Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Deepfm dcn dense #361

Draft
wants to merge 8 commits into
base: main
Choose a base branch
from
Draft

Deepfm dcn dense #361

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
77fab53
deepfm dense multiply
ShawnXuan Jul 20, 2022
17ee8dc
update
ShawnXuan Jul 20, 2022
a2baddf
update
ShawnXuan Jul 20, 2022
fe7ea2f
update
ShawnXuan Jul 20, 2022
4826144
update
ShawnXuan Jul 20, 2022
2c5c21f
deepfm ok
ShawnXuan Jul 20, 2022
3e7fd1f
update dcn
ShawnXuan Jul 20, 2022
caa7e55
update
ShawnXuan Jul 20, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
92 changes: 46 additions & 46 deletions RecommenderSystems/dcn/dcn_train_eval.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,9 +46,9 @@ def str_list(x):
)
parser.add_argument("--embedding_vec_size", type=int, default=16)
parser.add_argument("--batch_norm", type=bool, default=False)
parser.add_argument("--dnn_hidden_units", type=int_list, default="1000,1000,1000,1000,1000")
parser.add_argument("--crossing_layers", type=int, default=3)
parser.add_argument("--net_dropout", type=float, default=0.2)
parser.add_argument("--dnn_hidden_units", type=int_list, default="1024,1024")
parser.add_argument("--crossing_layers", type=int, default=6)
parser.add_argument("--net_dropout", type=float, default=0.5)
parser.add_argument("--embedding_regularizer", type=float, default=None)
parser.add_argument("--net_regularizer", type=float, default=None)

Expand Down Expand Up @@ -112,9 +112,9 @@ def __init__(
self,
parquet_file_url_list,
batch_size,
num_epochs=1,
num_epochs,
shuffle_row_groups=True,
shard_seed=2019,
shard_seed=1234,
shard_count=1,
cur_shard=0,
):
Expand All @@ -126,16 +126,17 @@ def __init__(
self.shard_count = shard_count
self.cur_shard = cur_shard

fields = ["Label"]
fields = ["label"]
fields += [f"I{i+1}" for i in range(num_dense_fields)]
self.I_end = len(fields)
fields += [f"C{i+1}" for i in range(num_sparse_fields)]
self.C_end = len(fields)
self.fields = fields
self.num_fields = len(fields)

def __enter__(self):
self.reader = make_batch_reader(
self.parquet_file_url_list,
workers_count=2,
workers_count=1,
shuffle_row_groups=self.shuffle_row_groups,
num_epochs=self.num_epochs,
shard_seed=self.shard_seed,
Expand All @@ -152,9 +153,7 @@ def __exit__(self, exc_type, exc_value, exc_traceback):
def get_batches(self, reader, batch_size=None):
if batch_size is None:
batch_size = self.batch_size

tail = None

for rg in reader:
rgdict = rg._asdict()
rglist = [rgdict[field] for field in self.fields]
Expand All @@ -164,24 +163,26 @@ def get_batches(self, reader, batch_size=None):
tail = list(
[
np.concatenate((tail[i], rglist[i][0 : (batch_size - len(tail[i]))]))
for i in range(self.num_fields)
for i in range(self.C_end)
]
)
if len(tail[0]) == batch_size:
label = tail[0]
features = tail[1 : self.num_fields]
dense = tail[1 : self.I_end]
sparse = tail[self.I_end : self.C_end]
tail = None
yield label, np.stack(features, axis=-1)
yield label, np.stack(dense, axis=-1), np.stack(sparse, axis=-1)
else:
pos = 0
continue
while (pos + batch_size) <= len(rglist[0]):
label = rglist[0][pos : pos + batch_size]
features = [rglist[j][pos : pos + batch_size] for j in range(1, self.num_fields)]
dense = [rglist[j][pos : pos + batch_size] for j in range(1, self.I_end)]
sparse = [rglist[j][pos : pos + batch_size] for j in range(self.I_end, self.C_end)]
pos += batch_size
yield label, np.stack(features, axis=-1)
yield label, np.stack(dense, axis=-1), np.stack(sparse, axis=-1)
if pos != len(rglist[0]):
tail = [rglist[i][pos:] for i in range(self.num_fields)]
tail = [rglist[i][pos:] for i in range(self.C_end)]


def make_criteo_dataloader(data_path, batch_size, shuffle=True, shard_seed=2022):
Expand Down Expand Up @@ -340,7 +341,7 @@ def __init__(
size_factor=size_factor,
)

input_dim = embedding_vec_size * (num_dense_fields + num_sparse_fields)
input_dim = embedding_vec_size * num_sparse_fields + num_dense_fields

self.dnn = (
DNN(
Expand All @@ -363,13 +364,12 @@ def __init__(

self.reset_parameters()

def forward(self, X):

feature_emb = self.embedding_layer(X)
flat_feature_emb = feature_emb.flatten(start_dim=1)
cross_out = self.crossnet(flat_feature_emb)
def forward(self, dense_fields, sparse_fields) -> flow.Tensor:
feature_emb = self.embedding_layer(sparse_fields)
feature = flow.cat([feature_emb.flatten(start_dim=1), dense_fields], dim=1)
cross_out = self.crossnet(feature)
if self.dnn is not None:
dnn_out = self.dnn(flat_feature_emb)
dnn_out = self.dnn(feature)
final_out = flow.cat([cross_out, dnn_out], dim=-1)
else:
final_out = cross_out
Expand Down Expand Up @@ -409,9 +409,9 @@ def __init__(self, dcn_module, amp=False):
if amp:
self.config.enable_amp(True)

def build(self, features):
predicts = self.module(features.to("cuda"))
return predicts
def build(self, dense_fields, sparse_fields):
predicts = self.module(dense_fields.to("cuda"), sparse_fields.to("cuda"))
return predicts.sigmoid() # TODO check sigmoid


class DCNTrainGraph(flow.nn.Graph):
Expand All @@ -429,14 +429,12 @@ def __init__(
self.config.enable_amp(True)
self.set_grad_scaler(grad_scaler)

def build(self, labels, features):

logits = self.module(features.to("cuda")).squeeze()
loss = self.loss(logits, labels.squeeze().to("cuda"))
def build(self, labels, dense_fields, sparse_fields):
logits = self.module(dense_fields.to("cuda"), sparse_fields.to("cuda"))
loss = self.loss(logits, labels.to("cuda"))
reduce_loss = flow.mean(loss)
reduce_loss.backward()

return reduce_loss.to("cpu")
return reduce_loss.to("cpu")# TODO: checkout to cpu


def make_lr_scheduler(args, optimizer):
Expand Down Expand Up @@ -535,8 +533,8 @@ def early_stop(
dcn_module.train()
last_step, last_time = 0, time.time()
for step in range(1, args.train_batches + 1):
labels, features = batch_to_global(*next(loader))
loss = train_graph(labels, features)
labels, dense_fields, sparse_fields = batch_to_global(*next(loader))
loss = train_graph(labels, dense_fields, sparse_fields)

if step % args.loss_print_interval == 0:
loss = loss.numpy()
Expand Down Expand Up @@ -586,28 +584,30 @@ def early_stop(
last_time = time.time()
if args.save_best_model:
load_model(f"{args.model_save_dir}/best_checkpoint")
if rank == 0:
print("================ Test Evaluation ================")
eval(args, eval_graph, tag="test", cur_step=step, epoch=epoch)
if args.num_test_samples > 0:
if rank == 0:
print("================ Test Evaluation ================")
eval(args, eval_graph, tag="test", cur_step=step, epoch=epoch)


def _np_to_global(np_array):
t = flow.from_numpy(np_array)
return t.to_global(placement=flow.env.all_device_placement("cpu"), sbp=flow.sbp.split(0))


def batch_to_global(np_label, np_features, is_train=True):
def batch_to_global(np_label, np_dense, np_sparse, is_train=True):
labels = _np_to_global(np_label.reshape(-1, 1)) if is_train else np_label.reshape(-1, 1)
features = _np_to_global(np_features)
return labels, features
dense_fields = _np_to_global(np_dense)
sparse_fields = _np_to_global(np_sparse)
return labels, dense_fields, sparse_fields


def prefetch_eval_batches(data_dir, batch_size, num_batches):
cached_eval_batches = []
with make_criteo_dataloader(data_dir, batch_size, shuffle=False) as loader:
for _ in range(num_batches):
label, features = batch_to_global(*next(loader), is_train=False)
cached_eval_batches.append((label, features))
label, dense_fields, sparse_fields = batch_to_global(*next(loader), is_train=False)
cached_eval_batches.append((label, dense_fields, sparse_fields))
return cached_eval_batches


Expand All @@ -627,14 +627,14 @@ def eval(args, eval_graph, tag="val", cur_step=0, epoch=0, cached_eval_batches=N
with make_criteo_dataloader(f"{args.data_dir}/{tag}", batch_size, shuffle=False) as loader:
eval_start_time = time.time()
for i in range(batches_per_epoch):
label, features = batch_to_global(*next(loader), is_train=False)
pred = eval_graph(features)
label, dense_fields, sparse_fields = batch_to_global(*next(loader), is_train=False)
pred = eval_graph(dense_fields, sparse_fields)
labels.append(label)
preds.append(pred.to_local())
else:
for i in range(batches_per_epoch):
label, features = cached_eval_batches[i]
pred = eval_graph(features)
label, dense_fields, sparse_fields = cached_eval_batches[i]
pred = eval_graph(dense_fields, sparse_fields)
labels.append(label)
preds.append(pred.to_local())

Expand Down
Loading