281 lines
14 KiB
Python
281 lines
14 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
import torch.optim as optim
|
|
from torchvision import datasets, models, transforms
|
|
import tqdm
|
|
from torch.utils.tensorboard import SummaryWriter
|
|
import torch.nn.init as init
|
|
import torch.nn.functional as F
|
|
|
|
def initialize_weights(net_l, scale=1):
|
|
if not isinstance(net_l, list):
|
|
net_l = [net_l]
|
|
for net in net_l:
|
|
for m in net.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
init.kaiming_normal_(m.weight, a=0, mode='fan_in')
|
|
m.weight.data *= scale # for residual block
|
|
if m.bias is not None:
|
|
m.bias.data.zero_()
|
|
elif isinstance(m, nn.Linear):
|
|
init.kaiming_normal_(m.weight, a=0, mode='fan_in')
|
|
m.weight.data *= scale
|
|
if m.bias is not None:
|
|
m.bias.data.zero_()
|
|
elif isinstance(m, nn.BatchNorm2d):
|
|
init.constant_(m.weight, 1)
|
|
init.constant_(m.bias.data, 0.0)
|
|
|
|
|
|
class ConvSwitch(nn.Module):
|
|
'''
|
|
Initializes the ConvSwitch.
|
|
nf_attention_basis: Number of filters provided to the attention_basis input of forward(). Must be divisible by two
|
|
and nf_attention_basis/2 >= num_convs.
|
|
num_convs: Number of elements that will appear in the conv_group() input. The attention mechanism will
|
|
select across this number.
|
|
att_kernel_size: The size of the attention mechanisms convolutional kernels.
|
|
att_stride: The stride of the attention mechanisms conv blocks.
|
|
att_pads: The padding of the attention mechanisms conv blocks.
|
|
att_interpolate_scale_factor:
|
|
The scale factor applied to the attention mechanism's outputs.
|
|
*** NOTE ***: Between stride, pads, and interpolation_scale_factor, the output of the attention mechanism MUST
|
|
have the same width/height as the conv_group inputs.
|
|
initial_temperature: The initial softmax temperature of the attention mechanism. For training from scratch, this
|
|
should be set to a high number, for example 30.
|
|
'''
|
|
def __init__(self, nf_attention_basis, num_convs=8, att_kernel_size=5, att_stride=1, att_pads=2,
|
|
att_interpolate_scale_factor=1, initial_temperature=1):
|
|
super(ConvSwitch, self).__init__()
|
|
|
|
# Requirements: input filter count is even, and there are more filters than there are sequences to attend to.
|
|
assert nf_attention_basis % 2 == 0
|
|
assert nf_attention_basis / 2 >= num_convs
|
|
|
|
self.num_convs = num_convs
|
|
self.interpolate_scale_factor = att_interpolate_scale_factor
|
|
self.attention_conv1 = nn.Conv2d(nf_attention_basis, int(nf_attention_basis / 2), att_kernel_size, att_stride, att_pads, bias=True)
|
|
self.att_bn1 = nn.BatchNorm2d(int(nf_attention_basis / 2))
|
|
self.attention_conv2 = nn.Conv2d(int(nf_attention_basis / 2), num_convs, att_kernel_size, 1, att_pads, bias=True)
|
|
self.softmax = nn.Softmax(dim=-1)
|
|
self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
|
|
self.temperature = initial_temperature
|
|
|
|
def set_attention_temperature(self, temp):
|
|
self.temperature = temp
|
|
|
|
# SwitchedConv.forward takes these arguments;
|
|
# attention_basis: The tensor to compute attention vectors from. Generally this should be the original inputs of
|
|
# your conv group.
|
|
# conv_group: A list of output tensors from the convolutional groups that the attention mechanism is selecting
|
|
# from. Each tensor in this list is expected to have the shape (batch, filters, width, height)
|
|
def forward(self, attention_basis, conv_group, output_attention_weights=False):
|
|
assert self.num_convs == len(conv_group)
|
|
|
|
# Stack up the conv_group input first and permute it to (batch, width, height, filter, groups)
|
|
conv_outputs = torch.stack(conv_outputs, dim=0).permute(1, 3, 4, 2, 0)
|
|
|
|
# Now calculate the attention across those convs.
|
|
conv_attention = self.lrelu(self.att_bn1(self.attention_conv1(attention_basis)))
|
|
conv_attention = self.attention_conv2(conv_attention).permute(0, 2, 3, 1)
|
|
conv_attention = self.softmax(conv_attention / self.temperature)
|
|
|
|
# conv_outputs shape: (batch, width, height, filters, groups)
|
|
# conv_attention shape: (batch, width, height, groups)
|
|
# We want to format them so that we can matmul them together to produce:
|
|
# desired shape: (batch, width, height, filters)
|
|
# Note: conv_attention will generally be cast to float32 regardless of the input type, so cast conv_outputs to
|
|
# float32 as well to match it.
|
|
attention_result = torch.einsum("...ij,...j->...i", [conv_outputs.to(dtype=torch.float32), conv_attention])
|
|
|
|
# Remember to shift the filters back into the expected slot.
|
|
if output_attention_weights:
|
|
return attention_result.permute(0, 3, 1, 2), conv_attention
|
|
else:
|
|
return attention_result.permute(0, 3, 1, 2)
|
|
|
|
|
|
class SwitchedConv2d(nn.Module):
|
|
def __init__(self, nf_in_per_conv, nf_out_per_conv, num_convs=8, att_kernel_size=5, att_stride=1, att_pads=2,
|
|
initial_temperature=1):
|
|
super(ConvSwitch, self).__init__()
|
|
|
|
# Requirements: input filter count is even, and there are more filters than there are sequences to attend to.
|
|
assert nf_in_per_conv % 2 == 0
|
|
assert nf_in_per_conv / 2 >= num_convs
|
|
|
|
self.nf = nf_out_per_conv
|
|
self.num_convs = num_convs
|
|
self.conv_list = nn.ModuleList([nn.Conv2d(nf_in_per_conv, nf_out_per_conv, kernel_size, att_stride, pads, bias=has_bias) for _ in range(num_convs)])
|
|
self.attention_conv1 = nn.Conv2d(nf_in_per_conv, int(nf_in_per_conv/2), att_kernel_size, att_stride, att_pads, bias=True)
|
|
self.att_bn1 = nn.BatchNorm2d(int(nf_in_per_conv/2))
|
|
self.attention_conv2 = nn.Conv2d(int(nf_in_per_conv/2), num_convs, att_kernel_size, 1, att_pads, bias=True)
|
|
self.softmax = nn.Softmax(dim=-1)
|
|
self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
|
|
self.temperature = initial_temperature
|
|
|
|
def set_attention_temperature(self, temp):
|
|
self.temperature = temp
|
|
|
|
def forward(self, x, output_attention_weights=False):
|
|
# Build up the individual conv components first.
|
|
conv_outputs = []
|
|
for conv in self.conv_list:
|
|
conv_outputs.append(conv.forward(x))
|
|
conv_outputs = torch.stack(conv_outputs, dim=0).permute(1, 3, 4, 2, 0)
|
|
|
|
# Now calculate the attention across those convs.
|
|
conv_attention = self.lrelu(self.att_bn1(self.attention_conv1(x)))
|
|
conv_attention = self.attention_conv2(conv_attention).permute(0, 2, 3, 1)
|
|
conv_attention = self.softmax(conv_attention / self.temperature)
|
|
|
|
# conv_outputs shape: (batch, width, height, filters, sequences)
|
|
# conv_attention shape: (batch, width, height, sequences)
|
|
# We want to format them so that we can matmul them together to produce:
|
|
# desired shape: (batch, width, height, filters)
|
|
# Note: conv_attention will generally be cast to float32 regardless of the input type, so cast conv_outputs to
|
|
# float32 as well to match it.
|
|
attention_result = torch.einsum("...ij,...j->...i", [conv_outputs.to(dtype=torch.float32), conv_attention])
|
|
|
|
# Remember to shift the filters back into the expected slot.
|
|
if output_attention_weights:
|
|
return attention_result.permute(0, 3, 1, 2), conv_attention
|
|
else:
|
|
return attention_result.permute(0, 3, 1, 2)
|
|
|
|
def compute_attention_specificity(att_weights, topk=3):
|
|
att = att_weights.detach()
|
|
vals, indices = torch.topk(att, topk, dim=-1)
|
|
avg = torch.sum(vals, dim=-1)
|
|
avg = avg.flatten().mean()
|
|
return avg.item(), indices.flatten().detach()
|
|
|
|
class DynamicConvTestModule(nn.Module):
|
|
def __init__(self):
|
|
super(DynamicConvTestModule, self).__init__()
|
|
self.init_conv = nn.Conv2d(3, 16, 3, 1, 1, bias=True)
|
|
self.conv1 = ConvSwitch(16, 32, 3, att_stride=2, pads=1, num_convs=4, initial_temperature=10)
|
|
self.bn1 = nn.BatchNorm2d(32)
|
|
self.conv2 = ConvSwitch(32, 64, 3, att_stride=2, pads=1, att_kernel_size=3, att_pads=1, num_convs=8, initial_temperature=10)
|
|
self.bn2 = nn.BatchNorm2d(64)
|
|
self.conv3 = ConvSwitch(64, 128, 3, att_stride=2, pads=1, att_kernel_size=3, att_pads=1, num_convs=16, initial_temperature=10)
|
|
self.bn3 = nn.BatchNorm2d(128)
|
|
self.relu = nn.ReLU()
|
|
self.dense1 = nn.Linear(128 * 4 * 4, 256)
|
|
self.dense2 = nn.Linear(256, 100)
|
|
self.softmax = nn.Softmax(-1)
|
|
|
|
def set_temp(self, temp):
|
|
self.conv1.set_attention_temperature(temp)
|
|
self.conv2.set_attention_temperature(temp)
|
|
self.conv3.set_attention_temperature(temp)
|
|
|
|
def forward(self, x):
|
|
x = self.init_conv(x)
|
|
x, att1 = self.conv1(x, output_attention_weights=True)
|
|
x = self.relu(self.bn1(x))
|
|
x, att2 = self.conv2(x, output_attention_weights=True)
|
|
x = self.relu(self.bn2(x))
|
|
x, att3 = self.conv3(x, output_attention_weights=True)
|
|
x = self.relu(self.bn3(x))
|
|
atts = [att1, att2, att3]
|
|
usage_hists = []
|
|
mean = 0
|
|
for a in atts:
|
|
m, u = compute_attention_specificity(a)
|
|
mean += m
|
|
usage_hists.append(u)
|
|
mean /= 3
|
|
|
|
x = x.flatten(1)
|
|
x = self.relu(self.dense1(x))
|
|
x = self.dense2(x)
|
|
# Compute metrics across attention weights.
|
|
|
|
return self.softmax(x), mean, usage_hists
|
|
|
|
def test_dynamic_conv():
|
|
writer = SummaryWriter()
|
|
dataset = datasets.ImageFolder("C:\\data\\cifar-100-python\\images\\train", transforms.Compose([
|
|
transforms.Resize(32, 32),
|
|
transforms.RandomHorizontalFlip(),
|
|
transforms.ToTensor(),
|
|
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
|
]))
|
|
batch_size = 256
|
|
temperature = 30
|
|
loader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=4)
|
|
device = torch.device("cuda:0")
|
|
net = DynamicConvTestModule()
|
|
net = net.to(device)
|
|
net.set_temp(temperature)
|
|
initialize_weights(net)
|
|
optimizer = optim.Adam(net.parameters(), lr=1e-3)
|
|
|
|
# Load state, where necessary.
|
|
'''
|
|
netstate, optimstate = torch.load("test_net.pth")
|
|
net.load_state_dict(netstate)
|
|
optimizer.load_state_dict(optimstate)
|
|
'''
|
|
|
|
criterion = nn.CrossEntropyLoss()
|
|
step = 0
|
|
running_corrects = 0
|
|
running_att_mean = 0
|
|
running_att_hist = None
|
|
for e in range(300):
|
|
tq = tqdm.tqdm(loader)
|
|
for batch, labels in tq:
|
|
batch = batch.to(device)
|
|
labels = labels.to(device)
|
|
optimizer.zero_grad()
|
|
logits, att_mean, att_usage_hist = net.forward(batch)
|
|
running_att_mean += att_mean
|
|
if running_att_hist is None:
|
|
running_att_hist = att_usage_hist
|
|
else:
|
|
for i in range(len(att_usage_hist)):
|
|
running_att_hist[i] = torch.cat([running_att_hist[i], att_usage_hist[i]]).flatten()
|
|
loss = criterion(logits, labels)
|
|
loss.backward()
|
|
|
|
if step % 50 == 0:
|
|
c1_grad_avg = sum([m.weight.grad.abs().mean().item() for m in net.conv1.conv_list._modules.values()]) / len(net.conv1.conv_list._modules)
|
|
c1a_grad_avg = (net.conv1.attention_conv1.weight.grad.abs().mean() + net.conv1.attention_conv2.weight.grad.abs().mean()) / 2
|
|
c2_grad_avg = sum([m.weight.grad.abs().mean().item() for m in net.conv2.conv_list._modules.values()]) / len(net.conv2.conv_list._modules)
|
|
c2a_grad_avg = (net.conv2.attention_conv1.weight.grad.abs().mean() + net.conv2.attention_conv2.weight.grad.abs().mean()) / 2
|
|
c3_grad_avg = sum([m.weight.grad.abs().mean().item() for m in net.conv3.conv_list._modules.values()]) / len(net.conv3.conv_list._modules)
|
|
c3a_grad_avg = (net.conv3.attention_conv1.weight.grad.abs().mean() + net.conv3.attention_conv2.weight.grad.abs().mean()) / 2
|
|
writer.add_scalar("c1_grad_avg", c1_grad_avg, global_step=step)
|
|
writer.add_scalar("c2_grad_avg", c2_grad_avg, global_step=step)
|
|
writer.add_scalar("c3_grad_avg", c3_grad_avg, global_step=step)
|
|
writer.add_scalar("c1a_grad_avg", c1a_grad_avg, global_step=step)
|
|
writer.add_scalar("c2a_grad_avg", c2a_grad_avg, global_step=step)
|
|
writer.add_scalar("c3a_grad_avg", c3a_grad_avg, global_step=step)
|
|
|
|
optimizer.step()
|
|
_, preds = torch.max(logits, 1)
|
|
running_corrects += torch.sum(preds == labels.data)
|
|
if step % 50 == 0:
|
|
print("Step: %i, Loss: %f, acc: %f, att_mean: %f" % (step, loss.item(), running_corrects / (50.0 * batch_size),
|
|
running_att_mean / 50.0))
|
|
writer.add_scalar("Loss", loss.item(), global_step=step)
|
|
writer.add_scalar("Accuracy", running_corrects / (50.0 * batch_size), global_step=step)
|
|
writer.add_scalar("Att Mean", running_att_mean / 50, global_step=step)
|
|
for i in range(len(running_att_hist)):
|
|
writer.add_histogram("Att Hist %i" % (i,), running_att_hist[i], global_step=step)
|
|
writer.flush()
|
|
running_corrects = 0
|
|
running_att_mean = 0
|
|
running_att_hist = None
|
|
if step % 1000 == 0:
|
|
temperature = max(temperature-1, 1)
|
|
net.set_temp(temperature)
|
|
print("Temperature drop. Now: %i" % (temperature,))
|
|
step += 1
|
|
torch.save((net.state_dict(), optimizer.state_dict()), "test_net.pth")
|
|
|
|
if __name__ == '__main__':
|
|
test_dynamic_conv()
|