3ce1a1878d
- Get rid of forwards(), it makes numeric_stability.py not work properly. - Do stability auditing across layers. - Upsample last instead of first, work in much higher dimensionality for transforms.
311 lines
15 KiB
Python
311 lines
15 KiB
Python
import torch
|
|
from torch import nn
|
|
from models.archs.SwitchedResidualGenerator_arch import ConvBnLelu, MultiConvBlock, initialize_weights
|
|
from switched_conv import BareConvSwitch, compute_attention_specificity
|
|
from switched_conv_util import save_attention_to_image
|
|
from functools import partial
|
|
import torch.nn.functional as F
|
|
import numpy as np
|
|
|
|
|
|
def conv3x3(in_planes, out_planes, stride=1):
|
|
"""3x3 convolution with padding"""
|
|
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
|
|
padding=1, bias=False)
|
|
|
|
|
|
def conv1x1(in_planes, out_planes, stride=1):
|
|
"""1x1 convolution"""
|
|
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
|
|
|
|
|
|
# Taken from Fixup resnet implementation https://github.com/hongyi-zhang/Fixup/blob/master/imagenet/models/fixup_resnet_imagenet.py
|
|
class FixupBottleneck(nn.Module):
|
|
expansion = 4
|
|
|
|
def __init__(self, inplanes, planes, stride=1, downsample=None):
|
|
super(FixupBottleneck, self).__init__()
|
|
# Both self.conv2 and self.downsample layers downsample the input when stride != 1
|
|
self.bias1a = nn.Parameter(torch.zeros(1))
|
|
self.conv1 = conv1x1(inplanes, planes)
|
|
self.bias1b = nn.Parameter(torch.zeros(1))
|
|
self.bias2a = nn.Parameter(torch.zeros(1))
|
|
self.conv2 = conv3x3(planes, planes, stride)
|
|
self.bias2b = nn.Parameter(torch.zeros(1))
|
|
self.bias3a = nn.Parameter(torch.zeros(1))
|
|
self.conv3 = conv1x1(planes, planes * self.expansion)
|
|
self.scale = nn.Parameter(torch.ones(1))
|
|
self.bias3b = nn.Parameter(torch.zeros(1))
|
|
self.relu = nn.ReLU(inplace=True)
|
|
self.downsample = downsample
|
|
self.stride = stride
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
|
|
out = self.conv1(x + self.bias1a)
|
|
out = self.relu(out + self.bias1b)
|
|
|
|
out = self.conv2(out + self.bias2a)
|
|
out = self.relu(out + self.bias2b)
|
|
|
|
out = self.conv3(out + self.bias3a)
|
|
out = out * self.scale + self.bias3b
|
|
|
|
if self.downsample is not None:
|
|
identity = self.downsample(x + self.bias1a)
|
|
|
|
out += identity
|
|
out = self.relu(out)
|
|
|
|
return out
|
|
|
|
|
|
class Switch(nn.Module):
|
|
def __init__(self, transform_block, transform_count, init_temp=20, pass_chain_forward=False, add_scalable_noise_to_transforms=False):
|
|
super(Switch, self).__init__()
|
|
|
|
self.transforms = nn.ModuleList([transform_block() for _ in range(transform_count)])
|
|
self.add_noise = add_scalable_noise_to_transforms
|
|
self.pass_chain_forward = pass_chain_forward
|
|
|
|
# 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))
|
|
|
|
if not self.pass_chain_forward:
|
|
self.parameterize = ConvBnLelu(64, 64, bn=False, lelu=False)
|
|
self.c_constric = MultiConvBlock(576, 256, 64, kernel_size=1, depth=3, bn=False)
|
|
self.c_process = ConvBnLelu(64, 64, kernel_size=1, lelu=False, bn=False)
|
|
|
|
# x is the input fed to the transform blocks.
|
|
# m is the output of the multiplexer which will be used to select from those transform blocks.
|
|
# chain is a chain of shared processing outputs used by the individual transforms.
|
|
def forward(self, x, m, chain):
|
|
if self.pass_chain_forward:
|
|
pcf = [t(x, chain) for t in self.transforms]
|
|
xformed = [o[0] for o in pcf]
|
|
atts = [o[1] for o in pcf]
|
|
else:
|
|
# These adjustments were determined statistically from numeric_stability.py and should start this context
|
|
# out in a normal distribution.
|
|
context = chain[-1]
|
|
context = F.interpolate(context, size=x.shape[2:], mode='nearest')
|
|
context = torch.cat([self.parameterize(x), context], dim=1)
|
|
context = self.c_constric(context) / 3
|
|
context = self.c_process(context)
|
|
context = x * context
|
|
|
|
if self.add_noise:
|
|
rand_feature = torch.randn_like(x)
|
|
xformed = [t(context, rand_feature) for t in self.transforms]
|
|
else:
|
|
xformed = [t(context) for t in self.transforms]
|
|
|
|
# Interpolate the multiplexer across the entire shape of the image.
|
|
m = F.interpolate(m, size=x.shape[2:], mode='nearest')
|
|
|
|
outputs, attention = self.switch(xformed, m, True)
|
|
outputs = outputs * self.scale + self.bias
|
|
|
|
if self.pass_chain_forward:
|
|
# Apply attention weights to collected [atts] and return the aggregate.
|
|
atts = torch.stack(atts, dim=3)
|
|
attention = atts * attention.unsqueeze(dim=-1)
|
|
attention = torch.flatten(attention, 3)
|
|
|
|
return outputs, attention
|
|
|
|
def set_temperature(self, temp):
|
|
self.switch.set_attention_temperature(temp)
|
|
if self.pass_chain_forward:
|
|
[t.set_temperature(temp) for t in self.transforms]
|
|
|
|
|
|
# Convolutional image processing block that optionally reduces image size by a factor of 2 using stride and performs a
|
|
# series of residual-block-like processing operations on it.
|
|
class Processor(nn.Module):
|
|
def __init__(self, base_filters, processing_depth, reduce=False):
|
|
super(Processor, self).__init__()
|
|
self.output_filter_count = base_filters * 2
|
|
|
|
# Downsample block used for bottleneck.
|
|
downsample = nn.Sequential(
|
|
nn.Conv2d(base_filters, self.output_filter_count, kernel_size=1, stride=2, bias=False),
|
|
nn.BatchNorm2d(self.output_filter_count),
|
|
)
|
|
# Bottleneck block outputs the requested filter sizex4, but we only want x2.
|
|
self.initial = FixupBottleneck(base_filters, base_filters // 2, stride=2 if reduce else 1, downsample=downsample)
|
|
|
|
self.res_blocks = nn.ModuleList([FixupBottleneck(self.output_filter_count, self.output_filter_count // 4) for _ in range(processing_depth)])
|
|
|
|
def forward(self, x):
|
|
x = (self.initial(x) - .4) / .6
|
|
for b in self.res_blocks:
|
|
r = (b(x) - .4) / .6
|
|
x = r + x
|
|
return x
|
|
|
|
|
|
# Convolutional image processing block that constricts an input image with a large number of filters to a small number
|
|
# of filters over a fixed number of layers.
|
|
class Constrictor(nn.Module):
|
|
def __init__(self, filters, output_filters):
|
|
super(Constrictor, self).__init__()
|
|
assert(filters > output_filters)
|
|
gap = filters - output_filters
|
|
gap_div_4 = int(gap / 4)
|
|
self.cbl1 = ConvBnLelu(filters, filters - (gap_div_4 * 2), kernel_size=1, bn=True)
|
|
self.cbl2 = ConvBnLelu(filters - (gap_div_4 * 2), filters - (gap_div_4 * 3), kernel_size=1, bn=True)
|
|
self.cbl3 = ConvBnLelu(filters - (gap_div_4 * 3), output_filters, kernel_size=1, lelu=False, bn=False)
|
|
|
|
def forward(self, x):
|
|
x = self.cbl1(x)
|
|
x = self.cbl2(x)
|
|
x = self.cbl3(x) / 4
|
|
return x
|
|
|
|
|
|
class RecursiveSwitchedTransform(nn.Module):
|
|
def __init__(self, transform_filters, filters_count_list, nesting_depth, transforms_at_leaf,
|
|
trans_kernel_size, trans_num_layers, trans_scale_init=1, initial_temp=20, add_scalable_noise_to_transforms=False):
|
|
super(RecursiveSwitchedTransform, self).__init__()
|
|
|
|
self.depth = nesting_depth
|
|
at_leaf = (self.depth == 0)
|
|
if at_leaf:
|
|
transform = partial(MultiConvBlock, transform_filters, transform_filters, transform_filters, kernel_size=trans_kernel_size, depth=trans_num_layers, scale_init=trans_scale_init)
|
|
else:
|
|
transform = partial(RecursiveSwitchedTransform, transform_filters, filters_count_list,
|
|
nesting_depth - 1, transforms_at_leaf, trans_kernel_size, trans_num_layers, trans_scale_init, initial_temp, add_scalable_noise_to_transforms)
|
|
selection_breadth = transforms_at_leaf if at_leaf else 2
|
|
self.switch = Switch(transform, selection_breadth, initial_temp, pass_chain_forward=not at_leaf, add_scalable_noise_to_transforms=add_scalable_noise_to_transforms)
|
|
self.multiplexer = Constrictor(filters_count_list[self.depth], selection_breadth)
|
|
|
|
def forward(self, x, processing_trunk_chain):
|
|
proc_out = processing_trunk_chain[self.depth]
|
|
m = self.multiplexer(proc_out)
|
|
return self.switch(x, m, processing_trunk_chain)
|
|
|
|
def set_temperature(self, temp):
|
|
self.switch.set_temperature(temp)
|
|
|
|
|
|
class NestedSwitchComputer(nn.Module):
|
|
def __init__(self, transform_filters, switch_base_filters, num_switch_processing_layers, nesting_depth, transforms_at_leaf,
|
|
trans_kernel_size, trans_num_layers, trans_scale_init, initial_temp=20, add_scalable_noise_to_transforms=False):
|
|
super(NestedSwitchComputer, self).__init__()
|
|
|
|
processing_trunk = []
|
|
filters = []
|
|
current_filters = switch_base_filters
|
|
for _ in range(nesting_depth):
|
|
processing_trunk.append(Processor(current_filters, num_switch_processing_layers, reduce=True))
|
|
current_filters = processing_trunk[-1].output_filter_count
|
|
filters.append(current_filters)
|
|
|
|
self.multiplexer_init_conv = ConvBnLelu(transform_filters, switch_base_filters, kernel_size=7, lelu=False, bn=False)
|
|
self.processing_trunk = nn.ModuleList(processing_trunk)
|
|
self.switch = RecursiveSwitchedTransform(transform_filters, filters, nesting_depth-1, transforms_at_leaf, trans_kernel_size, trans_num_layers-1, trans_scale_init, initial_temp=initial_temp, add_scalable_noise_to_transforms=add_scalable_noise_to_transforms)
|
|
self.anneal = ConvBnLelu(transform_filters, transform_filters, kernel_size=1, bn=False)
|
|
|
|
# Init the parameters in the trunk. Uses the fixup algorithm for residual conv initialization.
|
|
self.num_layers = nesting_depth + nesting_depth * num_switch_processing_layers
|
|
for m in self.processing_trunk.modules():
|
|
if isinstance(m, FixupBottleneck):
|
|
nn.init.normal_(m.conv1.weight, mean=0, std=np.sqrt(2 / (m.conv1.weight.shape[0] * np.prod(m.conv1.weight.shape[2:]))) * self.num_layers ** (-0.25))
|
|
nn.init.normal_(m.conv2.weight, mean=0, std=np.sqrt(2 / (m.conv2.weight.shape[0] * np.prod(m.conv2.weight.shape[2:]))) * self.num_layers ** (-0.25))
|
|
nn.init.constant_(m.conv3.weight, 0)
|
|
if m.downsample is not None:
|
|
nn.init.normal_(m.downsample[0].weight, mean=0, std=np.sqrt(2 / (m.downsample[0].weight.shape[0] * np.prod(m.downsample[0].weight.shape[2:]))))
|
|
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
|
|
nn.init.constant_(m.weight, 1)
|
|
nn.init.constant_(m.bias, 0)
|
|
|
|
def forward(self, x):
|
|
feed_forward = x
|
|
trunk = []
|
|
trunk_input = self.multiplexer_init_conv(x)
|
|
for m in self.processing_trunk:
|
|
trunk_input = (m(trunk_input) - 3.3) / 12.5
|
|
trunk.append(trunk_input)
|
|
|
|
self.trunk = trunk[-1]
|
|
x, att = self.switch(x, trunk)
|
|
x = x + feed_forward
|
|
return feed_forward + self.anneal(x) / .86, att
|
|
|
|
def set_temperature(self, temp):
|
|
self.switch.set_temperature(temp)
|
|
|
|
|
|
class NestedSwitchedGenerator(nn.Module):
|
|
def __init__(self, switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes,
|
|
trans_layers, transformation_filters, initial_temp=20, final_temperature_step=50000, heightened_temp_min=1,
|
|
heightened_final_step=50000, upsample_factor=1, add_scalable_noise_to_transforms=False):
|
|
super(NestedSwitchedGenerator, self).__init__()
|
|
self.initial_conv = ConvBnLelu(3, transformation_filters, kernel_size=7, lelu=False, bn=False)
|
|
self.proc_conv = ConvBnLelu(transformation_filters, transformation_filters, bn=False)
|
|
self.final_conv = ConvBnLelu(transformation_filters, 3, kernel_size=1, lelu=False, bn=False)
|
|
|
|
switches = []
|
|
for sw_reduce, sw_proc, trans_count, kernel, layers in zip(switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers):
|
|
switches.append(NestedSwitchComputer(transform_filters=transformation_filters, switch_base_filters=switch_filters, num_switch_processing_layers=sw_proc,
|
|
nesting_depth=sw_reduce, transforms_at_leaf=trans_count, trans_kernel_size=kernel, trans_num_layers=layers,
|
|
trans_scale_init=.2/len(switch_reductions), initial_temp=initial_temp, add_scalable_noise_to_transforms=add_scalable_noise_to_transforms))
|
|
self.switches = nn.ModuleList(switches)
|
|
|
|
self.transformation_counts = trans_counts
|
|
self.init_temperature = initial_temp
|
|
self.final_temperature_step = final_temperature_step
|
|
self.heightened_temp_min = heightened_temp_min
|
|
self.heightened_final_step = heightened_final_step
|
|
self.attentions = None
|
|
self.upsample_factor = upsample_factor
|
|
|
|
def forward(self, x):
|
|
x = self.initial_conv(x) / .2
|
|
|
|
self.attentions = []
|
|
for i, sw in enumerate(self.switches):
|
|
x, att = sw(x)
|
|
self.attentions.append(att)
|
|
|
|
if self.upsample_factor > 1:
|
|
x = F.interpolate(x, scale_factor=self.upsample_factor, mode="nearest")
|
|
|
|
x = self.proc_conv(x) / .85
|
|
x = self.final_conv(x) / 4.6
|
|
return x / 16,
|
|
|
|
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))
|
|
if temp == 1 and self.heightened_final_step and self.heightened_final_step != 1:
|
|
# Once the temperature passes (1) it enters an inverted curve to match the linear curve from above.
|
|
# without this, the attention specificity "spikes" incredibly fast in the last few iterations.
|
|
h_steps_total = self.heightened_final_step - self.final_temperature_step
|
|
h_steps_current = min(step - self.final_temperature_step, h_steps_total)
|
|
# The "gap" will represent the steps that need to be traveled as a linear function.
|
|
h_gap = 1 / self.heightened_temp_min
|
|
temp = h_gap * h_steps_current / h_steps_total
|
|
# Invert temperature to represent reality on this side of the curve
|
|
temp = 1 / temp
|
|
self.set_temperature(temp)
|
|
if step % 50 == 0:
|
|
[save_attention_to_image(experiments_path, self.attentions[i], self.transformation_counts[i], step, "a%i" % (i+1,)) for i in range(len(self.switches))]
|
|
|
|
def get_debug_values(self, step):
|
|
temp = self.switches[0].switch.switch.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 |