mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ResidualGen mods

Browse Source 
- Add filters_mid spec which allows a expansion->squeeze for the transformation layers.
- Add scale and bias AFTER the switch
- Remove identity transform (models were converging on this)
- Move attention image generation and temperature setting into new function which gets called every step with a save path
This commit is contained in:
James Betker 2020-06-17 17:18:28 -06:00
parent 6f8406fbdc
commit 645d0ca767
3 changed files with 32 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
codes/models/SRGAN_model.py
View File

@ -183,6 +183,10 @@ class SRGANModel(BaseModel):
self.pix = [t.to(self.device) for t in torch.chunk(data['PIX'], chunks=self.mega_batch_factor, dim=0)]
def optimize_parameters(self, step):
# Some generators have variants depending on the current step.
if hasattr(self.netG.module, "update_for_step"):
self.netG.module.update_for_step(step, os.path.join(self.opt['path']['models'], ".."))
# G
for p in self.netD.parameters():
p.requires_grad = False

47
codes/models/archs/SwitchedResidualGenerator_arch.py
View File

@ -29,11 +29,12 @@ class ConvBnLelu(nn.Module):
class ResidualBranch(nn.Module):
def __init__(self, filters_in, filters_out, kernel_size, depth):
def __init__(self, filters_in, filters_mid, filters_out, kernel_size, depth):
assert depth >= 2
super(ResidualBranch, self).__init__()
self.bnconvs = nn.ModuleList([ConvBnLelu(filters_in, filters_out, kernel_size)] +
[ConvBnLelu(filters_out, filters_out, kernel_size) for i in range(depth-2)] +
[ConvBnLelu(filters_out, filters_out, kernel_size, lelu=False)])
self.bnconvs = nn.ModuleList([ConvBnLelu(filters_in, filters_mid, kernel_size)] +
[ConvBnLelu(filters_mid, filters_mid, kernel_size) for i in range(depth-2)] +
[ConvBnLelu(filters_mid, filters_out, kernel_size, lelu=False)])
self.scale = nn.Parameter(torch.ones(1))
self.bias = nn.Parameter(torch.zeros(1))
@ -67,16 +68,15 @@ class SwitchComputer(nn.Module):
self.proc_switch_conv = ConvBnLelu(final_filters, proc_block_filters)
self.final_switch_conv = nn.Conv2d(proc_block_filters, transform_count, 1, 1, 0)
# Always include the identity transform (all zeros), hence transform_count-10
self.transforms = nn.ModuleList([transform_block() for i in range(transform_count-1)])
self.transforms = nn.ModuleList([transform_block() for i in range(transform_count)])
# And the switch itself
# And the switch itself, including learned scalars
self.switch = BareConvSwitch(initial_temperature=init_temp)
self.scale = nn.Parameter(torch.ones(1))
self.bias = nn.Parameter(torch.zeros(1))
def forward(self, x, output_attention_weights=False):
xformed = [t.forward(x) for t in self.transforms]
# Append the identity transform.
xformed.append(torch.zeros_like(xformed[0]))
multiplexer = self.filter_conv(x)
for block in self.reduction_blocks:
@ -88,18 +88,23 @@ class SwitchComputer(nn.Module):
# Interpolate the multiplexer across the entire shape of the image.
multiplexer = F.interpolate(multiplexer, size=x.shape[2:], mode='nearest')
return self.switch(xformed, multiplexer, output_attention_weights)
outputs, attention = self.switch(xformed, multiplexer, True)
outputs = outputs * self.scale + self.bias
if output_attention_weights:
return outputs, attention
else:
return outputs
def set_temperature(self, temp):
self.switch.set_attention_temperature(temp)
class ConfigurableSwitchedResidualGenerator(nn.Module):
def __init__(self, switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers, initial_temp=20, final_temperature_step=50000):
def __init__(self, switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers, trans_filters_mid, initial_temp=20, final_temperature_step=50000):
super(ConfigurableSwitchedResidualGenerator, self).__init__()
switches = []
for filters, sw_reduce, sw_proc, trans_count, kernel, layers in zip(switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers):
switches.append(SwitchComputer(3, filters, functools.partial(ResidualBranch, 3, 3, kernel_size=kernel, depth=layers), trans_count, sw_reduce, sw_proc, initial_temp))
for filters, sw_reduce, sw_proc, trans_count, kernel, layers, mid_filters in zip(switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers, trans_filters_mid):
switches.append(SwitchComputer(3, filters, functools.partial(ResidualBranch, 3, mid_filters, 3, kernel_size=kernel, depth=layers), trans_count, sw_reduce, sw_proc, initial_temp))
initialize_weights(switches, 1)
# Initialize the transforms with a lesser weight, since they are repeatedly added on to the resultant image.
initialize_weights([s.transforms for s in switches], .2 / len(switches))
@ -107,6 +112,7 @@ class ConfigurableSwitchedResidualGenerator(nn.Module):
self.transformation_counts = trans_counts
self.init_temperature = initial_temp
self.final_temperature_step = final_temperature_step
self.attentions = None
def forward(self, x):
self.attentions = []
@ -119,14 +125,15 @@ class ConfigurableSwitchedResidualGenerator(nn.Module):
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, int(self.init_temperature * (self.final_temperature_step - step) / self.final_temperature_step))
self.set_temperature(temp)
if step % 2 == 0:
[save_attention_to_image(experiments_path, self.attentions[i], self.transformation_counts[i], step, "a%i" % (i+1,), l_mult=float(self.transformation_counts[i]/4)) for i in range(len(self.switches))]
def get_debug_values(self, step):
# Take the chance to update the temperature here.
temp = max(1, int(self.init_temperature * (self.final_temperature_step - step) / self.final_temperature_step))
self.set_temperature(temp)
if step % 250 == 0:
[save_attention_to_image(self.attentions[i], self.transformation_counts[i], step, "a%i" % (i+1,), l_mult=float(self.transformation_counts[i]/4)) for i in range(len(self.switches))]
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]

1
codes/models/networks.py
View File

@ -66,6 +66,7 @@ def define_G(opt, net_key='network_G'):
netG = SwitchedGen_arch.ConfigurableSwitchedResidualGenerator(switch_filters=opt_net['switch_filters'], switch_reductions=opt_net['switch_reductions'],
switch_processing_layers=opt_net['switch_processing_layers'], trans_counts=opt_net['trans_counts'],
trans_kernel_sizes=opt_net['trans_kernel_sizes'], trans_layers=opt_net['trans_layers'],
trans_filters_mid=opt_net['trans_filters_mid'],
initial_temp=opt_net['temperature'], final_temperature_step=opt_net['temperature_final_step'])
# image corruption