Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
64 changes: 33 additions & 31 deletions hyperswap/src/dataset.py
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,8 @@
from configparser import ConfigParser
from typing import cast

import albumentations
import torch
from kornia import augmentation
from torch import Tensor
from torch.utils.data import Dataset
from torchvision import io, transforms
Expand Down Expand Up @@ -76,9 +77,8 @@ def compose_transforms(self) -> transforms:
return transforms.Compose(
[
AugmentTransform(),
transforms.ToPILImage(),
transforms.Resize((self.config_transform_size, self.config_transform_size), interpolation = transforms.InterpolationMode.BICUBIC),
transforms.ToTensor(),
transforms.ConvertImageDtype(torch.float32),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

Expand Down Expand Up @@ -119,36 +119,38 @@ def __init__(self) -> None:
self.transforms = self.compose_transforms()

def __call__(self, input_tensor : Tensor) -> Tensor:
temp_tensor = input_tensor.numpy().transpose(1, 2, 0)
return self.transforms(image = temp_tensor).get('image')
temp_tensor = transforms.functional.convert_image_dtype(input_tensor, torch.float32).unsqueeze(0)
temp_tensor = self.transforms(temp_tensor).squeeze(0).clamp(0, 1)
return transforms.functional.convert_image_dtype(temp_tensor, torch.uint8)

@staticmethod
def compose_transforms() -> albumentations.Compose:
return albumentations.Compose(
[
albumentations.HorizontalFlip(p = 0.5),
albumentations.OneOf(
[
albumentations.MotionBlur(),
albumentations.ZoomBlur(max_factor = (1.0, 1.2))
], p = 0.1),
albumentations.OneOf(
[
albumentations.RandomGamma(),
albumentations.RandomBrightnessContrast(),
albumentations.Illumination()
], p = 0.2),
albumentations.OneOf(
[
albumentations.ColorJitter(),
albumentations.RGBShift(),
albumentations.HueSaturationValue()
], p = 0.2),
albumentations.Affine(
translate_percent = (-0.05, 0.05),
def compose_transforms() -> augmentation.AugmentationSequential:
return augmentation.AugmentationSequential(
augmentation.RandomHorizontalFlip(p = 0.5),
augmentation.ImageSequential(
augmentation.RandomMotionBlur(kernel_size = (3, 7), angle = (-90, 90), direction = (-1, 1), p = 0.1),
augmentation.RandomGaussianBlur(kernel_size = (5, 5), sigma = (0.1, 2.0), p = 0.1),
random_apply = 1
),
augmentation.ImageSequential(
augmentation.RandomGamma(gamma = (0.8, 1.2), p = 0.2),
augmentation.RandomBrightness(brightness = (0.8, 1.2), p = 0.2),
augmentation.RandomContrast(contrast = (0.8, 1.2), p = 0.2),
augmentation.RandomLinearIllumination(gain = (0.01, 0.2), p = 0.2),
random_apply = 1
),
augmentation.ImageSequential(
augmentation.ColorJitter(brightness = 0.2, contrast = 0.2, saturation = 0.2, hue = 0.1, p = 0.2),
augmentation.RandomRGBShift(r_shift_limit = 0.08, g_shift_limit = 0.08, b_shift_limit = 0.08, p = 0.2),
augmentation.RandomHue(hue = (-0.1, 0.1), p = 0.2),
augmentation.RandomSaturation(saturation = (0.8, 1.2), p = 0.2),
random_apply = 1
),
augmentation.RandomAffine(
degrees = (-2, 2),
translate = (0.05, 0.05),
scale = (0.95, 1.05),
rotate = (-2, 2),
border_mode = 1,
padding_mode = 'border',
p = 0.2
)
])
)
13 changes: 0 additions & 13 deletions hyperswap/src/helper.py
Original file line number Diff line number Diff line change
Expand Up @@ -64,19 +64,6 @@ def overlay_mask(input_tensor : Tensor, input_mask : Mask) -> Tensor:
return output_tensor


def blur_tensor(input_tensor : Tensor, kernel_size : int, sigma : float) -> Tensor:
coordinate_tensor = torch.arange(kernel_size, dtype = input_tensor.dtype, device = input_tensor.device) - (kernel_size - 1) * 0.5
kernel_tensor = torch.exp(-coordinate_tensor ** 2 / (2 * sigma ** 2))
kernel_tensor = kernel_tensor / kernel_tensor.sum()
horizontal_kernel = kernel_tensor.view(1, 1, 1, kernel_size).repeat(input_tensor.shape[1], 1, 1, 1)
vertical_kernel = kernel_tensor.view(1, 1, kernel_size, 1).repeat(input_tensor.shape[1], 1, 1, 1)
padding = kernel_size // 2
temp_tensor = nn.functional.pad(input_tensor, (padding, padding, padding, padding), mode = 'reflect')
temp_tensor = nn.functional.conv2d(temp_tensor, horizontal_kernel, groups = input_tensor.shape[1])
output_tensor = nn.functional.conv2d(temp_tensor, vertical_kernel, groups = input_tensor.shape[1])
return output_tensor


def dilate_mask(input_tensor : Tensor, factor : float) -> Tensor:
padding = int(input_tensor.shape[2] * factor + 0.5)
kernel_size = 1 + 2 * padding
Expand Down
5 changes: 3 additions & 2 deletions hyperswap/src/models/loss.py
Original file line number Diff line number Diff line change
Expand Up @@ -2,10 +2,11 @@
from typing import List, Tuple

import torch
from kornia import filters
from pytorch_msssim import ssim
from torch import Tensor, nn

from ..helper import blur_tensor, calculate_face_embedding, dilate_mask, oddify_size
from ..helper import calculate_face_embedding, dilate_mask, oddify_size
from ..types import EmbedderModule, FaceAlignerModule, FaceMaskerModule, Feature, Loss, Mask


Expand Down Expand Up @@ -161,7 +162,7 @@ def forward(self, target_tensor : Tensor, output_tensor : Tensor) -> Tuple[Loss,
def calculate_pose(self, input_tensor : Tensor) -> Tensor:
temp_tensor = torch.nn.functional.interpolate(input_tensor, (256, 256), mode = 'area')
temp_tensor = (temp_tensor.clip(-1, 1) + 1) * 0.5
temp_tensor = blur_tensor(temp_tensor, 9, 1.7)
temp_tensor = filters.gaussian_blur2d(temp_tensor, (9, 9), (1.7, 1.7))

_, pose_tensor, _ = self.face_aligner(temp_tensor)
yaw_tensor = nn.functional.normalize(pose_tensor[:, 0:2], p = 2)
Expand Down
2 changes: 1 addition & 1 deletion requirements.txt
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
--extra-index-url https://download.pytorch.org/whl/cu128
albumentations==2.0.8
kornia==0.8.3
lightning==2.5.5
onnx==1.18.0
onnxruntime==1.22.0
Expand Down
Loading