def testIt():
allAccuracy = []
allWrongs = []
allWrongNames = []
for batch_idx, (data, target, batchNames) in enumerate(
dataLoaderCuda(tensors,
batchSize,
randomAxis=True,
shuffle=False,
names=names)):
data_temp = np.copy(data)
data, target = Variable(data), Variable(target)
data, target = data.cuda(), target.cuda()
#data = data.view(-1, 24*24*24*6)
data = data.permute(0, 4, 1, 2, 3)
net_out = net(data)
prediction = net_out.max(1)[1]
selector = (prediction != target).cpu().numpy().astype(np.bool)
wrongs = data_temp[selector]
wrongNames = [batchNames[i] for i, j in enumerate(selector) if j]
for wrong, n in zip(wrongs, wrongNames):
pass
# print(n)
allAccuracy.extend(list((prediction == target).cpu().numpy()))
allWrongs.extend(wrongs)
allWrongNames.extend(wrongNames)
#print(f"Test accuracy is {np.mean(allAccuracy)}")
print(f"Test accuracy is {np.mean(allAccuracy)}")
print("Wrong tensors:")
for n in allWrongNames:
print(n)
epochs = 1000
log_interval = 1
batchSize = 20
optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
optimizer = optim.Adam(net.parameters(), lr=learning_rate)
# criterion = nn.BCELoss()
# criterion = nn.MultiLabelSoftMarginLoss()
criterion = nn.NLLLoss()
test = tensors[:int(len(tensors)*0.2)]
train = tensors[int(len(tensors)*0.2):]
# Train the network
for epoch in range(epochs):
epochAcc = []
for batch_idx, (data, target, batchNames) in enumerate(dataLoaderCuda(train, batchSize, randomAxis=True, shuffle=False, names=names)):
data, target = Variable(data), Variable(target)
data, target = data, target.cuda()
data = data.view(-1, 24*24*24*6)
optimizer.zero_grad()
net_out = net(data)
prediction = net_out.max(1)[1]
loss = criterion(net_out, target)
#print(f"Loss is {loss}, accuracy is {np.mean((prediction == target).numpy())}")
loss.backward()
optimizer.step()
currAcc = (prediction == target).cpu().numpy()
if not sum(currAcc) == len(target):
print(prediction.cpu().numpy()[currAcc == 0])
print(target.cpu().numpy()[currAcc==0])
print('\n'.join([batchNames[i] for i in np.where(currAcc == 0)[0]]))
def trainTest(epochs):
for epoch in range(epochs):
epochAcc = []
for batch_idx, (data, target, batchNames) in enumerate(
dataLoaderCuda(train,
batchSize,
randomAxis=True,
shuffle=False,
names=names)):
data, target = Variable(data), Variable(target)
data, target = data, target.cuda()
#data = data.view(-1, 24*24*24*6)
data = data.permute(0, 4, 1, 2, 3)
optimizer.zero_grad()
net_out = net(data)
prediction = net_out.max(1)[1]
loss = criterion(net_out, target)
#print(f"Loss is {loss}, accuracy is {np.mean((prediction == target).numpy())}")
loss.backward()
optimizer.step()
currAcc = (prediction == target).cpu().numpy()
if not sum(currAcc) == len(target):
print(prediction.cpu().numpy()[currAcc == 0])
print(target.cpu().numpy()[currAcc == 0])
print('\n'.join(
[batchNames[i] for i in np.where(currAcc == 0)[0]]))
epochAcc.extend(list(currAcc))
# if batch_idx % log_interval == 0:
print(
f"\nTrain epoch: {epoch}, loss is {loss.data.item()}, accuracy is {np.mean(epochAcc)}\n"
)
allAccuracy = []
allWrongs = []
allWrongNames = []
for batch_idx, (data, target, batchNames) in enumerate(
dataLoaderCuda(test,
batchSize,
randomAxis=True,
shuffle=False,
names=names)):
data_temp = np.copy(data)
data, target = Variable(data), Variable(target)
data, target = data.cuda(), target.cuda()
#data = data.view(-1, 24*24*24*6)
data = data.permute(0, 4, 1, 2, 3)
net_out = net(data)
prediction = net_out.max(1)[1]
selector = (prediction != target).cpu().numpy().astype(np.bool)
wrongs = data_temp[selector]
wrongNames = [batchNames[i] for i, j in enumerate(selector) if j]
for wrong, n in zip(wrongs, wrongNames):
pass
# print(n)
allAccuracy.extend(list((prediction == target).cpu().numpy()))
allWrongs.extend(wrongs)
allWrongNames.extend(wrongNames)
#print(f"Test accuracy is {np.mean(allAccuracy)}")
print(f"Test accuracy is {np.mean(allAccuracy)}")
print("Wrong tensors:")
for n in allWrongNames:
print(n)
optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
optimizer = optim.Adam(net.parameters(), lr=learning_rate)
# criterion = nn.BCELoss()
# criterion = nn.MultiLabelSoftMarginLoss()
criterion = nn.NLLLoss()
test = tensors[:int(len(tensors) * 0.2)]
train = tensors[int(len(tensors) * 0.2):]
# Train the network
for epoch in range(epochs):
epochAcc = []
for batch_idx, (data, target, batchNames) in enumerate(
dataLoaderCuda(train,
batchSize,
randomAxis=True,
shuffle=True,
names=names,
maxShift=6)):
data, target = Variable(data), Variable(target)
data, target = data, target.cuda()
#data = data.view(-1, 24*24*24*6)
data = data.permute(0, 4, 1, 2, 3)
optimizer.zero_grad()
net_out = net(data)
prediction = net_out.max(1)[1]
loss = criterion(net_out, target)
#print(f"Loss is {loss}, accuracy is {np.mean((prediction == target).numpy())}")
loss.backward()
optimizer.step()
currAcc = (prediction == target).cpu().numpy()
if not sum(currAcc) == len(target):
optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
optimizer = optim.Adam(net.parameters(), lr=learning_rate)
# criterion = nn.BCELoss()
# criterion = nn.MultiLabelSoftMarginLoss()
criterion = nn.NLLLoss()
test = tensors[:int(len(tensors) * 0.2)]
train = tensors[int(len(tensors) * 0.2):]
# Train the network
for epoch in range(epochs):
epochAcc = []
for batch_idx, (data, target, batchNames) in enumerate(
dataLoaderCuda(train,
batchSize,
randomAxis=True,
shuffle=False,
names=names)):
data, target = Variable(data), Variable(target)
data, target = data, target.cuda()
#data = data.view(-1, 24*24*24*6)
data = data.permute(0, 4, 1, 2, 3)
optimizer.zero_grad()
net_out = net(data)
prediction = net_out.max(1)[1]
loss = criterion(net_out, target)
#print(f"Loss is {loss}, accuracy is {np.mean((prediction == target).numpy())}")
loss.backward()
optimizer.step()
currAcc = (prediction == target).cpu().numpy()
if not sum(currAcc) == len(target):