def pre_train(self, dataset, **kwargs):
train_x = []
train_y = []
for subj_data in dataset:
for task_data in subj_data:
syllogism = ccobra.syllogistic.Syllogism(task_data['item'])
# Encode the task input
task = onehot.onehot_syllogism_content(syllogism.encoded_task)
# Encode the response output
encoded_response = syllogism.encode_response(
task_data['response'])
resp = onehot.onehot_response(encoded_response)
train_x.append(task)
train_y.append(resp)
self.train_x = torch.from_numpy(np.array(train_x)).float()
self.train_y = torch.from_numpy(np.array(train_y)).float()
self.train_network(self.train_x,
self.train_y,
self.batch_size,
self.n_epochs,
verbose=True)
def pre_train(self, dataset):
# Prepare the data for training by converting it into a 64 x n_subj x 12
train_x = []
train_y = []
for subj_data in dataset:
subj_train_x = []
subj_train_y = []
for task_data in subj_data:
syllogism = ccobra.syllogistic.Syllogism(task_data['item'])
# Onehot encodings
onehot_task = onehot.onehot_syllogism_content(syllogism.encoded_task)
onehot_response = onehot.onehot_response(
syllogism.encode_response(task_data['response']))
subj_train_x.append(onehot_task)
subj_train_y.append(onehot_response)
train_x.append(subj_train_x)
train_y.append(subj_train_y)
self.train_x = torch.from_numpy(np.array(train_x)).float()
self.train_y = torch.from_numpy(np.array(train_y)).float()
self.train_network(self.train_x, self.train_y, self.n_epochs, verbose=True)
def predict(self, item, **kwargs):
syllogism = ccobra.syllogistic.Syllogism(item)
# Obtain the prediction
input = torch.from_numpy(onehot.onehot_syllogism_content(syllogism.encoded_task)).float()
output, self.hidden = self.net(input.view(1, 1, -1), self.hidden)
# Return maximum response
response = output.argmax().item()
enc_response = ccobra.syllogistic.RESPONSES[response]
return syllogism.decode_response(enc_response)
def predict(self, item, **kwargs):
# Encode the task
syllogism = ccobra.syllogistic.Syllogism(item)
# Query the model
inp = onehot.onehot_syllogism_content(syllogism.encoded_task)
inp_tensor = torch.from_numpy(inp).float()
output = self.net(inp_tensor)
# Return maximum response
response = output.argmax().item()
enc_response = ccobra.syllogistic.RESPONSES[response]
return syllogism.decode_response(enc_response)
def adapt(self, item, truth, **kwargs):
syllogism = ccobra.syllogistic.Syllogism(item)
# Onehot encoding
onehot_syl = onehot.onehot_syllogism_content(syllogism.encoded_task)
onehot_resp = onehot.onehot_response(syllogism.encode_response(truth))
adapt_x = torch.from_numpy(onehot_syl.reshape(1, -1)).float()
adapt_y = torch.from_numpy(onehot_resp.reshape(1, -1)).float()
self.train_network(adapt_x,
adapt_y,
1,
self.n_epochs_adapt,
verbose=False)