• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2024 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# ============================================================================
15
16import os
17import numpy as np
18import mindspore as ms
19from  mindspore import nn, ops, Tensor
20from mindspore.nn.layer.embedding_service import EmbeddingService
21from mindspore.nn.layer.embedding_service_layer import EsEmbeddingLookup
22from mindspore.communication import init, release, get_rank
23from mindspore import context
24
25
26class Net(nn.Cell):
27    """
28    EsNet
29    """
30    def __init__(self, embedding_dim, max_feature_count, table_id_dict=None, es_initializer=None,
31                 es_counter_filter=None):
32        super(Net, self).__init__()
33        self.table_id = table_id_dict["test"]
34        self.embedding = EsEmbeddingLookup(self.table_id, es_initializer[self.table_id], embedding_dim=[embedding_dim],
35                                           max_key_num=max_feature_count, optimizer_mode="adam",
36                                           optimizer_params=[0.0, 0.0],
37                                           es_filter=es_counter_filter[self.table_id])
38        self.w = ms.Parameter(Tensor([1.5], ms.float32), name="w", requires_grad=True)
39
40    def construct(self, keys, actual_keys_input=None, unique_indices=None, key_count=None):
41        if (actual_keys_input is not None) and (unique_indices is not None):
42            es_out = self.embedding(keys, actual_keys_input, unique_indices, key_count)
43        else:
44            es_out = self.embedding(keys)
45        output = es_out * self.w
46        return output
47
48
49class NetworkWithLoss(nn.Cell):
50    """
51    NetworkWithLoss
52    """
53    def __init__(self, network, loss):
54        super(NetworkWithLoss, self).__init__()
55        self.network = network
56        self.loss_fn = loss
57
58    def construct(self, x, label):
59        logits = self.network(x)
60        loss = self.loss_fn(logits, label)
61        return loss
62
63
64def train():
65    """
66    train net.
67    """
68    init()
69    vocab_size = 1000
70    embedding_dim = 12
71    feature_length = 16
72    context.set_context(mode=ms.GRAPH_MODE, device_target="Ascend")
73
74    es = EmbeddingService()
75    filter_option = es.counter_filter(filter_freq=2, default_value=10.0)
76    ev_option = es.embedding_variable_option(filter_option=filter_option)
77
78    table_id_dict, es_initializer, es_counter_filter = es.embedding_init("test", init_vocabulary_size=vocab_size,
79                                                                         embedding_dim=embedding_dim,
80                                                                         max_feature_count=feature_length,
81                                                                         optimizer="adam", ev_option=ev_option,
82                                                                         mode="train")
83    if "test" not in table_id_dict:
84        raise ValueError("embedding_init error, not contain test table!")
85    if len(es_initializer) != 1 or len(es_counter_filter) != 1:
86        raise ValueError("embedding_init error, table len should be 1!")
87
88    print("Succ do embedding_init: ", table_id_dict, es_initializer, es_counter_filter, flush=True)
89
90
91    net = Net(embedding_dim, feature_length, table_id_dict, es_initializer, es_counter_filter)
92    loss_fn = ops.SigmoidCrossEntropyWithLogits()
93    optimizer = nn.Adam(params=net.trainable_params(), learning_rate=1e-3)
94    net_with_loss = NetworkWithLoss(net, loss_fn)
95    train_network = nn.TrainOneStepCell(net_with_loss, optimizer=optimizer)
96    train_network.set_train()
97
98    data = Tensor(np.array(np.ones((2, 8)), dtype=np.float32))
99    label = Tensor(np.array(np.ones((2, 8, 12)), dtype=np.float32))
100
101    loss = train_network(data, label)
102    print("Succ do train, loss is: ", loss, flush=True)
103
104    rank = get_rank()
105    print("Succ get rank, rank is: ", rank, flush=True)
106    if rank == 0:
107        save_embedding_path = os.path.join(os.getcwd(), "embedding")
108        save_ckpt_path = os.path.join(os.getcwd(), "ckpt")
109        print("After get path is: ", save_embedding_path, save_ckpt_path, flush=True)
110        es.embedding_table_export(save_embedding_path)
111        print("Succ do export embedding.", flush=True)
112        es.embedding_ckpt_export(save_ckpt_path)
113        print("Succ do export ckpt.", flush=True)
114
115    es.embedding_ckpt_import(save_ckpt_path)
116    print("Succ do import embedding.", flush=True)
117
118    release()
119
120
121if __name__ == "__main__":
122    train()
123