float_keys,
shuffle=True)
valid_batches = batch_iter(valid_data,
len(valid_data),
embed_keys,
float_keys,
shuffle=False)
test_batches = batch_iter(test_data,
len(test_data),
embed_keys,
float_keys,
shuffle=False)
model = MultiNet(len(float_keys), embed_dims, embed_sizes, hidden_dim,
embed_keys, num_class)
model = set_cuda(model)
opt = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
start = time.time()
best_epoch, best_tst_ret, best_med_ret, best_std_ret = train(
model, num_batch, train_batches, valid_batches, test_batches, opt,
num_epochs, hidden_dim)
print(
'Best test return: %.3f, median: %.3f, std: %.3f in epoch %d. Finished in %.3f seconds'
% (best_tst_ret, best_med_ret, best_std_ret, best_epoch,
time.time() - start))
print('Training finished')
max_acc = acc
outfile = os.path.join(params.checkpoint_dir, 'best_model.tar')
torch.save({'epoch': epoch, 'state': model.state_dict()}, outfile)
if (epoch % params.save_freq == 0) or (epoch == stop_epoch - 1):
outfile = os.path.join(params.checkpoint_dir,
'{:d}.tar'.format(epoch))
torch.save({'epoch': epoch, 'state': model.state_dict()}, outfile)
return model
if __name__ == '__main__':
np.random.seed(10)
params = parse_args('train')
set_cuda(params.cuda)
if params.dataset == 'cross':
base_file = configs.data_dir['miniImagenet'] + 'all.json'
val_file = configs.data_dir['CUB'] + 'val.json'
elif params.dataset == 'cross_char':
base_file = configs.data_dir['omniglot'] + 'noLatin.json'
val_file = configs.data_dir['emnist'] + 'val.json'
else:
base_file = configs.data_dir[params.dataset] + 'base.json'
val_file = configs.data_dir[params.dataset] + 'val.json'
if 'Conv' in params.model:
if params.dataset in ['omniglot', 'cross_char']:
image_size = 28
else:
torch.save(model.state_dict(), 'em_model.pt')
return best_auc, best_acc
num_samples = X_train.shape[0]
batch_size = 1000
num_batch = int(num_samples / batch_size)
hidden_dim = 4
num_epochs = 50
learning_rate = 1e-2
weight_decay = 5e-5
rpt_step = 10000
test_step = 50000
ratio_0 = sum([1 for t in y_test if t == 0]) / y_train.shape[0]
ratio_1 = sum([1 for t in y_test if t == 1]) / y_train.shape[0]
weight = set_cuda(
torch.FloatTensor([1 / ratio_0, 1 / ratio_1, 1 / (1 - ratio_0 - ratio_1)]))
num_class = 3
train_batches = batch_iter(X_train,
y_train,
batch_size,
c1=GBC,
c2=rf_en,
shuffle=True)
test_batches = batch_iter(X_test,
y_test,
X_test.shape[0],
c1=GBC,
c2=rf_en,
shuffle=False)