1# Copyright 2020 Huawei Technologies Co., Ltd 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================ 15import pytest 16import numpy as np 17 18import mindspore as ms 19import mindspore.nn as nn 20from mindspore.common.api import _cell_graph_executor 21from mindspore.ops import operations as P 22from mindspore.ops import composite as C 23from mindspore import Tensor, context 24from mindspore.nn import TrainOneStepCell, Adam 25from tests.ut.python.ops.test_math_ops import VirtualLoss 26 27grad_all = C.GradOperation(get_all=True) 28 29 30@pytest.fixture(name="test_context") 31def _test_context(): 32 context.set_context(enable_sparse=True) 33 yield 34 context.set_context(enable_sparse=False) 35 context.reset_auto_parallel_context() 36 37 38class GradWrap(nn.Cell): 39 def __init__(self, network): 40 super(GradWrap, self).__init__() 41 self.network = network 42 43 def construct(self, x, y, z): 44 return grad_all(self.network)(x, y, z) 45 46 47class NetWithLoss(nn.Cell): 48 def __init__(self, network): 49 super(NetWithLoss, self).__init__() 50 self.loss = VirtualLoss() 51 self.network = network 52 53 def construct(self, x, y, z): 54 predict = self.network(x, y, z) 55 return self.loss(predict) 56 57 58class Net(nn.Cell): 59 def __init__(self, shape, field_size=10, slice_mode=nn.EmbeddingLookup.BATCH_SLICE, target="Device", 60 operator='SUM'): 61 super().__init__() 62 self.embedding = nn.MultiFieldEmbeddingLookup(vocab_size=32, embedding_size=64, target=target, 63 field_size=field_size, slice_mode=slice_mode, operator=operator) 64 self.reshape = P.Reshape() 65 self.batch_size = shape[0] 66 67 def construct(self, x, y, z): 68 out = self.embedding(x, y, z) 69 out = self.reshape(out, (self.batch_size, -1)) 70 return out 71 72 73def compile_net(net, shape): 74 x = Tensor(np.ones(shape), dtype=ms.int32) 75 y = Tensor(np.ones(shape), dtype=ms.float32) 76 z = Tensor(np.ones(shape), dtype=ms.int32) 77 optimizer = Adam(net.trainable_params(), learning_rate=0.1) 78 train_net = TrainOneStepCell(net, optimizer) 79 train_net.set_auto_parallel() 80 train_net.set_train() 81 _cell_graph_executor.compile(train_net, x, y, z) 82 context.reset_auto_parallel_context() 83 84 85def test_embeddinglookup_batch_parallel_sum(test_context): 86 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 87 shape = [64, 64] 88 net = NetWithLoss(Net(shape, field_size=10, target='DEVICE')) 89 compile_net(net, shape) 90 91 92def test_embeddinglookup_row_parallel_sum(test_context): 93 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 94 shape = [64, 64] 95 net = NetWithLoss(Net(shape, field_size=9, slice_mode=nn.EmbeddingLookup.TABLE_ROW_SLICE, target='DEVICE')) 96 compile_net(net, shape) 97 98 99def test_embeddinglookup_column_parallel_sum(test_context): 100 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 101 shape = [64, 64] 102 net = NetWithLoss(Net(shape, field_size=10, slice_mode=nn.EmbeddingLookup.TABLE_COLUMN_SLICE, target='DEVICE')) 103 compile_net(net, shape) 104 105 106def test_embeddinglookup_batch_parallel_mean(test_context): 107 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 108 shape = [64, 64] 109 net = NetWithLoss(Net(shape, field_size=1, target='DEVICE', operator='MEAN')) 110 compile_net(net, shape) 111 112 113def test_embeddinglookup_column_parallel_mean(test_context): 114 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 115 shape = [64, 64] 116 net = NetWithLoss(Net(shape, target='DEVICE', slice_mode=nn.EmbeddingLookup.TABLE_COLUMN_SLICE, operator='MEAN')) 117 compile_net(net, shape) 118 119 120def test_embeddinglookup_row_parallel_mean(test_context): 121 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 122 shape = [64, 64] 123 net = NetWithLoss(Net(shape, target='DEVICE', slice_mode=nn.EmbeddingLookup.TABLE_ROW_SLICE, operator='MEAN')) 124 compile_net(net, shape) 125 126 127def test_embeddinglookup_batch_parallel_max(test_context): 128 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 129 shape = [64, 64] 130 net = NetWithLoss(Net(shape, target='DEVICE', operator='MAX')) 131 compile_net(net, shape) 132 133 134def test_embeddinglookup_column_parallel_max(test_context): 135 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 136 shape = [64, 64] 137 net = NetWithLoss(Net(shape, target='DEVICE', slice_mode=nn.EmbeddingLookup.TABLE_COLUMN_SLICE, operator='MAX')) 138 compile_net(net, shape) 139 140 141def test_embeddinglookup_row_parallel_max(test_context): 142 context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel") 143 shape = [64, 64] 144 net = NetWithLoss(Net(shape, target='DEVICE', slice_mode=nn.EmbeddingLookup.TABLE_ROW_SLICE, operator='MAX')) 145 compile_net(net, shape) 146