Skip to content

P
PaddleRec

PaddlePaddle / PaddleRec

通知 68
Star 12
Fork 5
P
PaddleRec
未验证 提交 d190118a 编写于 作者: W wuzhihua 提交者： GitHub
浏览文件
操作

Merge pull request #21 from yaoxuefeng6/mod_yaml 

update rank yaml, model, and add infer option in rank models
上级 5a6abdf6 2ff6b226
显示空白变更内容
内联 并排
Showing with 507 addition and 402 deletion
models/rank/dcn/config.yaml
浏览文件 @ d190118a
......@@ -12,24 +12,30 @@
# See the License for the specific language governing permissions and
# limitations under the License.
train:
trainer:
# for cluster training
strategy: "async"
epochs: 10
workspace: "paddlerec.models.rank.dcn"
reader:
batch_size: 2
train_data_path: "{workspace}/data/sample_data/train"
feat_dict_name: "{workspace}/data/vocab"
# global settings
debug: false
workspace: "paddlerec.models.rank.dcn"
dataset:
- name: train_sample
type: QueueDataset
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C18 C19 C20 C21 C22 C23 C24 C25 C26"
dense_slots: "I1:1 I2:1 I3:1 I4:1 I5:1 I6:1 I7:1 I8:1 I9:1 I10:1 I11:1 I12:1 I13:1"
- name: infer_sample
type: QueueDataset
batch_size: 5
data_path: "{workspace}/data/sample_data/infer"
sparse_slots: "label C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C18 C19 C20 C21 C22 C23 C24 C25 C26"
dense_slots: "I1:1 I2:1 I3:1 I4:1 I5:1 I6:1 I7:1 I8:1 I9:1 I10:1 I11:1 I12:1 I13:1"
model:
models: "{workspace}/model.py"
hyper_parameters:
hyper_parameters:
optimizer:
class: Adam
learning_rate: 0.0001
# 用户自定义配置
cross_num: 2
dnn_hidden_units: [128, 128]
l2_reg_cross: 0.00005
......@@ -37,18 +43,35 @@ train:
clip_by_norm: 100.0
cat_feat_num: "{workspace}/data/sample_data/cat_feature_num.txt"
is_sparse: False
is_test: False
num_field: 39
learning_rate: 0.0001
act: "relu"
optimizer: adam
save:
increment:
dirname: "increment"
epoch_interval: 2
save_last: True
inference:
dirname: "inference"
epoch_interval: 4
save_last: True
mode: train_runner
# if infer, change mode to "infer_runner" and change phase to "infer_phase"
runner:
- name: train_runner
trainer_class: single_train
epochs: 1
device: cpu
init_model_path: ""
save_checkpoint_interval: 1
save_inference_interval: 1
save_checkpoint_path: "increment"
save_inference_path: "inference"
print_interval: 1
- name: infer_runner
trainer_class: single_infer
epochs: 1
device: cpu
init_model_path: "increment/0"
print_interval: 1
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: train_sample
thread_num: 1
#- name: infer_phase
# model: "{workspace}/model.py"
# dataset_name: infer_sample
# thread_num: 1
models/rank/dcn/data/sample_data/infer/infer_sample_data 0 → 100644
浏览文件 @ d190118a
label:0 I1:0.69314718056 I2:1.60943791243 I3:1.79175946923 I4:0.0 I5:7.23201033166 I6:1.60943791243 I7:2.77258872224 I8:1.09861228867 I9:5.20400668708 I10:0.69314718056 I11:1.09861228867 I12:0 I13:1.09861228867 C1:95 C2:398 C3:0 C4:0 C5:53 C6:1 C7:73 C8:71 C9:3 C10:1974 C11:832 C12:0 C13:875 C14:8 C15:1764 C16:0 C17:5 C18:390 C19:226 C20:1 C21:0 C22:0 C23:8 C24:1759 C25:1 C26:862
label:0 I1:1.09861228867 I2:1.38629436112 I3:3.80666248977 I4:0.69314718056 I5:4.63472898823 I6:2.19722457734 I7:1.09861228867 I8:1.09861228867 I9:1.60943791243 I10:0.69314718056 I11:0.69314718056 I12:0 I13:1.60943791243 C1:95 C2:200 C3:1184 C4:1929 C5:53 C6:4 C7:1477 C8:2 C9:3 C10:1283 C11:1567 C12:1048 C13:271 C14:6 C15:1551 C16:899 C17:1 C18:162 C19:226 C20:2 C21:575 C22:0 C23:8 C24:1615 C25:1 C26:659
label:0 I1:1.09861228867 I2:1.38629436112 I3:0.69314718056 I4:2.7080502011 I5:6.64378973315 I6:4.49980967033 I7:1.60943791243 I8:1.09861228867 I9:5.50533153593 I10:0.69314718056 I11:1.38629436112 I12:1.38629436112 I13:3.82864139649 C1:123 C2:378 C3:991 C4:197 C5:53 C6:1 C7:689 C8:2 C9:3 C10:245 C11:623 C12:1482 C13:887 C14:21 C15:106 C16:720 C17:3 C18:768 C19:0 C20:0 C21:1010 C22:1 C23:8 C24:720 C25:0 C26:0
label:0 I1:0 I2:6.79905586206 I3:0 I4:0 I5:8.38776764398 I6:0 I7:0.0 I8:0.0 I9:0.0 I10:0 I11:0.0 I12:0 I13:0 C1:95 C2:227 C3:0 C4:219 C5:53 C6:4 C7:3174 C8:2 C9:3 C10:569 C11:1963 C12:0 C13:1150 C14:21 C15:1656 C16:0 C17:6 C18:584 C19:0 C20:0 C21:0 C22:0 C23:8 C24:954 C25:0 C26:0
label:0 I1:1.38629436112 I2:1.09861228867 I3:0 I4:0.0 I5:1.09861228867 I6:0.0 I7:1.38629436112 I8:0.0 I9:0.0 I10:0.69314718056 I11:0.69314718056 I12:0 I13:0.0 C1:121 C2:147 C3:0 C4:1356 C5:53 C6:7 C7:2120 C8:2 C9:3 C10:703 C11:1678 C12:1210 C13:1455 C14:8 C15:538 C16:1276 C17:6 C18:346 C19:0 C20:0 C21:944 C22:0 C23:10 C24:355 C25:0 C26:0
label:0 I1:0 I2:1.09861228867 I3:0 I4:0 I5:9.45915167004 I6:0 I7:0.0 I8:0.0 I9:1.94591014906 I10:0 I11:0.0 I12:0 I13:0 C1:14 C2:75 C3:993 C4:480 C5:50 C6:6 C7:1188 C8:2 C9:3 C10:245 C11:1037 C12:1365 C13:1421 C14:21 C15:786 C16:5 C17:2 C18:555 C19:0 C20:0 C21:1408 C22:6 C23:7 C24:753 C25:0 C26:0
label:0 I1:0 I2:1.60943791243 I3:1.09861228867 I4:0 I5:8.06117135969 I6:0 I7:0.0 I8:0.69314718056 I9:1.09861228867 I10:0 I11:0.0 I12:0 I13:0 C1:139 C2:343 C3:553 C4:828 C5:50 C6:4 C7:0 C8:2 C9:3 C10:245 C11:2081 C12:260 C13:455 C14:21 C15:122 C16:1159 C17:2 C18:612 C19:0 C20:0 C21:1137 C22:0 C23:1 C24:1583 C25:0 C26:0
label:1 I1:0.69314718056 I2:2.07944154168 I3:1.09861228867 I4:0.0 I5:0.0 I6:0.0 I7:0.69314718056 I8:0.0 I9:0.0 I10:0.69314718056 I11:0.69314718056 I12:0 I13:0.0 C1:95 C2:227 C3:0 C4:1567 C5:21 C6:7 C7:2496 C8:71 C9:3 C10:1913 C11:2212 C12:0 C13:673 C14:21 C15:1656 C16:0 C17:5 C18:584 C19:0 C20:0 C21:0 C22:0 C23:10 C24:954 C25:0 C26:0
label:0 I1:0 I2:3.87120101091 I3:1.60943791243 I4:2.19722457734 I5:9.85277303799 I6:5.52146091786 I7:3.36729582999 I8:3.4657359028 I9:4.9558270576 I10:0 I11:0.69314718056 I12:0 I13:2.19722457734 C1:14 C2:14 C3:454 C4:197 C5:53 C6:1 C7:1386 C8:2 C9:3 C10:0 C11:1979 C12:205 C13:214 C14:6 C15:1837 C16:638 C17:5 C18:6 C19:0 C20:0 C21:70 C22:0 C23:10 C24:720 C25:0 C26:0
label:0 I1:0 I2:3.66356164613 I3:0 I4:0.69314718056 I5:10.4263800775 I6:3.09104245336 I7:0.69314718056 I8:1.09861228867 I9:1.38629436112 I10:0 I11:0.69314718056 I12:0 I13:0.69314718056 C1:14 C2:179 C3:120 C4:746 C5:53 C6:0 C7:1312 C8:2 C9:3 C10:1337 C11:1963 C12:905 C13:1150 C14:21 C15:1820 C16:328 C17:9 C18:77 C19:0 C20:0 C21:311 C22:0 C23:10 C24:89 C25:0 C26:0
models/rank/dcn/model.py
浏览文件 @ d190118a
......@@ -24,44 +24,21 @@ class Model(ModelBase):
def __init__(self, config):
ModelBase.__init__(self, config)
def init_network(self):
def _init_hyper_parameters(self):
self.cross_num = envs.get_global_env("hyper_parameters.cross_num",
None, self._namespace)
None)
self.dnn_hidden_units = envs.get_global_env(
"hyper_parameters.dnn_hidden_units", None, self._namespace)
"hyper_parameters.dnn_hidden_units", None)
self.l2_reg_cross = envs.get_global_env(
"hyper_parameters.l2_reg_cross", None, self._namespace)
"hyper_parameters.l2_reg_cross", None)
self.dnn_use_bn = envs.get_global_env("hyper_parameters.dnn_use_bn",
None, self._namespace)
None)
self.clip_by_norm = envs.get_global_env(
"hyper_parameters.clip_by_norm", None, self._namespace)
cat_feat_num = envs.get_global_env("hyper_parameters.cat_feat_num",
None, self._namespace)
self.sparse_inputs = self._sparse_data_var[1:]
self.dense_inputs = self._dense_data_var
self.target_input = self._sparse_data_var[0]
cat_feat_dims_dict = OrderedDict()
for line in open(cat_feat_num):
spls = line.strip().split()
assert len(spls) == 2
cat_feat_dims_dict[spls[0]] = int(spls[1])
self.cat_feat_dims_dict = cat_feat_dims_dict if cat_feat_dims_dict else OrderedDict(
)
"hyper_parameters.clip_by_norm", None)
self.cat_feat_num = envs.get_global_env(
"hyper_parameters.cat_feat_num", None)
self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse",
None, self._namespace)
self.dense_feat_names = [i.name for i in self.dense_inputs]
self.sparse_feat_names = [i.name for i in self.sparse_inputs]
# {feat_name: dims}
self.feat_dims_dict = OrderedDict(
[(feat_name, 1) for feat_name in self.dense_feat_names])
self.feat_dims_dict.update(self.cat_feat_dims_dict)
self.net_input = None
self.loss = None
None)
def _create_embedding_input(self):
# sparse embedding
......@@ -121,9 +98,29 @@ class Model(ModelBase):
def _l2_loss(self, w):
return fluid.layers.reduce_sum(fluid.layers.square(w))
def train_net(self):
self._init_slots()
self.init_network()
def net(self, inputs, is_infer=False):
self.sparse_inputs = self._sparse_data_var[1:]
self.dense_inputs = self._dense_data_var
self.target_input = self._sparse_data_var[0]
cat_feat_dims_dict = OrderedDict()
for line in open(self.cat_feat_num):
spls = line.strip().split()
assert len(spls) == 2
cat_feat_dims_dict[spls[0]] = int(spls[1])
self.cat_feat_dims_dict = cat_feat_dims_dict if cat_feat_dims_dict else OrderedDict(
)
self.dense_feat_names = [i.name for i in self.dense_inputs]
self.sparse_feat_names = [i.name for i in self.sparse_inputs]
# {feat_name: dims}
self.feat_dims_dict = OrderedDict(
[(feat_name, 1) for feat_name in self.dense_feat_names])
self.feat_dims_dict.update(self.cat_feat_dims_dict)
self.net_input = None
self.loss = None
self.net_input = self._create_embedding_input()
......@@ -146,6 +143,9 @@ class Model(ModelBase):
self._metrics["AUC"] = auc_var
self._metrics["BATCH_AUC"] = batch_auc_var
if is_infer:
self._infer_results["AUC"] = auc_var
# logloss
logloss = fluid.layers.log_loss(
self.prob, fluid.layers.cast(
......@@ -157,11 +157,7 @@ class Model(ModelBase):
self.loss = self.avg_logloss + l2_reg_cross_loss
self._cost = self.loss
def optimizer(self):
learning_rate = envs.get_global_env("hyper_parameters.learning_rate",
None, self._namespace)
optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
return optimizer
def infer_net(self):
self.train_net()
#def optimizer(self):
#
# optimizer = fluid.optimizer.Adam(self.learning_rate, lazy_mode=True)
# return optimizer
models/rank/deepfm/config.yaml
浏览文件 @ d190118a
......@@ -12,39 +12,65 @@
# See the License for the specific language governing permissions and
# limitations under the License.
train:
trainer:
# for cluster training
strategy: "async"
epochs: 10
workspace: "paddlerec.models.rank.deepfm"
reader:
batch_size: 2
train_data_path: "{workspace}/data/sample_data/train"
feat_dict_name: "{workspace}/data/sample_data/feat_dict_10.pkl2"
# global settings
debug: false
workspace: "paddlerec.models.rank.deepfm"
dataset:
- name: train_sample
type: QueueDataset
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label feat_idx"
dense_slots: "feat_value:39"
- name: infer_sample
type: QueueDataset
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label feat_idx"
dense_slots: "feat_value:39"
model:
models: "{workspace}/model.py"
hyper_parameters:
hyper_parameters:
optimizer:
class: SGD
learning_rate: 0.0001
sparse_feature_number: 1086460
sparse_feature_dim: 9
num_field: 39
fc_sizes: [400, 400, 400]
learning_rate: 0.0001
reg: 0.001
act: "relu"
optimizer: SGD
save:
increment:
dirname: "increment"
epoch_interval: 2
save_last: True
inference:
dirname: "inference"
epoch_interval: 4
save_last: True
mode: train_runner
# if infer, change mode to "infer_runner" and change phase to "infer_phase"
runner:
- name: train_runner
trainer_class: single_train
epochs: 2
device: cpu
init_model_path: ""
save_checkpoint_interval: 1
save_inference_interval: 1
save_checkpoint_path: "increment"
save_inference_path: "inference"
print_interval: 1
- name: infer_runner
trainer_class: single_infer
epochs: 1
device: cpu
init_model_path: "increment/0"
print_interval: 1
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: train_sample
thread_num: 1
#- name: infer_phase
# model: "{workspace}/model.py"
# dataset_name: infer_sample
# thread_num: 1
models/rank/deepfm/model.py
浏览文件 @ d190118a
......@@ -24,42 +24,46 @@ class Model(ModelBase):
def __init__(self, config):
ModelBase.__init__(self, config)
def deepfm_net(self):
def _init_hyper_parameters(self):
self.sparse_feature_number = envs.get_global_env(
"hyper_parameters.sparse_feature_number", None)
self.sparse_feature_dim = envs.get_global_env(
"hyper_parameters.sparse_feature_dim", None)
self.num_field = envs.get_global_env("hyper_parameters.num_field",
None)
self.reg = envs.get_global_env("hyper_parameters.reg", 1e-4)
self.layer_sizes = envs.get_global_env("hyper_parameters.fc_sizes",
None)
self.act = envs.get_global_env("hyper_parameters.act", None)
def net(self, inputs, is_infer=False):
init_value_ = 0.1
is_distributed = True if envs.get_trainer() == "CtrTrainer" else False
sparse_feature_number = envs.get_global_env(
"hyper_parameters.sparse_feature_number", None, self._namespace)
sparse_feature_dim = envs.get_global_env(
"hyper_parameters.sparse_feature_dim", None, self._namespace)
# ------------------------- network input --------------------------
num_field = envs.get_global_env("hyper_parameters.num_field", None,
self._namespace)
raw_feat_idx = self._sparse_data_var[1]
raw_feat_value = self._dense_data_var[0]
self.label = self._sparse_data_var[0]
feat_idx = raw_feat_idx
feat_value = fluid.layers.reshape(
raw_feat_value, [-1, num_field, 1]) # None * num_field * 1
raw_feat_value, [-1, self.num_field, 1]) # None * num_field * 1
reg = envs.get_global_env("hyper_parameters.reg", 1e-4,
self._namespace)
first_weights_re = fluid.embedding(
input=feat_idx,
is_sparse=True,
is_distributed=is_distributed,
dtype='float32',
size=[sparse_feature_number + 1, 1],
size=[self.sparse_feature_number + 1, 1],
padding_idx=0,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.TruncatedNormalInitializer(
loc=0.0, scale=init_value_),
regularizer=fluid.regularizer.L1DecayRegularizer(reg)))
regularizer=fluid.regularizer.L1DecayRegularizer(self.reg)))
first_weights = fluid.layers.reshape(
first_weights_re, shape=[-1, num_field, 1]) # None * num_field * 1
first_weights_re,
shape=[-1, self.num_field, 1]) # None * num_field * 1
y_first_order = fluid.layers.reduce_sum((first_weights * feat_value),
1)
......@@ -70,16 +74,17 @@ class Model(ModelBase):
is_sparse=True,
is_distributed=is_distributed,
dtype='float32',
size=[sparse_feature_number + 1, sparse_feature_dim],
size=[self.sparse_feature_number + 1, self.sparse_feature_dim],
padding_idx=0,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.TruncatedNormalInitializer(
loc=0.0,
scale=init_value_ / math.sqrt(float(sparse_feature_dim)))))
scale=init_value_ /
math.sqrt(float(self.sparse_feature_dim)))))
feat_embeddings = fluid.layers.reshape(
feat_embeddings_re,
shape=[-1, num_field,
sparse_feature_dim]) # None * num_field * embedding_size
shape=[-1, self.num_field, self.sparse_feature_dim
]) # None * num_field * embedding_size
feat_embeddings = feat_embeddings * feat_value # None * num_field * embedding_size
# sum_square part
......@@ -101,17 +106,13 @@ class Model(ModelBase):
# ------------------------- DNN --------------------------
layer_sizes = envs.get_global_env("hyper_parameters.fc_sizes", None,
self._namespace)
act = envs.get_global_env("hyper_parameters.act", None,
self._namespace)
y_dnn = fluid.layers.reshape(feat_embeddings,
[-1, num_field * sparse_feature_dim])
for s in layer_sizes:
y_dnn = fluid.layers.reshape(
feat_embeddings, [-1, self.num_field * self.sparse_feature_dim])
for s in self.layer_sizes:
y_dnn = fluid.layers.fc(
input=y_dnn,
size=s,
act=act,
act=self.act,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.TruncatedNormalInitializer(
loc=0.0, scale=init_value_ / math.sqrt(float(10)))),
......@@ -133,21 +134,12 @@ class Model(ModelBase):
self.predict = fluid.layers.sigmoid(y_first_order + y_second_order +
y_dnn)
def train_net(self):
self._init_slots()
self.deepfm_net()
# ------------------------- Cost(logloss) --------------------------
cost = fluid.layers.log_loss(
input=self.predict, label=fluid.layers.cast(self.label, "float32"))
avg_cost = fluid.layers.reduce_sum(cost)
self._cost = avg_cost
# ------------------------- Metric(Auc) --------------------------
predict_2d = fluid.layers.concat([1 - self.predict, self.predict], 1)
label_int = fluid.layers.cast(self.label, 'int64')
auc_var, batch_auc_var, _ = fluid.layers.auc(input=predict_2d,
......@@ -155,12 +147,5 @@ class Model(ModelBase):
slide_steps=0)
self._metrics["AUC"] = auc_var
self._metrics["BATCH_AUC"] = batch_auc_var
def optimizer(self):
learning_rate = envs.get_global_env("hyper_parameters.learning_rate",
None, self._namespace)
optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
return optimizer
def infer_net(self):
self.train_net()
if is_infer:
self._infer_results["AUC"] = auc_var
models/rank/din/config.yaml
浏览文件 @ d190118a
......@@ -12,40 +12,60 @@
# See the License for the specific language governing permissions and
# limitations under the License.
train:
trainer:
# for cluster training
strategy: "async"
# global settings
debug: false
workspace: "paddlerec.models.rank.din"
epochs: 10
workspace: "paddlerec.models.rank.din"
dataset:
- name: sample_1
type: DataLoader
batch_size: 5
data_path: "{workspace}/data/train_data"
data_converter: "{workspace}/reader.py"
- name: infer_sample
type: DataLoader
batch_size: 5
data_path: "{workspace}/data/train_data"
data_converter: "{workspace}/reader.py"
reader:
batch_size: 2
class: "{workspace}/reader.py"
train_data_path: "{workspace}/data/train_data"
dataset_class: "DataLoader"
model:
models: "{workspace}/model.py"
hyper_parameters:
hyper_parameters:
optimizer:
class: SGD
learning_rate: 0.0001
use_DataLoader: True
item_emb_size: 64
cat_emb_size: 64
is_sparse: False
config_path: "data/config.txt"
fc_sizes: [400, 400, 400]
learning_rate: 0.0001
reg: 0.001
item_count: 63001
cat_count: 801
act: "sigmoid"
optimizer: SGD
save:
increment:
dirname: "increment"
epoch_interval: 2
save_last: True
inference:
dirname: "inference"
epoch_interval: 4
save_last: True
mode: train_runner
runner:
- name: train_runner
trainer_class: single_train
epochs: 1
device: cpu
init_model_path: ""
save_checkpoint_interval: 1
save_inference_interval: 1
save_checkpoint_path: "increment"
save_inference_path: "inference"
print_interval: 1
- name: infer_runner
trainer_class: single_infer
epochs: 1
device: cpu
init_model_path: "increment/0"
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: sample_1
thread_num: 1
#- name: infer_phase
# model: "{workspace}/model.py"
# dataset_name: infer_sample
# thread_num: 1
models/rank/din/model.py
浏览文件 @ d190118a
......@@ -22,12 +22,58 @@ class Model(ModelBase):
def __init__(self, config):
ModelBase.__init__(self, config)
def config_read(self, config_path):
with open(config_path, "r") as fin:
user_count = int(fin.readline().strip())
item_count = int(fin.readline().strip())
cat_count = int(fin.readline().strip())
return user_count, item_count, cat_count
def _init_hyper_parameters(self):
self.item_emb_size = envs.get_global_env(
"hyper_parameters.item_emb_size", 64)
self.cat_emb_size = envs.get_global_env(
"hyper_parameters.cat_emb_size", 64)
self.act = envs.get_global_env("hyper_parameters.act", "sigmoid")
self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse",
False)
#significant for speeding up the training process
self.use_DataLoader = envs.get_global_env(
"hyper_parameters.use_DataLoader", False)
self.item_count = envs.get_global_env("hyper_parameters.item_count",
63001)
self.cat_count = envs.get_global_env("hyper_parameters.cat_count", 801)
def input_data(self, is_infer=False, **kwargs):
seq_len = -1
self.data_var = []
hist_item_seq = fluid.data(
name="hist_item_seq", shape=[None, seq_len], dtype="int64")
self.data_var.append(hist_item_seq)
hist_cat_seq = fluid.data(
name="hist_cat_seq", shape=[None, seq_len], dtype="int64")
self.data_var.append(hist_cat_seq)
target_item = fluid.data(
name="target_item", shape=[None], dtype="int64")
self.data_var.append(target_item)
target_cat = fluid.data(name="target_cat", shape=[None], dtype="int64")
self.data_var.append(target_cat)
label = fluid.data(name="label", shape=[None, 1], dtype="float32")
self.data_var.append(label)
mask = fluid.data(
name="mask", shape=[None, seq_len, 1], dtype="float32")
self.data_var.append(mask)
target_item_seq = fluid.data(
name="target_item_seq", shape=[None, seq_len], dtype="int64")
self.data_var.append(target_item_seq)
target_cat_seq = fluid.data(
name="target_cat_seq", shape=[None, seq_len], dtype="int64")
self.data_var.append(target_cat_seq)
train_inputs = [hist_item_seq] + [hist_cat_seq] + [target_item] + [
target_cat
] + [label] + [mask] + [target_item_seq] + [target_cat_seq]
return train_inputs
def din_attention(self, hist, target_expand, mask):
"""activation weight"""
......@@ -59,104 +105,58 @@ class Model(ModelBase):
out = fluid.layers.reshape(x=out, shape=[0, hidden_size])
return out
def train_net(self):
seq_len = -1
self.item_emb_size = envs.get_global_env(
"hyper_parameters.item_emb_size", 64, self._namespace)
self.cat_emb_size = envs.get_global_env(
"hyper_parameters.cat_emb_size", 64, self._namespace)
self.act = envs.get_global_env("hyper_parameters.act", "sigmoid",
self._namespace)
#item_emb_size = 64
#cat_emb_size = 64
self.is_sparse = envs.get_global_env("hyper_parameters.is_sparse",
False, self._namespace)
#significant for speeding up the training process
self.config_path = envs.get_global_env(
"hyper_parameters.config_path", "data/config.txt", self._namespace)
self.use_DataLoader = envs.get_global_env(
"hyper_parameters.use_DataLoader", False, self._namespace)
user_count, item_count, cat_count = self.config_read(self.config_path)
def net(self, inputs, is_infer=False):
hist_item_seq = inputs[0]
hist_cat_seq = inputs[1]
target_item = inputs[2]
target_cat = inputs[3]
label = inputs[4]
mask = inputs[5]
target_item_seq = inputs[6]
target_cat_seq = inputs[7]
item_emb_attr = fluid.ParamAttr(name="item_emb")
cat_emb_attr = fluid.ParamAttr(name="cat_emb")
hist_item_seq = fluid.data(
name="hist_item_seq", shape=[None, seq_len], dtype="int64")
self._data_var.append(hist_item_seq)
hist_cat_seq = fluid.data(
name="hist_cat_seq", shape=[None, seq_len], dtype="int64")
self._data_var.append(hist_cat_seq)
target_item = fluid.data(
name="target_item", shape=[None], dtype="int64")
self._data_var.append(target_item)
target_cat = fluid.data(name="target_cat", shape=[None], dtype="int64")
self._data_var.append(target_cat)
label = fluid.data(name="label", shape=[None, 1], dtype="float32")
self._data_var.append(label)
mask = fluid.data(
name="mask", shape=[None, seq_len, 1], dtype="float32")
self._data_var.append(mask)
target_item_seq = fluid.data(
name="target_item_seq", shape=[None, seq_len], dtype="int64")
self._data_var.append(target_item_seq)
target_cat_seq = fluid.data(
name="target_cat_seq", shape=[None, seq_len], dtype="int64")
self._data_var.append(target_cat_seq)
if self.use_DataLoader:
self._data_loader = fluid.io.DataLoader.from_generator(
feed_list=self._data_var,
capacity=10000,
use_double_buffer=False,
iterable=False)
hist_item_emb = fluid.embedding(
input=hist_item_seq,
size=[item_count, self.item_emb_size],
size=[self.item_count, self.item_emb_size],
param_attr=item_emb_attr,
is_sparse=self.is_sparse)
hist_cat_emb = fluid.embedding(
input=hist_cat_seq,
size=[cat_count, self.cat_emb_size],
size=[self.cat_count, self.cat_emb_size],
param_attr=cat_emb_attr,
is_sparse=self.is_sparse)
target_item_emb = fluid.embedding(
input=target_item,
size=[item_count, self.item_emb_size],
size=[self.item_count, self.item_emb_size],
param_attr=item_emb_attr,
is_sparse=self.is_sparse)
target_cat_emb = fluid.embedding(
input=target_cat,
size=[cat_count, self.cat_emb_size],
size=[self.cat_count, self.cat_emb_size],
param_attr=cat_emb_attr,
is_sparse=self.is_sparse)
target_item_seq_emb = fluid.embedding(
input=target_item_seq,
size=[item_count, self.item_emb_size],
size=[self.item_count, self.item_emb_size],
param_attr=item_emb_attr,
is_sparse=self.is_sparse)
target_cat_seq_emb = fluid.embedding(
input=target_cat_seq,
size=[cat_count, self.cat_emb_size],
size=[self.cat_count, self.cat_emb_size],
param_attr=cat_emb_attr,
is_sparse=self.is_sparse)
item_b = fluid.embedding(
input=target_item,
size=[item_count, 1],
size=[self.item_count, 1],
param_attr=fluid.initializer.Constant(value=0.0))
hist_seq_concat = fluid.layers.concat(
......@@ -195,12 +195,5 @@ class Model(ModelBase):
slide_steps=0)
self._metrics["AUC"] = auc_var
self._metrics["BATCH_AUC"] = batch_auc_var
def optimizer(self):
learning_rate = envs.get_global_env("hyper_parameters.learning_rate",
None, self._namespace)
optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
return optimizer
def infer_net(self, parameter_list):
self.deepfm_net()
if is_infer:
self._infer_results["AUC"] = auc_var
models/rank/din/reader.py
浏览文件 @ d190118a
......@@ -29,8 +29,8 @@ from paddlerec.core.utils import envs
class TrainReader(Reader):
def init(self):
self.train_data_path = envs.get_global_env("train_data_path", None,
"train.reader")
self.train_data_path = envs.get_global_env(
"dataset.sample_1.data_path", None)
self.res = []
self.max_len = 0
......@@ -46,7 +46,8 @@ class TrainReader(Reader):
fo = open("tmp.txt", "w")
fo.write(str(self.max_len))
fo.close()
self.batch_size = envs.get_global_env("batch_size", 32, "train.reader")
self.batch_size = envs.get_global_env("dataset.sample_1.batch_size",
32, "train.reader")
self.group_size = self.batch_size * 20
def _process_line(self, line):
......
models/rank/readme.md
浏览文件 @ d190118a
......@@ -56,7 +56,18 @@
<img align="center" src="../../doc/imgs/din.png">
<p>
## 使用教程
## 使用教程(快速开始)
使用样例数据快速开始，参考[训练](###训练) & [预测](###预测)
## 使用教程(复现论文)
为了方便使用者能够快速的跑通每一个模型，我们在每个模型下都提供了样例数据，并且调整了batch_size等超参以便在样例数据上更加友好的显示训练&测试日志。如果需要复现readme中的效果请按照如下表格调整batch_size等超参，并使用提供的脚本下载对应数据集以及数据预处理。
| 模型 | batch_size | thread_num | epoch_num |
| :------------------: | :--------------------: | :--------------------: | :--------------------: |
| DNN | 1000 | 10 | 1 |
| DCN | 512 | 20 | 2 |
| DeepFM | 100 | 10 | 30 |
| DIN | 32 | 10 | 100 |
| Wide&Deep | 40 | 1 | 40 |
| xDeepFM | 100 | 1 | 10 |
### 数据处理
参考每个模型目录数据下载&预处理脚本
......@@ -68,11 +79,21 @@ sh run.sh
### 训练
```
python -m paddlerec.run -m paddlerec.models.rank.dnn # 以DNN为例
cd modles/rank/dnn # 进入选定好的排序模型的目录 以DNN为例
python -m paddlerec.run -m paddlerec.models.rank.dnn # 使用内置配置
# 如果需要使用自定义配置，config.yaml中workspace需要使用改模型目录的绝对路径
# 自定义修改超参后，指定配置文件，使用自定义配置
python -m paddlerec.run -m ./config.yaml
```
### 预测
```
python -m paddlerec.run -m paddlerec.models.rank.dnn # 以DNN为例
# 修改对应模型的config.yaml，mode配置infer_runner
# 示例: mode: runner1 -> mode: infer_runner
# infer_runner中 class配置为 class: single_infer
# 如果训练阶段和预测阶段的模型输入一致，phase不需要改动，复用train的即可
# 修改完config.yaml后 执行:
python -m paddlerec.run -m ./config.yaml # 以DNN为例
```
## 效果对比
......@@ -87,6 +108,7 @@ python -m paddlerec.run -m paddlerec.models.rank.dnn # 以DNN为例
| Census-income Data | Wide&Deep | 0.76195 | 0.90577 | -- | -- |
| Amazon Product | DIN | 0.47005 | 0.86379 | -- | -- |
## 分布式
### 模型训练性能 (样本/s)
| 数据集 | 模型 | 单机 | 同步 (4节点) | 同步 (8节点) | 同步 (16节点) | 同步 (32节点) |
......
models/rank/wide_deep/config.yaml
浏览文件 @ d190118a
......@@ -12,37 +12,59 @@
# See the License for the specific language governing permissions and
# limitations under the License.
train:
trainer:
# for cluster training
strategy: "async"
# global settings
debug: false
workspace: "paddlerec.models.rank.wide_deep"
epochs: 10
workspace: "paddlerec.models.rank.wide_deep"
reader:
batch_size: 2
train_data_path: "{workspace}/data/sample_data/train"
dataset:
- name: sample_1
type: QueueDataset
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label"
dense_slots: "wide_input:8 deep_input:58"
- name: infer_sample
type: QueueDataset
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label"
dense_slots: "wide_input:8 deep_input:58"
model:
models: "{workspace}/model.py"
hyper_parameters:
hyper_parameters:
optimizer:
class: SGD
learning_rate: 0.0001
hidden1_units: 75
hidden2_units: 50
hidden3_units: 25
learning_rate: 0.0001
reg: 0.001
act: "relu"
optimizer: SGD
save:
increment:
dirname: "increment"
epoch_interval: 2
save_last: True
inference:
dirname: "inference"
epoch_interval: 4
save_last: True
mode: train_runner
# if infer, change mode to "infer_runner" and change phase to "infer_phase"
runner:
- name: train_runner
trainer_class: single_train
epochs: 1
device: cpu
init_model_path: ""
save_checkpoint_interval: 1
save_inference_interval: 1
save_checkpoint_path: "increment"
save_inference_path: "inference"
- name: infer_runner
trainer_class: single_infer
epochs: 1
device: cpu
init_model_path: "increment/0"
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: sample_1
thread_num: 1
#- name: infer_phase
# model: "{workspace}/model.py"
# dataset_name: infer_sample
# thread_num: 1
models/rank/wide_deep/model.py
浏览文件 @ d190118a
......@@ -24,6 +24,14 @@ class Model(ModelBase):
def __init__(self, config):
ModelBase.__init__(self, config)
def _init_hyper_parameters(self):
self.hidden1_units = envs.get_global_env(
"hyper_parameters.hidden1_units", 75)
self.hidden2_units = envs.get_global_env(
"hyper_parameters.hidden2_units", 50)
self.hidden3_units = envs.get_global_env(
"hyper_parameters.hidden3_units", 25)
def wide_part(self, data):
out = fluid.layers.fc(
input=data,
......@@ -56,21 +64,14 @@ class Model(ModelBase):
return l3
def train_net(self):
self._init_slots()
def net(self, inputs, is_infer=False):
wide_input = self._dense_data_var[0]
deep_input = self._dense_data_var[1]
label = self._sparse_data_var[0]
hidden1_units = envs.get_global_env("hyper_parameters.hidden1_units",
75, self._namespace)
hidden2_units = envs.get_global_env("hyper_parameters.hidden2_units",
50, self._namespace)
hidden3_units = envs.get_global_env("hyper_parameters.hidden3_units",
25, self._namespace)
wide_output = self.wide_part(wide_input)
deep_output = self.deep_part(deep_input, hidden1_units, hidden2_units,
hidden3_units)
deep_output = self.deep_part(deep_input, self.hidden1_units,
self.hidden2_units, self.hidden3_units)
wide_model = fluid.layers.fc(
input=wide_output,
......@@ -109,18 +110,12 @@ class Model(ModelBase):
self._metrics["AUC"] = auc_var
self._metrics["BATCH_AUC"] = batch_auc
self._metrics["ACC"] = acc
if is_infer:
self._infer_results["AUC"] = auc_var
self._infer_results["ACC"] = acc
cost = fluid.layers.sigmoid_cross_entropy_with_logits(
x=prediction, label=fluid.layers.cast(
label, dtype='float32'))
avg_cost = fluid.layers.mean(cost)
self._cost = avg_cost
def optimizer(self):
learning_rate = envs.get_global_env("hyper_parameters.learning_rate",
None, self._namespace)
optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
return optimizer
def infer_net(self):
self.train_net()
models/rank/xdeepfm/config.yaml
浏览文件 @ d190118a
......@@ -11,41 +11,61 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
debug: false
workspace: "paddlerec.models.rank.xdeepfm"
train:
trainer:
# for cluster training
strategy: "async"
epochs: 10
workspace: "paddlerec.models.rank.xdeepfm"
reader:
batch_size: 2
train_data_path: "{workspace}/data/sample_data/train"
dataset:
- name: sample_1
type: QueueDataset #或者DataLoader
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label feat_idx"
dense_slots: "feat_value:39"
- name: infer_sample
type: QueueDataset #或者DataLoader
batch_size: 5
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "label feat_idx"
dense_slots: "feat_value:39"
model:
models: "{workspace}/model.py"
hyper_parameters:
hyper_parameters:
optimizer:
class: SGD
learning_rate: 0.0001
layer_sizes_dnn: [10, 10, 10]
layer_sizes_cin: [10, 10]
sparse_feature_number: 1086460
sparse_feature_dim: 9
num_field: 39
fc_sizes: [400, 400, 400]
learning_rate: 0.0001
reg: 0.0001
act: "relu"
optimizer: SGD
save:
increment:
dirname: "increment"
epoch_interval: 2
save_last: True
inference:
dirname: "inference"
epoch_interval: 4
save_last: True
mode: train_runner
# if infer, change mode to "infer_runner" and change phase to "infer_phase"
runner:
- name: train_runner
trainer_class: single_train
epochs: 1
device: cpu
init_model_path: ""
save_checkpoint_interval: 1
save_inference_interval: 1
save_checkpoint_path: "increment"
save_inference_path: "inference"
- name: infer_runner
trainer_class: single_infer
epochs: 1
device: cpu
init_model_path: "increment/0"
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: sample_1
thread_num: 1
#- name: infer_phase
# model: "{workspace}/model.py"
# dataset_name: infer_sample
# thread_num: 1
models/rank/xdeepfm/model.py
浏览文件 @ d190118a
......@@ -22,38 +22,45 @@ class Model(ModelBase):
def __init__(self, config):
ModelBase.__init__(self, config)
def xdeepfm_net(self):
def _init_hyper_parameters(self):
self.sparse_feature_number = envs.get_global_env(
"hyper_parameters.sparse_feature_number", None)
self.sparse_feature_dim = envs.get_global_env(
"hyper_parameters.sparse_feature_dim", None)
self.num_field = envs.get_global_env("hyper_parameters.num_field",
None)
self.layer_sizes_cin = envs.get_global_env(
"hyper_parameters.layer_sizes_cin", None)
self.layer_sizes_dnn = envs.get_global_env(
"hyper_parameters.layer_sizes_dnn", None)
self.act = envs.get_global_env("hyper_parameters.act", None)
def net(self, inputs, is_infer=False):
raw_feat_idx = self._sparse_data_var[1]
raw_feat_value = self._dense_data_var[0]
self.label = self._sparse_data_var[0]
init_value_ = 0.1
initer = fluid.initializer.TruncatedNormalInitializer(
loc=0.0, scale=init_value_)
is_distributed = True if envs.get_trainer() == "CtrTrainer" else False
sparse_feature_number = envs.get_global_env(
"hyper_parameters.sparse_feature_number", None, self._namespace)
sparse_feature_dim = envs.get_global_env(
"hyper_parameters.sparse_feature_dim", None, self._namespace)
# ------------------------- network input --------------------------
num_field = envs.get_global_env("hyper_parameters.num_field", None,
self._namespace)
raw_feat_idx = self._sparse_data_var[1]
raw_feat_value = self._dense_data_var[0]
self.label = self._sparse_data_var[0]
feat_idx = raw_feat_idx
feat_value = fluid.layers.reshape(
raw_feat_value, [-1, num_field, 1]) # None * num_field * 1
raw_feat_value, [-1, self.num_field, 1]) # None * num_field * 1
feat_embeddings = fluid.embedding(
input=feat_idx,
is_sparse=True,
dtype='float32',
size=[sparse_feature_number + 1, sparse_feature_dim],
size=[self.sparse_feature_number + 1, self.sparse_feature_dim],
padding_idx=0,
param_attr=fluid.ParamAttr(initializer=initer))
feat_embeddings = fluid.layers.reshape(feat_embeddings, [
-1, num_field, sparse_feature_dim
-1, self.num_field, self.sparse_feature_dim
]) # None * num_field * embedding_size
feat_embeddings = feat_embeddings * feat_value # None * num_field * embedding_size
......@@ -63,11 +70,11 @@ class Model(ModelBase):
input=feat_idx,
is_sparse=True,
dtype='float32',
size=[sparse_feature_number + 1, 1],
size=[self.sparse_feature_number + 1, 1],
padding_idx=0,
param_attr=fluid.ParamAttr(initializer=initer))
weights_linear = fluid.layers.reshape(
weights_linear, [-1, num_field, 1]) # None * num_field * 1
weights_linear, [-1, self.num_field, 1]) # None * num_field * 1
b_linear = fluid.layers.create_parameter(
shape=[1],
dtype='float32',
......@@ -77,31 +84,30 @@ class Model(ModelBase):
# -------------------- CIN --------------------
layer_sizes_cin = envs.get_global_env(
"hyper_parameters.layer_sizes_cin", None, self._namespace)
Xs = [feat_embeddings]
last_s = num_field
for s in layer_sizes_cin:
last_s = self.num_field
for s in self.layer_sizes_cin:
# calculate Z^(k+1) with X^k and X^0
X_0 = fluid.layers.reshape(
fluid.layers.transpose(Xs[0], [0, 2, 1]),
[-1, sparse_feature_dim, num_field,
[-1, self.sparse_feature_dim, self.num_field,
1]) # None, embedding_size, num_field, 1
X_k = fluid.layers.reshape(
fluid.layers.transpose(Xs[-1], [0, 2, 1]),
[-1, sparse_feature_dim, 1,
[-1, self.sparse_feature_dim, 1,
last_s]) # None, embedding_size, 1, last_s
Z_k_1 = fluid.layers.matmul(
X_0, X_k) # None, embedding_size, num_field, last_s
# compresses Z^(k+1) to X^(k+1)
Z_k_1 = fluid.layers.reshape(Z_k_1, [
-1, sparse_feature_dim, last_s * num_field
-1, self.sparse_feature_dim, last_s * self.num_field
]) # None, embedding_size, last_s*num_field
Z_k_1 = fluid.layers.transpose(
Z_k_1, [0, 2, 1]) # None, s*num_field, embedding_size
Z_k_1 = fluid.layers.reshape(
Z_k_1, [-1, last_s * num_field, 1, sparse_feature_dim]
Z_k_1,
[-1, last_s * self.num_field, 1, self.sparse_feature_dim]
) # None, last_s*num_field, 1, embedding_size (None, channal_in, h, w)
X_k_1 = fluid.layers.conv2d(
Z_k_1,
......@@ -112,7 +118,8 @@ class Model(ModelBase):
param_attr=fluid.ParamAttr(
initializer=initer)) # None, s, 1, embedding_size
X_k_1 = fluid.layers.reshape(
X_k_1, [-1, s, sparse_feature_dim]) # None, s, embedding_size
X_k_1,
[-1, s, self.sparse_feature_dim]) # None, s, embedding_size
Xs.append(X_k_1)
last_s = s
......@@ -130,17 +137,13 @@ class Model(ModelBase):
# -------------------- DNN --------------------
layer_sizes_dnn = envs.get_global_env(
"hyper_parameters.layer_sizes_dnn", None, self._namespace)
act = envs.get_global_env("hyper_parameters.act", None,
self._namespace)
y_dnn = fluid.layers.reshape(feat_embeddings,
[-1, num_field * sparse_feature_dim])
for s in layer_sizes_dnn:
y_dnn = fluid.layers.reshape(
feat_embeddings, [-1, self.num_field * self.sparse_feature_dim])
for s in self.layer_sizes_dnn:
y_dnn = fluid.layers.fc(
input=y_dnn,
size=s,
act=act,
act=self.act,
param_attr=fluid.ParamAttr(initializer=initer),
bias_attr=None)
y_dnn = fluid.layers.fc(input=y_dnn,
......@@ -152,11 +155,6 @@ class Model(ModelBase):
# ------------------- xDeepFM ------------------
self.predict = fluid.layers.sigmoid(y_linear + y_cin + y_dnn)
def train_net(self):
self._init_slots()
self.xdeepfm_net()
cost = fluid.layers.log_loss(
input=self.predict,
label=fluid.layers.cast(self.label, "float32"),
......@@ -172,12 +170,5 @@ class Model(ModelBase):
slide_steps=0)
self._metrics["AUC"] = auc_var
self._metrics["BATCH_AUC"] = batch_auc_var
def optimizer(self):
learning_rate = envs.get_global_env("hyper_parameters.learning_rate",
None, self._namespace)
optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
return optimizer
def infer_net(self):
self.train_net()
if is_infer:
self._infer_results["AUC"] = auc_var
setup.py
浏览文件 @ d190118a
......@@ -62,7 +62,8 @@ def build(dirname):
models_copy = [
'data/*.txt', 'data/*/*.txt', '*.yaml', '*.sh', 'tree/*.npy',
'tree/*.txt', 'data/sample_data/*', 'data/sample_data/train/*'
'tree/*.txt', 'data/sample_data/*', 'data/sample_data/train/*',
'data/sample_data/infer/*'
]
engine_copy = ['*/*.sh']
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或