def train():
_, _, relation_data, profile_data, _ = utils.dataloader.load_data(
args.i, train=True)
agglo = relations_agglomerator()
agglo.train(relation_data, profile_data)
save_model("relation_agglo", args.o, agglo)
print("Model saved")
model = None
if MODEL_TOKEN == 'sahp':
with autograd.detect_anomaly():
params = args, process_dim, device, tmax, \
train_times_tensor, train_seq_types, train_seq_lengths, \
dev_times_tensor, dev_seq_types, dev_seq_lengths, \
test_times_tensor, test_seq_types, test_seq_lengths, \
BATCH_SIZE, EPOCHS, USE_CUDA
model = train_eval_sahp(params)
else:
exit()
if args.save_model:
# Model file dump
SAVED_MODELS_PATH = os.path.abspath('saved_models')
os.makedirs(SAVED_MODELS_PATH, exist_ok=True)
# print("Saved models directory: {}".format(SAVED_MODELS_PATH))
date_format = "%Y%m%d-%H%M%S"
now_timestamp = datetime.datetime.now().strftime(date_format)
extra_tag = "{}".format(args.task)
filename_base = "{}-{}_hidden{}-{}".format(MODEL_TOKEN, extra_tag,
hidden_size, now_timestamp)
from utils.save_model import save_model
save_model(model, chosen_file, extra_tag, hidden_size, now_timestamp,
MODEL_TOKEN)
print('Done! time elapsed %.2f sec for %d epoches' %
(time.time() - start_time, EPOCHS))
def save_model_(dir, method, model, timestamp):
folder = os.path.join(dir, method['name'] + '_' + timestamp)
save_model(model, folder)
create_dir(method_model_dir)
if SAVE_PREDS:
create_dir(method_pred_dir)
start_time = time()
classifier, hist = train(method, train_provider, val_provider)
train_time = (time() - start_time) / num_epochs
# save_train_results(method_results_dir, hist, train_time)
start_time = time()
acc, per_class_acc, mean_class_acc, mIoU, per_class_iou, mean_class_iou, conf_mat, data_, part_labels_, \
pred_part_labels_ = \
test(classifier, test_provider, method=method, dataset=dataset)
test_time = (time() - start_time)
save_test_results(method_results_dir, acc, per_class_acc,
mean_class_acc, mIoU, per_class_iou, mean_class_iou,
conf_mat, test_time)
# save_results(method_results_dir, hist, conf_mat, test_acc, train_time, test_time)
if SAVE_PREDS:
save_part_labels(method_pred_dir,
dataset['name'] + '_' + method['name'],
pred_part_labels_, part_labels_, data_)
if SAVE_MODELS:
save_model(method_model_dir, classifier)
elif args.model == 'lstm':
loss = train_lstm(
model,
optimizer,
seq_times,
seq_types,
seq_lengths,
-1, # tmax actually doesn't matter
BATCH_SIZE,
EPOCHS,
use_cuda=args.cuda)
# Model file dump
SAVED_MODELS_PATH = os.path.abspath('saved_models')
os.makedirs(SAVED_MODELS_PATH, exist_ok=True)
# print("Saved models directory: {}".format(SAVED_MODELS_PATH))
date_format = "%Y%m%d-%H%M%S"
now_timestamp = datetime.datetime.now().strftime(date_format)
extra_tag = "{}d".format(process_dim)
filename_base = "{}-{}_hidden{}-{}".format(MODEL_NAME, extra_tag, hidden_size,
now_timestamp)
save_model(model, data_files, extra_tag, hidden_size, now_timestamp,
MODEL_NAME)
# Plot loss and save
filename_loss_plot = "logs/loss_plot_" + filename_base + ".png"
fig = plot_loss(EPOCHS, loss)
fig.savefig(filename_loss_plot)