def val(args, zoomout, model, val_loader, epoch):
# modified from https://github.com/wkentaro/pytorch-fcn/blob/master/examples/voc/evaluate.py
model.eval()
print("Validating...")
label_trues, label_preds = [], []
count = 0
for batch_idx, (data, target) in enumerate(val_loader):
data, target, im_viz, lbl_viz = data.float(), target.float(
), data.float(), target.to('cpu')
score = model(zoomout(data))
_, pred = torch.max(score, dim=1) # changed dim from 0 to 1
lbl_pred = pred.data.numpy().astype(np.int64)
lbl_true = target.data.numpy().astype(np.int64)
for _, lt, lp in zip(_, lbl_true, lbl_pred):
label_trues.append(lt)
label_preds.append(lp)
if (batch_idx % 10 == 0) and (epoch in [0, 1, 4, 9]):
count = count + 1
"""
Visualization of results on val dataset.
epoch 0: only with weights learned from FCClassifier
epoch 1: after 2 epochs of training DenseClassifier
epoch 4: after 5 epochs of training DenseClassifier
epoch 9: after 10 epochs of training DenseClassifier
"""
pred = score[0, :, :, :]
gt = lbl_viz[0, :, :].data.numpy().squeeze()
im = im_viz[0, :, :, :].data.numpy().squeeze()
im = np.swapaxes(im, 0, 2)
im = np.swapaxes(im, 0, 1)
_, pred_mx = torch.max(pred, 0)
pred = pred_mx.data.numpy().squeeze()
image = Image.fromarray(im.astype(np.uint8), mode='RGB')
image.save("./imgs/val/im_" + str(count) + "_" + str(epoch) +
"_.png")
visualize(
"./lbls/val/pred_" + str(count) + "_" + str(epoch) + ".png",
pred)
visualize(
"./lbls/val/gt_" + str(count) + "_" + str(epoch) + ".png", gt)
n_class = 21
metrics = label_accuracy_score(label_trues, label_preds, n_class=n_class)
metrics = np.array(metrics)
metrics *= 100
print('''\
Accuracy: {0}
Accuracy Class: {1}
Mean IU: {2}
FWAV Accuracy: {3}'''.format(*metrics))
def train(model, train_loader, val_loader, epoch, num_epochs, loss_function, optimiser, scheduler, savename, highest_iou):
model.train()
losses = list()
gpu1 = 'cuda:0'
gpu2 = 'cuda:1'
ious = list()
max_iou = highest_iou
count = 0
savename2 = savename[ : -3] + '_opt.pt'
loop = tqdm(train_loader)
num_steps = len(loop)
for data, target in loop:
model.train()
model = model.to(gpu1)
data, target = data.float().to(gpu1), target.float().to(gpu1)
optimiser.zero_grad()
prediction = model(data)
prediction = prediction.squeeze(1)
loss = loss_function(prediction, target) + dice_loss(torch.sigmoid(prediction), target)
losses.append(loss.item())
loss.backward()
optimiser.step()
scheduler.step()
loop.set_description('Epoch {}/{}'.format(epoch + 1, num_epochs))
loop.set_postfix(loss = loss.item())
count += 1
if count % (num_steps // 3) == 0 :
model.eval()
for data, target in val_loader :
model = model.to(gpu1)
data, target = data.float().to(gpu1), target.float()
prediction = model(data)
prediction = prediction.squeeze(1)
ious.append(iou(target, prediction))
avg_iou = sum(ious) / len(ious)
if avg_iou > max_iou :
max_iou = avg_iou
torch.save(model.state_dict(), savename)
torch.save(optimiser.state_dict(), savename2)
print('new max_iou', max_iou)
print('avg_iou: ', avg_iou)
print('avg_loss: ', sum(losses) / len(losses))
return max_iou
def __getitem__(self, index):
sample = self.imgs[index]
if self.phase == 'train':
splits = sample.split(' ')
img_path = splits[0]
data = Image.open(img_path).convert(self.img_mode)
data = self.transforms(data)
label = np.int32(splits[1])
return data.float(), label
else:
data = Image.open(sample).convert(self.img_mode)
data = self.transforms(data)
name = self.imgs_name[index][:-1]
return data.float(), name
def train(model, loader, epoch):
optimizer = optim.Adam(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=30, gamma=0.5)
scheduler.step()
model.train()
torch.set_grad_enabled(True)
correct = 0
dataset_size = 0
for batch_idx, (data, target) in enumerate(loader):
dataset_size += data.shape[0]
data, target = data.float(), target.long().squeeze()
if args.cuda:
data, target = data.cuda(), target.cuda()
optimizer.zero_grad()
output, _ = model(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
pred = output.data.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).long().cpu().sum()
if batch_idx % args.log_interval == 0:
percentage = 100. * batch_idx * args.batch_size / len(
loader.dataset)
print(
f'Train Epoch: {epoch} [{batch_idx * args.batch_size}/{len(loader.dataset)} ({percentage}%)]'
f'\tLoss: {loss.item()}\t{args.tag}')
logger.add_scalar('train_loss',
loss.cpu().item(),
batch_idx + epoch * len(loader))
return float(correct) / float(dataset_size)
def test(model, device, test_loader):
# Testing
model.eval()
test_loss = 0
accuracy = 0
balanced_accuracy = 0
for sample in test_loader:
data, target = sample['image'], sample['label']
data, target = data.to(device), target.to(device)
outputs = net(data.float())
#loss = criterion(outputs, torch.max(target, 1)[1])
loss = criterion(outputs, target.squeeze(1).long())
test_loss += loss.item() * data.size(0)
acc, bacc, precision, recall, f1_score, rep = \
METRIX(target.squeeze(1).long(), outputs)
accuracy += acc
balanced_accuracy += bacc
# ----------------------
print('Acc:\t{:.3f}%\tBalanced Acc.:\t{:.3f}%\tPrecision:\t{:.3f}%\t'
'Recall:\t{:.3f}%\tF1 Score:\t{:.3f}%\t'.format(
acc * 100, bacc * 100, precision * 100, recall * 100,
f1_score * 100))
print('Report: \n', rep)
# ----------------------
return test_loss, accuracy, balanced_accuracy
def fit(self, path, mask):
if self.feature_length == None:
self.feature_length = len(self.completeData[0])
self.h_length = self.feature_length // 2
dataloader = torch.utils.data.DataLoader(self.completeData,
batch_size=self.batch_size,
shuffle=True,
num_workers=2)
encoder = Encoder(self.feature_length, self.h_length)
decoder = Decoder(self.feature_length, self.h_length)
ae = AE(encoder, decoder)
criterion = torch.nn.MSELoss()
optimizer = torch.optim.Adam(ae.parameters(), lr=self.learning_rate)
l = None
bestLoss = np.inf
for epoch in range(self.epochs):
for data in dataloader:
inputs = data.float()
inputs = inputs.resize_(self.batch_size, self.feature_length)
if bi['core'] == 'cuda':
inputs = inputs.cuda()
optimizer.zero_grad()
dec = ae(inputs)
loss = criterion(dec[mask], inputs[mask])
loss.backward()
optimizer.step()
l = loss.item()
print(epoch, l)
if np.isnan(l):
break
if epoch % 5 == 0:
torch.save(ae.state_dict(), path)
logger.info("{}'s loss is {}".format(path, l))
print("模型保存成功")
def __getitem__(self,index):
qid=self.val[index+self.len_end+opt.question_train_set_items-50000]
data=self.data[index]
label=self.label_file[qid]
data = data.float()
label_tensor = t.zeros(25556).scatter_(0,t.LongTensor(label),1).long()
return data,label_tensor
def val(args, zoomout, model, val_loader):
# modified from https://github.com/wkentaro/pytorch-fcn/blob/master/examples/voc/evaluate.py
USE_GPU = True
dtype = torch.float32 # we will be using float throughout this tutorial
if USE_GPU and torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
model.eval()
print("Validating...")
label_trues, label_preds = [], []
for batch_idx, (data, target) in enumerate(val_loader):
data, target = data.float(), target.float()
score = model(zoomout(data))
score = score.squeeze()
_, pred = torch.max(score, 0)
lbl_pred = pred.data.numpy().astype(np.int64)
lbl_true = target.data.numpy().astype(np.int64)
lbl_true = lbl_true.squeeze()
for _, lt, lp in zip(_, lbl_true, lbl_pred):
label_trues.append(lt)
label_preds.append(lp)
n_class = 21
metrics = label_accuracy_score(label_trues, label_preds, n_class=n_class)
metrics = np.array(metrics)
metrics *= 100
print('''\
Accuracy: {0}
Accuracy Class: {1}
Mean IU: {2}
FWAV Accuracy: {3}'''.format(*metrics))
def train_agent(self, epochs: int):
if not self.dataset:
raise Exception(
"No Training Data Set to Train Agent. Please set Training Data using ImmitationLearningAgent.injest_demonstrations"
)
self.logger.info("Starting Training of Agent ")
self.neural_network.train()
for epoch in range(epochs):
running_loss = 0.0
steps = 0
for batch_idx, (data, target) in enumerate(self.data_loader):
data, target = Variable(data), Variable(target)
data, target = data.to(device), target.to(device)
self.optimizer.zero_grad()
network_pred = self.neural_network(data.float()).to(device)
loss = self.loss_function(network_pred, target.float())
loss.backward()
if self.collect_gradients:
self.set_gradients(self.neural_network.named_parameters())
self.optimizer.step()
running_loss += loss.item()
steps += 1
self.logger.info('[%d] loss: %.6f' % (epoch + 1, running_loss /
(steps + 1)))
def __getitem__(self,index):
qid=self.val[index+self.len_end+2999967-200000]
data=self.data[index]
label=self.label_file[qid]
data = data.float()
label_tensor = t.zeros(1999).scatter_(0,t.LongTensor(label),1).long()
return data,label_tensor
def train(epoch, denoise=False):
model.train()
train_loss = 0
for batch_idx, data in enumerate(train_loader):
data = data.float().to(device)
# import pdb; pdb.set_trace()
optimizer.zero_grad()
if denoise:
noise = torch.bernoulli((torch.rand_like(data))).to(device)
noisy_data = data + noise
reconBatch, mu = model(noisy_data)
else:
reconBatch, mu = model(data)
loss = loss_function(data, reconBatch)
loss.backward()
train_loss += loss.item()
optimizer.step()
if(batch_idx % log_interval == 0):
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader),
loss.item() / len(data)))
train_loss /= len(train_loader.dataset)
print('====> Epoch: {} Average loss: {:.4f}'.format(
epoch, train_loss))
return train_loss
def train(epoch, criterion, optimizer):
model.train()
train_loss = 0
correct = 0
running_loss = 0.0
running_corrects = 0
for data, target in tqdm.tqdm(dataloaders['train'],
total=len(dataloaders['train']),
desc='Batch'):
data = data.to(device)
data = torch.squeeze(data)
data = data.float()
target = target.to(target)
if args.cuda:
data, target = data.cuda(), target.cuda()
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
train_loss += loss.data
sm = nn.Softmax(dim=1)
output_sm = sm(output)
_, preds = torch.max(output_sm, 1)
running_loss += loss.item() * data.size(0)
running_corrects += torch.sum(preds == target.data)
loss.backward()
optimizer.step()
epoch_loss = running_loss / dataset_sizes['train']
epoch_acc = running_corrects.double() / dataset_sizes['train']
line_to_save_train = 'Train set: Average loss: {:.4f} Accuracy: {}/{} {:.4f}\n'.format(
epoch_loss, running_corrects, len(train_loader.dataset), epoch_acc)
with open(args.outf + '/ACC_train.txt', 'a') as f:
f.write(line_to_save_train)
print(line_to_save_train)
def train(epoch):
model.train()
trainLoss = 0
cos_sims = 0
klds = 0
loss_divider = len(
train_loader.dataset) - (len(train_loader.dataset) % batch_size)
for batch_idx, data in enumerate(train_loader):
data = data.float().to(device)
optimizer.zero_grad()
reconBatch, mu, logvar = model(data)
loss, cos_sim, kld = loss_function(reconBatch, data, mu, logvar)
loss.backward()
trainLoss += loss.item()
cos_sims += cos_sim.item()
klds += kld.item()
optimizer.step()
weights = []
for i, f in enumerate(model.parameters()):
weights.append(f.cpu().data.numpy())
if (batch_idx % log_interval == 0):
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader),
loss.item() / len(data)))
print('====> Epoch: {} Average Loss: {:.4f}'.format(
epoch, trainLoss / loss_divider))
return trainLoss / loss_divider, cos_sims / loss_divider, \
klds / loss_divider, weights, mu
def one_epoch(self, mode, epoch_num):
if mode not in ['train', 'test']:
raise ValueError("Unknown value {} for mode".format(mode))
print("{}ing... epoch: {}".format(mode, epoch_num))
if mode == 'train':
self.model.train()
dl = self.train_data
one_iter_function = self.one_train_iteration
else:
self.model.eval()
dl = self.test_data
one_iter_function = self.one_test_iteration
acc_avg = RunningAverage()
loss_avg = RunningAverage()
with tqdm(total=len(dl)) as t:
for n, (data, label) in enumerate(dl):
if self.train_params['use_gpu']:
data, label = data.cuda(
self.train_params['gpu_id']), label.cuda(
self.train_params['gpu_id'])
data, label = Variable(data), Variable(label)
data = data.float()
loss, acc = one_iter_function(data, label)
loss_avg.update(loss)
acc_avg.update(acc)
t.set_postfix(
run_param="Epoch{} Loss:{:.2f} Acc:{:.2f}".format(
epoch_num, loss_avg(), acc_avg()))
t.update()
return acc_avg, loss_avg
def test(epoch):
model.eval()
test_loss = 0
for i, (data, _) in enumerate(test_loader):
# for i, data in enumerate(test_loader):
data = data.float()
# for x in data:
# scaler.partial_fit(x[0]) # 0 because there is only one dimension
# # # # normalizes the data
# for x in data:
# x[0] = torch.from_numpy(scaler.transform(x[0])) # new
data = (data - torch.mean(data)) / torch.std(data)
if args.cuda:
data = data.cuda()
data = Variable(data, volatile=True)
output = model(data)
loss = MSE(output, data)
test_loss += loss.data[0]
# if i % 100 == 0:
# n = min(data.size(0), 8)
# comparison = torch.cat([data[:n],
# output[:n]])
# save_image(comparison.data.cpu(),
# 'snapshots/conv_vae/reconstruction_' + str(epoch) +
# '.png', nrow=n)
test_loss /= len(test_loader.dataset)
print('====> Test set loss: {:.4f}'.format(test_loss))
def train(model, loader, epoch):
scheduler.step()
model.train()
torch.set_grad_enabled(True)
correct = 0
dataset_size = 0
for batch_idx, (data, target) in enumerate(loader):
dataset_size += data.shape[0]
data, target = data.float(), target.long().squeeze()
if args.cuda:
data, target = data.cuda(), target.cuda()
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
pred = output.data.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).long().cpu().sum()
if batch_idx % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\t{}'.format(
epoch, batch_idx * len(data), len(loader.dataset),
100. * batch_idx * len(data) / len(loader.dataset),
loss.item(), args.tag))
logger.add_scalar('train_loss',
loss.cpu().item(),
batch_idx + epoch * len(loader))
return float(correct) / float(dataset_size)
def __getitem__(self, index):
img_path = self.image_list[index]
img = Image.open(img_path)
data = img.convert('RGB')
data = self.transforms(data)
label = self.label_list[index]
return data.float(), label
def test(model, loader):
model.eval()
torch.set_grad_enabled(False)
test_loss = 0
correct = 0
dataset_size = 0
da = {}
db = {}
res = []
for batch_idx, (data, target, obj_name) in enumerate(loader):
dataset_size += data.shape[0]
data, target = data.float(), target.long().squeeze()
if args.cuda:
data, target = data.cuda(), target.cuda()
output = model(data) # N*C
test_loss += F.nll_loss(output, target,
size_average=False).cpu().item()
pred = output.data.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).long().cpu().sum()
for i, j, k in zip(obj_name,
pred.data.cpu().numpy(),
target.data.cpu().numpy()):
res.append((i, j[0], k))
test_loss /= len(loader.dataset)
acc = float(correct) / float(dataset_size)
return acc, test_loss
def class_probabilities_test(model, device, test_loader, class_num):
model.eval()
scores = np.empty((0, class_num))
y_target = np.empty((0, class_num))
test_correct = 0
test_total = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
outputs = model(data.float())
_, predicted = torch.max(outputs.data, 1)
test_total += target.size(0)
test_correct += (predicted == target).sum()
# print(test_correct)
out_result = outputs.cpu().numpy()
out_target = target.cpu()
prediction_prob = out_result[:, 1]
prediction_binary = np.where(prediction_prob > 0.5, 1, 0)
acc_mlp = accuracy_score(out_target, prediction_binary)
# print(acc_mlp)
scores = np.append(scores, out_result, axis=0)
out_target = out_target.tolist()
# print(out_target)
labels_onehot = to_onehot(out_target)
y_target = np.append(y_target, labels_onehot, axis=0)
test_acc = test_correct.item() / test_total
# print("Test Accuracy: %.4f " % (test_acc))
return scores, y_target, test_acc
def eval_accuracy(net_path, name):
path = 'UCRArchive_2018/' + name + '/' + name + '_TEST.tsv'
net = torch.load(net_path)
net.eval()
test_set, n_class = load_ucr(path)
# x_test = test_set[:, 1:]
# x_test = torch.from_numpy(x_test).unsqueeze(dim=-1).float()
# print(x_test.shape)
# y_true = test_set[:, 0]
# y_pred = net(x_test)
# y_pred = torch.argmax(y_pred, dim=-1)
# print(y_pred)
# y_pred = y_pred.detach().numpy()
dataset = UcrDataset(test_set, channel_last=opt.channel_last)
dataloader = UCR_dataloader(dataset, batch_size=128)
y_pred = []
y_true = []
with torch.no_grad():
for i, (data, label) in enumerate(dataloader):
data = data.float().to(device)
y_ = net(data).cpu()
y_ = torch.argmax(y_, dim=1)
y_pred.append(y_.detach().numpy())
y_true.append(label.view(-1).long())
y_pred = np.concatenate(y_pred)
y_true = np.concatenate(y_true)
res = calculate_metrics(y_true, y_pred)
return res
def test_model(model, test_load, criterion, testout, batch_size):
with torch.no_grad():
# indicate that we are evaluating model
model.eval()
# initialize errors to 0
running_loss = 0.0
total_predictions = 0.0
correct_predictions = 0.0
# start iterating
for batch_idx, (data, target) in enumerate(test_load):
data = data.to(device)
target = target.to(device)
# run forward pass and then compute loss
outputs = model(data.float())
loss = criterion(outputs, target.long()).detach()
# get predictions
_, predicted = torch.max(outputs.data, 1)
# calculate correct predictions / loss
total_predictions += target.size(0)
correct_predictions += (predicted == target).sum().item()
running_loss += loss.item()
# calculate average loss and accuract
running_loss /= len(test_load)
acc = (correct_predictions/total_predictions)*100.0
return running_loss, acc
def train_epoch(model, train_load, criterion, optimizer):
# indicate that we are training
model.train()
running_loss = 0.0
# start timer and start iterating
start_train = time.time()
for batch_idx, (data, target) in enumerate(train_load):
data = data.to(device)
target = target.to(device) # all data & model on same device
# forward, then backward, then step
outputs = model(data.float())
loss = criterion(outputs, target.long())
loss.backward()
optimizer.step()
optimizer.zero_grad() # .backward() accumulates gradients
# accumulate loss
running_loss += loss.item()
# end timer and take average loss
end_train = time.time()
running_loss /= len(train_load)
return end_train, start_train, running_loss
def test(epoch, data_loader, test_set=False, valid_set=False):
model.eval()
loss = 0
cos_sim = 0
kld = 0
loss_divider = len(
data_loader.dataset) - (len(data_loader.dataset) % batch_size)
with torch.no_grad():
for i, data in enumerate(data_loader):
data = data.float().to(device)
reconBatch, mu, logvar = model(data)
temp_loss, cos_sim_temp, kld_temp = loss_function(
reconBatch, data, mu, logvar)
loss += temp_loss.item()
cos_sim += cos_sim_temp.item()
kld += kld_temp.item()
loss /= loss_divider
if test_set:
print('====> Test set loss: {:.4f}'.format(loss))
elif valid_set:
print('====> Validation set loss: {:.4f}'.format(loss))
return loss, cos_sim / loss_divider, kld / loss_divider
def val(args, zoomout, model, val_loader):
# modified from https://github.com/wkentaro/pytorch-fcn/blob/master/examples/voc/evaluate.py
model.eval()
print("Validating...")
label_trues, label_preds = [], []
for batch_idx, (data, target) in enumerate(val_loader):
data, target = data.float(), target.float()
score = model(zoomout(data))
_, pred = torch.max(score, 0)
lbl_pred = pred.data.numpy().astype(np.int64)
lbl_true = target.data.numpy().astype(np.int64)
for _, lt, lp in zip(_, lbl_true, lbl_pred):
label_trues.append(lt)
label_preds.append(lp)
n_class = 21
metrics = label_accuracy_score(label_trues, label_preds, n_class=n_class)
metrics = np.array(metrics)
metrics *= 100
print('''\
Accuracy: {0}
Accuracy Class: {1}
Mean IU: {2}
FWAV Accuracy: {3}'''.format(*metrics))
def test(epoch):
model.eval()
test_loss = 0
correct = 0
running_loss = 0.0
running_corrects = 0
for data, target in tqdm.tqdm(dataloaders['validation'],
total=len(dataloaders['validation']),
desc='Batch'):
data = data.to(device)
data = data.float()
data = torch.squeeze(data)
target = target.to(target)
if args.cuda:
data, target = data.cuda(), target.cuda()
output = model(data)
sm = nn.Softmax(dim=1)
output_sm = sm(output)
_, preds = torch.max(output_sm, 1)
loss = criterion(output, target)
running_loss += loss.item() * data.size(0)
running_corrects += torch.sum(preds == target.data)
epoch_loss = running_loss / dataset_sizes['validation']
epoch_acc = running_corrects.double() / dataset_sizes['validation']
line_to_save_test = 'Test set: Average loss: {:.4f} Accuracy: {}/{} {:.4f}\n'.format(
epoch_loss, running_corrects, dataset_sizes['validation'], epoch_acc)
with open(args.outf + '/ACC_test.txt', 'a') as f:
f.write(line_to_save_test)
print(line_to_save_test)
return epoch_loss, epoch_acc
def train(epoch):
model.train()
train_loss = 0
train_distance = 0
criterion = nn.MSELoss()
accuracy_criterion = nn.CosineSimilarity()
for batch_idx, data in enumerate(train_loader):
optimizer.zero_grad()
#float byte tensor
data = data.float().to(device)
data = data.view(-1, 1, 96, 60)
#embed data with autoencoder
with torch.no_grad():
mu, logvar = autoencoder_model.encoder(data)
#prepare for input lstm
mu = mu.view(model.batch_size, model.seq_length, 100)
embedding = mu.double()
# Normalize to mean 0 and std 1
# mean_batch = torch.mean(embedding)
# std_batch = torch.std(embedding)
# embedding_norm = (embedding - mean_batch) / std_batch
g_truth = embedding[:, -1, :]
input_lstm = embedding[:, :-1, :]
_, output_lstm = model(input_lstm)
loss = criterion(output_lstm, g_truth)
loss.backward()
train_loss += loss.item()
# torch.nn.utils.clip_grad_value_(model.parameters(), 5)
optimizer.step()
# with torch.no_grad():
# prediction = autoencoder_model.decoder(output_lstm.float().view(-1, model.input_size))
# prediction = prediction.view(-1, half_seq_length, 1, 96, 60)
# train_distance += np.linalg.norm(data.view(-1, model.seq_length, 1,
# 96, 60)[:,half_seq_length:].cpu().numpy() - prediction.cpu().numpy())
gradients = []
weights = []
for i, f in enumerate(model.parameters()):
gradients.append(f.grad.cpu().data.numpy())
weights.append(f.cpu().data.numpy())
if (batch_idx % log_interval == 0):
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data),
len(train_loader.dataset) * model.seq_length,
100. * batch_idx / len(train_loader), loss.item()))
# average train loss
train_loss /= (batch_idx + 1)
train_distance /= (batch_idx + 1)
print('====> Epoch: {} Average Loss: {:.4f}'.format(epoch, train_loss))
return train_loss, train_distance, gradients, weights
def get_batch(self, train_data, bptt, evaluation=False):
data, target = bptt
data = data.float().cuda()
target = target.long().cuda()
if evaluation:
data.requires_grad_(False)
target.requires_grad_(False)
return data, target
def transform_ToTensor(data):
# transform numpy array to float tensor
data = torch.from_numpy(np.array(data, dtype='uint8'))
if isinstance(data, torch.ByteTensor):
return data.float().div(255)
else:
return data
def __getitem__(self, index):
sample = self.imgs[index]
splits = sample.split()
img_path = splits[0]
data = Image.open(img_path)
data = data.convert('L')
data = self.transforms(data)
label = np.int32(splits[1])
return data.float(), label
def train(self):
# load data from file
with h5py.File(
'C:/Users/Kai/GitHub/Forever_Machine_Learning/code/data_950.h5',
'r') as hf:
data_x = hf['x'][:]
data_y = hf['y'][:]
# 80/20 split
# data_tr, data_te = data_obs.split()
train_ind = int(len(data_x) * 0.8)
data_tr = Dataset(data_x[:train_ind], data_y[:train_ind])
data_te = Dataset(data_x[train_ind:], data_y[train_ind:])
train_loader = DataLoader(data_tr,
batch_size=self.batch_size,
shuffle=self.shuffle)
train_losses = []
for epoch in range(self.num_epochs):
start_time1 = time.time()
epoch_loss = torch.tensor(0.0, device=self.device)
for batch_idx, batch in enumerate(train_loader):
start_time2 = time.time()
data, target = batch
outputs = self.model(data.float().to(self.device))
loss_gen = self.criterion(outputs,
target.float().to(self.device))
loss_L1 = self.criterionL1(outputs, target) * self.lambda_L1
loss = loss_gen + loss_L1
# Aggregate loss across mini-batches (per epoch)
epoch_loss += loss
# Backprop and perform Adam optimisation
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
print(
f"Epoch: {epoch}\tBatch: {batch_idx}\tGen Loss: {loss_gen:.5f}\tL1 Loss: {loss_L1:.5f}"
f"\tTotal Time: {(time.time() - start_time2)/60:.3f} minutes"
)
print(
f"Epoch: {epoch}\tTrain Loss: {epoch_loss/len(train_loader):.5f}"
f"\tTotal Time: {(time.time() - start_time1)/60:.3f} minutes")
train_losses.append(epoch_loss / len(train_loader))
return train_losses