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():
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')
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
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()
args.distributed, args.world_size = init_distributed(args.local_rank)
log_args(args)
if args.seed is not None:
torch.manual_seed(args.seed)
print("Saving results to {}".format(args.checkpoint_dir))
if not os.path.exists(args.checkpoint_dir) and args.checkpoint_dir != '':
os.makedirs(args.checkpoint_dir, exist_ok=True)
# 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 workers before timing
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
main_start_time = time.time()
LOGGER.log(key=tags.RUN_START)
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)))
test_negs = torch.load(args.data + '/test_negatives.pt',
map_location=torch.device('cuda:{}'.format(
args.local_rank)))
valid_negative = test_negs.shape[1]
LOGGER.log(key=tags.PREPROC_HP_NUM_EVAL, value=valid_negative)
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item() + 1
nb_items = nb_maxs[1].item() + 1
LOGGER.log(key=tags.INPUT_SIZE, value=len(train_ratings))
all_test_users = test_ratings.shape[0]
test_users, test_items, dup_mask, real_indices = dataloading.create_test_data(
test_ratings, test_negs, args)
# 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
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
optimizer = FusedAdam(model.parameters(),
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
eps=args.eps)
criterion = nn.BCEWithLogitsLoss(
reduction='none'
) # use torch.mean() with dim later to avoid copy to host
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
if args.opt_level == "O2":
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.opt_level,
keep_batchnorm_fp32=False,
loss_scale='dynamic')
if args.distributed:
model = DDP(model)
local_batch = args.batch_size // args.world_size
traced_criterion = torch.jit.trace(
criterion.forward,
(torch.rand(local_batch, 1), torch.rand(local_batch, 1)))
print(model)
print("{} parameters".format(admm_utils.count_parameters(model)))
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)
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
state_dict = {
k.replace('module.', ''): v
for k, v in state_dict.items()
}
model.load_state_dict(state_dict)
if args.mode == 'test':
LOGGER.log(key=tags.EVAL_START, value=0)
start = time.time()
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
distributed=args.distributed)
print('HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(
K=args.topk, hit_rate=hr, ndcg=ndcg))
val_time = time.time() - start
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
LOGGER.log(key=tags.EVAL_ACCURACY, value={"epoch": 0, "value": hr})
LOGGER.log(key=tags.EVAL_STOP, value=0)
LOGGER.log(key='best_eval_throughput', value=eval_throughput)
return
success = False
max_hr = 0
train_throughputs, eval_throughputs = [], []
LOGGER.log(key=tags.TRAIN_LOOP)
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()
epoch_users, epoch_items, epoch_label = dataloading.prepare_epoch_train_data(
train_ratings, nb_items, args)
num_batches = len(epoch_users)
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[batch_idx]
item = epoch_items[batch_idx]
label = epoch_label[batch_idx].view(-1, 1)
outputs = model(user, item)
loss = traced_criterion(outputs, label).float()
loss = torch.mean(loss.view(-1), 0)
if args.opt_level == "O2":
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
for p in model.parameters():
p.grad = None
del epoch_users, epoch_items, epoch_label
train_time = time.time() - begin
begin = time.time()
epoch_samples = len(train_ratings) * (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_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
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 * (valid_negative + 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
save_checkpoint_path = os.path.join(args.checkpoint_dir,
'model.pth')
print("New best hr! Saving the model to: ", save_checkpoint_path)
torch.save(model.state_dict(), save_checkpoint_path)
best_model_timestamp = time.time()
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
success = True
break
if args.local_rank == 0:
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='time_to_best_model',
value=best_model_timestamp - main_start_time)
LOGGER.log(key=tags.RUN_STOP, value={"success": success})
LOGGER.log(key=tags.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)
# Your code here
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)
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
# Create model
model = NeuMF(2197225,
855776,
mf_dim=64,
mf_reg=0.,
mlp_layer_sizes=[256, 256, 128, 64],
mlp_layer_regs=[0. for i in [256, 256, 128, 64]])
print(model)
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
if args.load_ckp:
ckp = torch.load(args.load_ckp)
model.load_state_dict(ckp)
if args.quantize:
all_embeding = [
n for n, m in model.named_modules() if isinstance(m, nn.Embedding)
]
all_linear = [
n for n, m in model.named_modules() if isinstance(m, nn.Linear)
]
all_relu = [
n for n, m in model.named_modules() if isinstance(m, nn.ReLU)
]
all_relu6 = [
n for n, m in model.named_modules() if isinstance(m, nn.ReLU6)
]
# layers = all_relu + all_relu6 + all_linear
layers = all_embeding
replacement_factory = {
nn.ReLU: ActivationModuleWrapperPost,
nn.ReLU6: ActivationModuleWrapperPost,
nn.Linear: ParameterModuleWrapperPost,
nn.Embedding: ActivationModuleWrapperPost
}
mq = ModelQuantizer(model, args, layers, replacement_factory)
# mq.log_quantizer_state(ml_logger, -1)
test_users, test_items, dup_mask, real_indices, K, samples_per_user, num_user = data_loader(
args.data)
data = NcfData(test_users, test_items, dup_mask, real_indices, K,
samples_per_user, num_user)
hr, ndcg = val(model, data)
print('')
print('')
print('HR@{K} = {hit_rate:.4f}, NDCG@{K} = {ndcg:.4f}'.format(K=K,
hit_rate=hr,
ndcg=ndcg))
def main():
args = parse_args()
print("init distributed")
init_distributed(args)
if args.rank == 0:
wandb.init()
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
])
else:
dllogger.init(backends=[])
dllogger.log(data=vars(args), step='PARAMETER')
torch.manual_seed(1)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# sync workers before timing
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
main_start_time = time.time()
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)))
test_negs = torch.load(args.data + '/test_negatives.pt',
map_location=torch.device('cuda:{}'.format(
args.local_rank)))
valid_negative = test_negs.shape[1]
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item() + 1
nb_items = nb_maxs[1].item() + 1
all_test_users = test_ratings.shape[0]
test_users, test_items, dup_mask, real_indices = dataloading.create_test_data(
test_ratings, test_negs, args)
# make pytorch memory behavior more consistent later
torch.cuda.empty_cache()
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout).cuda()
optimizer = torch.optim.Adam(model.parameters(),
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
eps=args.eps)
criterion = nn.BCEWithLogitsLoss(reduction='none').cuda(
) # use torch.mean() with dim later to avoid copy to host
# Move model and loss to GPU
if args.distributed:
model = DDP(model, device_ids=[args.local_rank])
local_batch = args.batch_size // args.world_size
traced_criterion = torch.jit.trace(
criterion.forward,
(torch.rand(local_batch, 1), torch.rand(local_batch, 1)))
print(model)
print("{} parameters".format(utils.count_parameters(model)))
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
state_dict = {
k.replace('module.', ''): v
for k, v in state_dict.items()
}
model.load_state_dict(state_dict)
if args.mode == 'test':
start = time.time()
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
distributed=args.distributed)
val_time = time.time() - start
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
dllogger.log(step=tuple(),
data={
'best_eval_throughput': eval_throughput,
'hr@10': hr
})
return
max_hr = 0
best_epoch = 0
train_throughputs, eval_throughputs = [], []
for epoch in range(args.epochs):
begin = time.time()
epoch_users, epoch_items, epoch_label = dataloading.prepare_epoch_train_data(
train_ratings, nb_items, args)
num_batches = len(epoch_users)
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[batch_idx]
item = epoch_items[batch_idx]
label = epoch_label[batch_idx].view(-1, 1)
outputs = model(user, item)
loss = traced_criterion(outputs, label).float()
loss = torch.mean(loss.view(-1), 0)
loss.backward()
optimizer.step()
if args.rank == 0:
wandb.log({"Test loss": loss})
for p in model.parameters():
p.grad = None
del epoch_users, epoch_items, epoch_label
train_time = time.time() - begin
begin = time.time()
epoch_samples = len(train_ratings) * (args.negative_samples + 1)
train_throughput = epoch_samples / train_time
train_throughputs.append(train_throughput)
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
epoch=epoch,
distributed=args.distributed)
val_time = time.time() - begin
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
eval_throughputs.append(eval_throughput)
dllogger.log(step=(epoch, ),
data={
'train_throughput': train_throughput,
'hr@10': hr,
'train_epoch_time': train_time,
'validation_epoch_time': val_time,
'eval_throughput': eval_throughput
})
if args.rank == 0:
wandb.log({"Test hit rate": hr})
wandb.log({"Test train epoch time": train_time})
wandb.log({"Test train throughput": train_throughput})
if hr > max_hr and args.local_rank == 0:
max_hr = hr
best_epoch = epoch
# save_checkpoint_path = os.path.join(args.checkpoint_dir, 'model.pth')
print("New best hr!")
# torch.save(model.state_dict(), save_checkpoint_path)
best_model_timestamp = time.time()
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
break
if args.local_rank == 0:
dllogger.log(data={
'best_train_throughput':
max(train_throughputs),
'best_eval_throughput':
max(eval_throughputs),
'mean_train_throughput':
np.mean(train_throughputs),
'mean_eval_throughput':
np.mean(eval_throughputs),
'best_accuracy':
max_hr,
'best_epoch':
best_epoch,
'time_to_target':
time.time() - main_start_time,
'time_to_best_model':
best_model_timestamp - main_start_time
},
step=tuple())
def main():
args = parse_args()
init_distributed(args)
print(args.desc)
mlflow.start_run(run_name=args.desc)
mlflow.log_param('batch_size', args.batch_size)
mlflow.log_param('num_threads', args.threads)
mlflow.log_param('num_of_epochs', args.epochs)
torch.set_num_threads(args.threads)
print(f"{vars(args)} step='PARAMETER'")
if args.seed is not None:
torch.manual_seed(args.seed)
if not os.path.exists(args.checkpoint_dir) and args.checkpoint_dir:
print("Saving results to {}".format(args.checkpoint_dir))
os.makedirs(args.checkpoint_dir, exist_ok=True)
# sync workers before timing
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
#torch.cuda.synchronize()
main_start_time = time.time()
train_ratings = torch.load(args.data + '/train_ratings.pt',
map_location=torch.device('cpu'))
test_ratings = torch.load(args.data + '/test_ratings.pt',
map_location=torch.device('cpu'))
test_negs = torch.load(args.data + '/test_negatives.pt',
map_location=torch.device('cpu'))
valid_negative = test_negs.shape[1]
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item() + 1
nb_items = nb_maxs[1].item() + 1
all_test_users = test_ratings.shape[0]
test_users, test_items, dup_mask, real_indices = dataloading.create_test_data(
test_ratings, test_negs, args)
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
optimizer = Adam(model.parameters(),
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
eps=args.eps)
criterion = nn.BCEWithLogitsLoss(
reduction='none'
) # use torch.mean() with dim later to avoid copy to host
if args.amp:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
keep_batchnorm_fp32=False,
loss_scale='dynamic')
local_batch = args.batch_size // args.world_size
traced_criterion = torch.jit.trace(
criterion.forward,
(torch.rand(local_batch, 1), torch.rand(local_batch, 1)))
print(model)
print("{} parameters".format(utils.count_parameters(model)))
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
state_dict = {
k.replace('module.', ''): v
for k, v in state_dict.items()
}
model.load_state_dict(state_dict)
if args.mode == 'test':
start = time.time()
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
distributed=args.distributed)
val_time = time.time() - start
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
log_data = data = {
'best_eval_throughput': eval_throughput,
'test_hr_at_10': hr
}
print(log_data)
return
max_hr = 0
best_epoch = 0
train_throughputs, eval_throughputs = [], []
for epoch in range(args.epochs):
begin = time.time()
epoch_users, epoch_items, epoch_label = dataloading.prepare_epoch_train_data(
train_ratings, nb_items, args)
num_batches = len(epoch_users)
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[batch_idx]
item = epoch_items[batch_idx]
label = epoch_label[batch_idx].view(-1, 1)
outputs = model(user, item)
loss = traced_criterion(outputs, label).float()
loss = torch.mean(loss.view(-1), 0)
if args.amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
for p in model.parameters():
p.grad = None
del epoch_users, epoch_items, epoch_label
train_time = time.time() - begin
begin = time.time()
epoch_samples = len(train_ratings) * (args.negative_samples + 1)
train_throughput = epoch_samples / train_time
train_throughputs.append(train_throughput)
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
epoch=epoch,
distributed=args.distributed)
val_time = time.time() - begin
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
eval_throughputs.append(eval_throughput)
log_data = {
'train_throughput': train_throughput,
'hr_at_10': hr,
'train_epoch_time': train_time,
'validation_epoch_time': val_time,
'eval_throughput': eval_throughput
}
print(log_data)
mlflow.log_metrics(log_data, epoch)
if hr > max_hr and args.local_rank == 0:
max_hr = hr
best_epoch = epoch
print("New best hr!")
if args.checkpoint_dir:
save_checkpoint_path = os.path.join(args.checkpoint_dir,
'model.pth')
print("Saving the model to: ", save_checkpoint_path)
torch.save(model.state_dict(), save_checkpoint_path)
best_model_timestamp = time.time()
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
break
if args.local_rank == 0:
log_data = {
'best_train_throughput': max(train_throughputs),
'best_eval_throughput': max(eval_throughputs),
'mean_train_throughput': np.mean(train_throughputs),
'mean_eval_throughput': np.mean(eval_throughputs),
'best_accuracy': max_hr,
'best_epoch': best_epoch,
'time_to_target': time.time() - main_start_time
}
print(log_data)
mlflow.log_metrics(log_data)
mlflow.end_run()
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)
def main():
args = parse_args()
init_distributed(args)
if args.local_rank == 0:
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
])
else:
dllogger.init(backends=[])
dllogger.metadata('train_throughput', {
"name": 'train_throughput',
'format': ":.3e"
})
dllogger.metadata('hr@10', {"name": 'hr@10', 'format': ":.5f"})
dllogger.metadata('train_epoch_time', {
"name": 'train_epoch_time',
'format': ":.3f"
})
dllogger.metadata('validation_epoch_time', {
"name": 'validation_epoch_time',
'format': ":.3f"
})
dllogger.metadata('eval_throughput', {
"name": 'eval_throughput',
'format': ":.3e"
})
dllogger.log(data=vars(args), step='PARAMETER')
if args.seed is not None:
torch.manual_seed(args.seed)
if not os.path.exists(args.checkpoint_dir) and args.checkpoint_dir:
print("Saving results to {}".format(args.checkpoint_dir))
os.makedirs(args.checkpoint_dir, exist_ok=True)
# sync workers before timing
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
main_start_time = time.time()
feature_spec_path = os.path.join(args.data, args.feature_spec_file)
feature_spec = FeatureSpec.from_yaml(feature_spec_path)
trainset = dataloading.TorchTensorDataset(feature_spec,
mapping_name='train',
args=args)
testset = dataloading.TorchTensorDataset(feature_spec,
mapping_name='test',
args=args)
train_loader = dataloading.TrainDataloader(trainset, args)
test_loader = dataloading.TestDataLoader(testset, args)
# make pytorch memory behavior more consistent later
torch.cuda.empty_cache()
# Create model
user_feature_name = feature_spec.channel_spec[USER_CHANNEL_NAME][0]
item_feature_name = feature_spec.channel_spec[ITEM_CHANNEL_NAME][0]
label_feature_name = feature_spec.channel_spec[LABEL_CHANNEL_NAME][0]
model = NeuMF(
nb_users=feature_spec.feature_spec[user_feature_name]['cardinality'],
nb_items=feature_spec.feature_spec[item_feature_name]['cardinality'],
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
optimizer = FusedAdam(model.parameters(),
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
eps=args.eps)
criterion = nn.BCEWithLogitsLoss(
reduction='none'
) # use torch.mean() with dim later to avoid copy to host
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
if args.amp:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
keep_batchnorm_fp32=False,
loss_scale='dynamic')
if args.distributed:
model = DDP(model)
local_batch = args.batch_size // args.world_size
traced_criterion = torch.jit.trace(
criterion.forward,
(torch.rand(local_batch, 1), torch.rand(local_batch, 1)))
print(model)
print("{} parameters".format(utils.count_parameters(model)))
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
state_dict = {
k.replace('module.', ''): v
for k, v in state_dict.items()
}
model.load_state_dict(state_dict)
if args.mode == 'test':
start = time.time()
hr, ndcg = val_epoch(model,
test_loader,
args.topk,
distributed=args.distributed)
val_time = time.time() - start
eval_size = test_loader.raw_dataset_length
eval_throughput = eval_size / val_time
dllogger.log(step=tuple(),
data={
'best_eval_throughput': eval_throughput,
'hr@10': hr
})
return
# this should always be overridden if hr>0.
# It is theoretically possible for the hit rate to be zero in the first epoch, which would result in referring
# to an uninitialized variable.
max_hr = 0
best_epoch = 0
best_model_timestamp = time.time()
train_throughputs, eval_throughputs = [], []
for epoch in range(args.epochs):
begin = time.time()
batch_dict_list = train_loader.get_epoch_data()
num_batches = len(batch_dict_list)
for i in range(num_batches // args.grads_accumulated):
for j in range(args.grads_accumulated):
batch_idx = (args.grads_accumulated * i) + j
batch_dict = batch_dict_list[batch_idx]
user_features = batch_dict[USER_CHANNEL_NAME]
item_features = batch_dict[ITEM_CHANNEL_NAME]
user_batch = user_features[user_feature_name]
item_batch = item_features[item_feature_name]
label_features = batch_dict[LABEL_CHANNEL_NAME]
label_batch = label_features[label_feature_name]
outputs = model(user_batch, item_batch)
loss = traced_criterion(outputs, label_batch.view(-1,
1)).float()
loss = torch.mean(loss.view(-1), 0)
if args.amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
for p in model.parameters():
p.grad = None
del batch_dict_list
train_time = time.time() - begin
begin = time.time()
epoch_samples = train_loader.length_after_augmentation
train_throughput = epoch_samples / train_time
train_throughputs.append(train_throughput)
hr, ndcg = val_epoch(model,
test_loader,
args.topk,
distributed=args.distributed)
val_time = time.time() - begin
eval_size = test_loader.raw_dataset_length
eval_throughput = eval_size / val_time
eval_throughputs.append(eval_throughput)
dllogger.log(step=(epoch, ),
data={
'train_throughput': train_throughput,
'hr@10': hr,
'train_epoch_time': train_time,
'validation_epoch_time': val_time,
'eval_throughput': eval_throughput
})
if hr > max_hr and args.local_rank == 0:
max_hr = hr
best_epoch = epoch
print("New best hr!")
if args.checkpoint_dir:
save_checkpoint_path = os.path.join(args.checkpoint_dir,
'model.pth')
print("Saving the model to: ", save_checkpoint_path)
torch.save(model.state_dict(), save_checkpoint_path)
best_model_timestamp = time.time()
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
break
if args.local_rank == 0:
dllogger.log(data={
'best_train_throughput':
max(train_throughputs),
'best_eval_throughput':
max(eval_throughputs),
'mean_train_throughput':
np.mean(train_throughputs),
'mean_eval_throughput':
np.mean(eval_throughputs),
'best_accuracy':
max_hr,
'best_epoch':
best_epoch,
'time_to_target':
time.time() - main_start_time,
'time_to_best_model':
best_model_timestamp - main_start_time
},
step=tuple())
def main():
from grace_dl.dist.helper import timer, volume, tensor_bits
args = parse_args()
init_distributed(args)
if args.weak_scaling:
args.batch_size *= args.world_size
init_wandb(args)
init_grace(args)
if args.local_rank == 0:
dllogger.init(backends=[
dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)
])
else:
dllogger.init(backends=[])
dllogger.log(data=vars(args), step='PARAMETER')
if not os.path.exists(args.checkpoint_dir) and args.checkpoint_dir:
print("Saving results to {}".format(args.checkpoint_dir))
os.makedirs(args.checkpoint_dir, exist_ok=True)
# sync workers before timing
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
main_start_time = time.time()
if args.seed is not None:
torch.manual_seed(args.seed)
train_ratings = torch.load(args.data + '/train_ratings.pt',
map_location=torch.device('cuda:0'))
test_ratings = torch.load(args.data + '/test_ratings.pt',
map_location=torch.device('cuda:0'))
test_negs = torch.load(args.data + '/test_negatives.pt',
map_location=torch.device('cuda:0'))
valid_negative = test_negs.shape[1]
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item() + 1
nb_items = nb_maxs[1].item() + 1
all_test_users = test_ratings.shape[0]
test_users, test_items, dup_mask, real_indices = dataloading.create_test_data(
test_ratings, test_negs, args)
# make pytorch memory behavior more consistent later
torch.cuda.empty_cache()
# Create model
model = NeuMF(nb_users,
nb_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
optimizer = FusedAdam(model.parameters(),
lr=args.learning_rate,
betas=(args.beta1, args.beta2),
eps=args.eps)
criterion = nn.BCEWithLogitsLoss(
reduction='none'
) # use torch.mean() with dim later to avoid copy to host
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
if args.amp:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
keep_batchnorm_fp32=False,
loss_scale='dynamic')
# if args.distributed:
# model = DDP(model)
local_batch = args.batch_size // args.world_size
traced_criterion = torch.jit.trace(
criterion.forward,
(torch.rand(local_batch, 1), torch.rand(local_batch, 1)))
if args.local_rank == 0:
print(model)
print("{} parameters".format(utils.count_parameters(model)))
# [print(parameter) for parameter in model.parameters()]
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
state_dict = {
k.replace('module.', ''): v
for k, v in state_dict.items()
}
model.load_state_dict(state_dict)
if args.mode == 'test':
start = time.time()
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
distributed=args.distributed)
val_time = time.time() - start
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
dllogger.log(step=tuple(),
data={
'best_eval_throughput': eval_throughput,
'hr@10': hr
})
return
max_hr = 0
best_epoch = 0
train_throughputs, eval_throughputs = [], []
# broadcast model states from rank0 to other nodes !!! This is important!
[torch.distributed.broadcast(p.data, src=0) for p in model.parameters()]
# if args.local_rank == 0:
# save_initial_state_path = os.path.join(args.checkpoint_dir, 'model_init.pth')
# print("Saving the model to: ", save_initial_state_path)
# torch.save(model.state_dict(), save_initial_state_path)
for epoch in range(args.epochs):
begin = time.time()
train_time = 0
epoch_users, epoch_items, epoch_label = dataloading.prepare_epoch_train_data(
train_ratings, nb_items, args)
num_batches = len(epoch_users)
for i in range(num_batches // args.grads_accumulated):
batch_start = time.time()
for j in range(args.grads_accumulated):
batch_idx = (args.grads_accumulated * i) + j
user = epoch_users[batch_idx]
item = epoch_items[batch_idx]
label = epoch_label[batch_idx].view(-1, 1)
outputs = model(user, item)
loss = traced_criterion(outputs, label).float()
loss = torch.mean(loss.view(-1), 0)
if args.amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# check grad sparsity
if args.sparsity_check:
total_nonzero = 0
total_numel = 0
for index, (name, p) in enumerate(model.named_parameters()):
sparsity = 1.0 - torch.sum(
p.grad.data.abs() > 0).float() / p.grad.data.numel()
total_nonzero += torch.sum(p.grad.data.abs() > 0).float()
total_numel += p.grad.data.numel()
if args.local_rank == 0:
wandb.log(
{
f"{name}(sparsity)(numel={p.grad.data.numel()})":
sparsity,
},
commit=False)
if args.local_rank == 0:
wandb.log(
{
f"total_sparsity(numel={total_numel})":
1 - total_nonzero / total_numel,
},
commit=True)
# add grace just before optimizer.step()
torch.cuda.synchronize()
comm_start = time.time()
for index, (name, p) in enumerate(model.named_parameters()):
new_grad = args.grc.step(p.grad.data, name)
p.grad.data = new_grad
torch.cuda.synchronize()
timer['comm'] = time.time() - comm_start
# [torch.distributed.all_reduce(p.grad.data) for p in model.parameters()]
# for param in model.parameters():
# dist.all_reduce(param.grad.data)
# param.grad.data /= float(args.world_size)
optimizer.step()
for p in model.parameters():
p.grad = None
if args.throughput:
torch.cuda.synchronize()
if args.log_time and args.local_rank == 0:
timer['batch_time'] = time.time() - batch_start
timer['computation'] = timer['batch_time'] - timer['comm']
print("Timer:", timer, '\n')
timer['en/decoding'] = 0
timer['batch_time'] = 0
timer['computation'] = 0
timer['comm'] = 0
if args.log_volume and args.local_rank == 0:
ratio = volume['compress'] / volume['nocompress']
volume['ratio_acc'].append(ratio)
avg_ratio = sum(volume['ratio_acc']) / len(volume['ratio_acc'])
print(
f"Data volume:: compress {volume['compress']} no_compress {volume['nocompress']} ratio {ratio:.4f} avg_ratio {avg_ratio:.4f}"
)
volume['compress'] = 0
volume['nocompress'] = 0
batch_throughput = args.batch_size / (time.time() - batch_start
) # global throughput
train_time += time.time() - batch_start
if (args.throughput
or args.eval_at_every_batch) and args.local_rank == 0:
print(
f"Train :: Epoch [{epoch}/{args.epochs}] \t Batch [{i}/{num_batches}] \t "
f"Time {time.time()-batch_start:.5f} \t Throughput {batch_throughput:.2f}"
)
if args.throughput and i == 3:
break
if args.local_rank == 0:
print(
f"Train :: Epoch [{epoch}/{args.epochs}] \t Batch [{i}/{num_batches}] \t "
f"Time {time.time()-batch_start:.5f} \t Throughput {batch_throughput:.2f}"
)
if args.eval_at_every_batch:
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
epoch=epoch,
distributed=args.distributed)
if args.local_rank == 0:
wandb.log({
"eval/hr@10": hr,
})
del epoch_users, epoch_items, epoch_label
# train_time = time.time() - begin
begin = time.time()
epoch_samples = len(train_ratings) * (args.negative_samples + 1)
train_throughput = epoch_samples / train_time
if args.throughput:
train_throughput = batch_throughput
train_throughputs.append(train_throughput)
hr, ndcg = val_epoch(model,
test_users,
test_items,
dup_mask,
real_indices,
args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users,
epoch=epoch,
distributed=args.distributed)
val_time = time.time() - begin
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
eval_throughputs.append(eval_throughput)
dllogger.log(step=(epoch, ),
data={
'train_throughput': train_throughput,
'hr@10': hr,
'train_epoch_time': train_time,
'validation_epoch_time': val_time,
'eval_throughput': eval_throughput
})
if args.local_rank == 0:
wandb.log(
{
"train_epoch_time": train_time,
'validation_epoch_time': val_time,
'eval_throughput': eval_throughput,
'train_throughput': train_throughput,
},
commit=False)
if not args.eval_at_every_batch:
wandb.log({
"eval/hr@10": hr,
}, commit=False)
wandb.log({"epoch": epoch})
if hr > max_hr and args.local_rank == 0:
max_hr = hr
best_epoch = epoch
print("New best hr!")
if args.checkpoint_dir:
save_checkpoint_path = os.path.join(args.checkpoint_dir,
'model.pth')
print("Saving the model to: ", save_checkpoint_path)
torch.save(model.state_dict(), save_checkpoint_path)
best_model_timestamp = time.time()
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
break
if args.throughput:
break
if args.local_rank == 0:
dllogger.log(data={
'best_train_throughput':
max(train_throughputs),
'best_eval_throughput':
max(eval_throughputs),
'mean_train_throughput':
np.mean(train_throughputs),
'mean_eval_throughput':
np.mean(eval_throughputs),
'best_accuracy':
max_hr,
'best_epoch':
best_epoch,
'time_to_target':
time.time() - main_start_time,
'time_to_best_model':
best_model_timestamp - main_start_time
},
step=tuple())
wandb.log({
'best_train_throughput': max(train_throughputs),
'best_eval_throughput': max(eval_throughputs),
'mean_train_throughput': np.mean(train_throughputs),
'mean_eval_throughput': np.mean(eval_throughputs),
'best_accuracy': max_hr,
'best_epoch': best_epoch,
'time_to_target': time.time() - main_start_time,
'time_to_best_model': best_model_timestamp - main_start_time
})
def main(args, ml_logger):
# Fix the seed
random.seed(args.seed)
if not args.dont_fix_np_seed:
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Check that GPUs are actually available
use_cuda = not args.no_cuda and torch.cuda.is_available()
# Create model
model = NeuMF(2197225,
855776,
mf_dim=64,
mf_reg=0.,
mlp_layer_sizes=[256, 256, 128, 64],
mlp_layer_regs=[0. for i in [256, 256, 128, 64]])
print(model)
if use_cuda:
# Move model and loss to GPU
model = model.cuda()
model.device = torch.device('cuda:{}'.format(0))
if args.load_ckp:
ckp = torch.load(args.load_ckp)
model.load_state_dict(ckp)
all_embeding = [
n for n, m in model.named_modules() if isinstance(m, nn.Embedding)
]
all_linear = [
n for n, m in model.named_modules() if isinstance(m, nn.Linear)
]
all_relu = [n for n, m in model.named_modules() if isinstance(m, nn.ReLU)]
all_relu6 = [
n for n, m in model.named_modules() if isinstance(m, nn.ReLU6)
]
layers = all_relu + all_relu6 + all_linear + all_embeding
replacement_factory = {
nn.ReLU: ActivationModuleWrapperPost,
nn.ReLU6: ActivationModuleWrapperPost,
nn.Linear: ParameterModuleWrapperPost,
nn.Embedding: ActivationModuleWrapperPost
}
mq = ModelQuantizer(model, args, layers, replacement_factory)
# mq.log_quantizer_state(ml_logger, -1)
test_users, test_items, dup_mask, real_indices, K, samples_per_user, num_user = data_loader(
args.data)
data = NcfData(test_users, test_items, dup_mask, real_indices, K,
samples_per_user, num_user)
cal_data = CalibrationSet('ml-20mx16x32/cal_set').cuda()
cal_data.split(batch_size=10000)
criterion = nn.BCEWithLogitsLoss(reduction='mean')
criterion = criterion.cuda()
print("init_method: {}, qtype {}".format(args.init_method, args.qtype))
# evaluate to initialize dynamic clipping
loss = evaluate_calibration(model, cal_data, criterion)
print("Initial loss: {:.4f}".format(loss))
# get clipping values
init = get_clipping(mq)
# evaluate
hr, ndcg = validate(model, data)
ml_logger.log_metric('HR init', hr, step='auto')
# run optimizer
min_options = {}
if args.maxiter is not None:
min_options['maxiter'] = args.maxiter
if args.maxfev is not None:
min_options['maxfev'] = args.maxfev
_iter = count(0)
def local_search_callback(x):
it = next(_iter)
loss = run_inference_on_calibration(x, model, mq, cal_data, criterion)
print("\n[{}]: Local search callback".format(it))
print("loss: {:.4f}\n".format(loss))
res = opt.minimize(lambda scales: run_inference_on_calibration(
scales, model, mq, cal_data, criterion),
np.array(init),
method=args.min_method,
options=min_options,
callback=local_search_callback)
print(res)
scales = res.x
set_clipping(mq, scales, model.device)
# evaluate
hr, ndcg = validate(model, data)
ml_logger.log_metric('HR Powell', hr, step='auto')
def main():
args = parse_args()
init_distributed(args)
if args.local_rank == 0:
dllogger.init(backends=[dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=args.log_path),
dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE)])
else:
dllogger.init(backends=[])
dllogger.log(data=vars(args), step='PARAMETER')
if args.seed is not None:
torch.manual_seed(args.seed)
print("Saving results to {}".format(args.checkpoint_dir))
if not os.path.exists(args.checkpoint_dir) and args.checkpoint_dir != '':
os.makedirs(args.checkpoint_dir, exist_ok=True)
# sync workers before timing
if args.distributed:
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
main_start_time = time.time()
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)))
test_negs = torch.load(args.data+'/test_negatives.pt', map_location=torch.device('cuda:{}'.format(args.local_rank)))
valid_negative = test_negs.shape[1]
nb_maxs = torch.max(train_ratings, 0)[0]
nb_users = nb_maxs[0].item() + 1
nb_items = nb_maxs[1].item() + 1
all_test_users = test_ratings.shape[0]
test_users, test_items, dup_mask, real_indices = dataloading.create_test_data(test_ratings, test_negs, args)
# make pytorch memory behavior more consistent later
torch.cuda.empty_cache()
# Create model
model = NeuMF(nb_users, nb_items,
mf_dim=args.factors,
mlp_layer_sizes=args.layers,
dropout=args.dropout)
optimizer = FusedAdam(model.parameters(), lr=args.learning_rate,
betas=(args.beta1, args.beta2), eps=args.eps)
criterion = nn.BCEWithLogitsLoss(reduction='none') # use torch.mean() with dim later to avoid copy to host
# Move model and loss to GPU
model = model.cuda()
criterion = criterion.cuda()
local_batch = args.batch_size // args.world_size
#traced_criterion = torch.jit.trace(criterion.forward,
# (torch.rand(local_batch,1),torch.rand(local_batch,1)))
traced_criterion = criterion
pyprof.init()
#import importlib
#pyprof.wrap(importlib.import_module(__name__), "traced_criterion")
#pyprof.wrap(traced_criterion, "__call__")
if args.opt_level == "O2":
model, optimizer = amp.initialize(model, optimizer, opt_level=args.opt_level,
keep_batchnorm_fp32=False, loss_scale='dynamic')
if args.distributed:
model = DDP(model)
print(model)
print("{} parameters".format(utils.count_parameters(model)))
if args.load_checkpoint_path:
state_dict = torch.load(args.load_checkpoint_path)
state_dict = {k.replace('module.', '') : v for k,v in state_dict.items()}
model.load_state_dict(state_dict)
if args.mode == 'test':
start = time.time()
hr, ndcg = val_epoch(model, test_users, test_items, dup_mask, real_indices, args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users, distributed=args.distributed)
val_time = time.time() - start
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
dllogger.log(step=tuple(), data={'best_eval_throughput' : eval_throughput,
'hr@10' : hr})
return
max_hr = 0
best_epoch = 0
train_throughputs, eval_throughputs = [], []
with torch.autograd.profiler.emit_nvtx():
for epoch in range(args.epochs):
begin = time.time()
epoch_users, epoch_items, epoch_label = dataloading.prepare_epoch_train_data(train_ratings, nb_items, args)
num_batches = len(epoch_users)
for i in range(num_batches // args.grads_accumulated):
if i == 10:
profiler.start()
for j in range(args.grads_accumulated):
batch_idx = (args.grads_accumulated * i) + j
user = epoch_users[batch_idx]
item = epoch_items[batch_idx]
label = epoch_label[batch_idx].view(-1,1)
outputs = model(user, item)
nvtx.range_push("layer:Loss")
loss = traced_criterion(outputs, label).float()
nvtx.range_pop()
nvtx.range_push("layer:Mean")
loss = torch.mean(loss.view(-1), 0)
nvtx.range_pop()
if args.opt_level == "O2":
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
nvtx.range_push("layer:Adam")
optimizer.step()
nvtx.range_pop()
if i == 10:
profiler.stop()
for p in model.parameters():
p.grad = None
del epoch_users, epoch_items, epoch_label
train_time = time.time() - begin
begin = time.time()
epoch_samples = len(train_ratings) * (args.negative_samples + 1)
train_throughput = epoch_samples / train_time
train_throughputs.append(train_throughput)
hr, ndcg = val_epoch(model, test_users, test_items, dup_mask, real_indices, args.topk,
samples_per_user=valid_negative + 1,
num_user=all_test_users, epoch=epoch, distributed=args.distributed)
val_time = time.time() - begin
eval_size = all_test_users * (valid_negative + 1)
eval_throughput = eval_size / val_time
eval_throughputs.append(eval_throughput)
dllogger.log(step=(epoch,),
data = {'train_throughput': train_throughput,
'hr@10': hr,
'train_epoch_time': train_time,
'validation_epoch_time': val_time,
'eval_throughput': eval_throughput})
if hr > max_hr and args.local_rank == 0:
max_hr = hr
best_epoch = epoch
save_checkpoint_path = os.path.join(args.checkpoint_dir, 'model.pth')
print("New best hr! Saving the model to: ", save_checkpoint_path)
torch.save(model.state_dict(), save_checkpoint_path)
best_model_timestamp = time.time()
if args.threshold is not None:
if hr >= args.threshold:
print("Hit threshold of {}".format(args.threshold))
break
if args.local_rank == 0:
dllogger.log(data={'best_train_throughput': max(train_throughputs),
'best_eval_throughput': max(eval_throughputs),
'mean_train_throughput': np.mean(train_throughputs),
'mean_eval_throughput': np.mean(eval_throughputs),
'best_accuracy': max_hr,
'best_epoch': best_epoch,
'time_to_target': time.time() - main_start_time,
'time_to_best_model': best_model_timestamp - main_start_time},
step=tuple())