I have just got a new computer running Windows 10 which has a GPU so I wanted to see if I could sensibly use it for machine learning. So I tried running an old model which I previously trained on Google Colab. The answer is that it does do quite well, but I discovered that I could not use more than one worker in the Dataloader. Googling found that this is a known issue with PyTorch on Windows in Jupyter Notebooks so I tried running it in a normal Python program. I found that it did work but that the creation of the DataIterator took a very long time. Below are the times in seconds for 1, 2 and 6 workers, each done twice:
I note that 2 workers seems to be the fastest and there seems to be quite a lot of variation which surprised me as the machine was doing nothing else.
So the first question is:
- Is there a way to let PyTorch choose the most efficient number of workers to use?
The second question is:
- If I install a version of Linux will I be able to use Jupyter Notebooks with multiple workers which is what I would prefer to do in an ideal world.
The code I ran is below the relevant part is after if __name__ == "__main__":
# -*- coding: utf-8 -*-
import torch
import torchvision
import torchvision.transforms as transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset,DataLoader
import os
import numpy as np
#import gym
import pickle
import matplotlib.pyplot as plt
import time
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, in_planes, planes, stride=1):
super(BasicBlock, self).__init__()
self.conv1 = nn.Conv2d(
in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = F.relu(out)
return out
# create dataset
class C4Dataset(Dataset):
'''
the data for the first 12moves is held in a pickles list
as (key,val)
The key has to be converted to the pos amd mask which can then be converted to the ones,twos and zeros
Val is the value for the player playing so needs to be changed for wnen moves i odd to minus val
'''
fileName = r'C:\Users\alan\Desktop\Python\Python36\connect4\Layers\ListAllKeyVal19'
bottom_mask = 4432676798593
board_mask = bottom_mask * ((1 << 6) - 1)
bNos = 2**np.arange(49, dtype = np.uint64)
def getData(fileName):
with open(fileName,'rb') as inFile:
dict = pickle.load(inFile)
return dict
def oneHot(x):
return np.eye(37,dtype = np.float32)[x]
def getNoMoves(ones,twos) :
return np.sum(ones+twos)
def getPosMask(key):
binary = ('{:049b}'.format(key))[::-1]
arr = np.frombuffer(binary.encode(encoding='utf-8', errors='strict'),'u1') - ord('0')
outArr = np.reshape(arr,(7,7),order = 'F')
arr = np.flipud(outArr)
pos = arr.copy()
mask =arr.copy()
for col in range(7):
res = np.where(arr[:,col]==1)
topPos = res[0][0]
pos[topPos,col] = 0
mask[topPos,col] = 0
if topPos<6:
mask[topPos+1:,col] = 1
msk = np.flipud(mask)
msk = np.reshape(msk,(49),order = 'F')
maskNo = np.array(msk.dot(C4Dataset.bNos),dtype = np.uint64).item()
return pos.astype('float32'),(pos ^ mask).astype('float32'),(np.logical_not(mask)).astype('float32'),maskNo
def possible(mask) :
poss = (mask + C4Dataset.bottom_mask) & C4Dataset.board_mask
binary = ('{:049b}'.format(poss))[::-1]
arr = np.frombuffer(binary.encode(encoding='utf-8', errors='strict'),'u1') - ord('0')
outArr = np.reshape(arr,(7,7),order = 'F')
arr = np.flipud(outArr)
return arr
def __init__(self):
self.lst = C4Dataset.getData(C4Dataset.fileName)
def __len__(self):
return len(self.lst)
def __getitem__(self, idx):
key,val = self.lst[idx]
val = int(val)
ones,twos,zeros,mask = C4Dataset.getPosMask(key)
arr = np.zeros((5,7,7),dtype = np.float32)
arr[0,:6,:7] = ones[1:,:]
arr[1,:6,:7] = twos[1:,:]
arr[2,:6,:7] = zeros[1:,:]
moves = int(C4Dataset.getNoMoves(ones,twos))
p = (moves % 2) + 3
arr[p,:6,:7] = C4Dataset.possible(mask)[1:,:]
return arr,val+18 #C4Dataset.oneHot(val+18)
class C4Net(nn.Module):
def __init__(self, inFilters,outFilters):
super(C4Net, self).__init__()
self.conv1 = nn.Conv2d(inFilters, 32, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(32)
self.layer1 = BasicBlock(32,32)
self.layer2 = BasicBlock(32,32)
self.layer3 = BasicBlock(32,32)
self.layer4 = BasicBlock(32,32)
self.layer5 = BasicBlock(32,32)
self.layer6 = BasicBlock(32,32)
self.layer7 = BasicBlock(32,32)
self.layer8 = BasicBlock(32,32)
self.linear = nn.Linear(32*7*7,outFilters)#1568
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = self.layer5(out)
out = self.layer6(out)
out = self.layer7(out)
out = self.layer8(out)
#out = F.avg_pool2d(out, 2)
out = out.view(out.size(0), -1)
out = self.linear(out)
return out
# show some images
def show(img):
npimg = img.numpy()[:3,:,:]
plt.imshow(np.transpose(npimg, (1,2,0)), interpolation='nearest')
# get some random training images
if __name__ == "__main__":
dirName =r'C:\Users\alan\Desktop\Python\Python36\connect4\Layers'
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Assuming that we are on a CUDA machine, this should print a CUDA device:
print(device)
# create dataloader
max_epochs = 1
batchSize = 1024#512#256
learningRate = .00003
# Parameters
params = {'batch_size': batchSize,'shuffle': True,'num_workers': 2}
# Generators
dataset = C4Dataset()
start = time.time()
dataloader = DataLoader(dataset, **params)
middle = time.time()
print('create dataloader',middle-start)
dataiter = iter(dataloader)
end = time.time()
print('create data iterator',end-middle)
images, labels = dataiter.next()
final = time.time()
print('get one batch',final-end)
# show images
show(torchvision.utils.make_grid(images[:16]))
#create the weights
wts =np.array([59, 963, 12406, 148920, 62551, 47281, 55136, 54312, 44465, 31688,
27912, 37907, 114778, 242800, 394530, 495237, 582174, 163370, 480850,
201152, 690905, 633937, 721340, 372479, 193375, 84648, 76576, 91087, 130428,
154184, 157339, 156453, 227696, 1705325, 548155, 44315, 2082],dtype = np.float32)
maxwt = wts.max()
weights = wts/maxwt
weights = torch.from_numpy(weights)
weights.to(device)
# create the network
net = C4Net(5,37)
net.to(device)
PATH = r'C:\Users\alan\Desktop\Python\connectX\c4Net37Weights00003.pth'
net.load_state_dict(torch.load(PATH,map_location=torch.device(device)))
#create the loss function and optimiser
criterion = nn.CrossEntropyLoss(weight = weights.to(device) )
optimizer = optim.Adam(net.parameters(), lr=learningRate)
#train the network
import time
start = time.time()
for epoch in range(max_epochs): # loop over the dataset multiple times
running_loss = 0.0
for i, data in enumerate(dataloader, 0):
# get the inputs; data is a list of [inputs, labels]
inputs, labels = inputs, labels = data[0].to(device), data[1].to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
if i % 2000 == 1999: # print every 2000 mini-batches
print('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / 2000))
running_loss = 0.0
torch.save(net.state_dict(),r'C:\Users\alan\Desktop\Python\connectX\tempWeights')
print('Finished Training')
# save the weights
PATH = r'C:\Users\alan\Desktop\Python\connectX\c4Net37Weights00004.pth'
torch.save(net.state_dict(), PATH)
end = time.time()
print('elapsed time',end-start)
PS the machine is a Dell XPS 17 with Intel Core i9-10885H with 8 cores and the GPU is a NVIDIA GeForce RTX 2060 with Max-Q. In this one test it runs 4 times faster than on Google Colab but I do not know what GPU I was allocated.