def main(args):
if args.use_cuda:
device = 'cuda:' + str(args.gpu_id)
else:
device = 'cpu'
print('\nLoading dataset')
#LOAD DATASET
with open(args.predictors_path, 'rb') as f:
predictors = pickle.load(f)
with open(args.target_path, 'rb') as f:
target = pickle.load(f)
predictors = np.array(predictors)
target = np.array(target)
print('\nShapes:')
print('Predictors: ', predictors.shape)
#convert to tensor
predictors = torch.tensor(predictors).float()
target = torch.tensor(target).float()
#build dataset from tensors
dataset_ = utils.TensorDataset(predictors, target)
#build data loader from dataset
dataloader = utils.DataLoader(dataset_, 1, shuffle=False, pin_memory=True)
if not os.path.exists(args.results_path):
os.makedirs(args.results_path)
#LOAD MODEL
if args.architecture == 'fasnet':
model = FaSNet_origin(enc_dim=args.enc_dim,
feature_dim=args.feature_dim,
hidden_dim=args.hidden_dim,
layer=args.layer,
segment_size=args.segment_size,
nspk=args.nspk,
win_len=args.win_len,
context_len=args.context_len,
sr=args.sr)
elif args.architecture == 'tac':
model = FaSNet_TAC(enc_dim=args.enc_dim,
feature_dim=args.feature_dim,
hidden_dim=args.hidden_dim,
layer=args.layer,
segment_size=args.segment_size,
nspk=args.nspk,
win_len=args.win_len,
context_len=args.context_len,
sr=args.sr)
if args.use_cuda:
print("Moving model to gpu")
model = model.to(device)
#load checkpoint
state = load_model(model, None, args.model_path, args.use_cuda)
#COMPUTING METRICS
print("COMPUTING TASK 1 METRICS")
print('M: Final Task 1 metric')
print('W: Word Error Rate')
print('S: Stoi')
WER = 0.
STOI = 0.
METRIC = 0.
count = 0
model.eval()
with tqdm(total=len(dataloader) // 1) as pbar, torch.no_grad():
for example_num, (x, target) in enumerate(dataloader):
outputs = enhance_sound(x, model, device, args.segment_length,
args.segment_overlap)
outputs = np.squeeze(outputs)
target = np.squeeze(target)
outputs = outputs / np.max(outputs) * 0.9 #normalize prediction
metric, wer, stoi = task1_metric(target, outputs)
if metric is not None:
METRIC += (1. / float(example_num + 1)) * (metric - METRIC)
WER += (1. / float(example_num + 1)) * (wer - WER)
STOI += (1. / float(example_num + 1)) * (stoi - STOI)
#save sounds
if args.save_sounds_freq is not None:
sounds_dir = os.path.join(args.results_path, 'sounds')
if not os.path.exists(sounds_dir):
os.makedirs(sounds_dir)
if count % args.save_sounds_freq == 0:
sf.write(
os.path.join(sounds_dir,
str(example_num) + '.wav'), outputs,
16000, 'PCM_16')
print('metric: ', metric, 'wer: ', wer, 'stoi: ', stoi)
else:
print('No voice activity on this frame')
pbar.set_description('M:' + str(np.round(METRIC, decimals=3)) +
', W:' + str(np.round(WER, decimals=3)) +
', S: ' + str(np.round(STOI, decimals=3)))
pbar.update(1)
count += 1
#visualize and save results
results = {'word error rate': WER, 'stoi': STOI, 'task 1 metric': METRIC}
print('RESULTS')
for i in results:
print(i, results[i])
out_path = os.path.join(args.results_path, 'task1_metrics_dict.json')
np.save(out_path, results)
'''
data_get = torch.zeros(data.shape[0], size, size)
data = data.cpu()
for i in range(data.shape[0]):
# print(i)
temp = unloader(data[i])
temp = crop(temp)
temp = loader(temp)
data_get[i] = temp
data_get = torch.unsqueeze(data_get, dim=1).type(torch.FloatTensor).cuda()
return data_get
batch_size = 64
# 将训练数据的特征和标签组合
dataset = Data.TensorDataset(train_x, train_y)
# 把 dataset 放入 DataLoader
data_iter = Data.DataLoader(
dataset=dataset, # torch TensorDataset format
batch_size=batch_size, # mini batch size
shuffle=True, # 要不要打乱数据 (打乱比较好)
num_workers=0, # 多线程来读数据, 注意多线程需要在 if __name__ == '__main__': 函数中运行
)
# num_workers=0 表示不用额外的进程来加速读取数据
testset = Data.TensorDataset(test_x, test_y)
test_iter = Data.DataLoader(
dataset=testset, # torch TensorDataset format
batch_size=batch_size, # mini batch size
shuffle=False, # 要不要打乱数据
def train(
self,
training_batch_size: int = 50,
learning_rate: float = 5e-4,
validation_fraction: float = 0.1,
stop_after_epochs: int = 20,
max_num_epochs: Optional[int] = None,
clip_max_norm: Optional[float] = 5.0,
calibration_kernel: Optional[Callable] = None,
exclude_invalid_x: bool = True,
resume_training: bool = False,
discard_prior_samples: bool = False,
retrain_from_scratch_each_round: bool = False,
show_train_summary: bool = False,
dataloader_kwargs: Optional[dict] = None,
) -> DirectPosterior:
r"""
Return density estimator that approximates the distribution $p(\theta|x)$.
Args:
training_batch_size: Training batch size.
learning_rate: Learning rate for Adam optimizer.
validation_fraction: The fraction of data to use for validation.
stop_after_epochs: The number of epochs to wait for improvement on the
validation set before terminating training.
max_num_epochs: Maximum number of epochs to run. If reached, we stop
training even when the validation loss is still decreasing. If None, we
train until validation loss increases (see also `stop_after_epochs`).
clip_max_norm: Value at which to clip the total gradient norm in order to
prevent exploding gradients. Use None for no clipping.
calibration_kernel: A function to calibrate the loss with respect to the
simulations `x`. See Lueckmann, Gonçalves et al., NeurIPS 2017.
exclude_invalid_x: Whether to exclude simulation outputs `x=NaN` or `x=±∞`
during training. Expect errors, silent or explicit, when `False`.
resume_training: Can be used in case training time is limited, e.g. on a
cluster. If `True`, the split between train and validation set, the
optimizer, the number of epochs, and the best validation log-prob will
be restored from the last time `.train()` was called.
discard_prior_samples: Whether to discard samples simulated in round 1, i.e.
from the prior. Training may be sped up by ignoring such less targeted
samples.
retrain_from_scratch_each_round: Whether to retrain the conditional density
estimator for the posterior from scratch each round.
show_train_summary: Whether to print the number of epochs and validation
loss after the training.
dataloader_kwargs: Additional or updated kwargs to be passed to the training
and validation dataloaders (like, e.g., a collate_fn)
Returns:
Density estimator that approximates the distribution $p(\theta|x)$.
"""
# Calibration kernels proposed in Lueckmann, Gonçalves et al., 2017.
if calibration_kernel is None:
calibration_kernel = lambda x: ones([len(x)], device=self._device)
max_num_epochs = 2**31 - 1 if max_num_epochs is None else max_num_epochs
# Starting index for the training set (1 = discard round-0 samples).
start_idx = int(discard_prior_samples and self._round > 0)
# For non-atomic loss, we can not reuse samples from previous rounds as of now.
# SNPE-A can, by construction of the algorithm, only use samples from the last
# round. SNPE-A is the only algorithm that has an attribute `_ran_final_round`,
# so this is how we check for whether or not we are using SNPE-A.
if self.use_non_atomic_loss or hasattr(self, "_ran_final_round"):
start_idx = self._round
theta, x, prior_masks = self.get_simulations(start_idx,
exclude_invalid_x,
warn_on_invalid=True)
# Dataset is shared for training and validation loaders.
dataset = data.TensorDataset(
theta,
x,
prior_masks,
)
# Set the proposal to the last proposal that was passed by the user. For
# atomic SNPE, it does not matter what the proposal is. For non-atomic
# SNPE, we only use the latest data that was passed, i.e. the one from the
# last proposal.
proposal = self._proposal_roundwise[-1]
train_loader, val_loader = self.get_dataloaders(
dataset,
training_batch_size,
validation_fraction,
resume_training,
dataloader_kwargs=dataloader_kwargs,
)
# First round or if retraining from scratch:
# Call the `self._build_neural_net` with the rounds' thetas and xs as
# arguments, which will build the neural network.
# This is passed into NeuralPosterior, to create a neural posterior which
# can `sample()` and `log_prob()`. The network is accessible via `.net`.
if self._neural_net is None or retrain_from_scratch_each_round:
self._neural_net = self._build_neural_net(
theta[self.train_indices], x[self.train_indices])
# If data on training device already move net as well.
if (not self._device == "cpu"
and f"{x.device.type}:{x.device.index}" == self._device):
self._neural_net.to(self._device)
test_posterior_net_for_multi_d_x(self._neural_net, theta, x)
self._x_shape = x_shape_from_simulation(x)
# Move entire net to device for training.
self._neural_net.to(self._device)
if not resume_training:
self.optimizer = optim.Adam(
list(self._neural_net.parameters()),
lr=learning_rate,
)
self.epoch, self._val_log_prob = 0, float("-Inf")
while self.epoch <= max_num_epochs and not self._converged(
self.epoch, stop_after_epochs):
# Train for a single epoch.
self._neural_net.train()
train_log_prob_sum = 0
epoch_start_time = time.time()
for batch in train_loader:
self.optimizer.zero_grad()
# Get batches on current device.
theta_batch, x_batch, masks_batch = (
batch[0].to(self._device),
batch[1].to(self._device),
batch[2].to(self._device),
)
batch_loss = torch.mean(
self._loss(
theta_batch,
x_batch,
masks_batch,
proposal,
calibration_kernel,
))
train_log_prob_sum += batch_loss.sum().item()
batch_loss.backward()
if clip_max_norm is not None:
clip_grad_norm_(
self._neural_net.parameters(),
max_norm=clip_max_norm,
)
self.optimizer.step()
self.epoch += 1
train_log_prob_sum /= int(theta.shape[0] *
(1.0 - validation_fraction))
self._summary["train_log_probs"].append(train_log_prob_sum)
# Calculate validation performance.
self._neural_net.eval()
log_prob_sum = 0
with torch.no_grad():
for batch in val_loader:
theta_batch, x_batch, masks_batch = (
batch[0].to(self._device),
batch[1].to(self._device),
batch[2].to(self._device),
)
# Take negative loss here to get validation log_prob.
batch_log_prob = -self._loss(
theta_batch,
x_batch,
masks_batch,
proposal,
calibration_kernel,
)
log_prob_sum += batch_log_prob.sum().item()
# Take mean over all validation samples.
self._val_log_prob = log_prob_sum / (len(val_loader) *
val_loader.batch_size)
# Log validation log prob for every epoch.
self._summary["validation_log_probs"].append(self._val_log_prob)
self._summary["epoch_durations_sec"].append(time.time() -
epoch_start_time)
self._maybe_show_progress(self._show_progress_bars, self.epoch)
self._report_convergence_at_end(self.epoch, stop_after_epochs,
max_num_epochs)
# Update summary.
self._summary["epochs"].append(self.epoch)
self._summary["best_validation_log_probs"].append(
self._best_val_log_prob)
# Update tensorboard and summary dict.
self._summarize(
round_=self._round,
x_o=None,
theta_bank=theta,
x_bank=x,
)
# Update description for progress bar.
if show_train_summary:
print(self._describe_round(self._round, self._summary))
return deepcopy(self._neural_net)
# offline_archs = []
if (ARCH in offline_archs):
# construct the transformer
batch_size = 512
z = torch.FloatTensor(np.load('{}.z.npy'.format(ARCH)))
y = torch.LongTensor(np.load('{}.y.npy'.format(ARCH)))
# to do some truncating
batch_num = z.shape[0] // batch_size
print(batch_num)
ARGS.save_p += ".{:.1f}".format(TRUNCATE_RATIO)
current_batch_num = int(batch_num * TRUNCATE_RATIO)
print("Batch Size {}/{}".format(current_batch_num, batch_num))
#
xl_dataset = data_utils.TensorDataset(z, y)
xl_dataloader = data_utils.DataLoader(xl_dataset,
batch_size=batch_size,
shuffle=True)
xl_dataloader = [(z, y) for z, y in xl_dataloader]
print(len(xl_dataloader))
def explate(seq):
out = ""
for c in seq:
out = out + c + ' '
return out[:-1]
def extract_genomes(path):
import torch.utils.data as Data
import matplotlib.pyplot as plt
LR=0.01
BATCH_SIZE=32
EPOCH=12
x=torch.unsqueeze(torch.linspace(-1,1,100),dim=1)
y=x.pow(2)+0.1*torch.normal(torch.zeros(*x.size()))
# plt.scatter(x.numpy(),y.numpy())
# plt.show()
torch_dataset=Data.TensorDataset(data_tensor=x,target_tensor=y)
loader=Data.DataLoader(
dataset=torch_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2,
)
class Net(torch.nn.Module):
def __init__(self):
super(Net,self).__init()
self.hidden=torch.nn.Linear(1,20)
self.predict=torch.nn.Linear(20,1)
def forward(self, x):
x=F.relu(self.hidden(x))
x=self.predict(x)
target = []
folder = sorted(glob.glob('./dynamic_image/*'))
# %%
for path in folder:
a = sorted(glob.glob(path + '/*.jpeg'))
frames = [cv2.imread(f) for f in a]
tmp = torch.stack([torch.Tensor(i) for i in frames]).permute(0, 3, 1, 2)
tensor_x = torch.cat((tensor_x, tmp), 0)
for i in range(len(a)):
target.append(path.split("/")[-1])
# %%
le.fit(target)
tensor_y = torch.tensor(le.transform(target))
tensor_x = tensor_x / 255
train_set = utils.TensorDataset(tensor_x, tensor_y)
# %%
# image,label = next(iter(train_set))
# image = image.permute(1,2,0)
# plt.imshow(image)
# %%
display_loader = torch.utils.data.DataLoader(train_set,
batch_size=10,
shuffle=True)
batch = next(iter(display_loader))
print('len:', len(batch))
images, labels = batch
grid = torchvision.utils.make_grid(images, nrow=10)
plt.figure(figsize=(30, 30))
def main():
args = define_and_get_arguments()
hook = sy.TorchHook(torch)
# 가상작업자(시뮬레이션) 사용시 이곳으로 분기
if args.use_virtual:
alice = VirtualWorker(id="alice", hook=hook, verbose=args.verbose)
bob = VirtualWorker(id="bob", hook=hook, verbose=args.verbose)
charlie = VirtualWorker(id="charlie", hook=hook, verbose=args.verbose)
# 웹소켓작업자 사용시 이곳으로 분기
else:
a_kwargs_websocket = {"host": "192.168.0.57", "hook": hook}
b_kwargs_websocket = {"host": "192.168.0.58", "hook": hook}
c_kwargs_websocket = {"host": "192.168.0.59", "hook": hook}
baseport = 10002
alice = WebsocketClientWorker(id="alice",
port=baseport,
**a_kwargs_websocket)
bob = WebsocketClientWorker(id="bob",
port=baseport,
**b_kwargs_websocket)
charlie = WebsocketClientWorker(id="charlie",
port=baseport,
**c_kwargs_websocket)
# 워커 객체를 리스트로 묶음
workers = [alice, bob, charlie]
# 쿠다 사용 여부
use_cuda = args.cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {"num_workers": 1, "pin_memory": True} if use_cuda else {}
# 랜덤 시드 설정
torch.manual_seed(args.seed)
labels_resampled_factorized, obs_resampled_with_noise_2 = process_data()
# percentage of test/valid set to use for testing and validation from the test_valid_idx (to be called test_size)
test_size = 0.1
# obtain training indices that will be used for validation
num_train = len(obs_resampled_with_noise_2)
indices = list(range(num_train))
np.random.shuffle(indices)
split = int(np.floor(test_size * num_train))
train_idx, test_idx = indices[split:], indices[:split]
print(type(obs_resampled_with_noise_2[train_idx]),
type(labels_resampled_factorized[train_idx]))
print(obs_resampled_with_noise_2[train_idx].shape,
labels_resampled_factorized[train_idx].shape)
print(labels_resampled_factorized[train_idx])
federated_train_dataset = D.TensorDataset(
torch.tensor(obs_resampled_with_noise_2[train_idx]),
torch.tensor(labels_resampled_factorized[train_idx]))
federated_train_loader = sy.FederatedDataLoader(
federated_train_dataset.federate(tuple(workers)),
batch_size=args.batch_size,
shuffle=True,
iter_per_worker=True,
**kwargs,
)
test_dataset = D.TensorDataset(
torch.tensor(obs_resampled_with_noise_2[test_idx]),
torch.tensor(labels_resampled_factorized[test_idx]))
test_loader = D.DataLoader(test_dataset,
shuffle=True,
batch_size=args.batch_size,
num_workers=0,
drop_last=True)
model = Net(input_features=1, output_dim=5).to(device)
criterion = nn.NLLLoss()
for epoch in range(1, args.epochs + 1):
logger.info("Starting epoch %s/%s", epoch, args.epochs)
model = train(model,
device,
federated_train_loader,
args.lr,
args.federate_after_n_batches,
criterion=criterion)
test(model,
test_loader,
args.batch_size,
criterion=criterion,
train_on_gpu=use_cuda)
if args.save_model:
torch.save(model.state_dict(), "./Model/mnist_cnn.pt")
def read_idx3_ubyte(path):
with open_maybe_compressed_file(path) as f:
data = f.read()
assert get_int(data[:4]) == 8 * 256 + 3
length = get_int(data[4:8])
num_rows = get_int(data[8:12])
num_cols = get_int(data[12:16])
parsed = np.frombuffer(data, dtype=np.uint8, offset=16)
return torch.from_numpy(parsed).view(length, num_rows, num_cols)
qtrainimgs = read_idx3_ubyte("qmnist-train-images-idx3-ubyte")
qtrainlbls = read_idx2_int("qmnist-train-labels-idx2-int")
qtrain = data.TensorDataset(qtrainimgs.float() / 255, qtrainlbls[:, 0])
# let's go
def subset(dataset, digit):
temp = []
for i in range(len(dataset)):
(_, lbl) = dataset[i]
if lbl == digit:
temp += (i, )
return data.Subset(dataset, temp)
def collect_images(data):
with torch.no_grad():
linear_feature_columns + dnn_feature_columns)
print("feature_index ", feature_index)
target = ['label']
X = train_model_input
Y = train[target].values
if isinstance(X, dict):
X = [X[feature] for feature in feature_index]
for i in range(len(X)):
if len(X[i].shape) == 1:
X[i] = np.expand_dims(X[i], axis=1)
train_tensor_data = Data.TensorDataset(torch.from_numpy(np.concatenate(X, axis=-1)), torch.from_numpy(Y))
torchX = torch.from_numpy(np.concatenate(X, axis=-1))
print("torchX size: ", torchX.size())
print("train_model_input: ", len(X))
for feat in sparse_feature_columns:
print("feat" , feat)
print("feature_index[feat.name][0]" , feature_index[feat.name][0])
test= torchX[:, feature_index[feat.name][0]:feature_index[feat.name][1]].long()
#print("torchX range:", test)
#print("embedding retrieved: ", embedding_dict[feat.embedding_name](test))
sparse_embedding_list = [embedding_dict[feat.embedding_name]
(torchX[:, feature_index[feat.name][0]:feature_index[feat.name][1]].long()) for feat in sparse_feature_columns]
print("sparse_embedding_list:", len(sparse_embedding_list))
narray = np.load("E:\\处理数据及问题\\安捷暖通-5m-模型-末端.npy", allow_pickle=True)
X = narray[:, 0:85]
Y = narray[:, 85][:, np.newaxis]
# Y = minmax(Y)
train_x, test_x, train_y, test_y = train_test_split(X,
Y,
test_size=0.15,
random_state=5)
train_x = torch.from_numpy(train_x).float()
train_y = torch.from_numpy(train_y).float()
test_x = torch.from_numpy(test_x).float()
test_y = torch.from_numpy(test_y).float()
torch_train_dataset = Data.TensorDataset(train_x, train_y)
train_loader = Data.DataLoader(dataset=torch_train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2)
torch_test_dataset = Data.TensorDataset(test_x, test_y)
test_loader = Data.DataLoader(dataset=torch_test_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2)
def mape(y_true, y_pred):
return np.mean(np.abs((y_pred - y_true) / y_true)) * 100
device = 'cuda' if torch.cuda.is_available() else 'cpu'
#print(device)
# %% [code]
#Import datasets
dataset = pd.read_csv("/kaggle/input/Kannada-MNIST/train.csv")
validation_dataset = pd.read_csv("/kaggle/input/Kannada-MNIST/test.csv")
# %% [code]
#Separate labels and vectors an one hot encoding
y_dataset, X_dataset = dataset["label"], dataset.drop("label", axis=1)
y_dataset, X_dataset = torch.tensor(y_dataset.to_numpy(), dtype = torch.long), torch.tensor(X_dataset.to_numpy(), dtype = torch.float32)/255.
#changing to a torch.dataset
train = data_utils.TensorDataset(X_dataset, y_dataset)
train_dataset, test_dataset, trash = torch.utils.data.random_split(train, [55000, 5000,0])
train_loader = data_utils.DataLoader(train_dataset, batch_size = params["batch_size"], shuffle = True)
test_loader = data_utils.DataLoader(test_dataset, batch_size = params["batch_size"], shuffle = True)
#for i in range(len(X_dataset[0])):
# X_dataset[i,:] = (X_dataset[i,:] - statistics.mean(X_dataset[i,:]))/statistics.sdev(X_dataset[i,:])
# %% [code]
#Definition of the class to hold the net arquitecture
class FullyConnected(nn.Module):
def __init__(self, num_inputs, hidden_size, num_classes):
super().__init__()
self.h1 = nn.Linear(num_inputs,hidden_size)
def obtain_fake_features(vf_path, f_num, nos_len):
g = Generated_Fake_Features()
# noise length
features, labels = g.generate_fake_features(vf_path, f_num, nos_len)
dataset = Data.TensorDataset(features, labels)
return dataset
wr_idx_num = int(right_num.item() * 1.5)
wr_idx = random.sample(random_wrong_idx, wr_idx_num)
w_f = fake_features[wr_idx]
wrong_feature = w_f.detach()
[wrong_features.append(i.squeeze()) for i in wrong_feature]
wrong_features = np.array(wrong_features)
# right_features = torch.from_numpy(np.array(right_features).astype(float))
right_features = torch.from_numpy(np.array([item.cpu().numpy() for item in right_features])).cuda()
wrong_features = torch.from_numpy(np.array([item.cpu().numpy() for item in wrong_features])).cuda()
original_labels = torch.from_numpy(np.array([item.cpu().numpy() for item in original_labels])).cuda()
right_labels = torch.ones(right_features.shape[0])
wrong_labels = torch.zeros(wrong_features.shape[0])
train_dataset_g_to_f = Data.TensorDataset(right_features, original_labels)
new_vs_features = torch.cat([right_features, wrong_features], dim=0)
new_vs_labels = torch.cat([right_labels, wrong_labels], dim=0)
train_dataset_g_to_c = Data.TensorDataset(new_vs_features, new_vs_labels)
dataset_loader_g_to_c = Data.DataLoader(train_dataset_g_to_c, 40, shuffle=True, drop_last=True)
dataset_loader_g_to_f = Data.DataLoader(train_dataset_g_to_f, 20, shuffle=True, drop_last=True)
############################################# G-->C ####################################################
loss_c = 0
loss_f = 0
for i, batch in enumerate(dataset_loader_g_to_c, 0):
g_net.zero_grad()
c_net.zero_grad()
init_model_params(g_net, True)
def compute(self):
for i in range(self.partition):
start = time.time()
temptrain_X, temptrain_Y, tempval_X, tempval_Y = self.create_data(i)
self.train_X, self.train_Y = factorize(temptrain_X, temptrain_Y, self.augment_data_flag, self.batch_size, self.batch_length)
self.val_X, self.val_Y = factorize(tempval_X, tempval_Y, self.augment_data_flag, self.batch_size, self.batch_length)
train_set = data_utils.TensorDataset(self.train_X, self.train_Y)
train_loader=data_utils.DataLoader(dataset=train_set, batch_size=BATCH_SIZE, drop_last=True, shuffle=True)
print(i, "phase 1 completed.")
cur_model = deepcopy(self.model).to(device)
optimizer = optim.SGD(cur_model.parameters(), lr = self.lr, weight_decay = 1e-5)
criterion = nn.SmoothL1Loss()
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size = 1, gamma = self.gamma)
print_loss_total = 0
plot_loss_total = 0
for j in range(self.epochs):
for num, (train_X, train_Y)in enumerate(train_loader):
input_tensor = train_X.to(device).float()
target_tensor = train_Y.to(device).float()
#print(input_tensor.shape)
loss = train(input_tensor, target_tensor, cur_model, decoder_optimizer=optimizer, criterion= criterion)
#print(num, loss)
print_loss_total += loss
plot_loss_total += loss
if (num+1)%self.plot_every == 0:
plot_loss_avg = plot_loss_total / self.plot_every
plot_loss_total = 0
self.loss_history.append(plot_loss_avg)
acc, one_acc, score = validate(cur_model, self.val_X, self.val_Y)
self.precision_history.append(acc[:-1])
self.recall_history.append(one_acc[:-1])
print("validation accuracy:", acc)
print("validation prediction accuracy:", one_acc)
if (num+1) % self.print_every == 0:
print_loss_avg = print_loss_total / self.print_every
print_loss_total = 0
print("partition%i epoch %i"%(i,j))
p = self.timeSince(start, (num+j*len(train_loader)) / (self.epochs * len(train_loader)))
print('%s (%d %d%%) %.4f' % (p, num + 1, (num + 1) / (self.epochs * len(train_loader)) * self.print_every,
print_loss_avg))
"""if(score > self.best_acc):
# torch.save(cur_model.state_dict(), '/home/yiqin/2018summer_project/saved_model/Bi-LSTM-CNN_best(cv).pt')
self.best_acc = score
print("best_score:", self.best_acc)"""
scheduler.step()
#torch.save(cur_model.state_dict(), '/home/yiqin/2018summer_project/saved_model/Bi-LSTM-CNN(cv){}-{}.pt'.format(i,j))
return self.loss_history, self.precision_history, self.recall_history
def _train(
self,
num_atoms: int,
training_batch_size: int,
learning_rate: float,
validation_fraction: float,
stop_after_epochs: int,
max_num_epochs: int,
clip_max_norm: Optional[float],
exclude_invalid_x: bool,
discard_prior_samples: bool,
) -> None:
r"""
Trains the neural classifier.
Update the classifier weights by maximizing a Bernoulli likelihood which
distinguishes between jointly distributed $(\theta, x)$ pairs and randomly
chosen $(\theta, x)$ pairs.
Uses performance on a held-out validation set as a terminating condition (early
stopping).
"""
# Starting index for the training set (1 = discard round-0 samples).
start_idx = int(discard_prior_samples and self._round > 0)
theta, x, _ = self._get_from_data_bank(start_idx, exclude_invalid_x)
# Get total number of training examples.
num_examples = len(theta)
# Select random train and validation splits from (theta, x) pairs.
permuted_indices = torch.randperm(num_examples)
num_training_examples = int((1 - validation_fraction) * num_examples)
num_validation_examples = num_examples - num_training_examples
train_indices, val_indices = (
permuted_indices[:num_training_examples],
permuted_indices[num_training_examples:],
)
clipped_batch_size = min(training_batch_size, num_validation_examples)
num_atoms = clamp_and_warn(
"num_atoms", num_atoms, min_val=2, max_val=clipped_batch_size
)
# Dataset is shared for training and validation loaders.
dataset = data.TensorDataset(theta, x)
# Create neural net and validation loaders using a subset sampler.
train_loader = data.DataLoader(
dataset,
batch_size=clipped_batch_size,
drop_last=True,
sampler=SubsetRandomSampler(train_indices),
)
val_loader = data.DataLoader(
dataset,
batch_size=clipped_batch_size,
shuffle=False,
drop_last=False,
sampler=SubsetRandomSampler(val_indices),
)
optimizer = optim.Adam(
list(self._posterior.net.parameters()), lr=learning_rate,
)
epoch, self._val_log_prob = 0, float("-Inf")
while epoch <= max_num_epochs and not self._converged(epoch, stop_after_epochs):
# Train for a single epoch.
self._posterior.net.train()
for batch in train_loader:
optimizer.zero_grad()
theta_batch, x_batch = (
batch[0].to(self._device),
batch[1].to(self._device),
)
loss = self._loss(theta_batch, x_batch, num_atoms)
loss.backward()
if clip_max_norm is not None:
clip_grad_norm_(
self._posterior.net.parameters(), max_norm=clip_max_norm,
)
optimizer.step()
epoch += 1
# Calculate validation performance.
self._posterior.net.eval()
log_prob_sum = 0
with torch.no_grad():
for batch in val_loader:
theta_batch, x_batch = (
batch[0].to(self._device),
batch[1].to(self._device),
)
log_prob = self._loss(theta_batch, x_batch, num_atoms)
log_prob_sum -= log_prob.sum().item()
self._val_log_prob = log_prob_sum / num_validation_examples
# Log validation log prob for every epoch.
self._summary["validation_log_probs"].append(self._val_log_prob)
self._maybe_show_progress(self._show_progress_bars, epoch)
self._report_convergence_at_end(epoch, stop_after_epochs, max_num_epochs)
# Update summary.
self._summary["epochs"].append(epoch)
self._summary["best_validation_log_probs"].append(self._best_val_log_prob)
low_resolution_samples, index = utils.divide(input_file)
low_resolution_samples = np.minimum(HiC_max_value, low_resolution_samples)
batch_size = low_resolution_samples.shape[0]
# Reshape the high-quality Hi-C sample as the target value of the training.
sample_size = low_resolution_samples.shape[-1]
padding = conv2d1_filters_size + conv2d2_filters_size + conv2d3_filters_size - 3
half_padding = padding / 2
output_length = sample_size - padding
print low_resolution_samples.shape
lowres_set = data.TensorDataset(
torch.from_numpy(low_resolution_samples),
torch.from_numpy(np.zeros(low_resolution_samples.shape[0])))
lowres_loader = torch.utils.data.DataLoader(lowres_set,
batch_size=batch_size,
shuffle=False)
hires_loader = lowres_loader
model = model.Net(40, 28)
model.load_state_dict(torch.load('../model/pytorch_model_12000'))
if use_gpu:
model = model.cuda()
_loss = nn.MSELoss()
running_loss = 0.0
def load_array(data_arrays, batch_size, is_train=True):
dataset = data.TensorDataset(*data_arrays)
return data.DataLoader(dataset, batch_size, shuffle=is_train)
def sac(
env_name,
total_steps,
model,
env_steps=0,
min_steps_per_update=1,
iters_per_update=100,
replay_batch_size=64,
seed=0,
gamma=0.95,
polyak=0.995,
alpha=0.2,
sgd_batch_size=64,
sgd_lr=1e-3,
exploration_steps=100,
replay_buf_size=int(100000),
use_gpu=False,
reward_stop=None,
env_config={},
):
"""
Implements soft actor critic
Args:
env_name: name of the openAI gym environment to solve
total_steps: number of timesteps to run the PPO for
model: model from seagul.rl.models. Contains policy, value fn, q1_fn, q2_fn
min_steps_per_update: minimun number of steps to take before running updates, will finish episodes before updating
env_steps: number of steps the environment takes before finishing, if the environment emits a done signal before this we consider it a failure.
iters_per_update: how many update steps to make every time we update
replay_batch_size: how big a batch to pull from the replay buffer for each update
seed: random seed for all rngs
gamma: discount applied to future rewards, usually close to 1
polyak: term determining how fast the target network is copied from the value function
alpha: weighting term for the entropy. 0 corresponds to no penalty for deterministic policy
sgd_batch_size: minibatch size for policy updates
sgd_lr: initial learning rate for policy optimizer
val_lr: initial learning rate for value optimizer
q_lr: initial learning rate for q fn optimizer
exploration_steps: initial number of random actions to take, aids exploration
replay_buf_size: how big of a replay buffer to use
use_gpu: determines if we try to use a GPU or not
reward_stop: reward value to bail at
env_config: dictionary containing kwargs to pass to your the environment
Returns:
model: trained model
avg_reward_hist: list with the average reward per episode at each epoch
var_dict: dictionary with all locals, for logging/debugging purposes
Example:
from seagul.rl.algos.sac import sac
import torch.nn as nn
from seagul.nn import MLP
from seagul.rl.models import SACModel
input_size = 3
output_size = 1
layer_size = 64
num_layers = 2
activation = nn.ReLU
policy = MLP(input_size, output_size*2, num_layers, layer_size, activation)
value_fn = MLP(input_size, 1, num_layers, layer_size, activation)
q1_fn = MLP(input_size + output_size, 1, num_layers, layer_size, activation)
q2_fn = MLP(input_size + output_size, 1, num_layers, layer_size, activation)
model = SACModel(policy, value_fn, q1_fn, q2_fn, 1)
model, rews, var_dict = sac("Pendulum-v0", 10000, model)
"""
torch.set_num_threads(1)
env = gym.make(env_name, **env_config)
if isinstance(env.action_space, gym.spaces.Box):
act_size = env.action_space.shape[0]
act_dtype = env.action_space.sample().dtype
else:
raise NotImplementedError("trying to use unsupported action space",
env.action_space)
obs_size = env.observation_space.shape[0]
random_model = RandModel(model.act_limit, act_size)
replay_buf = ReplayBuffer(obs_size, act_size, replay_buf_size)
target_value_fn = dill.loads(dill.dumps(model.value_fn))
pol_opt = torch.optim.Adam(model.policy.parameters(), lr=sgd_lr)
val_opt = torch.optim.Adam(model.value_fn.parameters(), lr=sgd_lr)
q1_opt = torch.optim.Adam(model.q1_fn.parameters(), lr=sgd_lr)
q2_opt = torch.optim.Adam(model.q2_fn.parameters(), lr=sgd_lr)
# seed all our RNGs
env.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
# set defaults, and decide if we are using a GPU or not
use_cuda = torch.cuda.is_available() and use_gpu
device = torch.device("cuda:0" if use_cuda else "cpu")
raw_rew_hist = []
val_loss_hist = []
pol_loss_hist = []
q1_loss_hist = []
q2_loss_hist = []
progress_bar = tqdm.tqdm(total=total_steps)
cur_total_steps = 0
progress_bar.update(0)
early_stop = False
while cur_total_steps < exploration_steps:
ep_obs1, ep_obs2, ep_acts, ep_rews, ep_done = do_rollout(
env, random_model, env_steps)
replay_buf.store(ep_obs1, ep_obs2, ep_acts, ep_rews, ep_done)
ep_steps = ep_rews.shape[0]
cur_total_steps += ep_steps
progress_bar.update(cur_total_steps)
while cur_total_steps < total_steps:
cur_batch_steps = 0
# Bail out if we have met out reward threshold
if len(raw_rew_hist) > 2 and reward_stop:
if raw_rew_hist[-1] >= reward_stop and raw_rew_hist[
-2] >= reward_stop:
early_stop = True
break
# collect data with the current policy
# ========================================================================
while cur_batch_steps < min_steps_per_update:
ep_obs1, ep_obs2, ep_acts, ep_rews, ep_done = do_rollout(
env, model, env_steps)
replay_buf.store(ep_obs1, ep_obs2, ep_acts, ep_rews, ep_done)
ep_steps = ep_rews.shape[0]
cur_batch_steps += ep_steps
cur_total_steps += ep_steps
raw_rew_hist.append(torch.sum(ep_rews))
progress_bar.update(cur_batch_steps)
for _ in range(min(int(ep_steps), iters_per_update)):
# compute targets for Q and V
# ========================================================================
replay_obs1, replay_obs2, replay_acts, replay_rews, replay_done = replay_buf.sample_batch(
replay_batch_size)
q_targ = replay_rews + gamma * (
1 - replay_done) * target_value_fn(replay_obs2)
q_targ = q_targ.detach()
noise = torch.randn(replay_batch_size, act_size)
sample_acts, sample_logp = model.select_action(replay_obs1, noise)
q_in = torch.cat((replay_obs1, sample_acts), dim=1)
q_preds = torch.cat((model.q1_fn(q_in), model.q2_fn(q_in)), dim=1)
q_min, q_min_idx = torch.min(q_preds, dim=1)
q_min = q_min.reshape(-1, 1)
v_targ = q_min - alpha * sample_logp
v_targ = v_targ.detach()
# For training, transfer model to GPU
model.policy = model.policy.to(device)
model.value_fn = model.value_fn.to(device)
model.q1_fn = model.q1_fn.to(device)
model.q2_fn = model.q2_fn.to(device)
# q_fn update
# ========================================================================
training_data = data.TensorDataset(replay_obs1, replay_acts,
q_targ)
training_generator = data.DataLoader(training_data,
batch_size=sgd_batch_size,
shuffle=True,
num_workers=0,
pin_memory=False)
for local_obs, local_acts, local_qtarg in training_generator:
# Transfer to GPU (if GPU is enabled, else this does nothing)
local_obs, local_acts, local_qtarg = (
local_obs.to(device),
local_acts.to(device),
local_qtarg.to(device),
)
q_in = torch.cat((local_obs, local_acts), dim=1)
q1_preds = model.q1_fn(q_in)
q2_preds = model.q2_fn(q_in)
q1_loss = torch.pow(q1_preds - local_qtarg, 2).mean()
q2_loss = torch.pow(q2_preds - local_qtarg, 2).mean()
q_loss = q1_loss + q2_loss
q1_opt.zero_grad()
q2_opt.zero_grad()
q_loss.backward()
q1_opt.step()
q2_opt.step()
# val_fn update
# ========================================================================
training_data = data.TensorDataset(replay_obs1, v_targ)
training_generator = data.DataLoader(training_data,
batch_size=sgd_batch_size,
shuffle=True,
num_workers=0,
pin_memory=False)
for local_obs, local_vtarg in training_generator:
# Transfer to GPU (if GPU is enabled, else this does nothing)
local_obs, local_vtarg = (local_obs.to(device),
local_vtarg.to(device))
# predict and calculate loss for the batch
val_preds = model.value_fn(local_obs)
val_loss = torch.sum(torch.pow(val_preds - local_vtarg,
2)) / replay_batch_size
# do the normal pytorch update
val_opt.zero_grad()
val_loss.backward()
val_opt.step()
# policy_fn update
# ========================================================================
training_data = data.TensorDataset(replay_obs1)
training_generator = data.DataLoader(training_data,
batch_size=sgd_batch_size,
shuffle=True,
num_workers=0,
pin_memory=False)
for local_obs in training_generator:
# Transfer to GPU (if GPU is enabled, else this does nothing)
local_obs = local_obs[0].to(device)
noise = torch.randn(local_obs.shape[0], act_size).to(device)
local_acts, local_logp = model.select_action(local_obs, noise)
q_in = torch.cat((local_obs, local_acts), dim=1)
pol_loss = torch.sum(alpha * local_logp -
model.q1_fn(q_in)) / replay_batch_size
# do the normal pytorch update
pol_opt.zero_grad()
pol_loss.backward()
pol_opt.step()
# Update target value fn with polyak average
# ========================================================================
val_loss_hist.append(val_loss.item())
pol_loss_hist.append(pol_loss.item())
q1_loss_hist.append(q1_loss.item())
q2_loss_hist.append(q2_loss.item())
#
# model.policy.state_means = update_mean(replay_obs1, model.policy.state_means, cur_total_steps)
# model.policy.state_var = update_var(replay_obs1, model.policy.state_var, cur_total_steps)
# model.value_fn.state_means = model.policy.state_means
# model.policy.state_var = model.policy.state_var
#
# model.q1_fn.state_means = update_mean(torch.cat((replay_obs1, replay_acts.detach()), dim=1), model.q1_fn.state_means, cur_total_steps)
# model.q1_fn.state_var = update_var(torch.cat((replay_obs1, replay_acts.detach()), dim=1), model.q1_fn.state_var, cur_total_steps)
# model.q2_fn.state_means = model.q1_fn.state_means
# model.q2_fn.state_var = model.q1_fn.state_var
# Transfer back to CPU, which is faster for rollouts
model.policy = model.policy.to('cpu')
model.value_fn = model.value_fn.to('cpu')
model.q1_fn = model.q1_fn.to('cpu')
model.q2_fn = model.q2_fn.to('cpu')
val_sd = model.value_fn.state_dict()
tar_sd = target_value_fn.state_dict()
for layer in tar_sd:
tar_sd[layer] = polyak * tar_sd[layer] + (
1 - polyak) * val_sd[layer]
target_value_fn.load_state_dict(tar_sd)
return model, raw_rew_hist, locals()
def train(h5file, h5key, pklfile, validationh5, trainedlossplot, train_target):
# ******* input dataset from h5, then divide it into train dataset and test dataset(16:1)
net = Net(n_feature=75, n_output=1)
# pklfile6 = 'train6/NN_train_params_3975284924_2.pkl'
# net.load_state_dict(torch.load(pklfile6))
net.cuda()
net = net.double()
print(net)
logdir = Dir_training + 'NN_logs_' + h5key
if os.path.isdir(logdir):
shutil.rmtree(logdir)
logger = Logger(logdir)
# optimizer = torch.optim.SGD(net.parameters(), lr=LR, weight_decay=0.01,momentum=0.9)
# optimizer = torch.optim.SGD(net.parameters(), lr=LR, momentum=0.5)
# optimizer = torch.optim.Adagrad(net.parameters(), lr=LR, lr_decay=0.01)
optimizer = torch.optim.Adam(net.parameters(), lr=LR)
# optimizer = torch.optim.RMSprop(net.parameters(), lr=LR, weight_decay=5e-2)
loss_func = nn.MSELoss()
print("Let's use", torch.cuda.device_count(), "GPUs!")
plt.ion()
plt.figure(figsize=(10,4))
loss_list_train = []
loss_list_test = []
step_list = []
# par_np = net.parameters()
Step = 0
lri = LR
# ****** test dataset
mydf_test = pd.read_hdf(h5file, h5key, start=0, stop= 100000)
test_data_np = mydf_test.iloc[:,4:].replace(np.nan, 0.0).values
test_data_tensor = torch.from_numpy(test_data_np).double()
if train_target == 'phi':
test_labels_np = mydf_test.mcPhi.values.reshape((mydf_test.shape[0],1))
test_rec_np = mydf_test.phi.values.reshape((mydf_test.shape[0],1))
elif train_target == 'theta':
test_labels_np = mydf_test.mcTheta.values.reshape((mydf_test.shape[0],1))
test_rec_np = mydf_test.theta.values.reshape((mydf_test.shape[0],1))
else:
print("Wrong train target!")
test_labels_tensor = torch.from_numpy(test_labels_np).double()
test_rec_tensor = torch.from_numpy(test_rec_np).double()
test_dataset = Data.TensorDataset(test_data_tensor, test_labels_tensor)
test_loader = Data.DataLoader(test_dataset, batch_size=BATCH_SIZE_test )
# res = net(test_data_tensor.cuda())
# res = Variable(torch.rand(75,640))
# writer = SummaryWriter(logdir)
# writer.add_graph(net, res.cuda().double())
# writer.close()
for epoch in range(EPOCH):
print('EPOCH: ', epoch)
reader = pd.read_hdf(h5file, h5key, chunksize=BATCH_SIZE*2, start = 100000)
for mydf_readd5 in reader:
mydf_train = mydf_readd5
# mydf_train = mydf_readd5.iloc[: int(mydf_readd5.shape[0]*15/16)]
# mydf_test = mydf_readd5.iloc[int(mydf_readd5.shape[0]*15/16):]
# print(mydf_train.iloc[:,54:].head())
# print(mydf_test.iloc[:,54:].head())
# print(mydf_train.shape)
# ****** train dataset
train_data_np = mydf_train.iloc[:,4:].replace(np.nan, 0.0).values
train_data_tensor = torch.from_numpy(train_data_np).double()
if train_target == 'phi':
train_labels_np = mydf_train.mcPhi.values.reshape((mydf_train.shape[0],1))
elif train_target == 'theta':
train_labels_np = mydf_train.mcTheta.values.reshape((mydf_train.shape[0],1))
else:
print("Wrong train target!")
train_labels_tensor = torch.from_numpy(train_labels_np).double()
train_dataset = Data.TensorDataset(train_data_tensor, train_labels_tensor)
train_loader = Data.DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)
for step, data in enumerate(train_loader):
# b_x, b_y = data
b_X, b_Y = data
b_x = b_X.cuda()
b_y = b_Y.cuda()
# ****** L2 regularization
reg_lambda = torch.tensor(0.2)
l2_reg = torch.tensor(0.)
for param in net.parameters():
l2_reg += param.cpu().float().norm(2)
prediction = net(b_x).cuda()
loss = loss_func(prediction, b_y)
# loss += (reg_lambda*l2_reg).cuda().double()
optimizer.zero_grad()
loss.backward()
optimizer.step()
Step+=1
if (Step+1) % 100 == 0:
test_output = net(test_data_tensor.cuda())
test_pred_y = test_output.cpu().data.numpy()
# test_pred_y = test_output.data.numpy()
accuracy_test = sum(test_pred_y - test_labels_np)
loss_test = loss_func(test_output, test_labels_tensor.cuda())
# loss_rec = loss_func(test_rec_tensor.cuda(), test_labels_tensor.cuda())
print('Epoch:', epoch, '|step:', Step,
'|train loss:%.8f'%loss.item(), '|test loss:%.8f'%loss_test.item())
step_list.append(Step)
loss_list_train.append(loss.item())
loss_list_test.append(loss_test.item())
plt.subplot(131)
plt.cla()
plt.plot(step_list, loss_list_train, 'b-', lw=1, label='train')
plt.plot(step_list, loss_list_test, 'r-', lw=3, label='test')
plt.xlabel('step')
plt.ylabel('loss')
plt.text(10, 0.027, 'Loss_train=%.8f' % loss.item(), fontdict={'size': 10, 'color': 'blue'})
plt.text(10, 0.025, 'Loss_test=%.8f' % loss_test.item(), fontdict={'size': 10, 'color': 'red'})
# plt.text(10, 0.023, 'Loss_rec=%.8f' % loss_rec.data[0], fontdict={'size': 10, 'color': 'red'})
legend = plt.legend(loc="best")#(loc="best")
frame = legend.get_frame()
frame.set_facecolor('none') # 璁剧疆鍥句緥legend鑳屾櫙閫忔槑
if train_target == 'phi':
Range = [-3.2, 3.2]
elif train_target == 'theta':
Range = [0.4, 2.4]
plt.subplot(133)
plt.cla()
plt.hist(test_labels_np, bins=200,range=Range, color='red',alpha=0.7, fill=False,histtype='step', label='test_truth')
plt.hist(test_pred_y, bins=200,range=Range, color='blue',alpha=0.7, fill=False,histtype='step', label='test_pre')
plt.hist(test_rec_np, bins=200,range=Range, color='green',alpha=0.7, fill=False,histtype='step', label='test_rec')
plt.xlabel(r'$' + '\\'+ train_target + '$')
legend = plt.legend(loc="best")#(loc="best")
frame = legend.get_frame()
frame.set_facecolor('none') # 璁剧疆鍥句緥legend鑳屾櫙閫忔槑
plt.subplot(132)
plt.cla()
plt.hist(b_y.cpu().data.numpy(), bins=200,range=Range, color='red',alpha=0.7, fill=False,histtype='step', label='train_truth')
plt.hist(prediction.cpu().data.numpy(), bins=200,range=Range, color='blue',alpha=0.7, fill=False,histtype='step', label='train_pre')
plt.xlabel(r'$' + '\\'+ train_target + '$')
legend = plt.legend(loc="best")#(loc="best")
frame = legend.get_frame()
frame.set_facecolor('none') # 璁剧疆鍥句緥legend鑳屾櫙閫忔槑
plt.pause(0.1)
# ================================================================== #
# Tensorboard Logging #
# ================================================================== #
# 1. Log scalar values (scalar summary)
info = { 'loss': loss.item(), 'loss_test': loss_test.item(), 'accuracy': accuracy_test.item() }
for tag, value in info.items():
logger.scalar_summary(tag, value, Step+1)
# 2. Log values and gradients of the parameters (histogram summary)
for tag, value in net.named_parameters():
tag = tag.replace('.', '/')
logger.histo_summary(tag, value.data.cpu().numpy(), Step+1)
logger.histo_summary(tag+'/grad', value.grad.data.cpu().numpy(), Step+1)
# 3. Log training images (image summary)
info = { 'images': b_x.view(-1, 5, 5)[:10].cpu().numpy() }
for tag, images in info.items():
logger.image_summary(tag, images, Step+1)
lri = lri/(1 + 0.005)
print("lri: ",lri)
for param_group in optimizer.param_groups:
param_group['lr'] = lri
if (epoch+1) % 50 == 0:
pklfile_epoch = Dir_pkl + 'NN_train_params_epoch' + str(epoch) + '.pkl'
torch.save(net.state_dict(), pklfile_epoch)
plt.ioff()
plt.savefig(trainedlossplot,dpi=300)
plt.show()
loss_df = pd.DataFrame.from_dict({'step' : step_list, 'train' : loss_list_train, 'test' : loss_list_test})
loss_df.to_hdf(train_lossh5, key=h5key, mode='w')
test_output = net(test_data_tensor[:10].cuda())
test_pred_y = test_output.cpu().data.numpy()
# test_pred_y = test_output.data.numpy()
print('prediction number: ', test_pred_y )
print( 'real number: ', test_labels_np[:10])
# ****** The model after train
for name, param in net.state_dict().items():
print(name, param.size())
# ****** save the whole model
# torch.save(model_object, 'model.pkl')
# only save the parameters ((recommended))
torch.save(net.state_dict(), pklfile)
test_pred_y = np.empty((0,1))
for step, data in enumerate(test_loader):
t_X, t_Y = data
t_x = t_X.cuda()
t_y = t_Y.cuda()
test_output = net(t_x).cuda()
test_pred_y = np.vstack([test_pred_y, test_output.cpu().data.numpy()])
# test_pred_y = np.delete(test_pred_y, 0, 0)
print("shapes: ", test_pred_y.shape)
pred_df = pd.DataFrame(mydf_test[['mcPhi','phi', 'mcTheta', 'theta']])
print("shapes: ", test_pred_y.shape, pred_df.shape)
if train_target == 'phi':
pred_df['prePhi'] = test_pred_y
elif train_target == 'theta':
pred_df['preTheta'] = test_pred_y
pred_df.to_hdf(validationh5, key=h5key, mode='w')
def DataloadtoGAN(path,
mark=None,
label=False,
single_dataset=False,
hacking=False,
select=''):
"""
func: read normal data to train GAN defined as the Class(new gan code)
:param path:dataset url
:param mark:'validate' 'test' 'train'
:param label: whether deliver label
:param single_dataset: for whole normal dataset,not suitable for normal status dataset from hacking dataset
:param hacking: whole normal dataset or normal status dataset from hacking dataset
:return: dataloader,torch.Tensor
"""
if mark == None:
print('mark is None, please checks')
return
if hacking:
files = []
for d in os.listdir(path):
if select.title() == d.title():
if mark == 'validate':
f = os.path.join(path, d, 'pure_normal.pkl')
else:
f = os.path.join(path, d, 'pure_attack.pkl')
files.append(f)
# print(files[0])
break
elif select == '':
pass
else:
continue
if 'normal.pkl' in d:
files.append(os.path.join(path, d))
continue
elif '.' in d:
continue
else:
for f in os.listdir(os.path.join(path, d)):
if 'normal.pkl' in f:
files.append(os.path.join(path, d, f))
else:
files = [os.path.join(path, f) for f in os.listdir(path) if 'pkl' in f]
fl = []
if single_dataset:
files = [i for i in files if 'Attack_free_dataset2' in i]
data2 = np.empty((64, 21))
atta2 = np.empty((64, 21))
# read dataset
for i, f in enumerate(files):
print('address:%s' % f)
atta = np.empty(((64, 21)))
data1 = pd.read_pickle(f, compression='zip')
data = data1.values.astype(np.float64)
# file = os.path.basename(path)
rows = data.shape[0]
start = 0
end = rows
row = int(rows // 64)
row1 = row
file = os.path.splitext(os.path.basename(f))[0]
fl.append(file)
dirname = os.path.dirname(f).split('/')[-1]
if mark == 'test':
start = int(((rows * 0.99) // 64) * 64)
row = int((rows * 0.01) // 64)
if start % 64 == 0:
pass
else:
start = ((start // 64) + 1) * 64
# end = int(start+((rows-start)//64)*64)
end = int(start + row * 64)
elif mark == 'train':
print('get type:%s' % 'train')
# row = int((rows*0.01)//64)
# end = int(row * 64)
row = int((rows * 0.98) // 64)
end = int(row * 64)
elif mark == 'validate':
print('get type:%s,datatype:%s' % ('validate', dirname))
row = int((rows * 0.01) // 64)
start = int(((rows * 0.98) // 64) * 64)
if start % 64 == 0:
pass
else:
start = int(((start // 64) + 1) * 64)
end = int(start + row * 64)
# end =int(((rows*0.99)//64) * 64)
if hacking:
data = data[start:end, :-1].reshape((-1, 21))
if mark == 'validate' or mark == 'test':
url = os.path.dirname(f) + '/pure_attack.pkl'
atta = pd.read_pickle(url, compression='zip')
atta = pd.DataFrame(atta).to_numpy().reshape((-1, 64, 22))
print('{},shape:{}'.format('pure_attack', atta.shape), end=',')
print(
'start at:{},%64={},end:{},%64={},acquires row:{},percent:{}%,done read files!!!'
.format(start, start % 64, end, end % 64, row,
float(row / atta.shape[0])))
atta = atta[:row, :, :21]
# print('atta.shape---:',atta.shape)
if i > 0:
data2 = np.concatenate((data2, data), axis=0).reshape((-1, 21))
atta2 = np.concatenate((atta2, atta), axis=0)
# print('atta2.shape:',atta2.shape)
else:
data2 = data
atta2 = atta
else:
data = data[start:end, :].reshape((-1, 21))
if i > 0:
data2 = np.concatenate((data2, data), axis=0).reshape((-1, 21))
else:
data2 = data
print('{} shaped:{},trunked:{}'.format(file, data1.shape, data.shape),
end=',')
print('get|all:{}|{},blocks:{}'.format(row, row1, row % 64), end=',')
print(
'start at:{},%64={},end:{},%64={},percent:{}%,done read files!!!'.
format(start, start % 64, end, end % 64, float(row / row1)))
# exit()
if mark == 'validate' or mark == 'test':
atta2 = atta2.reshape((-1, 64, 21))
data2 = data2.reshape((-1, 64, 21))
label1 = np.ones((atta2.shape[0], 1))
label0 = np.zeros((data2.shape[0], 1))
data2 = np.concatenate((data2, atta2), axis=0)
labels = np.concatenate((label0, label1), axis=0)
TraindataM = torch.from_numpy(
data2).float() # transform to float torchTensor
TraindataM = torch.unsqueeze(TraindataM, 1)
Traindata_LabelM = torch.from_numpy(labels).float()
TorchDataset = Data.TensorDataset(TraindataM, Traindata_LabelM)
print('{},size:{} label:{},done read files!!!\n'.format(
'validate mix dataset', TraindataM.shape, label))
return Data.DataLoader(dataset=TorchDataset,
batch_size=BATCH_SIZE,
shuffle=True)
TraindataM = torch.from_numpy(data2.reshape(
(-1, 64, 21))).float() # transform to float torchTensor
TraindataM = torch.unsqueeze(TraindataM, 1)
if label:
# if mark == 'train' or mark == 'test':
if select == 'Normal':
labels = np.zeros((TraindataM.shape[0], 1))
else:
labels = np.ones((TraindataM.shape[0], 1))
Traindata_LabelM = torch.from_numpy(labels).float()
TorchDataset = Data.TensorDataset(TraindataM, Traindata_LabelM)
print('{},size:{} label:{},done read files!!!\n'.format(
fl, TraindataM.shape, label))
return Data.DataLoader(dataset=TorchDataset,
batch_size=BATCH_SIZE,
shuffle=True)
else:
# if mark == 'train' or mark == 'test':
# Data Loader for easy mini-batch return in training
TorchDataset = Data.TensorDataset(TraindataM)
print('{},size:{} label:{},done read files!!!\n'.format(
fl, TraindataM.shape, label))
return Data.DataLoader(dataset=TorchDataset,
batch_size=BATCH_SIZE,
shuffle=True)
def get_dataloaders(
cfg: DictConfig,
num_workers: int = 4,
dataset_name: str = "mnist") -> Tuple[DataLoader, DataLoader]:
"""Return training and validation dataloaders"""
##################################################################
# Change this. Demo dataset is MNIST
##################################################################
if dataset_name == "mnist":
data_transform = Compose(
[ToTensor(), Normalize((0.1307, ), (0.3081, ))])
dataset = torchvision.datasets.MNIST(
os.path.join(hydra.utils.get_original_cwd(), cfg.dirs.data),
download=True,
transform=data_transform,
)
train_dataloader = DataLoader(
dataset,
batch_size=cfg.mode.train.batch_size,
shuffle=cfg.mode.train.shuffle,
num_workers=num_workers,
)
val_dataloader = DataLoader(
dataset,
batch_size=cfg.mode.val.batch_size,
shuffle=cfg.mode.val.shuffle,
num_workers=num_workers,
)
elif dataset_name == "reunion":
(
list_of_training_inputs,
training_target_df,
list_of_testing_inputs,
testing_target_df,
) = process_uni_data(cfg)
inputs_train = Variable(torch.FloatTensor(list_of_training_inputs))
targets_train = Variable(torch.FloatTensor(training_target_df))
inputs_test = Variable(torch.FloatTensor(list_of_testing_inputs))
targets_test = Variable(torch.FloatTensor(testing_target_df))
training_samples = utils_data.TensorDataset(inputs_train,
targets_train)
train_dataloader = utils_data.DataLoader(training_samples,
batch_size=200,
drop_last=False,
shuffle=False)
validation_samples = utils_data.TensorDataset(inputs_test,
targets_test)
val_dataloader = utils_data.DataLoader(validation_samples,
batch_size=200,
drop_last=False,
shuffle=False)
return train_dataloader, val_dataloader
def read_dataset(root_path_=str,
target_type=str,
read_target=str,
usage=str,
res_num=int,
res_type='dataloader',
selected=None,
bias_dataset=str): #,label=True
"""
func:read dataset to nets,get data to Nets,satisfied multifunction
:param root_path_:basedir of dataset
:param target_type: 'csv','pkl','txt'
:param read_target:'all','select'
:param usage: 'train','validate','test','coding'
:param res_num:refine how many result return for call
:param label: default dataloarder with label
:param res_type: default 'dataloader'
:param selected:if read_target== select,select the selected to read,could be 'Dos Fuzzy,RPM,gear' dataset file name
:param bias_dataset: whether a requirement of dataset is only normal or attack,it could be 'Normal','Attack','Both'
:return: list of dataloarder or single dataloarder contained all read dataset
"""
print(
'-----------------------------------%s,%s-----------------------------'
% (read_dataset.__name__, usage))
print('data address:{}, sub-dataset:{}'.format(root_path_,
os.listdir(root_path_)))
# selected attack type to read data
if selected != None:
selected = list(map(title, selected))
if read_target == 'all':
files = [os.path.join(root_path_, f) for f in os.listdir(root_path_)]
elif read_target == 'select':
files = [
os.path.join(root_path_, f) for f in os.listdir(root_path_)
if f.title() in selected
]
else:
print('func read_dataset: arise error at param read_target')
# dataset_urls = []
results = []
for file in files:
flag = 0
try:
for i in os.listdir(file):
if target_type in i:
flag += 1
pass
if flag == 0:
print('{} has not {} file'.format(file, target_type),
'please check the file folder')
print(os.listdir(file))
# dataset_urls.append(os.path.join(file,i))
except:
print(files, '\n error for target file folder')
return
pool = mp.Pool(processes=len(files))
for i in files:
# print(i)
# results.append(pool.apply(testdata,(os.path.join(path,i),mark,)))#_async
results.append(
pool.apply_async(base_read, (
i,
usage,
target_type,
bias_dataset,
))) # _async
pool.close()
pool.join()
names = []
flags = []
row = 0
column = results[0].get()[3]
# print('column:',column)
if res_type == 'seperate' or res_num > 1:
data = []
else:
data = np.empty((64, column))
f2 = lambda x: len(x)
# ll = 0
for i, result in enumerate(results):
# result = result#.get()
result = result.get()
# print('i:', i,'file:',result[2])
# # print('%s,%d'%(result[2],len(flg)))
# flags.append(result[0])
# data.append(result[1])
label_ = []
# ll = 0
for flg in result[0]:
# print(flg.__class__,len(flg))
row += len(flg)
if res_type == 'seperate':
flags.append(flg)
else:
label_.extend(flg)
# older codes
# if ll == 0 and i == 0:
# flags = flg
# # print('fg:',flg,flg.__class__)
# ll+=1
# else:
# # print('flags:',flags)
# flags.extend(flg)
# # print('%s,%d'%(result[2],len(flg)))
# concat normal status and attack status or single data suche as normal status or attack status
# of all types of attack to one container
# concat all to one container
if res_num == 1:
flags.extend(label_)
# concat to res_num containers,res_num default equal to the number of target_type
else:
flags.append(label_)
la = 0
dat = np.empty((1, 64, column))
for dt in result[1]:
if res_type == 'seperate':
data.append(
np.array(dt).astype(np.float64).reshape((-1, 64, column)))
else:
dt = np.array(dt).reshape((-1, 64, column)).astype(np.float64)
if la == 0:
dat = dt
la += 1
else:
dat = np.concatenate((dat, dt), axis=0).reshape(
(-1, 64, column))
# older codes
# if la == 0 and i == 0:
# data = dt
# la += 1
# else:
# data = np.concatenate((data,dt)).reshape((-1,64,column))
# concat normal status and attack status or single data suche as normal status or attack status
# of all types of attack to one container
# concat all to one container
if res_num == 1:
if i == 0:
data = dat
else:
data = np.concatenate((data, dat)).reshape((-1, 64, column))
# concat to res_num containers,res_num default equal to the number of target_type
else:
data.append(dat)
names.append(result[2])
# row += sum(list(map(f2,result[0])))
print('-' * 20, 'total result', '-' * 20)
print(
'\n return {} blocks of data,{} blocks of label,all {} blocks'.format(
data.__len__(), len(flags), row),
end=',')
if res_type == 'seperate':
return data, flags, names
print(names, end=',')
if res_num == 1:
# data_array = np.array(data).reshape((-1,64.21))
data_array = data
labels = np.array(flags).reshape((-1, 1)).astype(np.float64)
TraindataM = torch.from_numpy(
data_array).float() # transform to float torchTensor
TraindataM = torch.unsqueeze(TraindataM, 1)
Traindata_LabelM = torch.from_numpy(labels).float()
TorchDataset = Data.TensorDataset(TraindataM, Traindata_LabelM)
dataloader = Data.DataLoader(dataset=TorchDataset,
batch_size=BATCH_SIZE,
shuffle=True)
print(
'return one dataloarder with {} tensors,data shape:{},label shape:{}'
.format(len(dataloader.dataset.tensors),
dataloader.dataset.tensors[0].size(),
dataloader.dataset.tensors[1].size()))
print(
'------------------------------------------------------------------'
)
return dataloader, names
else:
print('result len:', len(data), len(flags))
dataloaders = []
f1 = lambda x: x.dataset.tensors[0].size()
# for i,label,dat in enumerate(list(zip(flags,data))):
for label, dat in list(zip(flags, data)):
dat = np.array(dat).reshape((-1, 64, column))
label = np.array(label).reshape((-1, 1))
TraindataM = torch.from_numpy(
dat).float() # transform to float torchTensor
TraindataM = torch.unsqueeze(TraindataM, 1)
Traindata_LabelM = torch.from_numpy(label).float()
TorchDataset = Data.TensorDataset(TraindataM, Traindata_LabelM)
dataloaders.append(
Data.DataLoader(dataset=TorchDataset,
batch_size=BATCH_SIZE,
shuffle=True))
print(
'return list of dataloader has {} dataloarders,data shape respectively:{}'
.format(len(dataloaders), list(map(f1, dataloaders))))
print(
'------------------------------------------------------------------'
)
return dataloaders, names
train_x = torch.from_numpy(train_x).type(torch.FloatTensor).cuda()
train_x = train_x / 255.0
train_y = torch.from_numpy(train_y).type(torch.int64).cuda()
test_x = torch.from_numpy(test_x).type(torch.FloatTensor).cuda()
test_x = test_x / 255.0
test_y = torch.from_numpy(test_y).type(torch.int64).cuda()
num_inputs = 784
num_outputs = 10
batch_size = 32
# 将训练数据的特征和标签组合
dataset = Data.TensorDataset(train_x, train_y)
# 把 dataset 放入 DataLoader
data_iter = Data.DataLoader(
dataset=dataset, # torch TensorDataset format
batch_size=batch_size, # mini batch size
shuffle=True, # 要不要打乱数据 (打乱比较好)
num_workers=0, # 多线程来读数据, 注意多线程需要在 if __name__ == '__main__': 函数中运行
)
# num_workers=0 表示不用额外的进程来加速读取数据
net = nn.Sequential(nn.Linear(num_inputs, num_outputs)).cuda()
loss_func = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.001)
testDIR = "D:/TestWithMFCC39/small data sets in the form of npy/test"
modelDIR = "D:/TestWithMFCC39/small data sets in the form of npy"
modelFILE = "reload_model.pth"
# 制作训练集
trainDS = np.load(train_npy_FILE)
#print("The shape of trainDS is {}".format(trainDS.shape))
trainDS = trainDS[np.newaxis, 0:TIME_STEP, :]
#print("The shape of trainDS is {}".format(trainDS.shape))
x_train = trainDS[:, :, 1:]
y_train = trainDS[:, :, 0]
y_train = torch.from_numpy(y_train).float()
x_train = torch.from_numpy(x_train).float()
trainDataSet = Data.TensorDataset(x_train, y_train)
trainLoader = Data.DataLoader(
dataset=trainDataSet, # torch TensorDataset format
batch_size=BATCH_SIZE, # mini batch size
shuffle=True, # 要不要打乱数据 (打乱比较好)
#num_workers=1, # 多线程来读数据
)
# 制作训练集
trainDS = np.load(train_npy_FILE)
print("The shape of trainDS is {}".format(trainDS.shape))
trainDS = trainDS[np.newaxis, 0:TIME_STEP, :]
print("The shape of trainDS is {}".format(trainDS.shape))
x_train = trainDS[:, :, 1:]
trainy = pickle.load(open("./lib/trainR.pkl", 'rb'))
testx = pickle.load(open("./lib/testD.pkl", 'rb'))
testy = pickle.load(open("./lib/testR.pkl", 'rb'))
return trainx, trainy, testx, testy
if __name__ == "__main__":
# Trans_net = nn.Linear(10,1)
model = RNN(input_size, hidden_size, num_layers).to(device)
trainx, trainy, testx, testy = get_data()
train1, train2, std1 = trans(trainx, trainy)
test1, test2, std2 = trans(testx, testy)
print(train1.shape)
print(std1.shape)
train_set = Data.TensorDataset(train1, train2, std1)
train_loader = Data.DataLoader(dataset=train_set,
batch_size=batch_size,
shuffle=False,
num_workers=0)
test_set = Data.TensorDataset(test1, test2, std2)
test_loader = Data.DataLoader(dataset=test_set,
batch_size=batch_size,
shuffle=False,
num_workers=0)
# std = torch.Tensor([[3.8],[4.6]])
milestone_list = [50, 150]
optimizer = torch.optim.Adam(model.parameters(), lr=LR)
lr_scheduler = MultiStepLR(optimizer, milestones=milestone_list, gamma=0.1)
import numpy as np
import torch
from torch.utils import data
dataset = np.load('./expert.npz')
tensor_dataset = data.TensorDataset(torch.Tensor(dataset['obs']), torch.Tensor(dataset['action']))
dataloader = data.DataLoader(tensor_dataset, batch_size=50, shuffle=True)
TRAIN_X = VECTORIZER.transform(TRAIN_X_RAW).todense()
for index, row in enumerate(TRAIN_X):
if np.sum(row) < 1e-2:
TRAIN_X[index, :] = np.ones((1, VOCABULARY_SIZE),dtype=np.float32)
TRAIN_X = TRAIN_X / TRAIN_X.sum(axis=1)
TRAIN_X = np.matmul(EMBEDDINGS, TRAIN_X.T)
DEV_X = VECTORIZER.transform(DEV_X_RAW).T
DEV_X = DEV_X / DEV_X.sum(axis=0)
DEV_X = np.matmul(EMBEDDINGS, DEV_X)
TEST_X = VECTORIZER.transform(TEST_X_RAW).T
TEST_X = TEST_X / TEST_X.sum(axis=0)
TEST_X = np.matmul(EMBEDDINGS, TEST_X)
TRAIN_DATA = data_utils.TensorDataset(torch.from_numpy(TRAIN_X.T.astype(np.float32)), torch.from_numpy(TRAIN_Y))
TRAIN_LOADER = data_utils.DataLoader(TRAIN_DATA, batch_size=173,
shuffle=True, num_workers=2, drop_last=True)
DEV_DATA = data_utils.TensorDataset(torch.from_numpy(DEV_X.T.astype(np.float32)), torch.from_numpy(DEV_Y))
DEV_LOADER = data_utils.DataLoader(DEV_DATA, batch_size=1,
shuffle=False, num_workers=2)
TEST_DATA = data_utils.TensorDataset(torch.from_numpy(TEST_X.T.astype(np.float32)), torch.from_numpy(TEST_Y))
TEST_LOADER = data_utils.DataLoader(TEST_DATA, batch_size=1,
shuffle=False, num_workers=2)
##################################################################
# Define the network
##################################################################
class SentimentNet(nn.Module):
def __init__(self, hidden_dim):
super(SentimentNet, self).__init__()
self.linear1 = nn.Linear(EMBEDDING_SIZE, hidden_dim)
Y_train = pd.get_dummies(bdf['label']).values
print(Y_train)
print("=" * 80)
print(len(X_train))
print(len(Y_train))
# rf.fit(X_train, Y_train)
# Y_test = rf.predict(X_test)
# result = np.argmax(Y_test, axis = 1)
# import tensorflow as tf
import torch
import torch.utils.data as Data
torch_dataset = Data.TensorDataset(torch.from_numpy(X_train).double(),
torch.from_numpy(Y_train).double())
train_loader = Data.DataLoader(
dataset=torch_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
)
import torch.nn as nn
from math import sqrt
import torch
class NNmodel(nn.Module):
def __init__(self):
super().__init__()
self.W = nn.Parameter(torch.Tensor(1, 12))
# _*_ coding: utf-8 _*_
__author__ = 'LelandYan'
__date__ = '2019/7/21 16:10'
import torch
import torch.utils.data as Data
BATCH_SIZE = 8
x = torch.linspace(1,10,10)
y = torch.linspace(10,1,10)
torch_dataset = Data.TensorDataset(x,y)
loader = Data.DataLoader(dataset=torch_dataset,shuffle=True,batch_size=BATCH_SIZE,num_workers=2)
if __name__ == '__main__':
pass
for epoch in range(3):
for step,(batch_x,batch_y) in enumerate(loader):
print("Epoch: ",epoch,"| Step: ",step,"| batch x: ",batch_x.numpy(),"| batch_y: ",batch_y.numpy())
# In[47]:
model = RNN(input_size, hidden_size, num_layers, num_classes)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# In[54]:
input_train = torch.load("input_train.pt")
_label_train = torch.load("label_train.pt")
_input_test = torch.load("input_test.pt")
_label_test = torch.load("label_test.pt")
print(input_train.pop())
train = data_utils.TensorDataset(_input_train, _label_train)
train_loader = data_utils.DataLoader(train,
batch_size=batch_size,
shuffle=True)
# test = data_utils.TensorDataset(_input_test, _label_test)
# test_loader = data_utils.DataLoader(test, shuffle=True)
# total_step = len(train_loader)
epoch_start = time.time()
loss = 0
all_losses = []
# for epoch in range(num_epochs):
# i is the counter, ith batch, j is the value of batch
# for i,(feature, label) in enumerate(train_loader):
# feature = feature.reshape(-1, sequence_length, input_size)
# print (feature.shape)
# # Forward pass