Serializing a Trial¶
This guide will explain the two different ways to how to save and reload your results from a Trial.
Setting up a Mock Example¶
Let’s assume we have a basic binary classification task where we have 100-dimensional samples as input and a binary label as output. Let’s also assume that we would like to solve this problem with a 2-layer neural network. Finally, we also want to keep track of the sum of hidden outputs for some arbitrary reason. Therefore we use the state functionality of Torchbearer.
We create a state key for the mock sum we wanted to track using state.
MOCK = torchbearer.state_key('mock')
Here is our basic 2-layer neural network.
class BasicModel(nn.Module):
def __init__(self):
super(BasicModel, self).__init__()
self.linear1 = nn.Linear(100, 25)
self.linear2 = nn.Linear(25, 1)
def forward(self, x, state):
x = self.linear1(x)
# The following step is here to showcase a useless but simple of example a forward method that uses state
state[MOCK] = torch.sum(x)
x = self.linear2(x)
return torch.sigmoid(x)
We create some random training dataset and put them in a DataLoader.
n_sample = 100
X = torch.rand(n_sample, 100)
y = torch.randint(0, 2, [n_sample, 1]).float()
traingen = DataLoader(TensorDataset(X, y))
Let’s say we would like to save the model every time we get a better training loss. Torchbearer’s Best checkpoint callback is perfect for this job. We then run the model for 3 epochs.
model = BasicModel()
# Create a checkpointer that track val_loss and saves a model.pt whenever we get a better loss
checkpointer = torchbearer.callbacks.checkpointers.Best(filepath='model.pt', monitor='loss')
optimizer = optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=0.001)
torchbearer_trial = Trial(model, optimizer=optimizer, criterion=F.binary_cross_entropy, metrics=['loss'],
callbacks=[checkpointer])
torchbearer_trial.with_train_generator(traingen)
torchbearer_trial.run(epochs=3)
Reloading the Trial for More Epochs¶
Given we recreate the exact same Trial structure, we can easily resume our run from the last checkpoint. The following code block shows how it’s done.
Remember here that the epochs
parameter we pass to Trial acts cumulative. In other words, the following run will complement the entire training to
a total of 6 epochs.
state_dict = torch.load('model.pt')
model = BasicModel()
trial_reloaded = Trial(model, optimizer=optimizer, criterion=F.binary_cross_entropy, metrics=['loss'],
callbacks=[checkpointer])
trial_reloaded.load_state_dict(state_dict)
trial_reloaded.with_train_generator(traingen)
trial_reloaded.run(epochs=6)
Trying to Reload to a PyTorch Module¶
We try to load the state_dict
to a regular PyTorch Module, as described in PyTorch’s own documentation here:
model = BasicModel()
try:
model.load_state_dict(state_dict)
except AttributeError as e:
print("\n")
print(e)
We will get the following error:
'StateKey' object has no attribute 'startswith'
The reason is that the state_dict
has Trial related attributes that are unknown to a native PyTorch model. This is why we have the save_model_params_only
option for our checkpointers. We try again with that option
model = BasicModel()
checkpointer = torchbearer.callbacks.checkpointers.Best(filepath='model.pt', monitor='loss', save_model_params_only=True)
optimizer = optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=0.001)
torchbearer_trial = Trial(model, optimizer=optimizer, criterion=F.binary_cross_entropy, metrics=['loss'],
callbacks=[checkpointer])
torchbearer_trial.with_train_generator(traingen)
torchbearer_trial.run(epochs=3)
# Try once again to load the module, forward another random sample for testing
state_dict = torch.load('model.pt')
model = BasicModel()
model.load_state_dict(state_dict)
No errors this time, but we still have to test. Here is a test sample and we run it through the model.
X_test = torch.rand(5, 100)
try:
model(X_test)
except TypeError as e:
print("\n")
print(e)
forward() missing 1 required positional argument: 'state'
Now we get a different error, stating that we should also be passing state
as an argument to module’s forward. This should not be a surprise
as we defined state
parameter in the forward method of BasicModule
as a required argument.
Robust Signature for Module¶
We define the model with a better signature this time, so it gracefully handles the problem above.
class BetterSignatureModel(nn.Module):
def __init__(self):
super(BetterSignatureModel, self).__init__()
self.linear1 = nn.Linear(100, 25)
self.linear2 = nn.Linear(25, 1)
def forward(self, x, **state):
x = self.linear1(x)
# Using kwargs instead of state is safer from a serialization perspective
if state is not None:
state = state
state[MOCK] = torch.sum(x)
x = self.linear2(x)
return torch.sigmoid(x)
Finally, we wrap it up once again to test the new definition of the model.
model = BetterSignatureModel()
checkpointer = torchbearer.callbacks.checkpointers.Best(filepath='model.pt', monitor='loss', save_model_params_only=True)
optimizer = optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=0.001)
torchbearer_trial = Trial(model, optimizer=optimizer, criterion=F.binary_cross_entropy, metrics=['loss'],
callbacks=[checkpointer])
torchbearer_trial.with_train_generator(traingen)
torchbearer_trial.run(epochs=3)
# This time, the forward function should work without the need for a state argument
state_dict = torch.load('model.pt')
model = BetterSignatureModel()
model.load_state_dict(state_dict)
X_test = torch.rand(5, 100)
model(X_test)