=device_props['totalGlobalMem']))
|
||||
|
||||
_cupy_devices = devices
|
||||
return _cupy_devices
|
||||
|
||||
|
|
@ -7,33 +7,16 @@ import numpy as np
|
|||
|
||||
class ImageProcessor:
|
||||
"""
|
||||
Generic image processor for numpy or cupy images
|
||||
Generic image processor for numpy images
|
||||
|
||||
arguments
|
||||
|
||||
img np.ndarray|
|
||||
cp.ndarray
|
||||
HW (2 ndim)
|
||||
img np.ndarray HW (2 ndim)
|
||||
HWC (3 ndim)
|
||||
NHWC (4 ndim)
|
||||
|
||||
for cupy you should set device before using ImageProcessor
|
||||
"""
|
||||
def __init__(self, img : Union[np.ndarray,'cp.ndarray'], copy=False):
|
||||
|
||||
if img.__class__ == np.ndarray:
|
||||
self._xp = np
|
||||
import scipy
|
||||
import scipy.ndimage
|
||||
self._sp = scipy
|
||||
if copy:
|
||||
img = img.copy()
|
||||
else:
|
||||
import cupy as cp # BUG eats 1.8Gb paging file per process, so import on demand
|
||||
import cupyx.scipy.ndimage
|
||||
self._xp = cp
|
||||
self._sp = cupyx.scipy
|
||||
|
||||
def __init__(self, img : np.ndarray, copy=False):
|
||||
ndim = img.ndim
|
||||
if ndim not in [2,3,4]:
|
||||
raise ValueError(f'img.ndim must be 2,3,4, not {ndim}.')
|
||||
|
|
@ -55,8 +38,6 @@ class ImageProcessor:
|
|||
"""
|
||||
ip = ImageProcessor.__new__(ImageProcessor)
|
||||
ip._img = self._img
|
||||
ip._xp = self._xp
|
||||
ip._sp = self._sp
|
||||
return ip
|
||||
|
||||
def get_dims(self) -> Tuple[int,int,int,int]:
|
||||
|
|
@ -73,16 +54,11 @@ class ImageProcessor:
|
|||
def adjust_gamma(self, red : float, green : float, blue : float) -> 'ImageProcessor':
|
||||
dtype = self.get_dtype()
|
||||
self.to_ufloat32()
|
||||
|
||||
xp, img = self._xp , self._img,
|
||||
|
||||
xp.power(img, xp.array([1.0 / blue, 1.0 / green, 1.0 / red], xp.float32), out=img)
|
||||
xp.clip(img, 0, 1.0, out=img)
|
||||
|
||||
img = self._img
|
||||
np.power(img, np.array([1.0 / blue, 1.0 / green, 1.0 / red], np.float32), out=img)
|
||||
np.clip(img, 0, 1.0, out=img)
|
||||
self._img = img
|
||||
|
||||
self.to_dtype(dtype)
|
||||
|
||||
return self
|
||||
|
||||
|
||||
|
|
@ -124,7 +100,6 @@ class ImageProcessor:
|
|||
"""
|
||||
#if interpolation is None:
|
||||
# interpolation = ImageProcessor.Interpolation.LINEAR
|
||||
xp, sp = self._xp, self._sp
|
||||
img = self._img
|
||||
N,H,W,C = img.shape
|
||||
|
||||
|
|
@ -146,12 +121,7 @@ class ImageProcessor:
|
|||
|
||||
if scale != 1.0:
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
|
||||
if self._xp == np:
|
||||
img = cv2.resize (img, ( int(W*scale), int(H*scale) ), interpolation=ImageProcessor.Interpolation.LINEAR)
|
||||
else:
|
||||
img = sp.ndimage.zoom(img, (scale, scale, 1.0), order=1)
|
||||
|
||||
img = cv2.resize (img, ( int(W*scale), int(H*scale) ), interpolation=ImageProcessor.Interpolation.LINEAR)
|
||||
H,W = img.shape[0:2]
|
||||
img = img.reshape( (H,W,N,C) ).transpose( (2,0,1,3) )
|
||||
|
||||
|
|
@ -159,14 +129,13 @@ class ImageProcessor:
|
|||
w_pad = (TW-W) if TW is not None else 0
|
||||
h_pad = (TH-H) if TH is not None else 0
|
||||
if w_pad != 0 or h_pad != 0:
|
||||
img = xp.pad(img, ( (0,0), (0,h_pad), (0,w_pad), (0,0) ))
|
||||
img = np.pad(img, ( (0,0), (0,h_pad), (0,w_pad), (0,0) ))
|
||||
self._img = img
|
||||
|
||||
return scale
|
||||
|
||||
def clip(self, min, max) -> 'ImageProcessor':
|
||||
xp = self._xp
|
||||
xp.clip(self._img, min, max, out=self._img)
|
||||
np.clip(self._img, min, max, out=self._img)
|
||||
return self
|
||||
|
||||
def clip2(self, low_check, low_val, high_check, high_val) -> 'ImageProcessor':
|
||||
|
|
@ -188,22 +157,14 @@ class ImageProcessor:
|
|||
if interpolation is None:
|
||||
interpolation = ImageProcessor.Interpolation.LINEAR
|
||||
|
||||
xp, sp, img = self._xp, self._sp, self._img
|
||||
img = self._img
|
||||
|
||||
N,H,W,C = img.shape
|
||||
W_lr = max(4, int(W*(1.0-power)))
|
||||
H_lr = max(4, int(H*(1.0-power)))
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
|
||||
if xp == np:
|
||||
W_lr = max(4, int(W*(1.0-power)))
|
||||
H_lr = max(4, int(H*(1.0-power)))
|
||||
img = cv2.resize (img, (W_lr,H_lr), interpolation=_cv_inter[interpolation])
|
||||
img = cv2.resize (img, (W,H) , interpolation=_cv_inter[interpolation])
|
||||
else:
|
||||
W_lr = max(4, round(W*(1.0-power)))
|
||||
H_lr = max(4, round(H*(1.0-power)))
|
||||
img = sp.ndimage.zoom(img, (H_lr/H, W_lr/W, 1), order=_scipy_order[interpolation])
|
||||
img = sp.ndimage.zoom(img, (H/img.shape[0], W/img.shape[1], 1), order=_scipy_order[interpolation])
|
||||
|
||||
img = cv2.resize (img, (W_lr,H_lr), interpolation=_cv_inter[interpolation])
|
||||
img = cv2.resize (img, (W,H) , interpolation=_cv_inter[interpolation])
|
||||
img = img.reshape( (H,W,N,C) ).transpose( (2,0,1,3) )
|
||||
|
||||
self._img = img
|
||||
|
|
@ -223,17 +184,13 @@ class ImageProcessor:
|
|||
dtype = self.get_dtype()
|
||||
self.to_ufloat32()
|
||||
|
||||
xp, sp, img = self._xp, self._sp, self._img
|
||||
img = self._img
|
||||
N,H,W,C = img.shape
|
||||
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
|
||||
if xp == np:
|
||||
img_blur = cv2.medianBlur(img, size)
|
||||
img = ne.evaluate('img*(1.0-power) + img_blur*power')
|
||||
else:
|
||||
img_blur = sp.ndimage.median_filter(img, size=(size,size,1) )
|
||||
img = img*(1.0-power) + img_blur*power
|
||||
img_blur = cv2.medianBlur(img, size)
|
||||
img = ne.evaluate('img*(1.0-power) + img_blur*power')
|
||||
|
||||
img = img.reshape( (H,W,N,C) ).transpose( (2,0,1,3) )
|
||||
self._img = img
|
||||
|
|
@ -250,32 +207,23 @@ class ImageProcessor:
|
|||
fade_to_border(False) clip the image in order
|
||||
to fade smoothly to the border with specified blur amount
|
||||
"""
|
||||
xp, sp = self._xp, self._sp
|
||||
|
||||
erode, blur = int(erode), int(blur)
|
||||
|
||||
img = self._img
|
||||
dtype = img.dtype
|
||||
N,H,W,C = img.shape
|
||||
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
img = xp.pad (img, ( (H,H), (W,W), (0,0) ) )
|
||||
img = np.pad (img, ( (H,H), (W,W), (0,0) ) )
|
||||
|
||||
if erode > 0:
|
||||
el = xp.asarray(cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))
|
||||
el = np.asarray(cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))
|
||||
iterations = max(1,erode//2)
|
||||
if self._xp == np:
|
||||
img = cv2.erode(img, el, iterations = iterations )
|
||||
else:
|
||||
img = sp.ndimage.binary_erosion(img, el[...,None], iterations = iterations, brute_force=True ).astype(dtype)
|
||||
img = cv2.erode(img, el, iterations = iterations )
|
||||
|
||||
elif erode < 0:
|
||||
el = xp.asarray(cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))
|
||||
el = np.asarray(cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))
|
||||
iterations = max(1,-erode//2)
|
||||
if self._xp == np:
|
||||
img = cv2.dilate(img, el, iterations = iterations )
|
||||
else:
|
||||
img = sp.ndimage.binary_dilation(img, el[...,None], iterations = iterations, brute_force=True).astype(dtype)
|
||||
img = cv2.dilate(img, el, iterations = iterations )
|
||||
|
||||
if fade_to_border:
|
||||
h_clip_size = H + blur // 2
|
||||
|
|
@ -287,13 +235,8 @@ class ImageProcessor:
|
|||
|
||||
if blur > 0:
|
||||
sigma = blur * 0.125 * 2
|
||||
if self._xp == np:
|
||||
img = cv2.GaussianBlur(img, (0, 0), sigma)
|
||||
else:
|
||||
img = sp.ndimage.gaussian_filter(img, (sigma, sigma,0), mode='constant')
|
||||
img = cv2.GaussianBlur(img, (0, 0), sigma)
|
||||
|
||||
#if img.ndim == 2:
|
||||
# img = img[...,None]
|
||||
img = img[H:-H,W:-W]
|
||||
img = img.reshape( (H,W,N,C) ).transpose( (2,0,1,3) )
|
||||
|
||||
|
|
@ -301,15 +244,15 @@ class ImageProcessor:
|
|||
return self
|
||||
|
||||
def rotate90(self) -> 'ImageProcessor':
|
||||
self._img = self._xp.rot90(self._img, k=1, axes=(1,2) )
|
||||
self._img = np.rot90(self._img, k=1, axes=(1,2) )
|
||||
return self
|
||||
|
||||
def rotate180(self) -> 'ImageProcessor':
|
||||
self._img = self._xp.rot90(self._img, k=2, axes=(1,2) )
|
||||
self._img = np.rot90(self._img, k=2, axes=(1,2) )
|
||||
return self
|
||||
|
||||
def rotate270(self) -> 'ImageProcessor':
|
||||
self._img = self._xp.rot90(self._img, k=3, axes=(1,2) )
|
||||
self._img = np.rot90(self._img, k=3, axes=(1,2) )
|
||||
return self
|
||||
|
||||
def flip_horizontal(self) -> 'ImageProcessor':
|
||||
|
|
@ -330,11 +273,7 @@ class ImageProcessor:
|
|||
"""
|
||||
|
||||
"""
|
||||
xp = self._xp
|
||||
img = self._img
|
||||
img = xp.pad(img, ( (0,0), (t_h,b_h), (l_w,r_w), (0,0) ))
|
||||
|
||||
self._img = img
|
||||
self._img = np.pad(self._img, ( (0,0), (t_h,b_h), (l_w,r_w), (0,0) ))
|
||||
return self
|
||||
|
||||
def pad_to_next_divisor(self, dw=None, dh=None) -> 'ImageProcessor':
|
||||
|
|
@ -343,7 +282,6 @@ class ImageProcessor:
|
|||
|
||||
dw,dh int
|
||||
"""
|
||||
xp = self._xp
|
||||
img = self._img
|
||||
_,H,W,_ = img.shape
|
||||
|
||||
|
|
@ -360,24 +298,18 @@ class ImageProcessor:
|
|||
h_pad = dh - h_pad
|
||||
|
||||
if w_pad != 0 or h_pad != 0:
|
||||
img = xp.pad(img, ( (0,0), (0,h_pad), (0,w_pad), (0,0) ))
|
||||
img = np.pad(img, ( (0,0), (0,h_pad), (0,w_pad), (0,0) ))
|
||||
|
||||
self._img = img
|
||||
return self
|
||||
|
||||
def sharpen(self, factor : float, kernel_size=3) -> 'ImageProcessor':
|
||||
xp = self._xp
|
||||
img = self._img
|
||||
|
||||
N,H,W,C = img.shape
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
|
||||
if xp == np:
|
||||
blur = cv2.GaussianBlur(img, (kernel_size, kernel_size) , 0)
|
||||
img = cv2.addWeighted(img, 1.0 + (0.5 * factor), blur, -(0.5 * factor), 0)
|
||||
else:
|
||||
raise
|
||||
|
||||
blur = cv2.GaussianBlur(img, (kernel_size, kernel_size) , 0)
|
||||
img = cv2.addWeighted(img, 1.0 + (0.5 * factor), blur, -(0.5 * factor), 0)
|
||||
img = img.reshape( (H,W,N,C) ).transpose( (2,0,1,3) )
|
||||
|
||||
self._img = img
|
||||
|
|
@ -394,8 +326,6 @@ class ImageProcessor:
|
|||
|
||||
zero dim will be set to 1
|
||||
"""
|
||||
xp = self._xp
|
||||
|
||||
format = format.upper()
|
||||
img = self._img
|
||||
|
||||
|
|
@ -418,7 +348,7 @@ class ImageProcessor:
|
|||
transpose_order = [ d[s] for s in format ]
|
||||
img = img.transpose(transpose_order)
|
||||
|
||||
return xp.ascontiguousarray(img)
|
||||
return np.ascontiguousarray(img)
|
||||
|
||||
def ch(self, TC : int) -> 'ImageProcessor':
|
||||
"""
|
||||
|
|
@ -426,7 +356,6 @@ class ImageProcessor:
|
|||
|
||||
TC int >= 1
|
||||
"""
|
||||
xp = self._xp
|
||||
img = self._img
|
||||
N,H,W,C = img.shape
|
||||
|
||||
|
|
@ -436,7 +365,7 @@ class ImageProcessor:
|
|||
if TC > C:
|
||||
# Ch expand
|
||||
img = img[...,0:1] # Clip to single ch first.
|
||||
img = xp.repeat (img, TC, -1) # Expand by repeat
|
||||
img = np.repeat (img, TC, -1) # Expand by repeat
|
||||
elif TC < C:
|
||||
# Ch reduction clip
|
||||
img = img[...,:TC]
|
||||
|
|
@ -448,7 +377,7 @@ class ImageProcessor:
|
|||
"""
|
||||
Converts 3 ch bgr to grayscale.
|
||||
"""
|
||||
img, xp = self._img, self._xp
|
||||
img = self._img
|
||||
_,_,_,C = img.shape
|
||||
if C != 1:
|
||||
dtype = self.get_dtype()
|
||||
|
|
@ -458,7 +387,7 @@ class ImageProcessor:
|
|||
elif C >= 3:
|
||||
img = img[...,:3]
|
||||
|
||||
img = xp.dot(img, xp.array([0.1140, 0.5870, 0.2989], xp.float32)) [...,None]
|
||||
img = np.dot(img, np.array([0.1140, 0.5870, 0.2989], np.float32)) [...,None]
|
||||
img = img.astype(dtype)
|
||||
self._img = img
|
||||
|
||||
|
|
@ -468,8 +397,6 @@ class ImageProcessor:
|
|||
"""
|
||||
resize to (W,H)
|
||||
"""
|
||||
xp, sp = self._xp, self._sp
|
||||
|
||||
img = self._img
|
||||
N,H,W,C = img.shape
|
||||
|
||||
|
|
@ -479,12 +406,7 @@ class ImageProcessor:
|
|||
interpolation = ImageProcessor.Interpolation.LINEAR
|
||||
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
|
||||
if self._xp == np:
|
||||
img = cv2.resize (img, (TW, TH), interpolation=_cv_inter[interpolation])
|
||||
else:
|
||||
img = sp.ndimage.zoom(img, (TW/W, TH/H, 1), order=_scipy_order[interpolation])
|
||||
|
||||
img = cv2.resize (img, (TW, TH), interpolation=_cv_inter[interpolation])
|
||||
img = img.reshape( (TH,TW,N,C) ).transpose( (2,0,1,3) )
|
||||
|
||||
if new_ip:
|
||||
|
|
@ -498,25 +420,14 @@ class ImageProcessor:
|
|||
"""
|
||||
img HWC
|
||||
"""
|
||||
xp, sp, img = self._xp, self._sp, self._img
|
||||
img = self._img
|
||||
N,H,W,C = img.shape
|
||||
img = img.transpose( (1,2,0,3) ).reshape( (H,W,N*C) )
|
||||
|
||||
if interpolation is None:
|
||||
interpolation = ImageProcessor.Interpolation.LINEAR
|
||||
|
||||
if xp == np:
|
||||
img = cv2.warpAffine(img, mat, (out_width, out_height), flags=_cv_inter[interpolation] )
|
||||
else:
|
||||
# AffineMat inverse
|
||||
xp_mat = xp.get_array_module(mat)
|
||||
mat = xp_mat.linalg.inv(xp_mat.concatenate( ( mat, xp_mat.array([[0,0,1]], xp_mat.float32)), 0) )[0:2,:]
|
||||
|
||||
mx, my = xp.meshgrid( xp.arange(0, out_width, dtype=xp.float32), xp.arange(0, out_height, dtype=xp.float32) )
|
||||
coords = xp.concatenate( (mx[None,...], my[None,...], xp.ones( (1, out_height,out_width), dtype=xp.float32)), 0 )
|
||||
|
||||
mat_coords = xp.matmul (xp.asarray(mat), coords.reshape( (3,-1) ) ).reshape( (2,out_height,out_width))
|
||||
img = xp.concatenate([sp.ndimage.map_coordinates( img[...,c], mat_coords[::-1,...], order=_scipy_order[interpolation], mode='opencv' )[...,None] for c in range(N*C) ], -1)
|
||||
img = cv2.warpAffine(img, mat, (out_width, out_height), flags=_cv_inter[interpolation] )
|
||||
|
||||
img = img.reshape( (out_height,out_width,N,C) ).transpose( (2,0,1,3) )
|
||||
self._img = img
|
||||
|
|
@ -531,23 +442,20 @@ class ImageProcessor:
|
|||
"""
|
||||
change image format to float32
|
||||
"""
|
||||
xp = self._xp
|
||||
self._img = self._img.astype(xp.float32)
|
||||
self._img = self._img.astype(np.float32)
|
||||
return self
|
||||
|
||||
def as_uint8(self) -> 'ImageProcessor':
|
||||
"""
|
||||
change image format to uint8
|
||||
"""
|
||||
xp = self._xp
|
||||
self._img = self._img.astype(xp.uint8)
|
||||
self._img = self._img.astype(np.uint8)
|
||||
return self
|
||||
|
||||
def to_dtype(self, dtype) -> 'ImageProcessor':
|
||||
xp = self._xp
|
||||
if dtype == xp.float32:
|
||||
if dtype == np.float32:
|
||||
return self.to_ufloat32()
|
||||
elif dtype == xp.uint8:
|
||||
elif dtype == np.uint8:
|
||||
return self.to_uint8()
|
||||
else:
|
||||
raise ValueError('unsupported dtype')
|
||||
|
|
@ -558,9 +466,8 @@ class ImageProcessor:
|
|||
if current image dtype uint8, then image will be divided by / 255.0
|
||||
otherwise no operation
|
||||
"""
|
||||
xp = self._xp
|
||||
if self._img.dtype == xp.uint8:
|
||||
self._img = self._img.astype(xp.float32)
|
||||
if self._img.dtype == np.uint8:
|
||||
self._img = self._img.astype(np.float32)
|
||||
self._img /= 255.0
|
||||
|
||||
return self
|
||||
|
|
@ -571,17 +478,13 @@ class ImageProcessor:
|
|||
|
||||
if current image dtype is float32/64, then image will be multiplied by *255
|
||||
"""
|
||||
xp = self._xp
|
||||
img = self._img
|
||||
|
||||
if img.dtype in [xp.float32, xp.float64]:
|
||||
if img.dtype in [np.float32, np.float64]:
|
||||
img *= 255.0
|
||||
img[img < 0] = 0
|
||||
img[img > 255] = 255
|
||||
|
||||
img = img.astype(xp.uint8, copy=False)
|
||||
self._img = img
|
||||
np.clip(img, 0, 255, out=img)
|
||||
|
||||
self._img = img.astype(np.uint8, copy=False)
|
||||
return self
|
||||
|
||||
class Interpolation(IntEnum):
|
||||
|
|
@ -590,6 +493,3 @@ class ImageProcessor:
|
|||
|
||||
_cv_inter = { ImageProcessor.Interpolation.LINEAR : cv2.INTER_LINEAR,
|
||||
ImageProcessor.Interpolation.CUBIC : cv2.INTER_CUBIC }
|
||||
|
||||
_scipy_order = { ImageProcessor.Interpolation.LINEAR : 1,
|
||||
ImageProcessor.Interpolation.CUBIC : 3 }
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue