diff --git a/src/models/sdynet.py b/src/models/sdynet.py index e500c59..75f0fb3 100644 --- a/src/models/sdynet.py +++ b/src/models/sdynet.py @@ -4,7 +4,7 @@ import torch.nn.functional as F import numpy as np from common.utils import round_func from common import lut -from common.layers import PercievePattern, DenseConvUpscaleBlock +from common import layers from pathlib import Path from . import sdylut @@ -12,12 +12,12 @@ class SDYNetx1(nn.Module): def __init__(self, hidden_dim = 64, layers_count = 4, scale = 4): super(SDYNetx1, self).__init__() self.scale = scale - self._extract_pattern_S = PercievePattern(receptive_field_idxes=[[0,0],[0,1],[1,0],[1,1]], center=[0,0], window_size=3) - self._extract_pattern_D = PercievePattern(receptive_field_idxes=[[0,0],[2,0],[0,2],[2,2]], center=[0,0], window_size=3) - self._extract_pattern_Y = PercievePattern(receptive_field_idxes=[[0,0],[1,1],[1,2],[2,1]], center=[0,0], window_size=3) - self.stageS = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - self.stageD = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - self.stageY = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) + s_pattern = [[0,0],[0,1],[1,0],[1,1]] + d_pattern = [[0,0],[2,0],[0,2],[2,2]] + y_pattern = [[0,0],[1,1],[1,2],[2,1]] + self.stage1_S = layers.UpscaleBlock(receptive_field_idxes=s_pattern, center=[0,0], window_size=2, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) + self.stage1_D = layers.UpscaleBlock(receptive_field_idxes=d_pattern, center=[0,0], window_size=3, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) + self.stage1_Y = layers.UpscaleBlock(receptive_field_idxes=y_pattern, center=[0,0], window_size=3, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) def forward(self, x): b,c,h,w = x.shape @@ -25,99 +25,65 @@ class SDYNetx1(nn.Module): output = torch.zeros([b*c, 1, h*self.scale, w*self.scale], dtype=x.dtype, device=x.device) for rotations_count in range(4): rotated = torch.rot90(x, k=rotations_count, dims=[-2, -1]) - rb,rc,rh,rw = rotated.shape - - s = self.stageS(self._extract_pattern_S(rotated)) - s = s.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - s = torch.rot90(s, k=-rotations_count, dims=[-2, -1]) - output += s - - d = self.stageD(self._extract_pattern_D(rotated)) - d = d.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - d = torch.rot90(d, k=-rotations_count, dims=[-2, -1]) - output += d - - y = self.stageY(self._extract_pattern_Y(rotated)) - y = y.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - y = torch.rot90(y, k=-rotations_count, dims=[-2, -1]) - output += y + rb,rc,rh,rw = rotated.shape + output += torch.rot90(self.stage1_S(rotated), k=-rotations_count, dims=[-2, -1]) + output += torch.rot90(self.stage1_D(rotated), k=-rotations_count, dims=[-2, -1]) + output += torch.rot90(self.stage1_Y(rotated), k=-rotations_count, dims=[-2, -1]) output /= 4*3 - output = output.view(b, c, h*self.scale, w*self.scale) - return output + x = output + x = round_func(x) + x = x.view(b, c, h*self.scale, w*self.scale) + return x def get_lut_model(self, quantization_interval=16, batch_size=2**10): - stageS = lut.transfer_2x2_input_SxS_output(self.stageS, quantization_interval=quantization_interval, batch_size=batch_size) - stageD = lut.transfer_2x2_input_SxS_output(self.stageD, quantization_interval=quantization_interval, batch_size=batch_size) - stageY = lut.transfer_2x2_input_SxS_output(self.stageY, quantization_interval=quantization_interval, batch_size=batch_size) + stageS = lut.transfer_2x2_input_SxS_output(self.stage1_S, quantization_interval=quantization_interval, batch_size=batch_size) + stageD = lut.transfer_2x2_input_SxS_output(self.stage1_D, quantization_interval=quantization_interval, batch_size=batch_size) + stageY = lut.transfer_2x2_input_SxS_output(self.stage1_Y, quantization_interval=quantization_interval, batch_size=batch_size) lut_model = sdylut.SDYLutx1.init_from_lut(stageS, stageD, stageY) return lut_model - class SDYNetx2(nn.Module): def __init__(self, hidden_dim = 64, layers_count = 4, scale = 4): super(SDYNetx2, self).__init__() self.scale = scale - self._extract_pattern_S = PercievePattern(receptive_field_idxes=[[0,0],[0,1],[1,0],[1,1]], center=[0,0], window_size=3) - self._extract_pattern_D = PercievePattern(receptive_field_idxes=[[0,0],[2,0],[0,2],[2,2]], center=[0,0], window_size=3) - self._extract_pattern_Y = PercievePattern(receptive_field_idxes=[[0,0],[1,1],[1,2],[2,1]], center=[0,0], window_size=3) - self.stage1_S = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) - self.stage1_D = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) - self.stage1_Y = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) - self.stage2_S = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - self.stage2_D = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - self.stage2_Y = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) + s_pattern = [[0,0],[0,1],[1,0],[1,1]] + d_pattern = [[0,0],[2,0],[0,2],[2,2]] + y_pattern = [[0,0],[1,1],[1,2],[2,1]] + self.stage1_S = layers.UpscaleBlock(receptive_field_idxes=s_pattern, center=[0,0], window_size=2, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) + self.stage1_D = layers.UpscaleBlock(receptive_field_idxes=d_pattern, center=[0,0], window_size=3, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) + self.stage1_Y = layers.UpscaleBlock(receptive_field_idxes=y_pattern, center=[0,0], window_size=3, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=1) + self.stage2_S = layers.UpscaleBlock(receptive_field_idxes=s_pattern, center=[0,0], window_size=2, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) + self.stage2_D = layers.UpscaleBlock(receptive_field_idxes=d_pattern, center=[0,0], window_size=3, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) + self.stage2_Y = layers.UpscaleBlock(receptive_field_idxes=y_pattern, center=[0,0], window_size=3, hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) def forward(self, x): b,c,h,w = x.shape x = x.view(b*c, 1, h, w) output_1 = torch.zeros([b*c, 1, h, w], dtype=x.dtype, device=x.device) - for rotations_count in range(4): + output_1 += self.stage1_S(x) + output_1 += self.stage1_D(x) + output_1 += self.stage1_Y(x) + for rotations_count in range(1,4): rotated = torch.rot90(x, k=rotations_count, dims=[-2, -1]) - rb,rc,rh,rw = rotated.shape - - s = self.stage1_S(self._extract_pattern_S(rotated)) - s = s.view(rb*rc, 1, rh, rw, 1, 1).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh, rw) - s = torch.rot90(s, k=-rotations_count, dims=[-2, -1]) - output_1 += s - - d = self.stage1_D(self._extract_pattern_D(rotated)) - d = d.view(rb*rc, 1, rh, rw, 1, 1).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh, rw) - d = torch.rot90(d, k=-rotations_count, dims=[-2, -1]) - output_1 += d - - y = self.stage1_Y(self._extract_pattern_Y(rotated)) - y = y.view(rb*rc, 1, rh, rw, 1, 1).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh, rw) - y = torch.rot90(y, k=-rotations_count, dims=[-2, -1]) - output_1 += y - + output_1 += torch.rot90(self.stage1_S(rotated), k=-rotations_count, dims=[-2, -1]) + output_1 += torch.rot90(self.stage1_D(rotated), k=-rotations_count, dims=[-2, -1]) + output_1 += torch.rot90(self.stage1_Y(rotated), k=-rotations_count, dims=[-2, -1]) output_1 /= 4*3 - output_1 = output_1.view(b*c, 1, h, w) - x = output_1 - + x = round_func(output_1) output_2 = torch.zeros([b*c, 1, h*self.scale, w*self.scale], dtype=x.dtype, device=x.device) - for rotations_count in range(4): + output_2 += self.stage2_S(x) + output_2 += self.stage2_D(x) + output_2 += self.stage2_Y(x) + for rotations_count in range(1,4): rotated = torch.rot90(x, k=rotations_count, dims=[-2, -1]) - rb,rc,rh,rw = rotated.shape - - s = self.stage2_S(self._extract_pattern_S(rotated)) - s = s.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - s = torch.rot90(s, k=-rotations_count, dims=[-2, -1]) - output_2 += s - - d = self.stage2_D(self._extract_pattern_D(rotated)) - d = d.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - d = torch.rot90(d, k=-rotations_count, dims=[-2, -1]) - output_2 += d - - y = self.stage2_Y(self._extract_pattern_Y(rotated)) - y = y.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - y = torch.rot90(y, k=-rotations_count, dims=[-2, -1]) - output_2 += y - + output_2 += torch.rot90(self.stage2_S(rotated), k=-rotations_count, dims=[-2, -1]) + output_2 += torch.rot90(self.stage2_D(rotated), k=-rotations_count, dims=[-2, -1]) + output_2 += torch.rot90(self.stage2_Y(rotated), k=-rotations_count, dims=[-2, -1]) output_2 /= 4*3 - output_2 = output_2.view(b, c, h*self.scale, w*self.scale) - return output_2 + x = round_func(output_2) + x = x.view(b, c, h*self.scale, w*self.scale) + return x def get_lut_model(self, quantization_interval=16, batch_size=2**10): stage1_S = lut.transfer_2x2_input_SxS_output(self.stage1_S, quantization_interval=quantization_interval, batch_size=batch_size) @@ -127,51 +93,4 @@ class SDYNetx2(nn.Module): stage2_D = lut.transfer_2x2_input_SxS_output(self.stage2_D, quantization_interval=quantization_interval, batch_size=batch_size) stage2_Y = lut.transfer_2x2_input_SxS_output(self.stage2_Y, quantization_interval=quantization_interval, batch_size=batch_size) lut_model = sdylut.SDYLutx2.init_from_lut(stage1_S, stage1_D, stage1_Y, stage2_S, stage2_D, stage2_Y) - return lut_model - - - -class SDYNetCenteredx1(nn.Module): - def __init__(self, hidden_dim = 64, layers_count = 4, scale = 4): - super(SDYNetCenteredx1, self).__init__() - self.scale = scale - self._extract_pattern_S = PercievePattern(receptive_field_idxes=[[0,0],[0,1],[1,0],[1,1]], center=[1,1], window_size=3) - self._extract_pattern_D = PercievePattern(receptive_field_idxes=[[0,0],[2,0],[0,2],[2,2]], center=[1,1], window_size=3) - self._extract_pattern_Y = PercievePattern(receptive_field_idxes=[[0,0],[1,1],[1,2],[2,1]], center=[1,1], window_size=3) - self.stageS = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - self.stageD = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - self.stageY = DenseConvUpscaleBlock(hidden_dim=hidden_dim, layers_count=layers_count, upscale_factor=scale) - - def forward(self, x): - b,c,h,w = x.shape - x = x.view(b*c, 1, h, w) - output = torch.zeros([b*c, 1, h*self.scale, w*self.scale], dtype=x.dtype, device=x.device) - for rotations_count in range(4): - rotated = torch.rot90(x, k=rotations_count, dims=[-2, -1]) - rb,rc,rh,rw = rotated.shape - - s = self.stageS(self._extract_pattern_S(rotated)) - s = s.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - s = torch.rot90(s, k=-rotations_count, dims=[-2, -1]) - output += s - - d = self.stageD(self._extract_pattern_D(rotated)) - d = d.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - d = torch.rot90(d, k=-rotations_count, dims=[-2, -1]) - output += d - - y = self.stageY(self._extract_pattern_Y(rotated)) - y = y.view(rb*rc, 1, rh, rw, self.scale, self.scale).permute(0,1,2,4,3,5).reshape(rb*rc, 1, rh*self.scale, rw*self.scale) - y = torch.rot90(y, k=-rotations_count, dims=[-2, -1]) - output += y - - output /= 4*3 - output = output.view(b, c, h*self.scale, w*self.scale) - return output - - def get_lut_model(self, quantization_interval=16, batch_size=2**10): - stageS = lut.transfer_2x2_input_SxS_output(self.stageS, quantization_interval=quantization_interval, batch_size=batch_size) - stageD = lut.transfer_2x2_input_SxS_output(self.stageD, quantization_interval=quantization_interval, batch_size=batch_size) - stageY = lut.transfer_2x2_input_SxS_output(self.stageY, quantization_interval=quantization_interval, batch_size=batch_size) - lut_model = sdylut.SDYLutCenteredx1.init_from_lut(stageS, stageD, stageY) return lut_model \ No newline at end of file