clara/pytorch-ai
1
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions

Implemented batching for TicTacToe AI

Browse source
This commit is contained in:
Clemens Dautermann 2020-01-28 14:45:00 +01:00
parent 55cff9b18f
commit 56ee2635b5
96 changed files with 8426 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

136
TicTacToe_AI/Net/wandb/run-20200128_133032-5hk954l0/diff.patch Normal file
View file

@ -0,0 +1,136 @@
diff --git a/TicTacToe_AI/Net/pytorch_ai.py b/TicTacToe_AI/Net/pytorch_ai.py
index efea5ae..f3bb291 100644
--- a/TicTacToe_AI/Net/pytorch_ai.py
+++ b/TicTacToe_AI/Net/pytorch_ai.py
@@ -4,6 +4,11 @@ import torch.optim as optim
from torch import nn
import torch.nn.functional as F
from tqdm import tqdm
+import wandb
+
+wandb.init(project="tictactoe")
+
+BATCH_SIZE = 3
def to_set(raw_list):
@@ -35,6 +40,40 @@ def to_set(raw_list):
return out_set
+def to_batched_set(raw_list):
+ counter = 0
+ out_set = []
+ boardtensor = torch.zeros((BATCH_SIZE, 1, 9))
+ labeltensor = torch.zeros(BATCH_SIZE, dtype=torch.long)
+ for line in tqdm(raw_list):
+ line = line.replace('\n', '')
+ raw_board, raw_label = line.split('|')[0], line.split('|')[1]
+
+ if not (int(raw_label) is -1):
+ labeltensor[counter] = int(raw_label)
+ else:
+ labeltensor[counter] = 9
+
+ raw_board = raw_board.split(',')
+ for n, block in enumerate(raw_board):
+ if int(block) is -1:
+ boardtensor[counter][0][n] = 0
+ elif int(block) is 0:
+ boardtensor[counter][0][n] = 0.5
+ elif int(block) is 1:
+ boardtensor[counter][0][n] = 1
+
+ if counter == (BATCH_SIZE - 1):
+ out_set.append([boardtensor, labeltensor])
+ boardtensor = torch.zeros((BATCH_SIZE, 1, 9))
+ labeltensor = torch.zeros(BATCH_SIZE, dtype=torch.long)
+ counter = 0
+ else:
+ counter += 1
+
+ return out_set
+
+
def buildsets():
with open('boards.bds', 'r') as infile:
print('Loading file...')
@@ -43,10 +82,10 @@ def buildsets():
random.shuffle(alllines)
print('Generating testset...')
- testset = to_set(alllines[0:10000])
+ testset = to_batched_set(alllines[0:10000])
print('Generating trainset...')
- trainset = to_set(alllines[10001:200000])
+ trainset = to_batched_set(alllines[10001:20000])
return trainset, testset
@@ -60,6 +99,7 @@ def testnet(net, testset):
if torch.argmax(output) == label[0]:
correct += 1
total += 1
+ wandb.log({'test_accuracy': correct / total})
print("Accuracy: ", round(correct / total, 3))
@@ -79,21 +119,35 @@ class Net(torch.nn.Module):
return F.log_softmax(x, dim=1)
-net = torch.load('./nets/net_3.pt')
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+print('running on %s' % device)
+
+# net = torch.load('./nets/net_3.pt')
+
+net = Net()
+wandb.watch(net)
+
+net.to(device)
optimizer = optim.Adam(net.parameters(), lr=0.001)
+loss_function = nn.CrossEntropyLoss()
trainset, testset = buildsets()
for epoch in range(100):
print('Epoch: ' + str(epoch))
+ wandb.log({'epoch': epoch})
for X, label in tqdm(trainset):
net.zero_grad()
+ X.to(device)
output = net(X)
- loss = F.nll_loss(output.view(1, 10), label[0])
+ output.cpu()
+ print(output)
+ print(label) # LABEL NEEDS TO BE ONE HOT VECTOR!!!!
+ loss = loss_function(output.view(-1, 10), label)
loss.backward()
optimizer.step()
+ wandb.log({'loss': loss})
- print(loss)
- torch.save(net, './nets/net_' + str(epoch + 3) + '.pt')
+ torch.save(net, './nets/gpunets/net_' + str(epoch) + '.pt')
testnet(net, testset)
diff --git a/other_scripts/setcounter.py b/other_scripts/setcounter.py
index 9735f20..e9eb00c 100644
--- a/other_scripts/setcounter.py
+++ b/other_scripts/setcounter.py
@@ -7,9 +7,12 @@ data = datasets.MNIST('../datasets', train=True, download=True,
transforms.ToTensor()
]))
-loader = torch.utils.data.DataLoader(data, batch_size=1, shuffle=False)
+loader = torch.utils.data.DataLoader(data, batch_size=15, shuffle=False)
set = {'0': 0, '1': 0, '2': 0, '3': 0, '4': 0, '5': 0, '6': 0, '7': 0, '8': 0, '9': 0}
+for data in loader:
+ print(data[1].shape)
+
for _, label in tqdm(loader):
set[str(label[0].item())] += 1