deepfm.py 15.4 KB
Newer Older
Y
yangyongjie 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
# Copyright 2020 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.
# ============================================================================
""" test_training """
import os

import numpy as np
from sklearn.metrics import roc_auc_score
import mindspore.common.dtype as mstype
from mindspore.ops import functional as F
from mindspore.ops import composite as C
from mindspore.ops import operations as P
from mindspore.nn import Dropout
from mindspore.nn.optim import Adam
from mindspore.nn.metrics import Metric
from mindspore import nn, ParameterTuple, Parameter
from mindspore.common.initializer import Uniform, initializer, Normal
from mindspore.train.callback import ModelCheckpoint, CheckpointConfig

from .callback import EvalCallBack, LossCallBack


np_type = np.float32
ms_type = mstype.float32


class AUCMetric(Metric):
    """AUC metric for DeepFM model."""
    def __init__(self):
        super(AUCMetric, self).__init__()
        self.pred_probs = []
        self.true_labels = []

    def clear(self):
        """Clear the internal evaluation result."""
        self.pred_probs = []
        self.true_labels = []

    def update(self, *inputs):
        batch_predict = inputs[1].asnumpy()
        batch_label = inputs[2].asnumpy()
        self.pred_probs.extend(batch_predict.flatten().tolist())
        self.true_labels.extend(batch_label.flatten().tolist())

    def eval(self):
        if len(self.true_labels) != len(self.pred_probs):
            raise RuntimeError('true_labels.size() is not equal to pred_probs.size()')
        auc = roc_auc_score(self.true_labels, self.pred_probs)
        return auc


def init_method(method, shape, name, max_val=0.01):
    """
    The method of init parameters.

    Args:
        method (str): The method uses to initialize parameter.
        shape (list): The shape of parameter.
        name (str): The name of parameter.
        max_val (float): Max value in parameter when uses 'random' or 'uniform' to initialize parameter.

    Returns:
        Parameter.
    """
    if method in ['random', 'uniform']:
        params = Parameter(initializer(Uniform(max_val), shape, ms_type), name=name)
    elif method == "one":
        params = Parameter(initializer("ones", shape, ms_type), name=name)
    elif method == 'zero':
        params = Parameter(initializer("zeros", shape, ms_type), name=name)
    elif method == "normal":
        params = Parameter(initializer(Normal(max_val), shape, ms_type), name=name)
    return params


def init_var_dict(init_args, values):
    """
    Init parameter.

    Args:
        init_args (list): Define max and min value of parameters.
        values (list): Define name, shape and init method of parameters.

    Returns:
        dict, a dict ot Parameter.
    """
    var_map = {}
    _, _max_val = init_args
    for key, shape, init_flag in values:
        if key not in var_map.keys():
            if init_flag in ['random', 'uniform']:
                var_map[key] = Parameter(initializer(Uniform(_max_val), shape, ms_type), name=key)
            elif init_flag == "one":
                var_map[key] = Parameter(initializer("ones", shape, ms_type), name=key)
            elif init_flag == "zero":
                var_map[key] = Parameter(initializer("zeros", shape, ms_type), name=key)
            elif init_flag == 'normal':
                var_map[key] = Parameter(initializer(Normal(_max_val), shape, ms_type), name=key)
    return var_map


class DenseLayer(nn.Cell):
    """
    Dense Layer for Deep Layer of DeepFM Model;
    Containing: activation, matmul, bias_add;
    Args:
        input_dim (int): the shape of weight at 0-aixs;
        output_dim (int): the shape of weight at 1-aixs, and shape of bias
        weight_bias_init (list): weight and bias init method, "random", "uniform", "one", "zero", "normal";
        act_str (str): activation function method, "relu", "sigmoid", "tanh";
        keep_prob (float): Dropout Layer keep_prob_rate;
        scale_coef (float): input scale coefficient;
    """
    def __init__(self, input_dim, output_dim, weight_bias_init, act_str, keep_prob=0.9, scale_coef=1.0):
        super(DenseLayer, self).__init__()
        weight_init, bias_init = weight_bias_init
        self.weight = init_method(weight_init, [input_dim, output_dim], name="weight")
        self.bias = init_method(bias_init, [output_dim], name="bias")
        self.act_func = self._init_activation(act_str)
        self.matmul = P.MatMul(transpose_b=False)
        self.bias_add = P.BiasAdd()
        self.cast = P.Cast()
        self.dropout = Dropout(keep_prob=keep_prob)
        self.mul = P.Mul()
        self.realDiv = P.RealDiv()
        self.scale_coef = scale_coef

    def _init_activation(self, act_str):
        act_str = act_str.lower()
        if act_str == "relu":
            act_func = P.ReLU()
        elif act_str == "sigmoid":
            act_func = P.Sigmoid()
        elif act_str == "tanh":
            act_func = P.Tanh()
        return act_func

    def construct(self, x):
P
panbingao 已提交
150
        """Dense Layer for Deep Layer of DeepFM Model."""
Y
yangyongjie 已提交
151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
        x = self.act_func(x)
        if self.training:
            x = self.dropout(x)
        x = self.mul(x, self.scale_coef)
        x = self.cast(x, mstype.float16)
        weight = self.cast(self.weight, mstype.float16)
        wx = self.matmul(x, weight)
        wx = self.cast(wx, mstype.float32)
        wx = self.realDiv(wx, self.scale_coef)
        output = self.bias_add(wx, self.bias)
        return output


class DeepFMModel(nn.Cell):
    """
    From paper: "DeepFM: A Factorization-Machine based Neural Network for CTR Prediction"

    Args:
        batch_size (int):  smaple_number of per step in training; (int, batch_size=128)
        filed_size (int):  input filed number, or called id_feature number; (int, filed_size=39)
        vocab_size (int):  id_feature vocab size, id dict size;  (int, vocab_size=200000)
        emb_dim (int):  id embedding vector dim, id mapped to embedding vector; (int, emb_dim=100)
        deep_layer_args (list):  Deep Layer args, layer_dim_list, layer_activator;
                             (int, deep_layer_args=[[100, 100, 100], "relu"])
        init_args (list): init args for Parameter init; (list, init_args=[min, max, seeds])
        weight_bias_init (list): weight, bias init method for deep layers;
                            (list[str], weight_bias_init=['random', 'zero'])
        keep_prob (float): if dropout_flag is True, keep_prob rate to keep connect; (float, keep_prob=0.8)
    """
    def __init__(self, config):
        super(DeepFMModel, self).__init__()

        self.batch_size = config.batch_size
        self.field_size = config.data_field_size
        self.vocab_size = config.data_vocab_size
        self.emb_dim = config.data_emb_dim
        self.deep_layer_dims_list, self.deep_layer_act = config.deep_layer_args
        self.init_args = config.init_args
        self.weight_bias_init = config.weight_bias_init
        self.keep_prob = config.keep_prob
        init_acts = [('W_l2', [self.vocab_size, 1], 'normal'),
                     ('V_l2', [self.vocab_size, self.emb_dim], 'normal'),
                     ('b', [1], 'normal')]
        var_map = init_var_dict(self.init_args, init_acts)
        self.fm_w = var_map["W_l2"]
        self.fm_b = var_map["b"]
        self.embedding_table = var_map["V_l2"]
        # Deep Layers
        self.deep_input_dims = self.field_size * self.emb_dim + 1
        self.all_dim_list = [self.deep_input_dims] + self.deep_layer_dims_list + [1]
        self.dense_layer_1 = DenseLayer(self.all_dim_list[0], self.all_dim_list[1],
                                        self.weight_bias_init, self.deep_layer_act, self.keep_prob)
        self.dense_layer_2 = DenseLayer(self.all_dim_list[1], self.all_dim_list[2],
                                        self.weight_bias_init, self.deep_layer_act, self.keep_prob)
        self.dense_layer_3 = DenseLayer(self.all_dim_list[2], self.all_dim_list[3],
                                        self.weight_bias_init, self.deep_layer_act, self.keep_prob)
        self.dense_layer_4 = DenseLayer(self.all_dim_list[3], self.all_dim_list[4],
                                        self.weight_bias_init, self.deep_layer_act, self.keep_prob)
        # FM, linear Layers
        self.Gatherv2 = P.GatherV2()
        self.Mul = P.Mul()
        self.ReduceSum = P.ReduceSum(keep_dims=False)
        self.Reshape = P.Reshape()
        self.Square = P.Square()
        self.Shape = P.Shape()
        self.Tile = P.Tile()
        self.Concat = P.Concat(axis=1)
        self.Cast = P.Cast()

    def construct(self, id_hldr, wt_hldr):
        """
        Args:
            id_hldr: batch ids;   [bs, field_size]
            wt_hldr: batch weights;   [bs, field_size]
        """

        mask = self.Reshape(wt_hldr, (self.batch_size, self.field_size, 1))
        # Linear layer
        fm_id_weight = self.Gatherv2(self.fm_w, id_hldr, 0)
        wx = self.Mul(fm_id_weight, mask)
        linear_out = self.ReduceSum(wx, 1)
        # FM layer
        fm_id_embs = self.Gatherv2(self.embedding_table, id_hldr, 0)
        vx = self.Mul(fm_id_embs, mask)
        v1 = self.ReduceSum(vx, 1)
        v1 = self.Square(v1)
        v2 = self.Square(vx)
        v2 = self.ReduceSum(v2, 1)
        fm_out = 0.5 * self.ReduceSum(v1 - v2, 1)
        fm_out = self.Reshape(fm_out, (-1, 1))
        #  Deep layer
        b = self.Reshape(self.fm_b, (1, 1))
        b = self.Tile(b, (self.batch_size, 1))
        deep_in = self.Reshape(vx, (-1, self.field_size * self.emb_dim))
        deep_in = self.Concat((deep_in, b))
        deep_in = self.dense_layer_1(deep_in)
        deep_in = self.dense_layer_2(deep_in)
        deep_in = self.dense_layer_3(deep_in)
        deep_out = self.dense_layer_4(deep_in)
        out = linear_out + fm_out + deep_out
        return out, fm_id_weight, fm_id_embs


class NetWithLossClass(nn.Cell):
    """
    NetWithLossClass definition.
    """
    def __init__(self, network, l2_coef=1e-6):
        super(NetWithLossClass, self).__init__(auto_prefix=False)
        self.loss = P.SigmoidCrossEntropyWithLogits()
        self.network = network
        self.l2_coef = l2_coef
        self.Square = P.Square()
        self.ReduceMean_false = P.ReduceMean(keep_dims=False)
        self.ReduceSum_false = P.ReduceSum(keep_dims=False)

    def construct(self, batch_ids, batch_wts, label):
        predict, fm_id_weight, fm_id_embs = self.network(batch_ids, batch_wts)
        log_loss = self.loss(predict, label)
        mean_log_loss = self.ReduceMean_false(log_loss)
        l2_loss_w = self.ReduceSum_false(self.Square(fm_id_weight))
        l2_loss_v = self.ReduceSum_false(self.Square(fm_id_embs))
        l2_loss_all = self.l2_coef * (l2_loss_v + l2_loss_w) * 0.5
        loss = mean_log_loss + l2_loss_all
        return loss


class TrainStepWrap(nn.Cell):
    """
    TrainStepWrap definition
    """
    def __init__(self, network, lr=5e-8, eps=1e-8, loss_scale=1000.0):
        super(TrainStepWrap, self).__init__(auto_prefix=False)
        self.network = network
285
        self.network.set_grad()
Y
yangyongjie 已提交
286 287 288 289
        self.network.set_train()
        self.weights = ParameterTuple(network.trainable_params())
        self.optimizer = Adam(self.weights, learning_rate=lr, eps=eps, loss_scale=loss_scale)
        self.hyper_map = C.HyperMap()
P
panyifeng 已提交
290
        self.grad = C.GradOperation(get_by_list=True, sens_param=True)
Y
yangyongjie 已提交
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372
        self.sens = loss_scale

    def construct(self, batch_ids, batch_wts, label):
        weights = self.weights
        loss = self.network(batch_ids, batch_wts, label)
        sens = P.Fill()(P.DType()(loss), P.Shape()(loss), self.sens) #
        grads = self.grad(self.network, weights)(batch_ids, batch_wts, label, sens)
        return F.depend(loss, self.optimizer(grads))


class PredictWithSigmoid(nn.Cell):
    """
    Eval model with sigmoid.
    """
    def __init__(self, network):
        super(PredictWithSigmoid, self).__init__(auto_prefix=False)
        self.network = network
        self.sigmoid = P.Sigmoid()

    def construct(self, batch_ids, batch_wts, labels):
        logits, _, _, = self.network(batch_ids, batch_wts)
        pred_probs = self.sigmoid(logits)

        return logits, pred_probs, labels


class ModelBuilder:
    """
    Model builder for DeepFM.

    Args:
        model_config (ModelConfig): Model configuration.
        train_config (TrainConfig): Train configuration.
    """
    def __init__(self, model_config, train_config):
        self.model_config = model_config
        self.train_config = train_config

    def get_callback_list(self, model=None, eval_dataset=None):
        """
        Get callbacks which contains checkpoint callback, eval callback and loss callback.

        Args:
            model (Cell): The network is added callback (default=None).
            eval_dataset (Dataset): Dataset for eval (default=None).
        """
        callback_list = []
        if self.train_config.save_checkpoint:
            config_ck = CheckpointConfig(save_checkpoint_steps=self.train_config.save_checkpoint_steps,
                                         keep_checkpoint_max=self.train_config.keep_checkpoint_max)
            ckpt_cb = ModelCheckpoint(prefix=self.train_config.ckpt_file_name_prefix,
                                      directory=self.train_config.output_path,
                                      config=config_ck)
            callback_list.append(ckpt_cb)
        if self.train_config.eval_callback:
            if model is None:
                raise RuntimeError("train_config.eval_callback is {}; get_callback_list() args model is {}".format(
                                        self.train_config.eval_callback, model))
            if eval_dataset is None:
                raise RuntimeError("train_config.eval_callback is {}; get_callback_list() "
                                   "args eval_dataset is {}".format(self.train_config.eval_callback, eval_dataset))
            auc_metric = AUCMetric()
            eval_callback = EvalCallBack(model, eval_dataset, auc_metric,
                                         eval_file_path=os.path.join(self.train_config.output_path,
                                                                     self.train_config.eval_file_name))
            callback_list.append(eval_callback)
        if self.train_config.loss_callback:
            loss_callback = LossCallBack(loss_file_path=os.path.join(self.train_config.output_path,
                                                                     self.train_config.loss_file_name))
            callback_list.append(loss_callback)
        if callback_list:
            return callback_list
        return None

    def get_train_eval_net(self):
        deepfm_net = DeepFMModel(self.model_config)
        loss_net = NetWithLossClass(deepfm_net, l2_coef=self.train_config.l2_coef)
        train_net = TrainStepWrap(loss_net, lr=self.train_config.learning_rate,
                                  eps=self.train_config.epsilon,
                                  loss_scale=self.train_config.loss_scale)
        eval_net = PredictWithSigmoid(deepfm_net)
        return train_net, eval_net