def setup_models(io, do_load, cuda, trainer=True):
models = dict()
if trainer:
name = "training"
else:
name = "best"
if do_load:
models["emb"] = load_model("emb_{}".format(name), io)
models["readout"] = load_model("readout_{}".format(name), io)
models["policy"] = load_model("policy_{}".format(name), io)
models["backup"] = load_model("backup_{}".format(name), io)
else:
models["emb"] = None
models["readout"] = None
models["policy"] = None
models["backup"] = None
for model_name, model in models.items():
if model is None:
models[model_name] = create_net(model_name)
train_mode(models)
if cuda:
activate_cuda(models)
return models
def train(datasetType,
dataset_path,
modelType,
dataLoaderType,
trainingType,
max_epochs,
max_epochs_without_improvement,
learning_rate,
momentum,
batch_size,
device,
printProgress=False):
# Load Data
path = dataset_path + datasetType.name + ".hdf5"
trainLoader, testLoader, classes = datasets.create_dataloaders(
dataLoaderType, path, batch_size, device, dataset=datasetType)
# Create net and move to correct device
in_channels = 3 if datasetType is datasets.DataSet.CIFAR10 else 1
num_classes = 10
net = models.create_net(modelType, in_channels, num_classes)
net.to(device)
# Training
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(net.parameters(),
lr=learning_rate,
momentum=momentum)
trainObj = training.create_training(trainingType, net, criterion,
optimizer, trainLoader, testLoader,
max_epochs,
max_epochs_without_improvement,
printProgress)
trainObj.train()
return trainObj
cosine_annealing = DictConfig([cosine_annealing,cosine_annealing_ops])
plateau = NamedConfig(("scheduler",optim.lr_scheduler.ReduceLROnPlateau))
ca_T_max = NamedConfig(("patience",5))
plateau_ops = NamedConfig(("scheduler_params",DictConfig([ca_T_max])))
plateau = DictConfig([plateau,plateau_ops])
schedulers = ExclusiveConfigs([cosine_annealing,plateau])
evaluator_obj = Evaluator("./models/mnist_classifier",inference_fn=debug_fn)
training_modes = NamedConfig(("modes",[1]))
max_epochs = NamedConfig(("max_epochs",40))
training_objectives = NamedConfig((1,SupervisedMetricList([[nn.CrossEntropyLoss()]],[[1.0]])))
validation_objectives = NamedConfig((0,SupervisedMetricList([[Accuracy()]],[[1.0]])))
objective_fns = NamedConfig(("objective_fns",DictConfig([training_objectives,validation_objectives])))
constants = DictConfig([objective_fns,training_modes,max_epochs])
network = NamedConfig(("network",create_net(CustomNetClassification)))
growth_factor = NamedConfig(("growth_factor",ExclusiveConfigs([10])))
input_dim = NamedConfig(("input_dim",28))
final_dim = NamedConfig(("final_conv_dim",8))
initial_channels = NamedConfig(("initial_channels",1))
num_classes = NamedConfig(("num_classes",10))
conv_module = NamedConfig(("conv_module",ExclusiveConfigs([DoubleConvLayer,InceptionLayer])))
network_params = NamedConfig(("network_params",DictConfig([growth_factor,input_dim,final_dim,initial_channels,num_classes,conv_module])))
train_loader_options = NamedConfig((1,{"shuffle":True,"batch_size":1000,"num_workers":4}))
val_loader_options = NamedConfig((0,{"batch_size":1000,"num_workers":4}))
loader_options = NamedConfig(("loader_options",DictConfig(
[train_loader_options,val_loader_options])))
network_loader_set1 = DictConfig([network_params,loader_options])
schedulers = ExclusiveConfigs([cosine_annealing, plateau])
evaluator_obj = Evaluator("./models/mnist_classifier",
inference_fn=debug_fn)
training_modes = NamedConfig(("modes", [1]))
max_epochs = NamedConfig(("max_epochs", 1))
training_objectives = NamedConfig(
(1, SupervisedMetricList([[nn.CrossEntropyLoss()]], [[1.0]])))
validation_objectives = NamedConfig(
(0, SupervisedMetricList([[nn.CrossEntropyLoss()]], [[1.0]])))
objective_fns = NamedConfig(
("objective_fns",
DictConfig([training_objectives, validation_objectives])))
constants = DictConfig([objective_fns, training_modes, max_epochs])
network = NamedConfig(("network", create_net(PreTrainedClassifier)))
num_classes = NamedConfig(("num_classes", 120))
model_class = NamedConfig(("model_class", DensenetModels))
model = NamedConfig(("model", models.densenet121))
network_params = NamedConfig(
("network_params", DictConfig([model, model_class, num_classes])))
train_loader_options = NamedConfig((1, {
"shuffle": True,
"batch_size": 20,
"num_workers": 4
}))
val_loader_options = NamedConfig((0, {"batch_size": 20, "num_workers": 4}))
loader_options = NamedConfig(
("loader_options",