Spaces:
Running
on
Zero
Running
on
Zero
"""NamedCurves model with interactive functionality. This version builds upon model.py and bezier_control_point_estimator.py by incorporating additional parameters.""" | |
from models.attention_fusion import LocalFusion | |
from models.color_naming import ColorNaming | |
from models.backbone import Backbone | |
from torch import nn | |
from PIL import Image | |
from torchvision.transforms import functional as TF | |
import torch | |
class NamedCurves(nn.Module): | |
def __init__(self, configs: dict, device="cuda"): | |
super().__init__() | |
self.model_configs = configs | |
self.backbone = Backbone(**configs['backbone']['params']) | |
self.color_naming = ColorNaming(num_categories=configs['color_naming']['num_categories'], device=device) | |
self.bcpe = BCPE(**configs['bezier_control_points_estimator']['params']) | |
self.local_fusion = LocalFusion(**configs['local_fusion']['params']) | |
def forward(self, x, return_backbone=False, return_curves=False, control_points=None): | |
x_backbone = self.backbone(x) | |
cn_probs = self.color_naming(x_backbone) | |
if return_curves: | |
x_global, control_points = self.bcpe(x_backbone, cn_probs, return_control_points=return_curves, control_points=control_points) | |
else: | |
x_global = self.bcpe(x_backbone, cn_probs, control_points=control_points) | |
out = self.local_fusion(x_global, cn_probs, q=x_backbone) | |
if return_backbone: | |
return out, x_backbone | |
if return_curves: | |
return out, control_points | |
return out | |
class ContextualFeatureExtractor(nn.Module): | |
def __init__(self): | |
super().__init__() | |
self.main = nn.Sequential( | |
nn.Conv2d(3, 8, 3, 1, 1), | |
nn.ReLU(), | |
nn.Conv2d(8, 16, 3, 1, 1), | |
nn.ReLU(), | |
nn.Conv2d(16, 32, 3, 1, 1), | |
nn.ReLU(), | |
nn.Dropout(0.2), | |
nn.Conv2d(32, 64, 3, 1, 1), | |
nn.ReLU()) | |
def forward(self, x): | |
return self.main(x) | |
class BezierColorBranch(nn.Module): | |
def __init__(self, num_control_points=10): | |
super().__init__() | |
self.num_control_points = num_control_points # +1, (0, 0) point | |
self.color_branch = nn.Sequential( | |
nn.Conv2d(65, 64, 3, 1, 1), | |
nn.ReLU(), | |
nn.MaxPool2d(2, 2), | |
nn.Conv2d(64, 32, 3, 1, 1), | |
nn.ReLU(), | |
nn.MaxPool2d(2, 2), | |
nn.Conv2d(32, 32, 3, 1, 1), | |
nn.ReLU(), | |
nn.Conv2d(32, 3 * self.num_control_points, 3, 1, 1), | |
nn.AdaptiveAvgPool2d((1, 1))) | |
self.sigmoid = nn.Sigmoid() | |
def create_control_points(self, x): | |
x = torch.cumsum(torch.cat([torch.zeros_like(x[..., :1]), x], dim=-1), dim=-1) | |
x = torch.stack([x, torch.linspace(0, 1, steps=self.num_control_points+1).unsqueeze(0).repeat(x.shape[0], x.shape[1], 1).to(x.device)], dim=-1) | |
return x | |
def forward(self, x): | |
x = self.color_branch(x).view(x.size(0), 3, self.num_control_points) | |
x = self.sigmoid(x) | |
x = x / torch.sum(x, dim=2)[..., None] | |
x = self.create_control_points(x) | |
return x | |
class BCPE(nn.Module): | |
def __init__(self, num_categories=6, num_control_points=10): | |
super().__init__() | |
self.contextual_feature_extractor = ContextualFeatureExtractor() | |
self.color_branches = nn.ModuleList([BezierColorBranch(num_control_points) for _ in range(num_categories)]) | |
def binomial_coefficient(self, n, k): | |
""" | |
Calculate the binomial coefficient (n choose k). | |
""" | |
if k < 0 or k > n: | |
return 0.0 | |
result = 1.0 | |
for i in range(min(k, n - k)): | |
result *= (n - i) | |
result //= (i + 1) | |
return result | |
def apply_cubic_bezier(self, x, control_points): | |
n = control_points.shape[2] | |
output = torch.zeros_like(x) | |
for j in range(n): | |
output += control_points[..., j, 0].view(control_points.shape[0], control_points.shape[1], 1, 1) * self.binomial_coefficient(n - 1, j) * (1 - x) ** (n - 1 - j) * x ** j | |
return output | |
def forward(self, x, cn_probs, return_control_points=False, control_points=None): | |
feat = self.contextual_feature_extractor(x) | |
bezier_control_points = [color_branch(torch.cat((feat, color_probs.unsqueeze(1)), dim=1).float()) for color_branch, color_probs in zip(self.color_branches, cn_probs)] | |
if control_points is not None: | |
bezier_control_points = control_points | |
global_adjusted_images = torch.stack([self.apply_cubic_bezier(x, control_points) for control_points in bezier_control_points], dim=0) | |
if return_control_points: | |
return global_adjusted_images, bezier_control_points | |
return global_adjusted_images |