def _setup(self, config):
if config["dataset"] == "BavarianCrops":
region = "HOLL_2018_MT_pilot"
root = "/home/marc/data/BavarianCrops"
nsamples = None
traindataset = BavarianCropsDataset(root=root,
region=region,
partition="train",
nsamples=nsamples)
validdataset = BavarianCropsDataset(root=root,
region=region,
partition="valid",
nsamples=nsamples)
else:
traindataset = UCRDataset(config["dataset"],
partition="train",
ratio=.8,
randomstate=config["fold"],
silent=False,
augment_data_noise=0)
validdataset = UCRDataset(config["dataset"],
partition="valid",
ratio=.8,
randomstate=config["fold"],
silent=False)
nclasses = traindataset.nclasses
self.epochs = config["epochs"]
# handles multitxhreaded batching andconfig shuffling
self.traindataloader = torch.utils.data.DataLoader(
traindataset,
batch_size=config["batchsize"],
shuffle=True,
num_workers=config["workers"],
pin_memory=False)
self.validdataloader = torch.utils.data.DataLoader(
validdataset,
batch_size=config["batchsize"],
shuffle=False,
num_workers=config["workers"],
pin_memory=False)
self.model = DualOutputRNN(input_dim=traindataset.ndims,
nclasses=nclasses,
hidden_dims=config["hidden_dims"],
num_rnn_layers=config["num_layers"])
if torch.cuda.is_available():
self.model = self.model.cuda()
self.trainer = Trainer(self.model, self.traindataloader,
self.validdataloader, **config)
def test_Trainer_TwoPatterns(self):
try:
traindataset = UCRDataset("TwoPatterns",
partition="train",
ratio=.75,
randomstate=0,
augment_data_noise=.1)
validdataset = UCRDataset("TwoPatterns",
partition="valid",
ratio=.75,
randomstate=0)
nclasses = traindataset.nclasses
traindataloader = torch.utils.data.DataLoader(traindataset,
batch_size=8,
shuffle=True,
num_workers=0,
pin_memory=True)
validdataloader = torch.utils.data.DataLoader(validdataset,
batch_size=8,
shuffle=False,
num_workers=0,
pin_memory=True)
model = DualOutputRNN(input_dim=1,
nclasses=nclasses,
hidden_dims=20,
num_rnn_layers=1,
dropout=.2)
if torch.cuda.is_available():
model = model.cuda()
config = dict(
epochs=2,
learning_rate=1e-3,
earliness_factor=.75,
visdomenv="unittest",
switch_epoch=1,
loss_mode="twophase_linear_loss",
show_n_samples=0,
store="/tmp",
overwrite=True,
)
trainer = Trainer(model, traindataloader, validdataloader,
**config)
trainer.fit()
except Exception as e:
self.fail(logging.exception(e))
def getModel(args):
# Get Model
if args.model == "DualOutputRNN":
model = DualOutputRNN(input_dim=args.input_dims,
nclasses=args.nclasses,
hidden_dims=args.hidden_dims,
num_rnn_layers=args.num_layers,
dropout=args.dropout,
init_late=True)
elif args.model == "WaveNet":
model = WaveNetModel(layers=5,
blocks=4,
dilation_channels=32,
residual_channels=32,
skip_channels=256,
end_channels=args.nclasses,
classes=args.nclasses,
output_length=1,
kernel_size=2,
dtype=torch.FloatTensor,
input_dims=args.input_dims,
bias=False)
elif args.model == "Conv1D":
model = ConvShapeletModel(
num_layers=args.num_layers,
hidden_dims=args.hidden_dims,
ts_dim=args.input_dims,
n_classes=args.nclasses,
use_time_as_feature=False,
seqlength=args.seqlength,
scaleshapeletsize=args.shapelet_width_in_percent,
drop_probability=args.dropout,
shapelet_width_increment=args.shapelet_width_increment)
else:
raise ValueError(
"Invalid Model, Please insert either 'DualOutputRNN', or 'Conv1D'")
if torch.cuda.is_available():
model = model.cuda()
return model
class RayTrainerDualOutputRNN(ray.tune.Trainable):
def _setup(self, config):
if config["dataset"] == "BavarianCrops":
region = "HOLL_2018_MT_pilot"
root = "/home/marc/data/BavarianCrops"
nsamples = None
traindataset = BavarianCropsDataset(root=root,
region=region,
partition="train",
nsamples=nsamples)
validdataset = BavarianCropsDataset(root=root,
region=region,
partition="valid",
nsamples=nsamples)
else:
traindataset = UCRDataset(config["dataset"],
partition="train",
ratio=.8,
randomstate=config["fold"],
silent=False,
augment_data_noise=0)
validdataset = UCRDataset(config["dataset"],
partition="valid",
ratio=.8,
randomstate=config["fold"],
silent=False)
nclasses = traindataset.nclasses
self.epochs = config["epochs"]
# handles multitxhreaded batching andconfig shuffling
self.traindataloader = torch.utils.data.DataLoader(
traindataset,
batch_size=config["batchsize"],
shuffle=True,
num_workers=config["workers"],
pin_memory=False)
self.validdataloader = torch.utils.data.DataLoader(
validdataset,
batch_size=config["batchsize"],
shuffle=False,
num_workers=config["workers"],
pin_memory=False)
self.model = DualOutputRNN(input_dim=traindataset.ndims,
nclasses=nclasses,
hidden_dims=config["hidden_dims"],
num_rnn_layers=config["num_layers"])
if torch.cuda.is_available():
self.model = self.model.cuda()
self.trainer = Trainer(self.model, self.traindataloader,
self.validdataloader, **config)
def _train(self):
# epoch is used to distinguish training phases. epoch=None will default to (first) cross entropy phase
# train five epochs and then infer once. to avoid overhead on these small datasets
for i in range(self.epochs):
self.trainer.train_epoch(epoch=None)
return self.trainer.test_epoch(dataloader=self.validdataloader)
def _save(self, path):
path = path + ".pth"
torch.save(self.model.state_dict(), path)
return path
def _restore(self, path):
state_dict = torch.load(path, map_location="cpu")
self.model.load_state_dict(state_dict)
def prepare_model_and_optimizer(args, input_dims, sequence_length,
num_classes):
if args.model == "DualOutputRNN":
hparams = pd.read_csv(config.DUALOUTPUTRNN_HYPERPARAMETER_CSV,
index_col=0)
if args.dataset not in hparams.index:
raise ValueError(
f"No dataset {args.dataset} found in {config.DUALOUTPUTRNN_HYPERPARAMETER_CSV}"
)
hparam = hparams.loc[args.dataset]
model = DualOutputRNN(input_dim=int(input_dims),
nclasses=num_classes,
hidden_dims=int(hparam.hidden_dims),
num_rnn_layers=int(hparam.num_layers),
dropout=hparam.dropout,
init_late=True)
optimizer = torch.optim.Adam(model.parameters(),
lr=hparam.learning_rate,
weight_decay=hparam.weight_decay)
#optimizer = ScheduledOptim(optimizer, d_model=model.d_model, n_warmup_steps=500)
return model, optimizer
elif args.model == "WaveNet":
model = WaveNetModel(layers=5,
blocks=4,
dilation_channels=32,
residual_channels=32,
skip_channels=256,
end_channels=args.nclasses,
classes=args.nclasses,
output_length=1,
kernel_size=2,
dtype=torch.FloatTensor,
input_dims=args.input_dims,
bias=False)
elif args.model == "Conv1D":
hparams = pd.read_csv(config.CONV1D_HYPERPARAMETER_CSV, index_col=0)
if args.dataset not in hparams.index:
raise ValueError(
f"No dataset {args.dataset} found in {config.CONV1D_HYPERPARAMETER_CSV}"
)
hparam = hparams.loc[args.dataset]
model = ConvShapeletModel(num_layers=int(hparam.num_layers),
hidden_dims=int(hparam.hidden_dims),
ts_dim=int(input_dims),
n_classes=int(num_classes),
drop_probability=hparam.dropout,
scaleshapeletsize=False,
seqlength=int(sequence_length),
shapelet_width_increment=int(
hparam.shapelet_width_increment))
optimizer = torch.optim.Adam(model.parameters(),
lr=hparam.learning_rate,
weight_decay=hparam.weight_decay)
return model, optimizer
else:
raise ValueError(
"Invalid Model, Please insert either 'DualOutputRNN', or 'Conv1D'")
def eval(dataset,
batchsize,
workers,
num_rnn_layers,
dropout,
hidden_dims,
store="/tmp",
epochs=30,
switch_epoch=30,
learning_rate=1e-3,
visdomenv="run",
earliness_factor=.75,
show_n_samples=1,
modelname="DualOutputRNN",
loss_mode=None,
load_weights=None,
entropy_factor=0):
if dataset == "synthetic":
traindataset = SyntheticDataset(num_samples=2000, T=100)
validdataset = SyntheticDataset(num_samples=1000, T=100)
else:
traindataset = UCRDataset(dataset, partition="trainvalid")
validdataset = UCRDataset(dataset, partition="test")
nclasses = traindataset.nclasses
np.random.seed(0)
torch.random.manual_seed(0)
traindataloader = torch.utils.data.DataLoader(traindataset,
batch_size=batchsize,
shuffle=True,
num_workers=workers,
pin_memory=True)
np.random.seed(1)
torch.random.manual_seed(1)
validdataloader = torch.utils.data.DataLoader(validdataset,
batch_size=batchsize,
shuffle=False,
num_workers=workers,
pin_memory=True)
if modelname == "DualOutputRNN":
model = DualOutputRNN(input_dim=1,
nclasses=nclasses,
hidden_dim=hidden_dims,
num_rnn_layers=num_rnn_layers,
dropout=dropout)
else:
raise ValueError(
"Invalid Model, Please insert either 'DualOutputRNN' or 'AttentionRNN'"
)
if load_weights is not None:
model.load(load_weights)
# parse epoch 29 from filename e.g., 'models/TwoPatterns/run/model_29.pth'
start_epoch = int(
os.path.basename(load_weights).split("_")[-1].split(".")[0])
else:
start_epoch = 0
if torch.cuda.is_available():
model = model.cuda()
#if run is None:
# visdomenv = "{}_{}_{}".format(args.experiment, dataset,args.loss_mode.replace("_","-"))
# storepath = store
#else:
# visdomenv = run
#storepath = os.path.join(store, dataset)
if switch_epoch is None:
switch_epoch = int(epochs / 2)
config = dict(epochs=epochs,
learning_rate=learning_rate,
earliness_factor=earliness_factor,
visdomenv=visdomenv,
switch_epoch=switch_epoch,
loss_mode=loss_mode,
show_n_samples=show_n_samples,
store=store,
entropy_factor=entropy_factor)
trainer = Trainer(model, traindataloader, validdataloader, config=config)
logged_data = trainer.fit(start_epoch=start_epoch)
return logged_data