def __init__(self,
input_dim,
side_feat_dim,
gcn_hidden_dim,
side_hidden_dim,
encode_hidden_dim,
num_support=5,
num_classes=5,
num_basis=3):
super(GraphMatrixCompletion, self).__init__()
self.encoder = StackGCNEncoder(input_dim, gcn_hidden_dim, num_support,
DROPOUT_RATIO)
self.dense1 = FullyConnected(side_feat_dim,
side_hidden_dim,
dropout=0.,
use_bias=True)
self.dense2 = FullyConnected(gcn_hidden_dim + side_hidden_dim,
encode_hidden_dim,
dropout=DROPOUT_RATIO,
activation=lambda x: x)
self.decoder = Decoder(encode_hidden_dim,
num_basis,
num_classes,
dropout=DROPOUT_RATIO,
activation=lambda x: x)
def train():
# config
epochs = 20
batchsize = 256
# model
encoder = Encoder().cuda()
decoder = Decoder().cuda()
trainloader = DataLoader('region', batchsize)
params = list(encoder.parameters()) + list(decoder.parameters())
# optimizier
optimizer = optim.SGD(params, lr=0.01, momentum=0.9, weight_decay=1e-4)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1)
# loss function
loss_func = nn.MSELoss().cuda()
best_loss = 1000
for epoch in range(epochs):
epoch_loss = []
pbar = tqdm(trainloader)
for image, coord in pbar:
image = image.cuda()
output = decoder(encoder(image))
# compute loss
loss = loss_func(output, image)
optimizer.zero_grad()
loss.backward()
optimizer.step()
epoch_loss.append(loss.item())
fmt = 'Epoch[{:2d}]-Loss:{:.3f}'.format(epoch + 1, loss.item())
pbar.set_description(fmt)
avg_loss = sum(epoch_loss) / len(epoch_loss)
# optimizer update
scheduler.step()
if avg_loss < best_loss:
best_loss = avg_loss
torch.save(
encoder.state_dict(),
os.path.join(
checkpoints,
'Encoder-epoch-%d-loss:%.3f.pth' % (epoch + 1, avg_loss)))
torch.save(
decoder.state_dict(),
os.path.join(
checkpoints,
'Decoder-epoch-%d-loss:%.3f.pth' % (epoch + 1, avg_loss)))
print('Train Finished!!!')
def _build_model(self):
"""Build auto-encoder model.
"""
embed = EmbeddingLayer(self.args.d_model,
self.args.vocab_size,
dropout=self.args.dropout)
encoder = Encoder(self.args.d_model,
self.args.N,
self.args.head_num,
self.args.d_ff,
self.args.hidden_size,
dropout=self.args.dropout)
decoder = Decoder(self.args.d_model,
self.args.N,
self.args.head_num,
self.args.d_ff,
self.args.hidden_size,
dropout=self.args.dropout)
linear_softmax = LinearSoftmax(self.args.d_model, self.args.vocab_size)
model = EncoderDecoder(embed, encoder, decoder, linear_softmax)
return model
def train(train_data, test_data, load_pretrained=False, pretrained_path=None):
device = 'cpu'
if torch.cuda.is_available():
device = 'cuda'
train_data = torch.Tensor(train_data).to(device)
_test_data = torch.Tensor(test_data).to(device)
num_data, input_dimension = train_data.shape # input dimension = dictionary size ~ 700
hidden_dimensions = [1024, 728]
code_dimension = 512
total_steps = 500
learning_rate = 1e-4
iter_per_batch = 10
batch_size = num_data // iter_per_batch
config_dict = {
'input_dimension': input_dimension,
'hidden_dimensions': hidden_dimensions,
'code_dimension': code_dimension
}
config_file_path = osp.join(output_path, 'config.json')
with open(config_file_path, 'w') as f:
json.dump(config_dict, f, indent=4)
encoder = Encoder(input_dimension, hidden_dimensions, code_dimension)
encoder.to(device)
decoder = Decoder(input_dimension, hidden_dimensions, code_dimension)
decoder.to(device)
if load_pretrained == True:
if pretrained_path != None:
encoder_file_path, decoder_file_path = pretrained_path
encoder.load_state_dict(torch.load(encoder_file_path))
decoder.load_state_dict(torch.load(decoder_file_path))
else:
raise AssertionError("Pretrained path must be provided!!!")
loss_func = nn.MSELoss()
optimizer = torch.optim.Adam(list(encoder.parameters()) + list(decoder.parameters()), lr=learning_rate)
best_loss = float("Inf")
for i in tqdm(range(total_steps)):
# Get next batch
batch_index = i % iter_per_batch * batch_size
X = train_data[batch_index : min(num_data, batch_index + batch_size)]
out = encoder(X)
out = decoder(out)
loss = loss_func(out, X)
loss.backward()
optimizer.step()
encoder.zero_grad()
decoder.zero_grad()
# Evaluate autoencoder performance on test data
out = decoder(encoder((_test_data))).detach().cpu().numpy()
test_data_loss = np.mean(np.sqrt(np.sum((out - test_data)**2, axis=1)))
print("-- Train loss: {}".format(loss))
print("-- Test loss: {}".format(test_data_loss))
if test_data_loss < best_loss:
best_loss = test_data_loss
encoder_file_path = osp.join(output_path, 'encoder.pt')
decoder_file_path = osp.join(output_path, 'decoder.pt')
torch.save(encoder.state_dict(), encoder_file_path)
torch.save(decoder.state_dict(), decoder_file_path)
cuda=torch.device('cuda:0')
dataset_train = FirmaData_onesubject(data_folder_dir, subject_id_train, 0.5, 0.2, 0.3, window_size, subset='train')
dataset_test_normal=FirmaData_onesubject(data_folder_dir,subject_id_train,0.5,0.2,0.3,window_size,subset='test')
dataset_test_anomaly = FirmaData_onesubject(data_folder_dir, subject_id_anomaly, 0.7, 0.0, 0.3, window_size, subset='train')
loader_train = DataLoader(dataset_train, batch_size=b_size, shuffle=True)
loader_test_normal=DataLoader(dataset_test_normal, batch_size=b_size, shuffle=True)
loader_test_anomaly=DataLoader(dataset_test_anomaly,batch_size=b_size,shuffle=True)
#
# [window_size,data_dim] 564 in our case (remove the first timestamp dimension)
input_size = dataset_train[0].shape[1]
writer = SummaryWriter()
model_encoder=Encoder(input_size,latent_dim,embed_dim,num_layer)
model_decoder=Decoder(input_size,latent_dim,input_size,embed_dim,num_layer)
model = Seq2Seq(model_encoder,model_decoder,cuda).to(cuda)
#loss_function = nn.MSELoss(reduce=False)
#loss_function = nn.L1Loss(reduce=False)
loss_function = nn.CrossEntropyLoss(weight=torch.tensor([1.,4.]).cuda()) #
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
global_step=0
#writer.add_graph(model,loader_train)
for epoch in range(Max_epoch):
for step, seq_data in enumerate(loader_train):
# print(seq_data.shape) # [32,10,564], [batch,seq_len,data_dim]
global_step+=1
seq_data=seq_data.cuda()
seq_pred = model(seq_data.float())