converting to cuda pytorch

[ker2x]

Friendly greetings !

I'm just a sysadmin that happen to have a server with an absurdly powerful NVidia P100, i totally enjoy generative art but i'm not that much of a python programmer.

I tried my best to convert this notebook to cuda but, meh, not luck.

this is what i got but it come with warning and still doesn't seems to run on gpu : Do you think you can convert your cpu code to gpu ? :pray: thx <3

#!/usr/bin/env python
# coding: utf-8

# In[1]:


import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
from IPython import display
from matplotlib import colors
import os, copy
from PIL import Image


# In[2]:


print("Cuda is available : ", torch.cuda.is_available())
print("Current device #", torch.cuda.current_device(), ", Name : ", torch.cuda.get_device_name(torch.cuda.current_device()))
print("Current Device Memory allocated : ", torch.cuda.memory_allocated())


# In[3]:


def init_normal(m):
    if type(m) == nn.Linear:        
        nn.init.normal_(m.weight)


class NN(nn.Module):

    def __init__(self, activation=nn.Tanh, num_neurons=16, num_layers=9):
        """
        num_layers must be at least two
        """
        super(NN, self).__init__()
        layers = [nn.Linear(2, num_neurons, bias=True), activation()]
        for _ in range(num_layers - 1):
            layers += [nn.Linear(num_neurons, num_neurons, bias=False), activation()]
        layers += [nn.Linear(num_neurons, 3, bias=False), nn.Sigmoid()]
        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        return self.layers(x)


def gen_new_image(size_x, size_y, save=True, **kwargs):
    net = NN(**kwargs).cuda()
    net.apply(init_normal)
    colors = run_net(net, size_x, size_y)
    plot_colors(colors)
    if save is True:
        save_colors(colors)
    return net, colors


def run_net(net, size_x=128, size_y=128):
    x = torch.arange(0, size_x, 1)
    y = torch.arange(0, size_y, 1)
    colors = torch.zeros((size_x, size_y, 2))
    for i in x:
        for j in y:
            colors[i][j] = torch.tensor([float(i) / size_y - 0.5, float(j) / size_x - 0.5])
    colors = colors.reshape(size_x * size_y, 2)
    #img = net(torch.tensor(colors).type(torch.FloatTensor)).detach().numpy()
    img = net(torch.tensor(colors).type(torch.cuda.FloatTensor)).cuda()
    img2 = img.cpu().detach().numpy()
    return img2.reshape(size_x, size_y, 3)


def plot_colors(colors, fig_size=15):
    plt.figure(figsize=(fig_size, fig_size))
    plt.imshow(colors, interpolation='nearest', vmin=0, vmax=1)


def save_colors(colors):
    plt.imsave(str(np.random.randint(100000)) + ".png", colors)


def run_plot_save(net, size_x, size_y, fig_size=15):
    colors = run_net(net, size_x, size_y)
    plot_colors(colors, fig_size)
    save_colors(colors)


# In[4]:


n,c = gen_new_image(1024, 1024, save=False, num_neurons=32)


# In[5]:


run_plot_save(n, 1080, 720)


# Let's see how the images change if we increase the depth

# In[57]:


for num_layers in range(2, 30, 3):
    print(f"{num_layers} layers")
    n,c = gen_new_image(128, 128, save=False, num_layers=num_layers)


# And also the effect of increasing the width

# In[58]:


for i in range(1, 10, 2):
    print(f"{num_layers} layers")
    n,c = gen_new_image(128, 128, save=False, num_neurons=2**i)


# What happens if we use ReLUs?

# In[60]:


n,c = gen_new_image(128, 128, save=False, activation=nn.ReLU)


# In[ ]:

[paraschopra]

Sorry, I don't have easy CUDA availability. Perhaps someone else can patch it.

On Mon, 18 Mar 2019 at 21:16, Laurent Laborde [email protected] wrote:

Friendly greetings !

I'm just a sysadmin that happen to have a server with an absurdly powerful NVidia P100, i totally enjoy generative art but i'm not that much of a python programmer.

I tried my best to convert this notebook to cuda but, meh, not luck.

this is what i got but it come with warning and still doesn't seems to run on gpu : Do you think you can convert your cpu code to gpu ? 🙏 thx <3

#!/usr/bin/env python

coding: utf-8

In[1]:

import torch

import torch.nn as nn

import torch.nn.functional as F

import numpy as np

import matplotlib.pyplot as plt

from IPython import display

from matplotlib import colors

import os, copy

from PIL import Image

In[2]:

print("Cuda is available : ", torch.cuda.is_available())

print("Current device #", torch.cuda.current_device(), ", Name : ", torch.cuda.get_device_name(torch.cuda.current_device()))

print("Current Device Memory allocated : ", torch.cuda.memory_allocated())

In[3]:

def init_normal(m):

if type(m) == nn.Linear:

    nn.init.normal_(m.weight)

class NN(nn.Module):

def __init__(self, activation=nn.Tanh, num_neurons=16, num_layers=9):

    """

    num_layers must be at least two

    """

    super(NN, self).__init__()

    layers = [nn.Linear(2, num_neurons, bias=True), activation()]

    for _ in range(num_layers - 1):

        layers += [nn.Linear(num_neurons, num_neurons, bias=False), activation()]

    layers += [nn.Linear(num_neurons, 3, bias=False), nn.Sigmoid()]

    self.layers = nn.Sequential(*layers)



def forward(self, x):

    return self.layers(x)

def gen_new_image(size_x, size_y, save=True, **kwargs):

net = NN(**kwargs).cuda()

net.apply(init_normal)

colors = run_net(net, size_x, size_y)

plot_colors(colors)

if save is True:

    save_colors(colors)

return net, colors

def run_net(net, size_x=128, size_y=128):

x = torch.arange(0, size_x, 1)

y = torch.arange(0, size_y, 1)

colors = torch.zeros((size_x, size_y, 2))

for i in x:

    for j in y:

        colors[i][j] = torch.tensor([float(i) / size_y - 0.5, float(j) / size_x - 0.5])

colors = colors.reshape(size_x * size_y, 2)

#img = net(torch.tensor(colors).type(torch.FloatTensor)).detach().numpy()

img = net(torch.tensor(colors).type(torch.cuda.FloatTensor)).cuda()

img2 = img.cpu().detach().numpy()

return img2.reshape(size_x, size_y, 3)

def plot_colors(colors, fig_size=15):

plt.figure(figsize=(fig_size, fig_size))

plt.imshow(colors, interpolation='nearest', vmin=0, vmax=1)

def save_colors(colors):

plt.imsave(str(np.random.randint(100000)) + ".png", colors)

def run_plot_save(net, size_x, size_y, fig_size=15):

colors = run_net(net, size_x, size_y)

plot_colors(colors, fig_size)

save_colors(colors)

In[4]:

n,c = gen_new_image(1024, 1024, save=False, num_neurons=32)

In[5]:

run_plot_save(n, 1080, 720)

Let's see how the images change if we increase the depth

In[57]:

for num_layers in range(2, 30, 3):

print(f"{num_layers} layers")

n,c = gen_new_image(128, 128, save=False, num_layers=num_layers)

And also the effect of increasing the width

In[58]:

for i in range(1, 10, 2):

print(f"{num_layers} layers")

n,c = gen_new_image(128, 128, save=False, num_neurons=2**i)

What happens if we use ReLUs?

In[60]:

n,c = gen_new_image(128, 128, save=False, activation=nn.ReLU)

In[ ]:

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/paraschopra/abstract-art-neural-network/issues/4, or mute the thread https://github.com/notifications/unsubscribe-auth/AAugqYYo2rCY3LRB4xIcxEwE2y4ja_94ks5vX7TegaJpZM4b6HXb .