def generate_batch_state(self, batch_tensor: torch.FloatTensor) -> torch.FloatTensor:
batch_sz, channels, height, width = batch_tensor.size()
roughness = type(self).roughness(batch_sz, device=batch_tensor.device)
plasma_sz = (batch_sz, 1, height, width)
alpha_range = type(self).alpha_range(batch_sz, device=batch_tensor.device).view(-1,1,1,1)
alpha_mean = type(self).alpha_mean(batch_sz, device=batch_tensor.device).view(-1,1,1,1)
beta_range = type(self).beta_range(batch_sz, device=batch_tensor.device).view(-1,1,1,1)
beta_mean = type(self).beta_mean(batch_sz, device=batch_tensor.device).view(-1,1,1,1)
alpha_plasma = functional_diamond_square(plasma_sz, roughness=roughness, device=batch_tensor.device)
#print("RndAlpha:",alpha_plasma.min().item(),alpha_plasma.max().item())
alpha_plasma = (alpha_plasma * alpha_range) + (1-alpha_range) * alpha_mean
#print("RndAlpha2:",alpha_plasma.min().item(),alpha_plasma.max().item())
beta_plasma = functional_diamond_square(plasma_sz, roughness=roughness, device=batch_tensor.device)
#print("RndBeta:",beta_plasma.min().item(),beta_plasma.max().item())
beta_plasma = (beta_plasma * beta_range) + (1-beta_range) * beta_mean + beta_range * .5
#print("RndBeta2:",beta_plasma.min().item(),beta_plasma.max().item())
#beta_available = (1 - alpha_plasma)
#beta_alpha = beta_available * beta_alpha
#beta_beta = (beta_available - beta_alpha) * beta_beta
#beta_plasma = functional_diamond_square(plasma_sz, roughness=roughness, device=batch_tensor.device)
#beta_plasma = beta_plasma * beta_alpha + beta_beta
#print ("Ranges:",(alpha_plasma+beta_plasma).min().item(),(alpha_plasma+beta_plasma).max().item())
plasma_sum = (alpha_plasma+beta_plasma)
alpha_plasmam, beta_plasma = (alpha_plasma/plasma_sum,beta_plasma/plasma_sum)
return alpha_plasma, beta_plasma,
def generate_batch_state(self, batch_tensor: torch.FloatTensor) -> torch.FloatTensor:
batch_sz, channels, height, width = batch_tensor.size()
roughness = type(self).roughness(batch_sz, device=batch_tensor.device)
plasma_sz = (batch_sz, 3, height, width)
intensity = type(self).intensity(batch_sz, device=batch_tensor.device).view(-1, 1, 1, 1)
brightness_map = 2 * functional_diamond_square(plasma_sz, roughness=roughness, device=batch_tensor.device) - 1
return brightness_map * intensity,
def generate_batch_state(self, image_batch: torch.Tensor) -> AugmentationState:
batch_sz, _, width, height = image_batch.size()
roughness = type(self).roughness(batch_sz, device=image_batch.device)
plasma_sz = (batch_sz, 1, width, height)
plasma = functional_diamond_square(plasma_sz, roughness=roughness, device=image_batch.device)
erase_percentile = type(self).erase_percentile(batch_sz, device=image_batch.device)
return plasma, erase_percentile
def generate_batch_state(self, batch_tensor: torch.FloatTensor) -> torch.FloatTensor:
batch_sz, channels, height, width = batch_tensor.size()
roughness = type(self).roughness(batch_sz, device=batch_tensor.device)
tensor_sz = batch_tensor.size()
noise = type(self).noise(tensor_sz, device=batch_tensor.device)
plasma_sz = (batch_sz, 1, height, width)
noise_intencity_coefficient = functional_diamond_square(plasma_sz, roughness=roughness, device=batch_tensor.device)
return noise * noise_intencity_coefficient,
def generate_batch_state(self, sampling_tensors: SamplingField) -> SpatialAugmentationState:
batch_sz, width, height = sampling_tensors[0].size()
roughness = type(self).roughness(batch_sz, device=sampling_tensors[0].device)
plasma_sz = (batch_sz, 1, width, height)
plasma = functional_diamond_square(plasma_sz, roughness=roughness, device=sampling_tensors[0].device)
inside = type(self).inside(batch_sz, device=sampling_tensors[0].device).float()
erase_percentile = type(self).erase_percentile(batch_sz, device=sampling_tensors[0].device)
return plasma, inside, erase_percentile
def generate_batch_state(
self, batch_tensor: torch.FloatTensor) -> torch.FloatTensor:
batch_sz, channels, height, width = batch_tensor.size()
roughness = type(self).roughness(batch_sz, device=batch_tensor.device)
plasma_sz = (batch_sz, 1, height, width)
contrast_map = 4 * functional_diamond_square(
plasma_sz, roughness=roughness, device=batch_tensor.device)
return contrast_map,
def generate_batch_state(self, batch_tensor: torch.FloatTensor) -> torch.FloatTensor:
batch_sz, channels, height, width = batch_tensor.size()
roughness = type(self).roughness(batch_sz, device=batch_tensor.device)
shade_intensity = type(self).shade_intensity(batch_sz, device=batch_tensor.device).view(-1, 1, 1, 1)
shade_quantity = type(self).shade_quantity(batch_sz, device=batch_tensor.device).view(-1, 1, 1, 1)
plasma_sz = (batch_sz, 1, height, width)
shade_map = functional_diamond_square(plasma_sz, roughness=roughness, device=batch_tensor.device)
shade_map = (shade_map < shade_quantity).float() * shade_intensity
return shade_map,
def forward_batch_img(self, tensor_image):
batch_size = tensor_image.size(0)
roughness = type(self).roughness(batch_size)
pixel_means = type(self).pixel_means(batch_size, device=tensor_image.device).view(-1, 1, 1, 1)
pixel_ranges = type(self).pixel_ranges(batch_size, device=tensor_image.device).view(-1, 1, 1, 1)
plasma = functional_diamond_square(tensor_image.size(), roughness=roughness, device=tensor_image.device)
plasma_reshaped = plasma.reshape([batch_size, -1])
plasma_min = plasma_reshaped.min(dim=1)[0].view(-1, 1, 1, 1)
plasma_max = plasma_reshaped.max(dim=1)[0].view(-1, 1, 1, 1)
return (pixel_ranges * (plasma - plasma_min) / (plasma_max - plasma_min)) + pixel_means - pixel_ranges / 2
def generate_batch_state(self, sampling_tensors: SamplingField) -> AugmentationState:
batch_sz, height, width = sampling_tensors[0].size()
roughness = type(self).roughness(batch_sz, device=sampling_tensors[0].device)
intensity = type(self).intensity(batch_sz, device=sampling_tensors[0].device)
plasma_sz = (batch_sz, 1, height, width)
plasma_x = functional_diamond_square(plasma_sz, roughness=roughness, device=sampling_tensors[0].device) - .5
plasma_y = functional_diamond_square(plasma_sz, roughness=roughness, device=sampling_tensors[0].device) - .5
plasma_x, plasma_y = plasma_x[:,0,:,:], plasma_y[:, 0,:,:]
plasma_dx = plasma_x[:, :, 1:] - plasma_x[:, :, :-1]
plasma_dy = plasma_y[:, 1:, :] - plasma_y[:, :-1, :]
plasma_scale_x = torch.cat([abs(plasma_dx.view(batch_sz, -1).min(dim=1)[0]).view(1, -1), plasma_dx.view(batch_sz, -1).max(dim=1)[0].view(1, -1)], dim=0).max(dim=0)[0]
plasma_scale_y = torch.cat([abs(plasma_dy.view(batch_sz, -1).min(dim=1)[0]).view(1, -1), plasma_dy.view(batch_sz, -1).max(dim=1)[0].view(1, -1)], dim=0).max(dim=0)[0]
plasma_x /= ((plasma_scale_x.view(-1, 1, 1) * .25 * width) / intensity.view(-1, 1, 1))
plasma_y /= ((plasma_scale_y.view(-1, 1, 1) * .25 * height) / intensity.view(-1, 1, 1))
return plasma_x, plasma_y