def main():
args = parse_args()
if args.seed is not None:
print("Using seed = {}".format(args.seed))
torch.manual_seed(args.seed)
np.random.seed(seed=args.seed)
# Save configuration to file
config = {k: v for k, v in args.__dict__.items()}
config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp())
config['local_timestamp'] = str(datetime.now())
run_dir = "./run/neumf/{}".format(config['timestamp'])
print("Saving config and results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
utils.save_config(config, run_dir)
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
t1 = time.time()
# Load Data
print('Loading data')
train_dataset = CFTrainDataset(
os.path.join(args.data, TRAIN_RATINGS_FILENAME), args.negative_samples)
train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
test_ratings = load_test_ratings(
os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME))
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d' %
(time.time() - t1, nb_users, nb_items, train_dataset.mat.nnz,
len(test_ratings)))
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mf_reg=0.,
mlp_layer_sizes=args.layers,
mlp_layer_regs=[0. for i in args.layers])
print(model)
print("{} parameters".format(utils.count_parameters(model)))
# Save model text description
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
# Add optimizer and loss to graph
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
criterion = nn.BCEWithLogitsLoss()
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
processes=args.processes)
print('Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(
K=args.topk, hit_rate=np.mean(hits), ndcg=np.mean(ndcgs)))
for epoch in range(args.epochs):
model.train()
losses = utils.AverageMeter()
begin = time.time()
loader = tqdm.tqdm(train_dataloader)
length = len(loader)
if length < 101:
print(
'Exiting, cannot profile the required 100 iterations. Please re-run with a larger batch size.'
)
cuda.profile_stop()
exit()
for batch_index, (user, item, label) in enumerate(loader):
if batch_index == length // 2 and epoch == 0:
print('Starting profiling for 100 iterations.')
cuda.profile_start()
if batch_index == length // 2 + 100 and epoch == 0:
print(
'Profiling completed, stopping profiling and continuing training.'
)
cuda.profile_stop()
user = torch.autograd.Variable(user, requires_grad=False)
item = torch.autograd.Variable(item, requires_grad=False)
label = torch.autograd.Variable(label, requires_grad=False)
if use_cuda:
user = user.cuda(async=True)
item = item.cuda(async=True)
label = label.cuda(async=True)
outputs = model(user, item)
loss = criterion(outputs, label)
losses.update(loss.data.item(), user.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Save stats to file
description = (
'Epoch {} Loss {loss.val:.4f} ({loss.avg:.4f})'.format(
epoch, loss=losses))
loader.set_description(description)
train_time = time.time() - begin
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
output=valid_results_file,
epoch=epoch,
processes=args.processes)
val_time = time.time() - begin
print(
'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f},'
' train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format(
epoch=epoch,
K=args.topk,
hit_rate=np.mean(hits),
ndcg=np.mean(ndcgs),
train_time=train_time,
val_time=val_time))
if args.threshold is not None:
if np.mean(hits) >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
return 0
def main():
# Note: The run start is in convert.py
args = parse_args()
if args.seed is not None:
print("Using seed = {}".format(args.seed))
torch.manual_seed(args.seed)
np.random.seed(seed=args.seed)
# Save configuration to file
config = {k: v for k, v in args.__dict__.items()}
config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp())
config['local_timestamp'] = str(datetime.now())
run_dir = "./run/neumf/{}".format(config['timestamp'])
print("Saving config and results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
utils.save_config(config, run_dir)
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
t1 = time.time()
# Load Data
print('Loading data')
train_dataset = CFTrainDataset(
os.path.join(args.data, TRAIN_RATINGS_FILENAME), args.negative_samples)
mlperf_log.ncf_print(key=mlperf_log.INPUT_BATCH_SIZE,
value=args.batch_size)
mlperf_log.ncf_print(
key=mlperf_log.INPUT_ORDER) # set shuffle=True in DataLoader
train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
test_ratings = load_test_ratings(
os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME))
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d' %
(time.time() - t1, nb_users, nb_items, train_dataset.mat.nnz,
len(test_ratings)))
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mf_reg=0.,
mlp_layer_sizes=args.layers,
mlp_layer_regs=[0. for i in args.layers])
print(model)
print("{} parameters".format(utils.count_parameters(model)))
# Save model text description
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
# Add optimizer and loss to graph
mlperf_log.ncf_print(key=mlperf_log.TRAIN_LEARN_RATE,
value=args.learning_rate)
beta1, beta2, epsilon = 0.9, 0.999, 1e-8
mlperf_log.ncf_print(key=mlperf_log.OPT_NAME, value="Adam")
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_BETA1, value=beta1)
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_BETA2, value=beta2)
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_EPSILON, value=epsilon)
optimizer = torch.optim.Adam(model.parameters(),
betas=(beta1, beta2),
lr=args.learning_rate,
eps=epsilon)
mlperf_log.ncf_print(key=mlperf_log.MODEL_HP_LOSS_FN, value=mlperf_log.BCE)
criterion = nn.BCEWithLogitsLoss()
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
processes=args.processes)
print('Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(
K=args.topk, hit_rate=np.mean(hits), ndcg=np.mean(ndcgs)))
success = False
mlperf_log.ncf_print(key=mlperf_log.TRAIN_LOOP)
for epoch in range(args.epochs):
mlperf_log.ncf_print(key=mlperf_log.TRAIN_EPOCH, value=epoch)
model.train()
losses = utils.AverageMeter()
mlperf_log.ncf_print(key=mlperf_log.INPUT_HP_NUM_NEG,
value=train_dataset.nb_neg)
mlperf_log.ncf_print(key=mlperf_log.INPUT_STEP_TRAIN_NEG_GEN)
begin = time.time()
loader = tqdm.tqdm(train_dataloader)
for batch_index, (user, item, label) in enumerate(loader):
user = torch.autograd.Variable(user, requires_grad=False)
item = torch.autograd.Variable(item, requires_grad=False)
label = torch.autograd.Variable(label, requires_grad=False)
if use_cuda:
user = user.cuda(async=True)
item = item.cuda(async=True)
label = label.cuda(async=True)
outputs = model(user, item)
loss = criterion(outputs, label)
losses.update(loss.data.item(), user.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Save stats to file
description = (
'Epoch {} Loss {loss.val:.4f} ({loss.avg:.4f})'.format(
epoch, loss=losses))
loader.set_description(description)
train_time = time.time() - begin
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
output=valid_results_file,
epoch=epoch,
processes=args.processes)
mlperf_log.ncf_print(key=mlperf_log.EVAL_ACCURACY,
value={
"epoch": epoch,
"value": float(np.mean(hits))
})
mlperf_log.ncf_print(key=mlperf_log.EVAL_TARGET, value=args.threshold)
mlperf_log.ncf_print(key=mlperf_log.EVAL_STOP)
val_time = time.time() - begin
print(
'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f},'
' train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format(
epoch=epoch,
K=args.topk,
hit_rate=np.mean(hits),
ndcg=np.mean(ndcgs),
train_time=train_time,
val_time=val_time))
if args.threshold is not None:
if np.mean(hits) >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
success = True
break
mlperf_log.ncf_print(key=mlperf_log.RUN_STOP, value={"success": success})
mlperf_log.ncf_print(key=mlperf_log.RUN_FINAL)
args = parse_args()
print('Loading data')
train_dataset = CFTrainDataset(os.path.join('ml-20m', TRAIN_RATINGS_FILENAME),
4)
train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=2048,
shuffle=True,
num_workers=0,
pin_memory=True)
test_ratings = load_test_ratings(os.path.join(
'ml-20m', TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join('ml-20m', TEST_NEG_FILENAME))
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
# Create model
layers = [256, 256, 128, 64]
model = NeuMF(nb_users,
nb_items,
mf_dim=64,
mf_reg=0.,
mlp_layer_sizes=layers,
mlp_layer_regs=[0. for i in layers])
model.load_state_dict(torch.load(args.path))
model = model.cuda
print(model)
def main():
# Note: The run start is in data_preprocess.py
args = parse_args()
if args.seed is not None:
print("Using seed = {}".format(args.seed))
torch.manual_seed(args.seed)
np.random.seed(seed=args.seed)
# Save configuration to file
config = {k: v for k, v in args.__dict__.items()}
config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp())
config['local_timestamp'] = str(datetime.now())
run_dir = "./run/MGPM/{}/{}".format(
os.path.basename(os.path.normpath(args.data)), config['timestamp'])
print("Saving config and results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
utils.save_config(config, run_dir)
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
if use_cuda:
print("Using cuda ...")
else:
print("Using CPU ...")
t1 = time.time()
best_hit, best_ndcg = 0., 0.
start_epoch = 0 # start from epoch 0 or last checkpoint epoch
# Load Data
print('Loading data')
print(os.path.join(args.data, TRAIN_RATINGS_FILENAME))
train_dataset = CFTrainDataset(
os.path.join(args.data, TRAIN_RATINGS_FILENAME),
os.path.join(args.data, DATA_SUMMARY_FILENAME), args.negative_samples)
mlperf_log.ncf_print(key=mlperf_log.INPUT_BATCH_SIZE,
value=args.batch_size)
mlperf_log.ncf_print(
key=mlperf_log.INPUT_ORDER) # set shuffle=True in DataLoader
train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
test_ratings = load_test_ratings(
os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME))
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d' %
(time.time() - t1, nb_users, nb_items, train_dataset.mat.nnz,
len(test_ratings)))
# Create model
model = Multi_Preference_Model(nb_users=nb_users,
nb_items=nb_items,
embed_dim=32,
history_size=9)
print(model)
print("{} parameters".format(utils.count_parameters(model)))
# Save model text description
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
# Add optimizer and loss to graph
mlperf_log.ncf_print(key=mlperf_log.OPT_LR, value=args.learning_rate)
beta1, beta2, epsilon = 0.9, 0.999, 1e-8
mlperf_log.ncf_print(key=mlperf_log.OPT_NAME, value="Adam")
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_BETA1, value=beta1)
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_BETA2, value=beta2)
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_EPSILON, value=epsilon)
optimizer = torch.optim.Adam(model.parameters(),
betas=(beta1, beta2),
lr=args.learning_rate,
eps=epsilon)
mlperf_log.ncf_print(key=mlperf_log.MODEL_HP_LOSS_FN, value=mlperf_log.BCE)
# optimizer = torch.optim.SGD(model.parameters(),lr=args.learning_rate,momentum=0.9)
criterion = nn.BCEWithLogitsLoss()
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
if args.resume:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir(
'checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/' + model._get_name() + '.pd')
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
start_epoch = checkpoint['epoch']
best_hit = checkpoint['hit']
best_ndcg = checkpoint['ndcg']
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
if start_epoch == 0:
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
processes=args.processes)
print('Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(
K=args.topk, hit_rate=np.mean(hits), ndcg=np.mean(ndcgs)))
mlperf_log.ncf_print(key=mlperf_log.TRAIN_LOOP)
for epoch in range(start_epoch, args.epochs):
mlperf_log.ncf_print(key=mlperf_log.TRAIN_EPOCH, value=epoch)
model.train()
losses = utils.AverageMeter()
mlperf_log.ncf_print(key=mlperf_log.INPUT_HP_NUM_NEG,
value=train_dataset.nb_neg)
mlperf_log.ncf_print(key=mlperf_log.INPUT_STEP_TRAIN_NEG_GEN)
begin = time.time()
loader = tqdm.tqdm(train_dataloader)
for batch_index, (user, item, history, label) in enumerate(loader):
user = torch.autograd.Variable(user, requires_grad=False)
item = torch.autograd.Variable(item, requires_grad=False)
history = torch.autograd.Variable(history, requires_grad=False)
label = torch.autograd.Variable(label, requires_grad=False)
if use_cuda:
user = user.cuda()
item = item.cuda()
history = history.cuda()
label = label.cuda()
# outputs, _ = model(user, item,history)
outputs = model(user, item, history)
loss = criterion(outputs, label)
losses.update(loss.data.item(), user.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Save stats to file
description = (
'Epoch {} Loss {loss.val:.4f} ({loss.avg:.4f})'.format(
epoch, loss=losses))
loader.set_description(description)
train_time = time.time() - begin
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
output=valid_results_file,
epoch=epoch,
processes=args.processes)
mlperf_log.ncf_print(key=mlperf_log.EVAL_ACCURACY,
value={
"epoch": epoch,
"value": float(np.mean(hits))
})
mlperf_log.ncf_print(key=mlperf_log.EVAL_STOP)
val_time = time.time() - begin
print(
'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f},'
' train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format(
epoch=epoch,
K=args.topk,
hit_rate=np.mean(hits),
ndcg=np.mean(ndcgs),
train_time=train_time,
val_time=val_time))
if np.mean(hits) >= best_hit or np.mean(ndcgs) >= best_ndcg:
best_hit = np.mean(hits)
best_ndcg = np.mean(ndcgs)
# Save checkpoint.
print('Saving checkpoint..')
state = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'hit': best_hit,
'ndcg': best_ndcg,
}
if not os.path.isdir('checkpoint'):
os.mkdir('checkpoint')
torch.save(state, './checkpoint/' + model._get_name() + '.pd')
print("Best hit: ", best_hit)
print("Best_ndcg: ", best_ndcg)
mlperf_log.ncf_print(key=mlperf_log.RUN_STOP)
mlperf_log.ncf_print(key=mlperf_log.RUN_FINAL)
def main():
args = parse_args()
if args.seed is not None:
print("Using seed = {}".format(args.seed))
torch.manual_seed(args.seed)
np.random.seed(seed=args.seed)
# Save configuration to file
config = {k: v for k, v in args.__dict__.items()}
config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp())
config['local_timestamp'] = str(datetime.now())
run_dir = "./run/neumf/{}".format(config['timestamp'])
print("Saving config and results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
utils.save_config(config, run_dir)
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
t1 = time.time()
# Load Data
# TODO: Reading CSVs is slow. Could use HDF or Apache Arrow
train_dataset = CFTrainDataset(
os.path.join(args.data, TRAIN_RATINGS_FILENAME), args.negative_samples)
print('batchsize=%d' % args.batch_size)
train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
test_ratings = load_test_ratings(
os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME))
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d' %
(time.time() - t1, nb_users, nb_items, train_dataset.mat.nnz,
len(test_ratings)))
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mf_reg=0.,
mlp_layer_sizes=args.layers,
mlp_layer_regs=[0. for i in args.layers])
print(model)
print("{} parameters".format(utils.count_parameters(model)))
# Save model text description
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
# Add optimizer and loss to graph
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
criterion = nn.BCEWithLogitsLoss()
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda)
print('Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(
K=args.topk, hit_rate=np.mean(hits), ndcg=np.mean(ndcgs)))
for epoch in range(args.epochs):
model.train()
losses = utils.AverageMeter()
begin = time.time()
loader = tqdm.tqdm(train_dataloader)
counting_data = 0
counting_forward = 0
counting_zerograd = 0
counting_backward = 0
counting_updateweight = 0
counting_des = 0
for batch_index, (user, item, label) in enumerate(loader):
start0 = time.time()
user = torch.autograd.Variable(user, requires_grad=False)
item = torch.autograd.Variable(item, requires_grad=False)
label = torch.autograd.Variable(label, requires_grad=False)
if use_cuda:
user = user.cuda(async=True)
item = item.cuda(async=True)
label = label.cuda(async=True)
start1 = time.time()
outputs = model(user, item)
loss = criterion(outputs, label)
losses.update(loss.data.item(), user.size(0))
start2 = time.time()
optimizer.zero_grad()
start3 = time.time()
loss.backward()
start4 = time.time()
optimizer.step()
start5 = time.time()
# Save stats to file
description = (
'Epoch {} Loss {loss.val:.4f} ({loss.avg:.4f})'.format(
epoch, loss=losses))
loader.set_description(description)
start6 = time.time()
counting_data += start1 - start0
counting_forward += start2 - start1
counting_zerograd += start3 - start2
counting_backward += start4 - start3
counting_updateweight += start5 - start3
counting_des += start6 - start5
train_time = time.time() - begin
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
output=valid_results_file,
epoch=epoch)
val_time = time.time() - begin
print(
'data: {data:.f4}, forward: {ft:.4f}, zerograd: {zg:.4f}, backward: {bw:.4f},'
' adam: {adam:.4f}, description: {des:.4f}'.format(
data=counting_data,
ft=counting_forward,
zg=counting_zerograd,
bw=counting_backward,
adam=counting_updateweight,
des=counting_des))
print(
'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f},'
' train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format(
epoch=epoch,
K=args.topk,
hit_rate=np.mean(hits),
ndcg=np.mean(ndcgs),
train_time=train_time,
val_time=val_time))
if args.threshold is not None:
if np.mean(hits) >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
return 0
def main():
args = parse_args()
if args.seed is not None:
print("Using seed = {}".format(args.seed))
# torch.manual_seed(args.seed)
mx.random.seed(seed_state=args.seed)
np.random.seed(seed=args.seed)
# Save configuration to file
config = {k: v for k, v in args.__dict__.items()}
config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp())
config['local_timestamp'] = str(datetime.now())
run_dir = "./run/neumf_" + args.data + "/{}".format(config['timestamp'])
print("Saving config and results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
utils.save_config(config, run_dir) #defined in utils.py
# Check that GPUs are actually available
use_cuda = not args.no_cuda and mx.test_utils.list_gpus()
t1 = time.time()
# Load Data
print('Loading data')
train_dataset = CFTrainDataset(
os.path.join(args.data, TRAIN_RATINGS_FILENAME), args.negative_samples)
#in original file, use 8 core as defaul
# the parameter：shuffle means random the samples
train_dataloader = mx.gluon.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers)
test_ratings = load_test_ratings(
os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME))
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d' %
(time.time() - t1, nb_users, nb_items, train_dataset.mat.nnz,
len(test_ratings)))
if (use_cuda):
ctx = mx.gpu(0)
# default to use NO.1 gpu can use docker to select a nvidia
else:
ctx = mx.cpu(0)
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mf_reg=0.,
mlp_layer_sizes=args.layers,
mlp_layer_regs=[0. for i in args.layers],
ctx=ctx)
model.initialize(ctx=ctx)
model.hybridize()
print(model)
# todo 9: to change the function in utils
# print("{} parameters".format(utils.count_parameters(model)))
# model.collect_params()
# Save model text description
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
# model.save_parameters(os.path.join("/home/net.params", 'net.params'))
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
processes=args.processes,
ctx=ctx)
print('Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(
K=args.topk, hit_rate=np.mean(hits), ndcg=np.mean(ndcgs)))
############# hyperparameters
# Add optimizer and loss to graph
lr = args.learning_rate
bs = args.batch_size
trainer = mx.gluon.Trainer(model.collect_params(), 'adam',
{'learning_rate': lr})
mxnet_criterion = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss(
) # equivalent to lossfunction
# training
for epoch in range(args.epochs):
begin = time.time()
# tqdm shows the percentage of the process
loader = tqdm.tqdm(train_dataloader)
for batch_index, (user, item, label) in enumerate(loader):
# TODO 7: search the autograd in mxnet
# todo : let user act in gpu
user = nd.array(user, ctx=ctx)
item = nd.array(item, ctx=ctx)
label = nd.array(label, ctx=ctx)
# compute the gradient automatically
with autograd.record():
outputs = model(user, item)
loss = mxnet_criterion(outputs, label.T)
loss.backward()
trainer.step(bs)
for x in loss.mean().asnumpy().tolist():
loss_number = x
description = ('Epoch {} Loss {:.4f}'.format(epoch, loss_number))
loader.set_description(description)
train_time = time.time() - begin
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
output=valid_results_file,
epoch=epoch,
processes=args.processes,
ctx=ctx)
val_time = time.time() - begin
print(
'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f},'
' train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format(
epoch=epoch,
K=args.topk,
hit_rate=np.mean(hits),
ndcg=np.mean(ndcgs),
train_time=train_time,
val_time=val_time))
if args.threshold is not None:
if np.mean(hits) >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
# Save model text description after modelling
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
# model.save_parameters(os.path.join("/home/net.params",'net.params'))
return 0
def main():
global msglogger
script_dir = os.path.dirname(__file__)
args = parse_args()
# Distiller loggers
msglogger = apputils.config_pylogger('logging.conf',
args.name,
output_dir=args.output_dir)
tflogger = TensorBoardLogger(msglogger.logdir)
# tflogger.log_gradients = True
# pylogger = PythonLogger(msglogger)
if args.seed is not None:
msglogger.info("Using seed = {}".format(args.seed))
torch.manual_seed(args.seed)
np.random.seed(seed=args.seed)
args.qe_mode = str(args.qe_mode).split('.')[1]
args.qe_clip_acts = str(args.qe_clip_acts).split('.')[1]
apputils.log_execution_env_state(sys.argv)
if args.gpus is not None:
try:
args.gpus = [int(s) for s in args.gpus.split(',')]
except ValueError:
msglogger.error(
'ERROR: Argument --gpus must be a comma-separated list of integers only'
)
exit(1)
if len(args.gpus) > 1:
msglogger.error('ERROR: Only single GPU supported for NCF')
exit(1)
available_gpus = torch.cuda.device_count()
for dev_id in args.gpus:
if dev_id >= available_gpus:
msglogger.error(
'ERROR: GPU device ID {0} requested, but only {1} devices available'
.format(dev_id, available_gpus))
exit(1)
# Set default device in case the first one on the list != 0
torch.cuda.set_device(args.gpus[0])
# Save configuration to file
config = {k: v for k, v in args.__dict__.items()}
config['timestamp'] = "{:.0f}".format(datetime.utcnow().timestamp())
config['local_timestamp'] = str(datetime.now())
run_dir = msglogger.logdir
msglogger.info("Saving config and results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
utils.save_config(config, run_dir)
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
t1 = time.time()
# Load Data
training = not (args.eval or args.qe_calibration
or args.activation_histograms)
msglogger.info('Loading data')
if training:
train_dataset = CFTrainDataset(
os.path.join(args.data, TRAIN_RATINGS_FILENAME),
args.negative_samples)
train_dataloader = torch.utils.data.DataLoader(
dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
nb_users, nb_items = train_dataset.nb_users, train_dataset.nb_items
else:
train_dataset = None
train_dataloader = None
nb_users, nb_items = (138493, 26744)
test_ratings = load_test_ratings(
os.path.join(args.data, TEST_RATINGS_FILENAME)) # noqa: E501
test_negs = load_test_negs(os.path.join(args.data, TEST_NEG_FILENAME))
msglogger.info(
'Load data done [%.1f s]. #user=%d, #item=%d, #train=%s, #test=%d' %
(time.time() - t1, nb_users, nb_items,
str(train_dataset.mat.nnz) if training else 'N/A', len(test_ratings)))
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mf_reg=0.,
mlp_layer_sizes=args.layers,
mlp_layer_regs=[0. for i in args.layers],
split_final=args.split_final)
if use_cuda:
model = model.cuda()
msglogger.info(model)
msglogger.info("{} parameters".format(utils.count_parameters(model)))
# Save model text description
with open(os.path.join(run_dir, 'model.txt'), 'w') as file:
file.write(str(model))
compression_scheduler = None
start_epoch = 0
optimizer = None
if args.load:
if training:
model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint(
model, args.load)
if args.reset_optimizer:
start_epoch = 0
optimizer = None
else:
model = apputils.load_lean_checkpoint(model, args.load)
# Add loss to graph
criterion = nn.BCEWithLogitsLoss()
if use_cuda:
criterion = criterion.cuda()
if training and optimizer is None:
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
msglogger.info('Optimizer Type: %s', type(optimizer))
msglogger.info('Optimizer Args: %s', optimizer.defaults)
if args.compress:
compression_scheduler = distiller.file_config(model, optimizer,
args.compress)
model.cuda()
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
if args.qe_calibration or args.activation_histograms:
calib = {
'portion':
args.qe_calibration,
'desc_str':
'quantization calibration stats',
'collect_func':
partial(distiller.data_loggers.collect_quant_stats,
inplace_runtime_check=True,
disable_inplace_attrs=True)
}
hists = {
'portion':
args.activation_histograms,
'desc_str':
'activation histograms',
'collect_func':
partial(distiller.data_loggers.collect_histograms,
activation_stats=None,
nbins=2048,
save_hist_imgs=True)
}
d = calib if args.qe_calibration else hists
distiller.utils.assign_layer_fq_names(model)
num_users = int(np.floor(len(test_ratings) * d['portion']))
msglogger.info(
"Generating {} based on {:.1%} of the test-set ({} users)".format(
d['desc_str'], d['portion'], num_users))
test_fn = partial(val_epoch,
ratings=test_ratings,
negs=test_negs,
K=args.topk,
use_cuda=use_cuda,
processes=args.processes,
num_users=num_users)
d['collect_func'](model=model,
test_fn=test_fn,
save_dir=run_dir,
classes=None)
return 0
if args.eval:
if args.quantize_eval and args.qe_calibration is None:
model.cpu()
quantizer = quantization.PostTrainLinearQuantizer.from_args(
model, args)
dummy_input = (torch.tensor([1]), torch.tensor([1]),
torch.tensor([True], dtype=torch.bool))
quantizer.prepare_model(dummy_input)
model.cuda()
distiller.utils.assign_layer_fq_names(model)
if args.eval_fp16:
model = model.half()
# Calculate initial Hit Ratio and NDCG
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
processes=args.processes)
val_time = time.time() - begin
hit_rate = np.mean(hits)
msglogger.info(
'Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}, val_time = {val_time:.2f}'
.format(K=args.topk,
hit_rate=hit_rate,
ndcg=np.mean(ndcgs),
val_time=val_time))
hit_rate = 0
if args.quantize_eval:
checkpoint_name = 'quantized'
apputils.save_checkpoint(0,
'NCF',
model,
optimizer=None,
extras={'quantized_hr@10': hit_rate},
name='_'.join([args.name, 'quantized'])
if args.name else checkpoint_name,
dir=msglogger.logdir)
return 0
total_samples = len(train_dataloader.sampler)
steps_per_epoch = math.ceil(total_samples / args.batch_size)
best_hit_rate = 0
best_epoch = 0
for epoch in range(start_epoch, args.epochs):
msglogger.info('')
model.train()
losses = utils.AverageMeter()
begin = time.time()
if compression_scheduler:
compression_scheduler.on_epoch_begin(epoch, optimizer)
loader = tqdm.tqdm(train_dataloader)
for batch_index, (user, item, label) in enumerate(loader):
user = torch.autograd.Variable(user, requires_grad=False)
item = torch.autograd.Variable(item, requires_grad=False)
label = torch.autograd.Variable(label, requires_grad=False)
if use_cuda:
user = user.cuda(async=True)
item = item.cuda(async=True)
label = label.cuda(async=True)
if compression_scheduler:
compression_scheduler.on_minibatch_begin(
epoch, batch_index, steps_per_epoch, optimizer)
outputs = model(user, item, torch.tensor([False],
dtype=torch.bool))
loss = criterion(outputs, label)
if compression_scheduler:
compression_scheduler.before_backward_pass(
epoch,
batch_index,
steps_per_epoch,
loss,
optimizer,
return_loss_components=False)
losses.update(loss.data.item(), user.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
if compression_scheduler:
compression_scheduler.on_minibatch_end(epoch, batch_index,
steps_per_epoch,
optimizer)
# Save stats to file
description = (
'Epoch {} Loss {loss.val:.4f} ({loss.avg:.4f})'.format(
epoch, loss=losses))
loader.set_description(description)
steps_completed = batch_index + 1
if steps_completed % args.log_freq == 0:
stats_dict = OrderedDict()
stats_dict['Loss'] = losses.avg
stats = ('Performance/Training/', stats_dict)
params = model.named_parameters(
) if args.log_params_histograms else None
distiller.log_training_progress(stats, params, epoch,
steps_completed,
steps_per_epoch, args.log_freq,
[tflogger])
tflogger.log_model_buffers(model,
['tracked_min', 'tracked_max'],
'Quant/Train/Acts/TrackedMinMax',
epoch, steps_completed,
steps_per_epoch, args.log_freq)
train_time = time.time() - begin
begin = time.time()
hits, ndcgs = val_epoch(model,
test_ratings,
test_negs,
args.topk,
use_cuda=use_cuda,
output=valid_results_file,
epoch=epoch,
processes=args.processes)
val_time = time.time() - begin
if compression_scheduler:
compression_scheduler.on_epoch_end(epoch, optimizer)
hit_rate = np.mean(hits)
mean_ndcgs = np.mean(ndcgs)
stats_dict = OrderedDict()
stats_dict['HR@{0}'.format(args.topk)] = hit_rate
stats_dict['NDCG@{0}'.format(args.topk)] = mean_ndcgs
stats = ('Performance/Validation/', stats_dict)
distiller.log_training_progress(stats,
None,
epoch,
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[tflogger])
msglogger.info(
'Epoch {epoch}: HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}, AvgTrainLoss = {loss.avg:.4f}, '
'train_time = {train_time:.2f}, val_time = {val_time:.2f}'.format(
epoch=epoch,
K=args.topk,
hit_rate=hit_rate,
ndcg=mean_ndcgs,
loss=losses,
train_time=train_time,
val_time=val_time))
is_best = False
if hit_rate > best_hit_rate:
best_hit_rate = hit_rate
is_best = True
best_epoch = epoch
extras = {
'current_hr@10': hit_rate,
'best_hr@10': best_hit_rate,
'best_epoch': best_epoch
}
apputils.save_checkpoint(epoch,
'NCF',
model,
optimizer,
compression_scheduler,
extras,
is_best,
dir=run_dir)
if args.threshold is not None:
if np.mean(hits) >= args.threshold:
msglogger.info("Hit threshold of {}".format(args.threshold))
break
