def train_mse_accuracy(self):
all_mse = []
correct = 0
total = 0
for x, y in self.train_test_loader:
y = y.view(-1, 1, 1)
N = x.shape[0]
dim = 1
for d in x.shape[1:]:
dim *= d
x_transform = x.view(N, 1, dim)
y_hat = self.forward(x_transform)
y_hat_decoded = torch.argmax(y_hat, dim=2, keepdim=True)
correct += torch.sum(y_hat_decoded == y).item()
total += N
y_hat = y_hat.view(N, -1)
y_onehot = oneHotEncodeOneCol(y.view(-1, 1), 10)
all_mse.append(y_mse(y_onehot, y_hat))
mse = sum(all_mse) / len(all_mse)
return (mse, correct / total)
def train(self, epochs=10):
optimizer = optim.SGD(self.parameters(), lr=0.1)
criterion = nn.MSELoss()
it = 0
i = 0
train_mse = []
train_accuracy = []
test_mse = []
test_accuracy = []
r = []
while it < epochs:
for x, y in self.train_loader:
N = x.shape[0]
y_onehot = oneHotEncodeOneCol(y.view(-1, 1), 10).view(N, 1, -1)
dim = 1
for d in x.shape[1:]:
dim *= d
x_transform = x.view(N, 1, dim)
optimizer.zero_grad()
y_hat = self.forward(x_transform)
loss = criterion(y_hat, y_onehot)
loss.backward()
# put in closure as stopping point
optimizer.step(closure=None)
if (i % 1000 == 0):
print(i)
tm, ta = self.test_mse_accuracy()
# trm, tra = self.train_mse_accuracy()
trm = y_mse(y_onehot.view(N, -1), y_hat.view(N, -1))
tra = 1
test_mse.append(tm)
test_accuracy.append(ta)
train_mse.append(trm)
train_accuracy.append(tra)
r.append(i / self.train_size)
i += 1
it += 1
if (i % 1000 == 0):
print(i)
tm, ta = self.test_mse_accuracy()
trm, tra = self.train_mse_accuracy()
# trm = y_mse(y_onehot.view(N, -1), y_hat.view(N, -1))
# tra = 1
test_mse.append(tm)
test_accuracy.append(ta)
train_mse.append(trm)
train_accuracy.append(tra)
r.append(i / self.train_size)
return ((train_mse, train_accuracy), (test_mse, test_accuracy), r)
def train(self, epochs=10):
it = 0
i = 0
train_mse = []
train_accuracy = []
test_mse = []
test_accuracy = []
r = []
while it < epochs:
# print(f'epoch {it + 1}/{epochs}')
for x, y in self.train_loader:
# 10 for the number of digits in MNIST
y_onehot = oneHotEncodeOneCol(y.view(-1, 1), 10)
dim = 1
for d in x.shape[1:]:
dim *= d
x_transform = x.view(x.shape[0], dim)
y_hat = self.nn.forward(x_transform)
self.nn.backward(y_onehot)
self.nn.updateParams()
if (i % 1000 == 0):
print(i)
tm, ta = self.test_mse_accuracy()
# trm, tra = self.train_mse_accuracy()
trm = y_mse(y_onehot, y_hat)
tra = 1
test_mse.append(tm)
test_accuracy.append(ta)
train_mse.append(trm)
train_accuracy.append(tra)
r.append(i / self.train_size)
i += 1
it += 1
if (i % 1000 == 0):
tm, ta = self.test_mse_accuracy()
trm, tra = self.train_mse_accuracy()
# trm = y_mse(y_onehot, y_hat)
# tra = 1
test_mse.append(tm)
test_accuracy.append(ta)
train_mse.append(trm)
train_accuracy.append(tra)
r.append(i / self.train_size)
return ((train_mse, train_accuracy), (test_mse, test_accuracy), r)
def train_mse(self):
all_mse = []
for x, y in self.train_test_loader:
N = x.shape[0]
dim = 1
for d in x.shape[1:]:
dim *= d
x_transform = x.view(N, 1, dim)
y_hat = self.forward(x_transform).view(N, -1)
y_onehot = oneHotEncodeOneCol(y.view(-1, 1), 10)
all_mse.append(y_mse(y_onehot, y_hat))
mse = sum(all_mse) / len(all_mse)
return mse