def main():
args = parse_args()
state_dict = torch.load(args.model_checkpoint_path,
map_location=torch.device("cpu"))
layers = [state_dict["mlp.0.weight"].size()[1]]
layer_keys = sorted([
k for k in state_dict.keys()
if (k.startswith("mlp") and k.endswith("bias"))
])
layers.extend([state_dict[k].size()[0] for k in layer_keys])
n_items = state_dict["mf_item_embed.weight"].size()[0]
n_users = state_dict["mf_user_embed.weight"].size()[0]
model = NeuMF(
nb_users=n_users,
nb_items=n_items,
mf_dim=state_dict["mf_item_embed.weight"].size()[1],
mlp_layer_sizes=layers,
dropout=0.5,
)
model.load_state_dict(state_dict)
model.eval()
users = torch.LongTensor(np.full(n_items, n_users - 1))
items = torch.LongTensor(np.arange(n_items, dtype=np.int64))
predictions = model(users, items, sigmoid=True)
predictions = predictions.detach().numpy().squeeze()
if args.output_dir != "./":
try:
os.makedirs(args.output_dir, exist_ok=True)
except OSError as err:
print("Failed to create output directory: {}".format(err))
sys.exit(-1)
names = []
with open(args.movie_db_file) as fh:
for line in fh:
mid, _, name = line.partition(",")
name, _, tail = name.rpartition(",")
name = name.strip('"')
names.append(name)
predictions_file = os.path.join(args.output_dir, "predictions.csv")
with open(predictions_file, "wt") as fh:
fh.write("Movie, Predicted Rating\n")
argsorts = np.argsort(predictions)[::-1]
for idx in argsorts:
fh.write("{}, {:0.2f}\n".format(names[idx], predictions[idx] * 5))
print("Predictions saved to {}".format(predictions_file))
return
def main():
args = parse_args()
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
])
dllogger.log(data=vars(args), step='PARAMETER')
model = NeuMF(nb_users=args.n_users,
nb_items=args.n_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
model = model.cuda()
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
model.load_state_dict(state_dict)
if args.opt_level == "O2":
model = amp.initialize(model,
opt_level=args.opt_level,
keep_batchnorm_fp32=False,
loss_scale='dynamic')
model.eval()
users = torch.cuda.LongTensor(args.batch_size).random_(0, args.n_users)
items = torch.cuda.LongTensor(args.batch_size).random_(0, args.n_items)
latencies = []
for _ in range(args.num_batches):
torch.cuda.synchronize()
start = time.time()
predictions = model(users, items, sigmoid=True)
torch.cuda.synchronize()
latencies.append(time.time() - start)
dllogger.log(data={
'batch_size':
args.batch_size,
'best_inference_throughput':
args.batch_size / min(latencies),
'best_inference_latency':
min(latencies),
'mean_inference_throughput':
args.batch_size / np.mean(latencies),
'mean_inference_latency':
np.mean(latencies),
'inference_latencies':
latencies
},
step=tuple())
dllogger.flush()
return
def main():
args = parse_args()
dllogger.init(backends=[dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)])
dllogger.log(data=vars(args), step='PARAMETER')
model = NeuMF(nb_users=args.n_users, nb_items=args.n_items, mf_dim=args.factors,
mlp_layer_sizes=args.layers, dropout=args.dropout)
model = model.cuda()
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
model.load_state_dict(state_dict)
if args.fp16:
model.half()
model.eval()
batch_sizes = args.batch_sizes.split(',')
batch_sizes = [int(s) for s in batch_sizes]
result_data = {}
for batch_size in batch_sizes:
print('benchmarking batch size: ', batch_size)
users = torch.cuda.LongTensor(batch_size).random_(0, args.n_users)
items = torch.cuda.LongTensor(batch_size).random_(0, args.n_items)
latencies = []
for _ in range(args.num_batches):
torch.cuda.synchronize()
start = time.time()
_ = model(users, items, sigmoid=True)
torch.cuda.synchronize()
latencies.append(time.time() - start)
result_data[f'batch_{batch_size}_mean_throughput'] = batch_size / np.mean(latencies)
result_data[f'batch_{batch_size}_mean_latency'] = np.mean(latencies)
result_data[f'batch_{batch_size}_p90_latency'] = np.percentile(latencies, 0.90)
result_data[f'batch_{batch_size}_p95_latency'] = np.percentile(latencies, 0.95)
result_data[f'batch_{batch_size}_p99_latency'] = np.percentile(latencies, 0.99)
dllogger.log(data=result_data, step=tuple())
dllogger.flush()
return
def main():
log_hardware()
args = parse_args()
log_args(args)
model = NeuMF(nb_users=args.n_users,
nb_items=args.n_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
model = model.cuda()
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
model.load_state_dict(state_dict)
if args.opt_level == "O2":
model = amp.initialize(model,
opt_level=args.opt_level,
keep_batchnorm_fp32=False,
loss_scale='dynamic')
model.eval()
users = torch.cuda.LongTensor(args.batch_size).random_(0, args.n_users)
items = torch.cuda.LongTensor(args.batch_size).random_(0, args.n_items)
latencies = []
for _ in range(args.num_batches):
torch.cuda.synchronize()
start = time.time()
predictions = model(users, items, sigmoid=True)
torch.cuda.synchronize()
latencies.append(time.time() - start)
LOGGER.log(key='batch_size', value=args.batch_size)
LOGGER.log(key='best_inference_throughput',
value=args.batch_size / min(latencies))
LOGGER.log(key='best_inference_latency', value=min(latencies))
LOGGER.log(key='inference_latencies', value=latencies)
return
