diff --git a/codes/models/ExtensibleTrainer.py b/codes/models/ExtensibleTrainer.py index 20130840..64940b06 100644 --- a/codes/models/ExtensibleTrainer.py +++ b/codes/models/ExtensibleTrainer.py @@ -213,6 +213,8 @@ class ExtensibleTrainer(BaseModel): for v in self.opt['logger']['visuals']: if step % self.opt['logger']['visual_debug_rate'] == 0: for i, dbgv in enumerate(state[v]): + if dbgv.shape[1] > 3: + dbgv = dbgv[:,:3,:,:] os.makedirs(os.path.join(sample_save_path, v), exist_ok=True) utils.save_image(dbgv, os.path.join(sample_save_path, v, "%05i_%02i.png" % (step, i))) diff --git a/codes/models/archs/SPSR_arch.py b/codes/models/archs/SPSR_arch.py index 93daa7f5..8485f53e 100644 --- a/codes/models/archs/SPSR_arch.py +++ b/codes/models/archs/SPSR_arch.py @@ -356,147 +356,12 @@ class SwitchedSpsr(nn.Module): return val -class SwitchedSpsrWithRef(nn.Module): - def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, init_temperature=10): - super(SwitchedSpsrWithRef, self).__init__() - n_upscale = int(math.log(upscale, 2)) - - # switch options - transformation_filters = nf - switch_filters = nf - self.transformation_counts = xforms - self.reference_processor = ReferenceImageBranch(transformation_filters) - multiplx_fn = functools.partial(ReferencingConvMultiplexer, transformation_filters, switch_filters, self.transformation_counts) - pretransform_fn = functools.partial(AdaInConvBlock, 512, transformation_filters, transformation_filters) - transform_fn = functools.partial(MultiConvBlock, transformation_filters, int(transformation_filters * 1.5), - transformation_filters, kernel_size=3, depth=3, - weight_init_factor=.1) - - # Feature branch - self.model_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False) - self.sw1 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, - pre_transform_block=pretransform_fn, transform_block=transform_fn, - attention_norm=True, - transform_count=self.transformation_counts, init_temp=init_temperature, - add_scalable_noise_to_transforms=True) - self.sw2 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, - pre_transform_block=pretransform_fn, transform_block=transform_fn, - attention_norm=True, - transform_count=self.transformation_counts, init_temp=init_temperature, - add_scalable_noise_to_transforms=True) - self.feature_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False) - self.feature_hr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False) - - # Grad branch - self.get_g_nopadding = ImageGradientNoPadding() - self.b_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False, bias=False) - mplex_grad = functools.partial(ReferencingConvMultiplexer, nf * 2, nf * 2, self.transformation_counts // 2) - self.sw_grad = ConfigurableSwitchComputer(transformation_filters, mplex_grad, - pre_transform_block=pretransform_fn, transform_block=transform_fn, - attention_norm=True, - transform_count=self.transformation_counts // 2, init_temp=init_temperature, - add_scalable_noise_to_transforms=True) - # Upsampling - self.grad_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=True, bias=False) - self.grad_hr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False) - # Conv used to output grad branch shortcut. - self.grad_branch_output_conv = ConvGnLelu(nf, out_nc, kernel_size=1, norm=False, activation=False, bias=False) - - # Conjoin branch. - transform_fn_cat = functools.partial(MultiConvBlock, transformation_filters * 2, int(transformation_filters * 1.5), - transformation_filters, kernel_size=3, depth=4, - weight_init_factor=.1) - pretransform_fn_cat = functools.partial(AdaInConvBlock, 512, transformation_filters * 2, transformation_filters * 2) - self._branch_pretrain_sw = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, - pre_transform_block=pretransform_fn_cat, transform_block=transform_fn_cat, - attention_norm=True, - transform_count=self.transformation_counts, init_temp=init_temperature, - add_scalable_noise_to_transforms=True) - self.upsample = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False) for _ in range(n_upscale)]) - self.upsample_grad = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False) for _ in range(n_upscale)]) - self.final_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False) - self.final_hr_conv1 = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=True, bias=False) - self.final_hr_conv2 = ConvGnLelu(nf, out_nc, kernel_size=3, norm=False, activation=False, bias=False) - self.switches = [self.sw1, self.sw2, self.sw_grad, self._branch_pretrain_sw] - self.attentions = None - self.init_temperature = init_temperature - self.final_temperature_step = 10000 - - def forward(self, x, ref, center_coord): - x_grad = self.get_g_nopadding(x) - ref = self.reference_processor(ref, center_coord) - x = self.model_fea_conv(x) - - x1, a1 = self.sw1((x, ref), True) - x2, a2 = self.sw2((x1, ref), True) - x_fea = self.feature_lr_conv(x2) - x_fea = self.feature_hr_conv2(x_fea) - - x_b_fea = self.b_fea_conv(x_grad) - x_grad, a3 = self.sw_grad((x_b_fea, ref), att_in=(torch.cat([x1, x_b_fea], dim=1), ref), output_attention_weights=True) - x_grad = self.grad_lr_conv(x_grad) - x_grad = self.grad_hr_conv(x_grad) - x_out_branch = self.upsample_grad(x_grad) - x_out_branch = self.grad_branch_output_conv(x_out_branch) - - x__branch_pretrain_cat = torch.cat([x_grad, x_fea], dim=1) - x__branch_pretrain_cat, a4 = self._branch_pretrain_sw((x__branch_pretrain_cat, ref), att_in=(x_fea, ref), identity=x_fea, output_attention_weights=True) - x_out = self.final_lr_conv(x__branch_pretrain_cat) - x_out = self.upsample(x_out) - x_out = self.final_hr_conv1(x_out) - x_out = self.final_hr_conv2(x_out) - - self.attentions = [a1, a2, a3, a4] - - return x_out_branch, x_out, x_grad - - def set_temperature(self, temp): - [sw.set_temperature(temp) for sw in self.switches] - - def update_for_step(self, step, experiments_path='.'): - if self.attentions: - temp = max(1, 1 + self.init_temperature * - (self.final_temperature_step - step) / self.final_temperature_step) - self.set_temperature(temp) - if step % 200 == 0: - output_path = os.path.join(experiments_path, "attention_maps", "a%i") - prefix = "attention_map_%i_%%i.png" % (step,) - [save_attention_to_image_rgb(output_path % (i,), self.attentions[i], self.transformation_counts, prefix, step) for i in range(len(self.attentions))] - - def get_debug_values(self, step): - temp = self.switches[0].switch.temperature - mean_hists = [compute_attention_specificity(att, 2) for att in self.attentions] - means = [i[0] for i in mean_hists] - hists = [i[1].clone().detach().cpu().flatten() for i in mean_hists] - val = {"switch_temperature": temp} - for i in range(len(means)): - val["switch_%i_specificity" % (i,)] = means[i] - val["switch_%i_histogram" % (i,)] = hists[i] - return val - - -class MultiplexerWithReducer(nn.Module): - def __init__(self, base_filters, multiplx_create_fn, transform_count): - super(MultiplexerWithReducer, self).__init__() - self.proc1 = ConvGnSilu(base_filters*2, base_filters*2, bias=False) - self.proc2 = ConvGnSilu(base_filters*2, base_filters*2, bias=False) - self.reduce = ConvGnSilu(base_filters*2, base_filters, activation=False, norm=False, bias=True) - self.conjoin = ConjoinBlock(base_filters) - self.mplex = multiplx_create_fn(transform_count) - - def forward(self, x, ref): - x = self.proc1(x) - x = self.proc2(x) - x = self.reduce(x) - return self.mplex(x, ref) - - class RefJoiner(nn.Module): def __init__(self, nf): super(RefJoiner, self).__init__() self.lin1 = nn.Linear(512, 256) self.lin2 = nn.Linear(256, nf) - self.join = ReferenceJoinBlock(nf, residual_weight_init_factor=.1, norm=False) + self.join = ReferenceJoinBlock(nf, residual_weight_init_factor=.05, norm=False) def forward(self, x, ref): ref = self.lin1(ref) @@ -526,7 +391,7 @@ class SwitchedSpsrWithRef2(nn.Module): # Feature branch self.model_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False) - self.noise_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.1, norm=False) + self.noise_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.01, norm=False) self.ref_join1 = RefJoiner(nf) self.sw1 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, pre_transform_block=pretransform_fn, transform_block=transform_fn, @@ -545,6 +410,7 @@ class SwitchedSpsrWithRef2(nn.Module): # Grad branch. Note - groupnorm on this branch is REALLY bad. Avoid it like the plague. self.get_g_nopadding = ImageGradientNoPadding() self.grad_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False, bias=False) + self.noise_ref_join_grad = ReferenceJoinBlock(nf, residual_weight_init_factor=.01, norm=False) self.ref_join3 = RefJoiner(nf) self.grad_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.2, norm=False, final_norm=False) self.sw_grad = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, @@ -559,6 +425,7 @@ class SwitchedSpsrWithRef2(nn.Module): # Join branch (grad+fea) self.ref_join4 = RefJoiner(nf) + self.noise_ref_join_conjoin = ReferenceJoinBlock(nf, residual_weight_init_factor=.01, norm=False) self.conjoin_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.2, norm=False) self.conjoin_sw = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, pre_transform_block=pretransform_fn, transform_block=transform_fn, @@ -579,16 +446,19 @@ class SwitchedSpsrWithRef2(nn.Module): ref = self.reference_processor(ref, center_coord) x = self.model_fea_conv(x) - x1 = self.noise_ref_join(x, torch.randn_like(x)) + x1 = x x1 = self.ref_join1(x1, ref) x1, a1 = self.sw1(x1, True, identity=x) x2 = x1 + x2 = self.noise_ref_join(x2, torch.randn_like(x2)) x2 = self.ref_join2(x2, ref) x2, a2 = self.sw2(x2, True, identity=x1) - x_grad_identity = self.grad_conv(x_grad) - x_grad = self.ref_join3(x_grad_identity, ref) + x_grad = self.grad_conv(x_grad) + x_grad_identity = x_grad + x_grad = self.noise_ref_join_grad(x_grad, torch.randn_like(x_grad)) + x_grad = self.ref_join3(x_grad, ref) x_grad = self.grad_ref_join(x_grad, x1) x_grad, a3 = self.sw_grad(x_grad, True, identity=x_grad_identity) x_grad = self.grad_lr_conv(x_grad) @@ -596,7 +466,9 @@ class SwitchedSpsrWithRef2(nn.Module): x_grad_out = self.upsample_grad(x_grad) x_grad_out = self.grad_branch_output_conv(x_grad_out) - x_out = self.ref_join4(x2, ref) + x_out = x2 + x_out = self.noise_ref_join_conjoin(x_out, torch.randn_like(x_out)) + x_out = self.ref_join4(x_out, ref) x_out = self.conjoin_ref_join(x_out, x_grad) x_out, a4 = self.conjoin_sw(x_out, True, identity=x2) x_out = self.final_lr_conv(x_out) diff --git a/codes/models/archs/arch_util.py b/codes/models/archs/arch_util.py index 4c84fc99..9e808e40 100644 --- a/codes/models/archs/arch_util.py +++ b/codes/models/archs/arch_util.py @@ -456,7 +456,7 @@ class ConjoinBlock(nn.Module): # Designed explicitly to join a mainline trunk with reference data. Implemented as a residual branch. class ReferenceJoinBlock(nn.Module): - def __init__(self, nf, residual_weight_init_factor=1, norm=False, block=ConvGnLelu, final_norm=True): + def __init__(self, nf, residual_weight_init_factor=1, norm=False, block=ConvGnLelu, final_norm=False): super(ReferenceJoinBlock, self).__init__() self.branch = MultiConvBlock(nf * 2, nf + nf // 2, nf, kernel_size=3, depth=3, scale_init=residual_weight_init_factor, norm=norm,