def main():
params = dict()
params['batch_size'] = 1
params['data_dir'] = args.path_to_train_data
params['major'] = 'users'
params['itemIdInd'] = 1
params['userIdInd'] = 0
params['extension'] = '.csv'
params['delimiter'] = ','
params['header'] = 1
print("Loading training data")
data_layer = input_layer.UserItemRecDataProvider(params=params)
print("Data loaded")
print("Total items found: {}".format(len(data_layer.data.keys())))
print("Vector dim: {}".format(data_layer.vector_dim))
print("Loading eval data")
eval_params = copy.deepcopy(params)
# must set eval batch size to 1 to make sure no examples are missed
eval_params['batch_size'] = 1
eval_params['data_dir'] = args.path_to_eval_data
eval_data_layer = input_layer.UserItemRecDataProvider(params=eval_params,
user_id_map=data_layer.userIdMap,
item_id_map=data_layer.itemIdMap)
rencoder = model.AutoEncoder(layer_sizes=[data_layer.vector_dim] + [int(l) for l in args.hidden_layers.split(',')],
nl_type=args.non_linearity_type,
is_constrained=args.constrained,
dp_drop_prob=args.drop_prob,
last_layer_activations=not args.skip_last_layer_nl)
path_to_model = Path(args.save_path)
if path_to_model.is_file():
print("Loading model from: {}".format(path_to_model))
rencoder.load_state_dict(torch.load(args.save_path))
print('######################################################')
print('######################################################')
print('############# AutoEncoder Model: #####################')
print(rencoder)
print('######################################################')
print('######################################################')
rencoder.eval()
if use_gpu: rencoder = rencoder.cuda()
inv_userIdMap = {v: k for k, v in data_layer.userIdMap.items()}
inv_itemIdMap = {v: k for k, v in data_layer.itemIdMap.items()}
eval_data_layer.src_data = data_layer.data
with open(args.predictions_path, 'w') as outf:
for i, ((out, src), majorInd) in enumerate(eval_data_layer.iterate_one_epoch_eval(for_inf=True)):
inputs = Variable(src.cuda().to_dense() if use_gpu else src.to_dense())
targets_np = out.to_dense().numpy()[0, :]
outputs = rencoder(inputs).cpu().data.numpy()[0, :]
non_zeros = targets_np.nonzero()[0].tolist()
major_key = inv_userIdMap[majorInd]
for ind in non_zeros:
outf.write("{}\t{}\t{}\t{}\n".format(major_key, inv_itemIdMap[ind], outputs[ind], targets_np[ind]))
if i % 10000 == 0:
print("Done: {}".format(i))
def load_recommender(vector_dim, hidden, activation, dropout, weights_path):
rencoder_api = model.AutoEncoder(layer_sizes=[vector_dim] +
[int(l) for l in hidden.split(',')],
nl_type=activation,
is_constrained=False,
dp_drop_prob=dropout,
last_layer_activations=True)
load_model_weights(rencoder_api, weights_path)
rencoder_api.eval()
if USE_GPU: rencoder_api = rencoder_api.cuda()
return rencoder_api
def main():
logger = Logger(args.logdir)
params = dict()
params['batch_size'] = args.batch_size
params['data_dir'] = args.path_to_train_data
params['major'] = 'users'
params['itemIdInd'] = 1
params['userIdInd'] = 0
print("Loading training data")
data_layer = input_layer.UserItemRecDataProvider(params=params)
print("Data loaded")
print("Total items found: {}".format(len(data_layer.data.keys())))
print("Vector dim: {}".format(data_layer.vector_dim))
print("Loading eval data")
eval_params = copy.deepcopy(params)
# must set eval batch size to 1 to make sure no examples are missed
eval_params['data_dir'] = args.path_to_eval_data
eval_data_layer = input_layer.UserItemRecDataProvider(
params=eval_params,
user_id_map=data_layer.userIdMap, # the mappings are provided
item_id_map=data_layer.itemIdMap)
eval_data_layer.src_data = data_layer.data
rencoder = model.AutoEncoder(
layer_sizes=[data_layer.vector_dim] +
[int(l) for l in args.hidden_layers.split(',')],
nl_type=args.non_linearity_type,
is_constrained=args.constrained,
dp_drop_prob=args.drop_prob,
last_layer_activations=not args.skip_last_layer_nl)
model_checkpoint = args.logdir + "/model"
path_to_model = Path(model_checkpoint)
if path_to_model.is_file():
print("Loading model from: {}".format(model_checkpoint))
rencoder.load_state_dict(torch.load(model_checkpoint))
print('######################################################')
print('######################################################')
print('############# AutoEncoder Model: #####################')
print(rencoder)
print('######################################################')
print('######################################################')
gpu_ids = [int(g) for g in args.gpu_ids.split(',')]
print('Using GPUs: {}'.format(gpu_ids))
if len(gpu_ids) > 1:
rencoder = nn.DataParallel(rencoder, device_ids=gpu_ids)
rencoder = rencoder.cuda()
if args.optimizer == "adam":
optimizer = optim.Adam(rencoder.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif args.optimizer == "adagrad":
optimizer = optim.Adagrad(rencoder.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif args.optimizer == "momentum":
optimizer = optim.SGD(rencoder.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
scheduler = MultiStepLR(optimizer,
milestones=[24, 36, 48, 66, 72],
gamma=0.5)
elif args.optimizer == "rmsprop":
optimizer = optim.RMSprop(rencoder.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
else:
raise ValueError('Unknown optimizer kind')
t_loss = 0.0
t_loss_denom = 0.0
global_step = 0
if args.noise_prob > 0.0:
dp = nn.Dropout(p=args.noise_prob)
for epoch in range(args.num_epochs):
print('Doing epoch {} of {}'.format(epoch, args.num_epochs))
e_start_time = time.time()
rencoder.train()
total_epoch_loss = 0.0
denom = 0.0
if args.optimizer == "momentum":
scheduler.step()
for i, mb in enumerate(data_layer.iterate_one_epoch()):
inputs = Variable(mb.cuda().to_dense())
optimizer.zero_grad()
outputs = rencoder(inputs)
loss, num_ratings = model.MSEloss(outputs, inputs)
loss = loss / num_ratings
loss.backward()
optimizer.step()
global_step += 1
t_loss += loss.data[0]
t_loss_denom += 1
if i % args.summary_frequency == 0:
print('[%d, %5d] RMSE: %.7f' %
(epoch, i, sqrt(t_loss / t_loss_denom)))
logger.scalar_summary("Training_RMSE",
sqrt(t_loss / t_loss_denom), global_step)
t_loss = 0
t_loss_denom = 0.0
log_var_and_grad_summaries(logger, rencoder.encode_w,
global_step, "Encode_W")
log_var_and_grad_summaries(logger, rencoder.encode_b,
global_step, "Encode_b")
if not rencoder.is_constrained:
log_var_and_grad_summaries(logger, rencoder.decode_w,
global_step, "Decode_W")
log_var_and_grad_summaries(logger, rencoder.decode_b,
global_step, "Decode_b")
total_epoch_loss += loss.data[0]
denom += 1
#if args.aug_step > 0 and i % args.aug_step == 0 and i > 0:
if args.aug_step > 0:
# Magic data augmentation trick happen here
for t in range(args.aug_step):
inputs = Variable(outputs.data)
if args.noise_prob > 0.0:
inputs = dp(inputs)
optimizer.zero_grad()
outputs = rencoder(inputs)
loss, num_ratings = model.MSEloss(outputs, inputs)
loss = loss / num_ratings
loss.backward()
optimizer.step()
e_end_time = time.time()
print(
'Total epoch {} finished in {} seconds with TRAINING RMSE loss: {}'
.format(epoch, e_end_time - e_start_time,
sqrt(total_epoch_loss / denom)))
logger.scalar_summary("Training_RMSE_per_epoch",
sqrt(total_epoch_loss / denom), epoch)
logger.scalar_summary("Epoch_time", e_end_time - e_start_time, epoch)
if epoch % 3 == 0 or epoch == args.num_epochs - 1:
eval_loss = do_eval(rencoder, eval_data_layer)
print('Epoch {} EVALUATION LOSS: {}'.format(epoch, eval_loss))
logger.scalar_summary("EVALUATION_RMSE", eval_loss, epoch)
print("Saving model to {}".format(model_checkpoint + ".epoch_" +
str(epoch)))
torch.save(rencoder.state_dict(),
model_checkpoint + ".epoch_" + str(epoch))
print("Saving model to {}".format(model_checkpoint + ".last"))
torch.save(rencoder.state_dict(), model_checkpoint + ".last")
