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():
args = parse_args()
args.distributed, args.world_size = init_distributed(args.local_rank)
if args.seed is not None:
print("Using seed = {}".format(args.seed))
torch.manual_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'],args.local_rank)
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()
# more like load trigger timmer now
mlperf_log.ncf_print(key=mlperf_log.PREPROC_HP_NUM_EVAL, value=args.valid_negative)
# The default of np.random.choice is replace=True, so does pytorch random_()
mlperf_log.ncf_print(key=mlperf_log.PREPROC_HP_SAMPLE_EVAL_REPLACEMENT, value=True)
mlperf_log.ncf_print(key=mlperf_log.INPUT_HP_SAMPLE_TRAIN_REPLACEMENT)
mlperf_log.ncf_print(key=mlperf_log.INPUT_STEP_EVAL_NEG_GEN)
# sync worker before timing.
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
#===========================================================================
#== The clock starts on loading the preprocessed data. =====================
#===========================================================================
mlperf_log.ncf_print(key=mlperf_log.RUN_START)
run_start_time = time.time()
# load not converted data, just seperate one for test
train_ratings = torch.load(args.data+'/train_ratings.pt', map_location=torch.device('cuda:{}'.format(args.local_rank)))
test_ratings = torch.load(args.data+'/test_ratings.pt', map_location=torch.device('cuda:{}'.format(args.local_rank)))
# get input data
# get dims
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item()+1
nb_items = nb_maxs[1].item()+1
train_users = train_ratings[:,0]
train_items = train_ratings[:,1]
del nb_maxs, train_ratings
mlperf_log.ncf_print(key=mlperf_log.INPUT_SIZE, value=len(train_users))
# produce things not change between epoch
# mask for filtering duplicates with real sample
# note: test data is removed before create mask, same as reference
mat = torch.cuda.ByteTensor(nb_users, nb_items).fill_(1)
mat[train_users, train_items] = 0
# create label
train_label = torch.ones_like(train_users, dtype=torch.float32)
neg_label = torch.zeros_like(train_label, dtype=torch.float32)
neg_label = neg_label.repeat(args.negative_samples)
train_label = torch.cat((train_label,neg_label))
del neg_label
if args.fp16:
train_label = train_label.half()
# produce validation negative sample on GPU
all_test_users = test_ratings.shape[0]
test_users = test_ratings[:,0]
test_pos = test_ratings[:,1].reshape(-1,1)
test_negs = generate_neg(test_users, mat, nb_items, args.valid_negative, True)[1]
# create items with real sample at last position
test_users = test_users.reshape(-1,1).repeat(1,1+args.valid_negative)
test_items = torch.cat((test_negs.reshape(-1,args.valid_negative), test_pos), dim=1)
del test_ratings, test_negs
# generate dup mask and real indice for exact same behavior on duplication compare to reference
# here we need a sort that is stable(keep order of duplicates)
# this is a version works on integer
sorted_items, indices = torch.sort(test_items) # [1,1,1,2], [3,1,0,2]
sum_item_indices = sorted_items.float()+indices.float()/len(indices[0]) #[1.75,1.25,1.0,2.5]
indices_order = torch.sort(sum_item_indices)[1] #[2,1,0,3]
stable_indices = torch.gather(indices, 1, indices_order) #[0,1,3,2]
# produce -1 mask
dup_mask = (sorted_items[:,0:-1] == sorted_items[:,1:])
dup_mask = torch.cat((torch.zeros_like(test_pos, dtype=torch.uint8), dup_mask),dim=1)
dup_mask = torch.gather(dup_mask,1,stable_indices.sort()[1])
# produce real sample indices to later check in topk
sorted_items, indices = (test_items != test_pos).sort()
sum_item_indices = sorted_items.float()+indices.float()/len(indices[0])
indices_order = torch.sort(sum_item_indices)[1]
stable_indices = torch.gather(indices, 1, indices_order)
real_indices = stable_indices[:,0]
del sorted_items, indices, sum_item_indices, indices_order, stable_indices, test_pos
if args.distributed:
test_users = torch.chunk(test_users, args.world_size)[args.local_rank]
test_items = torch.chunk(test_items, args.world_size)[args.local_rank]
dup_mask = torch.chunk(dup_mask, args.world_size)[args.local_rank]
real_indices = torch.chunk(real_indices, args.world_size)[args.local_rank]
# make pytorch memory behavior more consistent later
torch.cuda.empty_cache()
mlperf_log.ncf_print(key=mlperf_log.INPUT_BATCH_SIZE, value=args.batch_size)
mlperf_log.ncf_print(key=mlperf_log.INPUT_ORDER) # we shuffled later with randperm
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d'
% (time.time()-run_start_time, nb_users, nb_items, len(train_users),
nb_users))
# 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])
if args.fp16:
model = model.half()
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
if args.fp16:
fp_optimizer = Fp16Optimizer(model, args.loss_scale)
params = fp_optimizer.fp32_params
else:
params = model.parameters()
#optimizer = torch.optim.Adam(params, lr=args.learning_rate, betas=(args.beta1, args.beta2), eps=args.eps)
# optimizer = AdamOpt(params, lr=args.learning_rate, betas=(args.beta1, args.beta2), eps=args.eps)
optimizer = FusedAdam(params, lr=args.learning_rate, betas=(args.beta1, args.beta2), eps=args.eps, eps_inside_sqrt=False)
criterion = nn.BCEWithLogitsLoss(reduction = 'none') # use torch.mean() with dim later to avoid copy to host
mlperf_log.ncf_print(key=mlperf_log.OPT_LR, value=args.learning_rate)
mlperf_log.ncf_print(key=mlperf_log.OPT_NAME, value="Adam")
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_BETA1, value=args.beta1)
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_BETA2, value=args.beta2)
mlperf_log.ncf_print(key=mlperf_log.OPT_HP_ADAM_EPSILON, value=args.eps)
mlperf_log.ncf_print(key=mlperf_log.MODEL_HP_LOSS_FN, value=mlperf_log.BCE)
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
if args.distributed:
model = DDP(model)
local_batch = args.batch_size // int(os.environ['WORLD_SIZE'])
else:
local_batch = args.batch_size
traced_criterion = torch.jit.trace(criterion.forward, (torch.rand(local_batch,1),torch.rand(local_batch,1)))
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
test_x = test_users.view(-1).split(args.valid_batch_size)
test_y = test_items.view(-1).split(args.valid_batch_size)
hr, ndcg = val_epoch(model, test_x, test_y, dup_mask, real_indices, args.topk, samples_per_user=test_items.size(1),
num_user=all_test_users, distributed=args.distributed)
print('Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'
.format(K=args.topk, hit_rate=hr, ndcg=ndcg))
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)
mlperf_log.ncf_print(key=mlperf_log.INPUT_HP_NUM_NEG, value=args.negative_samples)
mlperf_log.ncf_print(key=mlperf_log.INPUT_STEP_TRAIN_NEG_GEN)
begin = time.time()
# prepare data for epoch
neg_users, neg_items = generate_neg(train_users, mat, nb_items, args.negative_samples)
epoch_users = torch.cat((train_users,neg_users))
epoch_items = torch.cat((train_items,neg_items))
del neg_users, neg_items
# shuffle prepared data and split into batches
epoch_indices = torch.randperm(len(epoch_users), device='cuda:{}'.format(args.local_rank))
epoch_users = epoch_users[epoch_indices]
epoch_items = epoch_items[epoch_indices]
epoch_label = train_label[epoch_indices]
if args.distributed:
epoch_users = torch.chunk(epoch_users, args.world_size)[args.local_rank]
epoch_items = torch.chunk(epoch_items, args.world_size)[args.local_rank]
epoch_label = torch.chunk(epoch_label, args.world_size)[args.local_rank]
epoch_users_list = epoch_users.split(local_batch)
epoch_items_list = epoch_items.split(local_batch)
epoch_label_list = epoch_label.split(local_batch)
# only print progress bar on rank 0
num_batches = (len(epoch_indices) + args.batch_size - 1) // args.batch_size
if args.local_rank == 0:
qbar = tqdm.tqdm(range(num_batches))
else:
qbar = range(num_batches)
# handle extremely rare case where last batch size < number of worker
if len(epoch_users_list) < num_batches:
print("epoch_size % batch_size < number of worker!")
exit(1)
for i in qbar:
# selecting input from prepared data
user = epoch_users_list[i]
item = epoch_items_list[i]
label = epoch_label_list[i].view(-1,1)
for p in model.parameters():
p.grad = None
outputs = model(user, item)
loss = traced_criterion(outputs, label).float()
loss = torch.mean(loss.view(-1), 0)
if args.fp16:
fp_optimizer.step(loss, optimizer)
else:
loss.backward()
optimizer.step()
del epoch_users, epoch_items, epoch_label, epoch_users_list, epoch_items_list, epoch_label_list, user, item, label
train_time = time.time() - begin
begin = time.time()
mlperf_log.ncf_print(key=mlperf_log.EVAL_START)
hr, ndcg = val_epoch(model, test_x, test_y, dup_mask, real_indices, args.topk, samples_per_user=test_items.size(1),
num_user=all_test_users, output=valid_results_file, epoch=epoch, distributed=args.distributed)
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=hr,
ndcg=ndcg, train_time=train_time,
val_time=val_time))
mlperf_log.ncf_print(key=mlperf_log.EVAL_ACCURACY, value={"epoch": epoch, "value": hr})
mlperf_log.ncf_print(key=mlperf_log.EVAL_TARGET, value=args.threshold)
mlperf_log.ncf_print(key=mlperf_log.EVAL_STOP)
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
success = True
break
mlperf_log.ncf_print(key=mlperf_log.RUN_STOP, value={"success": success})
run_stop_time = time.time()
mlperf_log.ncf_print(key=mlperf_log.RUN_FINAL)
# easy way of tracking mlperf score
if success:
print("mlperf_score", run_stop_time - run_start_time)
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
def main():
log_hardware()
args = parse_args()
args.distributed, args.world_size = init_distributed(args.local_rank)
log_args(args)
main_start_time = time.time()
if args.seed is not None:
torch.manual_seed(args.seed)
# Save configuration to file
timestamp = "{:.0f}".format(datetime.utcnow().timestamp())
run_dir = "./run/neumf/{}.{}".format(timestamp, args.local_rank)
print("Saving results to {}".format(run_dir))
if not os.path.exists(run_dir) and run_dir != '':
os.makedirs(run_dir)
# more like load trigger timer now
LOGGER.log(key=tags.PREPROC_HP_NUM_EVAL, value=args.valid_negative)
# The default of np.random.choice is replace=True, so does pytorch random_()
LOGGER.log(key=tags.PREPROC_HP_SAMPLE_EVAL_REPLACEMENT, value=True)
LOGGER.log(key=tags.INPUT_HP_SAMPLE_TRAIN_REPLACEMENT, value=True)
LOGGER.log(key=tags.INPUT_STEP_EVAL_NEG_GEN)
# sync worker before timing.
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
LOGGER.log(key=tags.RUN_START)
run_start_time = time.time()
# load not converted data, just seperate one for test
train_ratings = torch.load(args.data + '/train_ratings.pt',
map_location=torch.device('cuda:{}'.format(
args.local_rank)))
test_ratings = torch.load(args.data + '/test_ratings.pt',
map_location=torch.device('cuda:{}'.format(
args.local_rank)))
# get input data
# get dims
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item() + 1
nb_items = nb_maxs[1].item() + 1
train_users = train_ratings[:, 0]
train_items = train_ratings[:, 1]
del nb_maxs, train_ratings
LOGGER.log(key=tags.INPUT_SIZE, value=len(train_users))
# produce things not change between epoch
# mask for filtering duplicates with real sample
# note: test data is removed before create mask, same as reference
mat = torch.cuda.ByteTensor(nb_users, nb_items).fill_(1)
mat[train_users, train_items] = 0
# create label
train_label = torch.ones_like(train_users, dtype=torch.float32)
neg_label = torch.zeros_like(train_label, dtype=torch.float32)
neg_label = neg_label.repeat(args.negative_samples)
train_label = torch.cat((train_label, neg_label))
del neg_label
if args.fp16:
train_label = train_label.half()
# produce validation negative sample on GPU
all_test_users = test_ratings.shape[0]
test_users = test_ratings[:, 0]
test_pos = test_ratings[:, 1].reshape(-1, 1)
test_negs = generate_neg(test_users, mat, nb_items, args.valid_negative,
True)[1]
# create items with real sample at last position
test_users = test_users.reshape(-1, 1).repeat(1, 1 + args.valid_negative)
test_items = torch.cat(
(test_negs.reshape(-1, args.valid_negative), test_pos), dim=1)
del test_ratings, test_negs
# generate dup mask and real indice for exact same behavior on duplication compare to reference
# here we need a sort that is stable(keep order of duplicates)
# this is a version works on integer
sorted_items, indices = torch.sort(test_items) # [1,1,1,2], [3,1,0,2]
sum_item_indices = sorted_items.float() + indices.float() / len(
indices[0]) #[1.75,1.25,1.0,2.5]
indices_order = torch.sort(sum_item_indices)[1] #[2,1,0,3]
stable_indices = torch.gather(indices, 1, indices_order) #[0,1,3,2]
# produce -1 mask
dup_mask = (sorted_items[:, 0:-1] == sorted_items[:, 1:])
dup_mask = torch.cat(
(torch.zeros_like(test_pos, dtype=torch.uint8), dup_mask), dim=1)
dup_mask = torch.gather(dup_mask, 1, stable_indices.sort()[1])
# produce real sample indices to later check in topk
sorted_items, indices = (test_items != test_pos).sort()
sum_item_indices = sorted_items.float() + indices.float() / len(indices[0])
indices_order = torch.sort(sum_item_indices)[1]
stable_indices = torch.gather(indices, 1, indices_order)
real_indices = stable_indices[:, 0]
del sorted_items, indices, sum_item_indices, indices_order, stable_indices, test_pos
if args.distributed:
test_users = torch.chunk(test_users, args.world_size)[args.local_rank]
test_items = torch.chunk(test_items, args.world_size)[args.local_rank]
dup_mask = torch.chunk(dup_mask, args.world_size)[args.local_rank]
real_indices = torch.chunk(real_indices,
args.world_size)[args.local_rank]
# make pytorch memory behavior more consistent later
torch.cuda.empty_cache()
LOGGER.log(key=tags.INPUT_BATCH_SIZE, value=args.batch_size)
LOGGER.log(key=tags.INPUT_ORDER) # we shuffled later with randperm
print('Load data done [%.1f s]. #user=%d, #item=%d, #train=%d, #test=%d' %
(time.time() - run_start_time, nb_users, nb_items, len(train_users),
nb_users))
# 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],
dropout=args.dropout)
if args.fp16:
model = model.half()
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
if args.fp16:
fp_optimizer = Fp16Optimizer(model, args.loss_scale)
params = fp_optimizer.fp32_params
else:
params = model.parameters()
optimizer = FusedAdam(params,
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
eps=args.eps,
eps_inside_sqrt=False)
criterion = nn.BCEWithLogitsLoss(
reduction='none'
) # use torch.mean() with dim later to avoid copy to host
LOGGER.log(key=tags.OPT_LR, value=args.learning_rate)
LOGGER.log(key=tags.OPT_NAME, value="Adam")
LOGGER.log(key=tags.OPT_HP_ADAM_BETA1, value=args.beta1)
LOGGER.log(key=tags.OPT_HP_ADAM_BETA2, value=args.beta2)
LOGGER.log(key=tags.OPT_HP_ADAM_EPSILON, value=args.eps)
LOGGER.log(key=tags.MODEL_HP_LOSS_FN, value=tags.VALUE_BCE)
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
if args.distributed:
model = DDP(model)
local_batch = args.batch_size // int(os.environ['WORLD_SIZE'])
else:
local_batch = args.batch_size
traced_criterion = torch.jit.trace(
criterion.forward,
(torch.rand(local_batch, 1), torch.rand(local_batch, 1)))
train_users_per_worker = len(train_label) / int(os.environ['WORLD_SIZE'])
train_users_begin = int(train_users_per_worker * args.local_rank)
train_users_end = int(train_users_per_worker * (args.local_rank + 1))
# Create files for tracking training
valid_results_file = os.path.join(run_dir, 'valid_results.csv')
# Calculate initial Hit Ratio and NDCG
test_x = test_users.view(-1).split(args.valid_batch_size)
test_y = test_items.view(-1).split(args.valid_batch_size)
if args.mode == 'test':
state_dict = torch.load(args.checkpoint_path)
model.load_state_dict(state_dict)
begin = time.time()
LOGGER.log(key=tags.EVAL_START, value=-1)
hr, ndcg = val_epoch(model,
test_x,
test_y,
dup_mask,
real_indices,
args.topk,
samples_per_user=test_items.size(1),
num_user=all_test_users,
distributed=args.distributed)
val_time = time.time() - begin
print(
'Initial HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}, valid_time: {val_time:.4f}'
.format(K=args.topk, hit_rate=hr, ndcg=ndcg, val_time=val_time))
LOGGER.log(key=tags.EVAL_ACCURACY, value={"epoch": -1, "value": hr})
LOGGER.log(key=tags.EVAL_TARGET, value=args.threshold)
LOGGER.log(key=tags.EVAL_STOP, value=-1)
if args.mode == 'test':
return
success = False
max_hr = 0
LOGGER.log(key=tags.TRAIN_LOOP)
train_throughputs = []
eval_throughputs = []
for epoch in range(args.epochs):
LOGGER.log(key=tags.TRAIN_EPOCH_START, value=epoch)
LOGGER.log(key=tags.INPUT_HP_NUM_NEG, value=args.negative_samples)
LOGGER.log(key=tags.INPUT_STEP_TRAIN_NEG_GEN)
begin = time.time()
# prepare data for epoch
neg_users, neg_items = generate_neg(train_users, mat, nb_items,
args.negative_samples)
epoch_users = torch.cat((train_users, neg_users))
epoch_items = torch.cat((train_items, neg_items))
del neg_users, neg_items
# shuffle prepared data and split into batches
epoch_indices = torch.randperm(train_users_end - train_users_begin,
device='cuda:{}'.format(
args.local_rank))
epoch_indices += train_users_begin
epoch_users = epoch_users[epoch_indices]
epoch_items = epoch_items[epoch_indices]
epoch_label = train_label[epoch_indices]
epoch_users_list = epoch_users.split(local_batch)
epoch_items_list = epoch_items.split(local_batch)
epoch_label_list = epoch_label.split(local_batch)
# only print progress bar on rank 0
num_batches = len(epoch_users_list)
# handle extremely rare case where last batch size < number of worker
if len(epoch_users) % args.batch_size < args.world_size:
print("epoch_size % batch_size < number of worker!")
exit(1)
for i in range(num_batches // args.grads_accumulated):
for j in range(args.grads_accumulated):
batch_idx = (args.grads_accumulated * i) + j
user = epoch_users_list[batch_idx]
item = epoch_items_list[batch_idx]
label = epoch_label_list[batch_idx].view(-1, 1)
outputs = model(user, item)
loss = traced_criterion(outputs, label).float()
loss = torch.mean(loss.view(-1), 0)
if args.fp16:
fp_optimizer.backward(loss)
else:
loss.backward()
if args.fp16:
fp_optimizer.step(optimizer)
else:
optimizer.step()
for p in model.parameters():
p.grad = None
del epoch_users, epoch_items, epoch_label, epoch_users_list, epoch_items_list, epoch_label_list, user, item, label
train_time = time.time() - begin
begin = time.time()
epoch_samples = len(train_users) * (args.negative_samples + 1)
train_throughput = epoch_samples / train_time
train_throughputs.append(train_throughput)
LOGGER.log(key='train_throughput', value=train_throughput)
LOGGER.log(key=tags.TRAIN_EPOCH_STOP, value=epoch)
LOGGER.log(key=tags.EVAL_START, value=epoch)
hr, ndcg = val_epoch(model,
test_x,
test_y,
dup_mask,
real_indices,
args.topk,
samples_per_user=test_items.size(1),
num_user=all_test_users,
output=valid_results_file,
epoch=epoch,
distributed=args.distributed)
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=hr,
ndcg=ndcg,
train_time=train_time,
val_time=val_time))
LOGGER.log(key=tags.EVAL_ACCURACY, value={"epoch": epoch, "value": hr})
LOGGER.log(key=tags.EVAL_TARGET, value=args.threshold)
LOGGER.log(key=tags.EVAL_STOP, value=epoch)
eval_size = all_test_users * test_items.size(1)
eval_throughput = eval_size / val_time
eval_throughputs.append(eval_throughput)
LOGGER.log(key='eval_throughput', value=eval_throughput)
if hr > max_hr and args.local_rank == 0:
max_hr = hr
print("New best hr! Saving the model to: ", args.checkpoint_path)
torch.save(model.state_dict(), args.checkpoint_path)
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
success = True
break
LOGGER.log(key='best_train_throughput', value=max(train_throughputs))
LOGGER.log(key='best_eval_throughput', value=max(eval_throughputs))
LOGGER.log(key='best_accuracy', value=max_hr)
LOGGER.log(key='time_to_target', value=time.time() - main_start_time)
LOGGER.log(key=tags.RUN_STOP, value={"success": success})
LOGGER.log(key=tags.RUN_FINAL)