mindspore/tests/st/dynamic_shape/test_dcn_dynamic_gpu.py

236 lines
9.4 KiB
Python

# Copyright 2022 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
# ============================================================================
from collections import namedtuple
import numpy as np
import pytest
from mindspore import ops, nn, Tensor, Parameter, ParameterTuple, context, set_seed
from mindspore.common.initializer import initializer, XavierUniform
import mindspore.dataset as ds
from mindspore.train import Callback, Model
from mindspore.common import dtype as mstype
import mindspore as ms
class CrossNet(nn.Cell):
def __init__(self, hidden_size, num_layer, l2_reg=0):
super(CrossNet, self).__init__()
self.l2_reg = l2_reg
self.num_layers = num_layer
kernels = []
bias_list = []
for i in range(self.num_layers):
kernel = Parameter(initializer(XavierUniform(0.02), (hidden_size, 1), mstype.float32),
requires_grad=True, name="kernerl" + str(i))
kernels.append(kernel)
bias = Parameter(Tensor(np.zeros((hidden_size, 1)), mstype.float32),
requires_grad=True, name="bias" + str(i))
bias_list.append(bias)
self.kernels = ParameterTuple(kernels)
self.bias = ParameterTuple(bias_list)
self.expand_dim = ops.ExpandDims()
self.squeeze = ops.Squeeze(2)
self.matmul = ops.MatMul()
def construct(self, x):
x_0 = self.expand_dim(x, 2)
x_l = x_0
for i in range(self.num_layers):
xl_w = ops.tensor_dot(x_l, self.kernels[i], axes=(1, 0))
dot = ops.matmul(x_0, xl_w)
x_l = dot + self.bias[i] + x_l
x_l = self.squeeze(x_l)
return x_l
class DNN(nn.Cell):
def __init__(self, input_size, hidden_units, activation='relu', l2_reg=0, dropout_rate=0):
super(DNN, self).__init__()
self.input_size = input_size
self.num_layers = len(hidden_units)
self.hidden_units = [self.input_size] + list(hidden_units)
self.activation = activation
self.l2_reg = l2_reg
self.dropout_rate = dropout_rate
dense_layers = []
drop_layers = []
for i in range(self.num_layers):
dense_layer = nn.Dense(in_channels=self.hidden_units[i], out_channels=self.hidden_units[i + 1],
activation=self.activation, weight_init="heUniform")
dense_layers.append(dense_layer)
drop_layer = nn.Dropout(p=self.dropout_rate)
drop_layers.append(drop_layer)
self.dense_layers = nn.CellList(dense_layers)
self.drop_layers = nn.CellList(drop_layers)
def construct(self, x):
for i in range(self.num_layers):
x = self.dense_layers[i](x)
x = self.drop_layers[i](x)
return x
class DCN(nn.Cell):
def __init__(self, numeric_size, sparse_list, hidden_units, cross_layer, output_num=1):
super(DCN, self).__init__()
self.embed_list = nn.CellList()
for sparse_feature in sparse_list:
embed = nn.Embedding(
sparse_feature.voc_size, sparse_feature.embed_size, embedding_table='xavierUniform')
self.embed_list.append(embed)
self.input_size = sum(
sparse_feature.embed_size for sparse_feature in sparse_list) + numeric_size
self.hidden_units = hidden_units
self.cross_layer = cross_layer
self.cross_net = CrossNet(self.input_size, self.cross_layer)
self.dense_net = DNN(self.input_size, self.hidden_units)
self.output_num = output_num
self.in_channels = self.input_size + self.hidden_units[-1]
self.out_dense = nn.Dense(in_channels=self.in_channels, out_channels=self.output_num,
has_bias=False, weight_init="xavierUniform")
self.split = ops.Split(1, len(sparse_list))
self.squeeze = ops.Squeeze(1)
self.transpose = ops.Transpose()
self.reduce_sum = ops.ReduceSum(keep_dims=False)
self.cast = ops.Cast()
self.shape = ops.Shape()
self.concat = ops.Concat(1)
def construct(self, x, cellist):
inputs = []
cellist = self.transpose(cellist, (1, 0))
for i, ele in enumerate(self.split(cellist)):
embed = self.embed_list[i]
inputs.append(embed(self.squeeze(ele)))
inputs.append(x)
concat_x = self.concat(inputs)
cross_x = self.cross_net(concat_x)
dense_x = self.dense_net(concat_x)
concat_x = self.concat((cross_x, dense_x))
x = self.out_dense(concat_x)
return x
class PairWiseLoss(nn.Cell):
def __init__(self):
super(PairWiseLoss, self).__init__()
self.sub = ops.Sub()
self.mul = ops.Mul()
self.relu = ops.ReLU()
self.expandim = ops.ExpandDims()
self.cast = ops.Cast()
self.greater = ops.Greater()
self.ones = Tensor(np.ones(1), mstype.float32)
self.reduce_sum = ops.ReduceSum()
def construct(self, y_pred, y_true):
pairwise_label_diff = self.sub(self.expandim(y_true, 1), y_true)
pairwise_logits = self.sub(self.expandim(y_pred, 1), y_pred)
pairwise_labels = self.cast(self.greater(
pairwise_label_diff, 0), mstype.float32)
losses = self.mul(pairwise_labels, self.relu(
self.ones - pairwise_logits))
loss = self.reduce_sum(losses)
return loss
class MyWithLossCell(nn.Cell):
def __init__(self, backbone, loss_fn):
super(MyWithLossCell, self).__init__(auto_prefix=False)
self._backbone = backbone
self._loss_fn = loss_fn
self._squeeze = ops.Squeeze(1)
@property
def backbone_network(self):
return self._backbone
def construct(self, x, y, label):
out = self._backbone(x, y)
out = self._squeeze(out)
return self._loss_fn(out, label)
class LossCallback(Callback):
def __init__(self):
super(LossCallback, self).__init__()
self.loss_list = []
def epoch_end(self, run_context):
cb_params = run_context.original_args()
result = cb_params.net_outputs
self.loss_list.append(result.asnumpy().mean())
def gen_data(numeric_columns, sparse_columns, batch_size_list):
np.random.seed(0)
data_list = []
for batch in batch_size_list:
numeric_values = np.random.randn(
batch, numeric_columns[0].size).astype(np.float32)
sparse_values = []
for sparse_column in sparse_columns:
voc_size = sparse_column.voc_size
sparse_value = np.random.randint(
0, voc_size, (1, batch), dtype=np.int32)
sparse_values.append(sparse_value)
sparse_values = np.concatenate(sparse_values)
label_values = np.random.randint(0, 10, batch).astype(np.float32)
data_list.append((numeric_values, sparse_values, label_values))
return data_list
def get_train_loss(numeric_columns, sparse_columns, data_list, mode):
context.set_context(mode=mode, device_target="GPU")
dataset = ds.GeneratorDataset(
data_list, ["dense", "category", "label"], shuffle=False)
numeric_size = numeric_columns[0].size
net = DCN(numeric_size, sparse_columns, hidden_units=(32, 32), cross_layer=2, output_num=1)
loss_fn = PairWiseLoss()
loss_net = MyWithLossCell(net, loss_fn)
train_net = nn.TrainOneStepCell(loss_net, nn.Adam(net.trainable_params(), learning_rate=1e-3, weight_decay=1e-5))
train_net.set_train()
train_net.set_inputs(Tensor(shape=[None, numeric_size], dtype=ms.float32),
Tensor(shape=[len(sparse_columns), None], dtype=ms.int32),
Tensor(shape=[None], dtype=ms.float32))
loss_callback = LossCallback()
model = Model(train_net)
sink_step = dataset.get_dataset_size()
model.train(sink_step, dataset, callbacks=loss_callback, sink_size=1, dataset_sink_mode=True)
loss_list = loss_callback.loss_list
return loss_list
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
def test_train():
"""
Feature: Test the dcn_dynamic network with small shape.
Description: The batch of inputs is dynamic.
Expectation: Assert that results of GRAPH_MODE(static graph) are consistent with expected result.
"""
batch_size_list = [6, 70, 123]
DenseFeature = namedtuple("DenseFeature", ['name', 'size'])
numeric_columns = [DenseFeature("dense", 32)]
SparseFeature = namedtuple("SparseFeature", ['name', 'voc_size', 'embed_size'])
sparse_columns = [SparseFeature('a', 7, 6), SparseFeature('b', 136, 18), SparseFeature('c', 3, 6)]
data_list = gen_data(numeric_columns, sparse_columns, batch_size_list)
# GRAPH_MODE is temporarily not supported due to some new features that are not completely complete
set_seed(0)
graph_loss = get_train_loss(numeric_columns, sparse_columns, data_list, context.GRAPH_MODE)
expect_loss = [6.687461, 2928.5852, 8715.267]
assert np.allclose(graph_loss, expect_loss, 1e-3, 1e-3)