Skip to content

Instantly share code, notes, and snippets.

@ProGamerGov

ProGamerGov/image_cppn.py

Last active May 22, 2020 00:35
Show Gist options
  • Save ProGamerGov/3bad855e1ebe84b4ea3e4db2349ef823 to your computer and use it in GitHub Desktop.
Save ProGamerGov/3bad855e1ebe84b4ea3e4db2349ef823 to your computer and use it in GitHub Desktop.
Download ZIP
tensorflow/lucid CPPN (X,Y) --> (R,G,B) Differentiable Image Parameterization in PyTorch
Raw
image_cppn.py
# tensorflow/lucid CPPN (X,Y) --> (R,G,B) Differentiable Image Parameterization in PyTorch
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
from inception5h import Inception5h
from PIL import Image
def main():
# Setup params
image_size=(224,224)
iterations = 512
lr = 0.005
use_device = 'cuda:0'
# Uncomment to use seed
#set_seed(876)
# Target layer and channel
layer_name='mixed4b_3x3_pre_relu_conv'
channel=77
# Load InceptionV1
cnn = Inception5h()
cnn.load_state_dict(torch.load('inception5h.pth'))
cnn.add_layers() # Add layers not found in model file
net = cnn.to(use_device).eval()
for param in net.parameters():
param.requires_grad = False
# Uncomment to print all hookable layers
#print(get_hookable_layers(cnn))
# Create instance of CPPN
img_cppn = CPPN_Conv(size=image_size, num_channels=24, num_layers=8, use_device=use_device, normalize=False)
print('CPPN Params Count', sum(p.numel() for p in img_cppn.parameters() if p.requires_grad))
loss_modules = register_hook_module(net, layer_name, channel)
# Create 224x224 image
output_tensor, img_cppn = dream_cppn(net, img_cppn, iterations, lr, loss_modules, use_device)
simple_deprocess(output_tensor, name='out.png')
# Create 720x720 image using multiscale generation
image_size=(720,720)
img_cppn.remake_input(image_size, use_device=use_device)
output_tensor, img_cppn = dream_cppn(net, img_cppn, iterations, lr, loss_modules, use_device)
simple_deprocess(output_tensor, name='out.png')
# Simple Deprocess
def simple_deprocess(output_tensor, name):
output_tensor = output_tensor.squeeze(0).cpu() /255
output_tensor.clamp_(0, 1)
Image2PIL = transforms.ToPILImage()
image = Image2PIL(output_tensor.cpu())
image.save(name)
# Function to maximize CNN activations
def dream_cppn(net, img_cppn, iterations, lr, loss_modules, use_device):
# Setup optimizer to optimize CPPN instance
optimizer = torch.optim.Adam(img_cppn.parameters(), lr=lr)
# Training loop
for i in range(iterations):
optimizer.zero_grad()
img = img_cppn() * 255 # Create RGB image with CPPN
net(img) # Create loss values
loss = 0
for mod in loss_modules: # Collect loss values
loss += mod.loss
loss.backward()
# Uncomment to save iterations
#if i % 25 == 0:
# simple_deprocess(img.detach(), 'out_'+str(i)+'.png')
print('Iteration', str(i+1), 'Loss', str(loss.item()))
optimizer.step()
img = img_cppn() * 255
return img, img_cppn
# Activation function for CPPN
class CompositeActivation(nn.Module):
def forward(self, input):
input = torch.atan(input)
return torch.cat([input / 0.67, (input * input) / 0.6], 1)
# Compositional pattern-producing network (CPPN) with Conv2d layers
class CPPN_Conv(nn.Module):
def __init__(self, size=(224, 224), num_channels=24, num_layers=8, activ_func=CompositeActivation(), use_device='cpu', normalize=False):
super(CPPN_Conv, self).__init__()
self.input_size = size
self.n_channels = num_channels
self.net = self.create_net(num_channels, num_layers, activ_func, use_device, normalize)
self.cppn_input = self.create_input(use_device)
# Create CPPN (X,Y) --> (R,G,B)
def create_net(self, num_channels, num_layers, activ_func, use_device, instance_norm, affine=True, bias=True):
net = nn.Sequential()
net.add_module(str(len(net)), nn.Conv2d(in_channels=2, out_channels=num_channels, kernel_size=1, bias=bias))
if instance_norm:
net.add_module(str(len(net)), nn.InstanceNorm2d(num_channels, affine=affine))
net.add_module(str(len(net)), activ_func)
for l in range(num_layers - 1):
net.add_module(str(len(net)), nn.Conv2d(in_channels=num_channels*2, out_channels=num_channels, kernel_size=1, bias=bias))
if instance_norm:
net.add_module(str(len(net)), nn.InstanceNorm2d(num_channels, affine=affine))
net.add_module(str(len(net)), activ_func)
net.add_module(str(len(net)), nn.Conv2d(in_channels=num_channels*2, out_channels=3, kernel_size=1, bias=bias))
net.add_module(str(len(net)), nn.Sigmoid())
net.apply(self.cppn_normal)
return net.to(use_device)
# Create X,Y input for CPPN
def create_input(self, use_device):
if type(self.input_size) is not tuple and type(self.input_size) is not list:
self.input_size = (self.input_size, self.input_size)
w = torch.arange(0, self.input_size[1]).to(use_device)
h = torch.arange(0, self.input_size[0]).to(use_device)
w_exp = w.unsqueeze(1).expand((self.input_size[1], self.input_size[0])).true_divide(self.input_size[0]) - 0.5
h_exp = h.unsqueeze(0).expand((self.input_size[1], self.input_size[0])).true_divide(self.input_size[1]) - 0.5
return torch.stack((w_exp, h_exp), -1).permute(2,1,0).unsqueeze(0)
# Replace input with different sized input
def remake_input(self, image_size, use_device):
self.input_size = image_size
self.cppn_input = self.create_input(use_device)
# Normalize Conv2d weights
def cppn_normal(self, l):
if type(l) == nn.Conv2d:
l.weight.data.normal_(0, (1.0/self.n_channels)**(1/2))
if l.bias is not None:
nn.init.zeros_(l.bias)
def forward(self):
return self.net(self.cppn_input)
# Create loss module and hook
def register_hook_module(net, layer_name, channel=-1):
loss_module = SimpleDreamLossHook(channel)
getattr(net, layer_name).register_forward_hook(loss_module)
return [loss_module]
# Define a simple forward hook to collect DeepDream loss
class SimpleDreamLossHook(nn.Module):
def __init__(self, channel=-1):
super(SimpleDreamLossHook, self).__init__()
self.get_loss = self.mean_loss
self.channel = channel
def mean_loss(self, input):
return input.mean()
def forward(self, module, input, output):
if self.channel > -1:
self.loss = -self.get_loss(output[:,self.channel,:,:])
else:
self.loss = -self.get_loss(output)
# Set global seeds to output reproducible
def set_seed(seed):
import random
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic=True
random.seed(seed)
# Get layers that can be hooked
def get_hookable_layers(cnn):
return [l[0] for l in list(cnn.named_children())]
if __name__ == "__main__":
main()
Raw
inception5h.py
# Model donwload: https://github.com/ProGamerGov/pytorch-old-tensorflow-models/blob/master/inception5h.pth
import torch
import torch.nn as nn
import torch.nn.functional as F
class ReluLayer(nn.Module):
def forward(self, input):
return F.relu(input)
class CatLayer(nn.Module):
def forward(self, input_list, dim=1):
return torch.cat(input_list, dim)
class MaxPool2dLayer(nn.Module):
def forward(self, input, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False):
return F.max_pool2d(input, kernel_size, stride=stride, padding=padding, ceil_mode=ceil_mode)
class SoftMaxLayer(nn.Module):
def forward(self, input, dim=1):
return F.softmax(input, dim=dim)
class Inception5h(nn.Module):
def __init__(self):
super(Inception5h, self).__init__()
self.conv2d0_pre_relu_conv = nn.Conv2d(in_channels=3, out_channels=64, kernel_size=(7, 7), stride=(2, 2), groups=1, bias=True)
self.conv2d1_pre_relu_conv = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.conv2d2_pre_relu_conv = nn.Conv2d(in_channels=64, out_channels=192, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed3a_1x1_pre_relu_conv = nn.Conv2d(in_channels=192, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3a_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=192, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3a_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=192, out_channels=16, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3a_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=192, out_channels=32, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3a_3x3_pre_relu_conv = nn.Conv2d(in_channels=96, out_channels=128, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed3a_5x5_pre_relu_conv = nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed3b_1x1_pre_relu_conv = nn.Conv2d(in_channels=256, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3b_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=256, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3b_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=256, out_channels=32, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3b_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=256, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed3b_3x3_pre_relu_conv = nn.Conv2d(in_channels=128, out_channels=192, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed3b_5x5_pre_relu_conv = nn.Conv2d(in_channels=32, out_channels=96, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed4a_1x1_pre_relu_conv = nn.Conv2d(in_channels=480, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4a_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=480, out_channels=96, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4a_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=480, out_channels=16, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4a_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=480, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4a_3x3_pre_relu_conv = nn.Conv2d(in_channels=96, out_channels=204, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed4a_5x5_pre_relu_conv = nn.Conv2d(in_channels=16, out_channels=48, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed4b_1x1_pre_relu_conv = nn.Conv2d(in_channels=508, out_channels=160, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4b_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=508, out_channels=112, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4b_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=508, out_channels=24, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4b_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=508, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4b_3x3_pre_relu_conv = nn.Conv2d(in_channels=112, out_channels=224, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed4b_5x5_pre_relu_conv = nn.Conv2d(in_channels=24, out_channels=64, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed4c_1x1_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4c_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4c_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=24, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4c_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4c_3x3_pre_relu_conv = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed4c_5x5_pre_relu_conv = nn.Conv2d(in_channels=24, out_channels=64, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed4d_1x1_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=112, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4d_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=144, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4d_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=32, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4d_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=512, out_channels=64, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4d_3x3_pre_relu_conv = nn.Conv2d(in_channels=144, out_channels=288, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed4d_5x5_pre_relu_conv = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed4e_1x1_pre_relu_conv = nn.Conv2d(in_channels=528, out_channels=256, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4e_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=528, out_channels=160, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4e_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=528, out_channels=32, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4e_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=528, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed4e_3x3_pre_relu_conv = nn.Conv2d(in_channels=160, out_channels=320, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed4e_5x5_pre_relu_conv = nn.Conv2d(in_channels=32, out_channels=128, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed5a_1x1_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=256, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5a_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=160, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5a_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=48, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5a_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5a_3x3_pre_relu_conv = nn.Conv2d(in_channels=160, out_channels=320, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed5a_5x5_pre_relu_conv = nn.Conv2d(in_channels=48, out_channels=128, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.mixed5b_1x1_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=384, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5b_3x3_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=192, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5b_5x5_bottleneck_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=48, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5b_pool_reduce_pre_relu_conv = nn.Conv2d(in_channels=832, out_channels=128, kernel_size=(1, 1), stride=(1, 1), groups=1, bias=True)
self.mixed5b_3x3_pre_relu_conv = nn.Conv2d(in_channels=192, out_channels=384, kernel_size=(3, 3), stride=(1, 1), groups=1, bias=True)
self.mixed5b_5x5_pre_relu_conv = nn.Conv2d(in_channels=48, out_channels=128, kernel_size=(5, 5), stride=(1, 1), groups=1, bias=True)
self.softmax2_pre_activation_matmul = nn.Linear(in_features = 1024, out_features = 1008, bias = True)
def add_layers(self):
self.conv2d0 = ReluLayer()
self.maxpool0 = MaxPool2dLayer()
self.conv2d1 = ReluLayer()
self.conv2d2 = ReluLayer()
self.maxpool1 = MaxPool2dLayer()
self.mixed3a_pool = MaxPool2dLayer()
self.mixed3a_1x1 = ReluLayer()
self.mixed3a_3x3_bottleneck = ReluLayer()
self.mixed3a_5x5_bottleneck = ReluLayer()
self.mixed3a_pool_reduce = ReluLayer()
self.mixed3a_3x3 = ReluLayer()
self.mixed3a_5x5 = ReluLayer()
self.mixed3a = CatLayer()
self.mixed3b_pool = MaxPool2dLayer()
self.mixed3b_1x1 = ReluLayer()
self.mixed3b_3x3_bottleneck = ReluLayer()
self.mixed3b_5x5_bottleneck = ReluLayer()
self.mixed3b_pool_reduce = ReluLayer()
self.mixed3b_3x3 = ReluLayer()
self.mixed3b_5x5 = ReluLayer()
self.mixed3b = CatLayer()
self.maxpool4 = MaxPool2dLayer()
self.mixed4a_pool = MaxPool2dLayer()
self.mixed4a_1x1 = ReluLayer()
self.mixed4a_3x3_bottleneck = ReluLayer()
self.mixed4a_5x5_bottleneck = ReluLayer()
self.mixed4a_pool_reduce = ReluLayer()
self.mixed4a_3x3 = ReluLayer()
self.mixed4a_5x5 = ReluLayer()
self.mixed4a = CatLayer()
self.mixed4b_pool = MaxPool2dLayer()
self.mixed4b_1x1 = ReluLayer()
self.mixed4b_3x3_bottleneck = ReluLayer()
self.mixed4b_5x5_bottleneck = ReluLayer()
self.mixed4b_pool_reduce = ReluLayer()
self.mixed4b_3x3 = ReluLayer()
self.mixed4b_5x5 = ReluLayer()
self.mixed4b = CatLayer()
self.mixed4c_pool = MaxPool2dLayer()
self.mixed4c_1x1 = ReluLayer()
self.mixed4c_3x3_bottleneck = ReluLayer()
self.mixed4c_5x5_bottleneck = ReluLayer()
self.mixed4c_pool_reduce = ReluLayer()
self.mixed4c_3x3 = ReluLayer()
self.mixed4c_5x5 = ReluLayer()
self.mixed4c = CatLayer()
self.mixed4d_pool = MaxPool2dLayer()
self.mixed4d_1x1 = ReluLayer()
self.mixed4d_3x3_bottleneck = ReluLayer()
self.mixed4d_5x5_bottleneck = ReluLayer()
self.mixed4d_pool_reduce = ReluLayer()
self.mixed4d_3x3 = ReluLayer()
self.mixed4d_5x5 = ReluLayer()
self.mixed4d = CatLayer()
self.mixed4e_pool = MaxPool2dLayer()
self.mixed4e_1x1 = ReluLayer()
self.mixed4e_3x3_bottleneck = ReluLayer()
self.mixed4e_5x5_bottleneck = ReluLayer()
self.mixed4e_pool_reduce = ReluLayer()
self.mixed4e_3x3 = ReluLayer()
self.mixed4e_5x5 = ReluLayer()
self.mixed4e = CatLayer()
self.maxpool10 = MaxPool2dLayer()
self.mixed5a_pool = MaxPool2dLayer()
self.mixed5a_1x1 = ReluLayer()
self.mixed5a_3x3_bottleneck = ReluLayer()
self.mixed5a_5x5_bottleneck = ReluLayer()
self.mixed5a_pool_reduce = ReluLayer()
self.mixed5a_3x3 = ReluLayer()
self.mixed5a_5x5 = ReluLayer()
self.mixed5a = CatLayer()
self.mixed5b_pool = MaxPool2dLayer()
self.mixed5b_1x1 = ReluLayer()
self.mixed5b_3x3_bottleneck = ReluLayer()
self.mixed5b_5x5_bottleneck = ReluLayer()
self.mixed5b_pool_reduce = ReluLayer()
self.mixed5b_3x3 = ReluLayer()
self.mixed5b_5x5 = ReluLayer()
self.mixed5b = CatLayer()
self.softmax2 = SoftMaxLayer()
def forward(self, x):
conv2d0_pre_relu_conv_pad = F.pad(x, (2, 3, 2, 3))
conv2d0_pre_relu_conv = self.conv2d0_pre_relu_conv(conv2d0_pre_relu_conv_pad)
conv2d0 = self.conv2d0(conv2d0_pre_relu_conv)
maxpool0_pad = F.pad(conv2d0, (0, 1, 0, 1), value=float('-inf'))
maxpool0 = self.maxpool0(maxpool0_pad, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False)
localresponsenorm0 = F.local_response_norm(maxpool0, size=9, alpha=9.99999974738e-05, beta=0.5, k=1)
conv2d1_pre_relu_conv = self.conv2d1_pre_relu_conv(localresponsenorm0)
conv2d1 = self.conv2d1(conv2d1_pre_relu_conv)
conv2d2_pre_relu_conv_pad = F.pad(conv2d1, (1, 1, 1, 1))
conv2d2_pre_relu_conv = self.conv2d2_pre_relu_conv(conv2d2_pre_relu_conv_pad)
conv2d2 = self.conv2d2(conv2d2_pre_relu_conv)
localresponsenorm1 = F.local_response_norm(conv2d2, size=9, alpha=9.99999974738e-05, beta=0.5, k=1)
maxpool1_pad = F.pad(localresponsenorm1, (0, 1, 0, 1), value=float('-inf'))
maxpool1 = self.maxpool1(maxpool1_pad, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False)
mixed3a_1x1_pre_relu_conv = self.mixed3a_1x1_pre_relu_conv(maxpool1)
mixed3a_3x3_bottleneck_pre_relu_conv = self.mixed3a_3x3_bottleneck_pre_relu_conv(maxpool1)
mixed3a_5x5_bottleneck_pre_relu_conv = self.mixed3a_5x5_bottleneck_pre_relu_conv(maxpool1)
mixed3a_pool_pad = F.pad(maxpool1, (1, 1, 1, 1), value=float('-inf'))
mixed3a_pool = self.mixed3a_pool(mixed3a_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed3a_1x1 = self.mixed3a_1x1(mixed3a_1x1_pre_relu_conv)
mixed3a_3x3_bottleneck = self.mixed3a_3x3_bottleneck(mixed3a_3x3_bottleneck_pre_relu_conv)
mixed3a_5x5_bottleneck = self.mixed3a_5x5_bottleneck(mixed3a_5x5_bottleneck_pre_relu_conv)
mixed3a_pool_reduce_pre_relu_conv = self.mixed3a_pool_reduce_pre_relu_conv(mixed3a_pool)
mixed3a_3x3_pre_relu_conv_pad = F.pad(mixed3a_3x3_bottleneck, (1, 1, 1, 1))
mixed3a_3x3_pre_relu_conv = self.mixed3a_3x3_pre_relu_conv(mixed3a_3x3_pre_relu_conv_pad)
mixed3a_5x5_pre_relu_conv_pad = F.pad(mixed3a_5x5_bottleneck, (2, 2, 2, 2))
mixed3a_5x5_pre_relu_conv = self.mixed3a_5x5_pre_relu_conv(mixed3a_5x5_pre_relu_conv_pad)
mixed3a_pool_reduce = self.mixed3a_pool_reduce(mixed3a_pool_reduce_pre_relu_conv)
mixed3a_3x3 = self.mixed3a_3x3(mixed3a_3x3_pre_relu_conv)
mixed3a_5x5 = self.mixed3a_5x5(mixed3a_5x5_pre_relu_conv)
mixed3a = self.mixed3a((mixed3a_1x1, mixed3a_3x3, mixed3a_5x5, mixed3a_pool_reduce), 1)
mixed3b_1x1_pre_relu_conv = self.mixed3b_1x1_pre_relu_conv(mixed3a)
mixed3b_3x3_bottleneck_pre_relu_conv = self.mixed3b_3x3_bottleneck_pre_relu_conv(mixed3a)
mixed3b_5x5_bottleneck_pre_relu_conv = self.mixed3b_5x5_bottleneck_pre_relu_conv(mixed3a)
mixed3b_pool_pad = F.pad(mixed3a, (1, 1, 1, 1), value=float('-inf'))
mixed3b_pool = self.mixed3b_pool(mixed3b_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed3b_1x1 = self.mixed3b_1x1(mixed3b_1x1_pre_relu_conv)
mixed3b_3x3_bottleneck = self.mixed3b_3x3_bottleneck(mixed3b_3x3_bottleneck_pre_relu_conv)
mixed3b_5x5_bottleneck = self.mixed3b_5x5_bottleneck(mixed3b_5x5_bottleneck_pre_relu_conv)
mixed3b_pool_reduce_pre_relu_conv = self.mixed3b_pool_reduce_pre_relu_conv(mixed3b_pool)
mixed3b_3x3_pre_relu_conv_pad = F.pad(mixed3b_3x3_bottleneck, (1, 1, 1, 1))
mixed3b_3x3_pre_relu_conv = self.mixed3b_3x3_pre_relu_conv(mixed3b_3x3_pre_relu_conv_pad)
mixed3b_5x5_pre_relu_conv_pad = F.pad(mixed3b_5x5_bottleneck, (2, 2, 2, 2))
mixed3b_5x5_pre_relu_conv = self.mixed3b_5x5_pre_relu_conv(mixed3b_5x5_pre_relu_conv_pad)
mixed3b_pool_reduce = self.mixed3b_pool_reduce(mixed3b_pool_reduce_pre_relu_conv)
mixed3b_3x3 = self.mixed3b_3x3(mixed3b_3x3_pre_relu_conv)
mixed3b_5x5 = self.mixed3b_5x5(mixed3b_5x5_pre_relu_conv)
mixed3b = self.mixed3b((mixed3b_1x1, mixed3b_3x3, mixed3b_5x5, mixed3b_pool_reduce), 1)
maxpool4_pad = F.pad(mixed3b, (0, 1, 0, 1), value=float('-inf'))
maxpool4 = self.maxpool4(maxpool4_pad, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False)
mixed4a_1x1_pre_relu_conv = self.mixed4a_1x1_pre_relu_conv(maxpool4)
mixed4a_3x3_bottleneck_pre_relu_conv = self.mixed4a_3x3_bottleneck_pre_relu_conv(maxpool4)
mixed4a_5x5_bottleneck_pre_relu_conv = self.mixed4a_5x5_bottleneck_pre_relu_conv(maxpool4)
mixed4a_pool_pad = F.pad(maxpool4, (1, 1, 1, 1), value=float('-inf'))
mixed4a_pool = self.mixed4a_pool(mixed4a_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed4a_1x1 = self.mixed4a_1x1(mixed4a_1x1_pre_relu_conv)
mixed4a_3x3_bottleneck = self.mixed4a_3x3_bottleneck(mixed4a_3x3_bottleneck_pre_relu_conv)
mixed4a_5x5_bottleneck = self.mixed4a_5x5_bottleneck(mixed4a_5x5_bottleneck_pre_relu_conv)
mixed4a_pool_reduce_pre_relu_conv = self.mixed4a_pool_reduce_pre_relu_conv(mixed4a_pool)
mixed4a_3x3_pre_relu_conv_pad = F.pad(mixed4a_3x3_bottleneck, (1, 1, 1, 1))
mixed4a_3x3_pre_relu_conv = self.mixed4a_3x3_pre_relu_conv(mixed4a_3x3_pre_relu_conv_pad)
mixed4a_5x5_pre_relu_conv_pad = F.pad(mixed4a_5x5_bottleneck, (2, 2, 2, 2))
mixed4a_5x5_pre_relu_conv = self.mixed4a_5x5_pre_relu_conv(mixed4a_5x5_pre_relu_conv_pad)
mixed4a_pool_reduce = self.mixed4a_pool_reduce(mixed4a_pool_reduce_pre_relu_conv)
mixed4a_3x3 = self.mixed4a_3x3(mixed4a_3x3_pre_relu_conv)
mixed4a_5x5 = self.mixed4a_5x5(mixed4a_5x5_pre_relu_conv)
mixed4a = self.mixed4a((mixed4a_1x1, mixed4a_3x3, mixed4a_5x5, mixed4a_pool_reduce), 1)
mixed4b_1x1_pre_relu_conv = self.mixed4b_1x1_pre_relu_conv(mixed4a)
mixed4b_3x3_bottleneck_pre_relu_conv = self.mixed4b_3x3_bottleneck_pre_relu_conv(mixed4a)
mixed4b_5x5_bottleneck_pre_relu_conv = self.mixed4b_5x5_bottleneck_pre_relu_conv(mixed4a)
mixed4b_pool_pad = F.pad(mixed4a, (1, 1, 1, 1), value=float('-inf'))
mixed4b_pool = self.mixed4b_pool(mixed4b_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed4b_1x1 = self.mixed4b_1x1(mixed4b_1x1_pre_relu_conv)
mixed4b_3x3_bottleneck = self.mixed4b_3x3_bottleneck(mixed4b_3x3_bottleneck_pre_relu_conv)
mixed4b_5x5_bottleneck = self.mixed4b_5x5_bottleneck(mixed4b_5x5_bottleneck_pre_relu_conv)
mixed4b_pool_reduce_pre_relu_conv = self.mixed4b_pool_reduce_pre_relu_conv(mixed4b_pool)
mixed4b_3x3_pre_relu_conv_pad = F.pad(mixed4b_3x3_bottleneck, (1, 1, 1, 1))
mixed4b_3x3_pre_relu_conv = self.mixed4b_3x3_pre_relu_conv(mixed4b_3x3_pre_relu_conv_pad)
mixed4b_5x5_pre_relu_conv_pad = F.pad(mixed4b_5x5_bottleneck, (2, 2, 2, 2))
mixed4b_5x5_pre_relu_conv = self.mixed4b_5x5_pre_relu_conv(mixed4b_5x5_pre_relu_conv_pad)
mixed4b_pool_reduce = self.mixed4b_pool_reduce(mixed4b_pool_reduce_pre_relu_conv)
mixed4b_3x3 = self.mixed4b_3x3(mixed4b_3x3_pre_relu_conv)
mixed4b_5x5 = self.mixed4b_5x5(mixed4b_5x5_pre_relu_conv)
mixed4b = self.mixed4b((mixed4b_1x1, mixed4b_3x3, mixed4b_5x5, mixed4b_pool_reduce), 1)
mixed4c_1x1_pre_relu_conv = self.mixed4c_1x1_pre_relu_conv(mixed4b)
mixed4c_3x3_bottleneck_pre_relu_conv = self.mixed4c_3x3_bottleneck_pre_relu_conv(mixed4b)
mixed4c_5x5_bottleneck_pre_relu_conv = self.mixed4c_5x5_bottleneck_pre_relu_conv(mixed4b)
mixed4c_pool_pad = F.pad(mixed4b, (1, 1, 1, 1), value=float('-inf'))
mixed4c_pool = self.mixed4c_pool(mixed4c_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed4c_1x1 = self.mixed4c_1x1(mixed4c_1x1_pre_relu_conv)
mixed4c_3x3_bottleneck = self.mixed4c_3x3_bottleneck(mixed4c_3x3_bottleneck_pre_relu_conv)
mixed4c_5x5_bottleneck = self.mixed4c_5x5_bottleneck(mixed4c_5x5_bottleneck_pre_relu_conv)
mixed4c_pool_reduce_pre_relu_conv = self.mixed4c_pool_reduce_pre_relu_conv(mixed4c_pool)
mixed4c_3x3_pre_relu_conv_pad = F.pad(mixed4c_3x3_bottleneck, (1, 1, 1, 1))
mixed4c_3x3_pre_relu_conv = self.mixed4c_3x3_pre_relu_conv(mixed4c_3x3_pre_relu_conv_pad)
mixed4c_5x5_pre_relu_conv_pad = F.pad(mixed4c_5x5_bottleneck, (2, 2, 2, 2))
mixed4c_5x5_pre_relu_conv = self.mixed4c_5x5_pre_relu_conv(mixed4c_5x5_pre_relu_conv_pad)
mixed4c_pool_reduce = self.mixed4c_pool_reduce(mixed4c_pool_reduce_pre_relu_conv)
mixed4c_3x3 = self.mixed4c_3x3(mixed4c_3x3_pre_relu_conv)
mixed4c_5x5 = self.mixed4c_5x5(mixed4c_5x5_pre_relu_conv)
mixed4c = self.mixed4c((mixed4c_1x1, mixed4c_3x3, mixed4c_5x5, mixed4c_pool_reduce), 1)
mixed4d_1x1_pre_relu_conv = self.mixed4d_1x1_pre_relu_conv(mixed4c)
mixed4d_3x3_bottleneck_pre_relu_conv = self.mixed4d_3x3_bottleneck_pre_relu_conv(mixed4c)
mixed4d_5x5_bottleneck_pre_relu_conv = self.mixed4d_5x5_bottleneck_pre_relu_conv(mixed4c)
mixed4d_pool_pad = F.pad(mixed4c, (1, 1, 1, 1), value=float('-inf'))
mixed4d_pool = self.mixed4d_pool(mixed4d_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed4d_1x1 = self.mixed4d_1x1(mixed4d_1x1_pre_relu_conv)
mixed4d_3x3_bottleneck = self.mixed4d_3x3_bottleneck(mixed4d_3x3_bottleneck_pre_relu_conv)
mixed4d_5x5_bottleneck = self.mixed4d_5x5_bottleneck(mixed4d_5x5_bottleneck_pre_relu_conv)
mixed4d_pool_reduce_pre_relu_conv = self.mixed4d_pool_reduce_pre_relu_conv(mixed4d_pool)
mixed4d_3x3_pre_relu_conv_pad = F.pad(mixed4d_3x3_bottleneck, (1, 1, 1, 1))
mixed4d_3x3_pre_relu_conv = self.mixed4d_3x3_pre_relu_conv(mixed4d_3x3_pre_relu_conv_pad)
mixed4d_5x5_pre_relu_conv_pad = F.pad(mixed4d_5x5_bottleneck, (2, 2, 2, 2))
mixed4d_5x5_pre_relu_conv = self.mixed4d_5x5_pre_relu_conv(mixed4d_5x5_pre_relu_conv_pad)
mixed4d_pool_reduce = self.mixed4d_pool_reduce(mixed4d_pool_reduce_pre_relu_conv)
mixed4d_3x3 = self.mixed4d_3x3(mixed4d_3x3_pre_relu_conv)
mixed4d_5x5 = self.mixed4d_5x5(mixed4d_5x5_pre_relu_conv)
mixed4d = self.mixed4d((mixed4d_1x1, mixed4d_3x3, mixed4d_5x5, mixed4d_pool_reduce), 1)
mixed4e_1x1_pre_relu_conv = self.mixed4e_1x1_pre_relu_conv(mixed4d)
mixed4e_3x3_bottleneck_pre_relu_conv = self.mixed4e_3x3_bottleneck_pre_relu_conv(mixed4d)
mixed4e_5x5_bottleneck_pre_relu_conv = self.mixed4e_5x5_bottleneck_pre_relu_conv(mixed4d)
mixed4e_pool_pad = F.pad(mixed4d, (1, 1, 1, 1), value=float('-inf'))
mixed4e_pool = self.mixed4e_pool(mixed4e_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed4e_1x1 = self.mixed4e_1x1(mixed4e_1x1_pre_relu_conv)
mixed4e_3x3_bottleneck = self.mixed4e_3x3_bottleneck(mixed4e_3x3_bottleneck_pre_relu_conv)
mixed4e_5x5_bottleneck = self.mixed4e_5x5_bottleneck(mixed4e_5x5_bottleneck_pre_relu_conv)
mixed4e_pool_reduce_pre_relu_conv = self.mixed4e_pool_reduce_pre_relu_conv(mixed4e_pool)
mixed4e_3x3_pre_relu_conv_pad = F.pad(mixed4e_3x3_bottleneck, (1, 1, 1, 1))
mixed4e_3x3_pre_relu_conv = self.mixed4e_3x3_pre_relu_conv(mixed4e_3x3_pre_relu_conv_pad)
mixed4e_5x5_pre_relu_conv_pad = F.pad(mixed4e_5x5_bottleneck, (2, 2, 2, 2))
mixed4e_5x5_pre_relu_conv = self.mixed4e_5x5_pre_relu_conv(mixed4e_5x5_pre_relu_conv_pad)
mixed4e_pool_reduce = self.mixed4e_pool_reduce(mixed4e_pool_reduce_pre_relu_conv)
mixed4e_3x3 = self.mixed4e_3x3(mixed4e_3x3_pre_relu_conv)
mixed4e_5x5 = self.mixed4e_5x5(mixed4e_5x5_pre_relu_conv)
mixed4e = self.mixed4e((mixed4e_1x1, mixed4e_3x3, mixed4e_5x5, mixed4e_pool_reduce), 1)
maxpool10_pad = F.pad(mixed4e, (0, 1, 0, 1), value=float('-inf'))
maxpool10 = self.maxpool10(maxpool10_pad, kernel_size=(3, 3), stride=(2, 2), padding=0, ceil_mode=False)
mixed5a_1x1_pre_relu_conv = self.mixed5a_1x1_pre_relu_conv(maxpool10)
mixed5a_3x3_bottleneck_pre_relu_conv = self.mixed5a_3x3_bottleneck_pre_relu_conv(maxpool10)
mixed5a_5x5_bottleneck_pre_relu_conv = self.mixed5a_5x5_bottleneck_pre_relu_conv(maxpool10)
mixed5a_pool_pad = F.pad(maxpool10, (1, 1, 1, 1), value=float('-inf'))
mixed5a_pool = self.mixed5a_pool(mixed5a_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed5a_1x1 = self.mixed5a_1x1(mixed5a_1x1_pre_relu_conv)
mixed5a_3x3_bottleneck = self.mixed5a_3x3_bottleneck(mixed5a_3x3_bottleneck_pre_relu_conv)
mixed5a_5x5_bottleneck = self.mixed5a_5x5_bottleneck(mixed5a_5x5_bottleneck_pre_relu_conv)
mixed5a_pool_reduce_pre_relu_conv = self.mixed5a_pool_reduce_pre_relu_conv(mixed5a_pool)
mixed5a_3x3_pre_relu_conv_pad = F.pad(mixed5a_3x3_bottleneck, (1, 1, 1, 1))
mixed5a_3x3_pre_relu_conv = self.mixed5a_3x3_pre_relu_conv(mixed5a_3x3_pre_relu_conv_pad)
mixed5a_5x5_pre_relu_conv_pad = F.pad(mixed5a_5x5_bottleneck, (2, 2, 2, 2))
mixed5a_5x5_pre_relu_conv = self.mixed5a_5x5_pre_relu_conv(mixed5a_5x5_pre_relu_conv_pad)
mixed5a_pool_reduce = self.mixed5a_pool_reduce(mixed5a_pool_reduce_pre_relu_conv)
mixed5a_3x3 = self.mixed5a_3x3(mixed5a_3x3_pre_relu_conv)
mixed5a_5x5 = self.mixed5a_5x5(mixed5a_5x5_pre_relu_conv)
mixed5a = self.mixed5a((mixed5a_1x1, mixed5a_3x3, mixed5a_5x5, mixed5a_pool_reduce), 1)
mixed5b_1x1_pre_relu_conv = self.mixed5b_1x1_pre_relu_conv(mixed5a)
mixed5b_3x3_bottleneck_pre_relu_conv = self.mixed5b_3x3_bottleneck_pre_relu_conv(mixed5a)
mixed5b_5x5_bottleneck_pre_relu_conv = self.mixed5b_5x5_bottleneck_pre_relu_conv(mixed5a)
mixed5b_pool_pad = F.pad(mixed5a, (1, 1, 1, 1), value=float('-inf'))
mixed5b_pool = self.mixed5b_pool(mixed5b_pool_pad, kernel_size=(3, 3), stride=(1, 1), padding=0, ceil_mode=False)
mixed5b_1x1 = self.mixed5b_1x1(mixed5b_1x1_pre_relu_conv)
mixed5b_3x3_bottleneck = self.mixed5b_3x3_bottleneck(mixed5b_3x3_bottleneck_pre_relu_conv)
mixed5b_5x5_bottleneck = self.mixed5b_5x5_bottleneck(mixed5b_5x5_bottleneck_pre_relu_conv)
mixed5b_pool_reduce_pre_relu_conv = self.mixed5b_pool_reduce_pre_relu_conv(mixed5b_pool)
mixed5b_3x3_pre_relu_conv_pad = F.pad(mixed5b_3x3_bottleneck, (1, 1, 1, 1))
mixed5b_3x3_pre_relu_conv = self.mixed5b_3x3_pre_relu_conv(mixed5b_3x3_pre_relu_conv_pad)
mixed5b_5x5_pre_relu_conv_pad = F.pad(mixed5b_5x5_bottleneck, (2, 2, 2, 2))
mixed5b_5x5_pre_relu_conv = self.mixed5b_5x5_pre_relu_conv(mixed5b_5x5_pre_relu_conv_pad)
mixed5b_pool_reduce = self.mixed5b_pool_reduce(mixed5b_pool_reduce_pre_relu_conv)
mixed5b_3x3 = self.mixed5b_3x3(mixed5b_3x3_pre_relu_conv)
mixed5b_5x5 = self.mixed5b_5x5(mixed5b_5x5_pre_relu_conv)
mixed5b = self.mixed5b((mixed5b_1x1, mixed5b_3x3, mixed5b_5x5, mixed5b_pool_reduce), 1)
avgpool0 = F.avg_pool2d(mixed5b, kernel_size=(7, 7), stride=(1, 1), padding=(0,), ceil_mode=False, count_include_pad=False)
avgpool0_reshape = torch.reshape(input = avgpool0, shape = (-1,1024))
softmax2_pre_activation_matmul = self.softmax2_pre_activation_matmul(avgpool0_reshape)
softmax2 = self.softmax2(softmax2_pre_activation_matmul)
return softmax2
@ProGamerGov
Copy link
Author

ProGamerGov commented May 7, 2020 •
edited
Loading

This will let you run the above code on a test image:

from PIL import Image
def preprocess(image_name, image_size):
    Loader = transforms.Compose([transforms.Resize(image_size), transforms.ToTensor()])
    Normalize = transforms.Compose([transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[1,1,1])])
    return Normalize(Loader(Image.open(image_name).convert('RGB'))).unsqueeze(0)
def deprocess(output_tensor, name='out.png'):
    Normalize = transforms.Compose([transforms.Normalize(mean=[-0.485, -0.456, -0.406], std=[1,1,1])])
    output_tensor = Normalize(output_tensor.squeeze(0).cpu())
    Image2PIL = transforms.ToPILImage()
    image = Image2PIL(output_tensor.clamp_(0, 1).cpu()).save(name)


# Use test image instead of CPPN
def dream_cppn(net, image_size, iterations, lr, use_device):
    # Create random image 
    img = nn.Parameter(preprocess('test.jpg', image_size).to(use_device)) 
    # Setup optimizer to optimize image
    optimizer = torch.optim.Adam([img], lr=lr)

    # Training loop
    for i in range(iterations):
        optimizer.zero_grad()
        out = net(img) # Get activations from CNN

        loss = -out.pow(2).sum().sqrt()
        loss.backward()

        # Uncomment to save iterations		
        #simple_deprocess(img.detach(), 'iters/out_'+str(i)+'.png')

        print('Iteration', str(i+1), 'Loss', str(loss.item()))

        optimizer.step()
    return img

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment