Train model on Natural Language Inference task
"""
epoch = 1
if params.load_model:
encoder_net.load_state_dict(
torch.load('../encoder_' + params.load_id + 'saved.pt'))
if params.load_classifier:
nli_net.load_state_dict(
torch.load('../classifier_' + params.load_id + 'saved.pt'))
while (not stop_training
or epoch <= params.min_epoch) and epoch <= params.n_epochs:
train_acc = trainepoch(epoch)
eval_acc, adv_models, bias_models, bias_loss_functions, bias_list, bias_optimizers = evaluate(
epoch, adv_models, bias_models, bias_loss_functions, bias_list,
bias_optimizers, optim_params, 'valid')
epoch += 1
print("Evaluating on test-set")
evaluate(0, adv_models, bias_models, bias_loss_functions, bias_list,
bias_optimizers, optim_params, 'test', True)
torch.save(encoder_net.state_dict(),
'../encoder_' + params.save_id + 'saved.pt')
torch.save(nli_net.state_dict(),
'../classifier_' + params.save_id + 'saved.pt')
}
# model
encoder_types = ['BLSTMEncoder', 'BLSTMprojEncoder', 'BGRUlastEncoder',
'InnerAttentionMILAEncoder', 'InnerAttentionYANGEncoder',
'InnerAttentionNAACLEncoder', 'ConvNetEncoder', 'LSTMEncoder']
assert params.encoder_type in encoder_types, "encoder_type must be in " + \
str(encoder_types)
nli_net = NLINet(config_nli_model)
if params.pre_trained_model:
print "Pre_trained_model: " + params.pre_trained_model
pre_trained_model = torch.load(params.pre_trained_model)
nli_net_params = nli_net.state_dict()
pre_trained_params = pre_trained_model.state_dict()
assert nli_net_params.keys() == pre_trained_params.keys(), "load model has different parameter state names that NLI_HYPOTHS_NET"
for key, parameters in nli_net_params.items():
if parameters.size() == pre_trained_params[key].size():
nli_net_params[key] = pre_trained_params[key]
nli_net.load_state_dict(nli_net_params)
print(nli_net)
# loss
weight = torch.FloatTensor(params.n_classes).fill_(1)
loss_fn = nn.CrossEntropyLoss(weight=weight)
loss_fn.size_average = False
# optimizer