# -*- coding: utf-8 -*-
# Author: Zhi Qiao <mingshan_ai@163.com>
# License: BSD 2 clause
import os
import torch
import torch.nn as nn
import pickle
import warnings
from ._loss import callLoss
from ._dlbase import BaseControler
warnings.filterwarnings('ignore')
[docs]class LocationAttention(nn.Module):
def __init__(self, hidden_size, device):
super(LocationAttention, self).__init__()
self.hidden_size = hidden_size
self.attention_value_ori_func = nn.Linear(self.hidden_size, 1)
self.device = device
[docs] def forward(self, input_data):
# shape of input_data: <n_batch, n_seq, hidden_size>
n_batch, n_seq, hidden_size = input_data.shape
# shape of reshape_feat: <n_batch*n_seq, hidden_size>
reshape_feat = input_data.reshape(n_batch*n_seq, hidden_size)
# shape of attention_value_ori: <n_batch*n_seq, 1>
attention_value_ori = torch.exp(self.attention_value_ori_func(reshape_feat))
# shape of attention_value_format: <n_batch, 1, n_seq>
attention_value_format = attention_value_ori.reshape(n_batch, n_seq).unsqueeze(1)
# shape of ensemble flag format: <1, n_seq, n_seq>
# if n_seq = 3, ensemble_flag_format can get below flag data
# [[[ 0 0 0
# 1 0 0
# 1 1 0 ]]]
ensemble_flag_format = torch.triu(torch.ones([n_seq, n_seq]), diagonal = 1).permute(1, 0).unsqueeze(0).to(self.device)
# shape of accumulate_attention_value: <n_batch, n_seq, 1>
accumulate_attention_value = torch.sum(attention_value_format * ensemble_flag_format, -1).unsqueeze(-1) + 1e-9
# shape of each_attention_value: <n_batch, n_seq, n_seq>
each_attention_value = attention_value_format * ensemble_flag_format
# shape of attention_weight_format: <n_batch, n_seq, n_seq>
attention_weight_format = each_attention_value/accumulate_attention_value
# shape of _extend_attention_weight_format: <n_batch, n_seq, n_seq, 1>
_extend_attention_weight_format = attention_weight_format.unsqueeze(-1)
# shape of _extend_input_data: <n_batch, 1, n_seq, hidden_size>
_extend_input_data = input_data.unsqueeze(1)
# shape of _weighted_input_data: <n_batch, n_seq, n_seq, hidden_size>
_weighted_input_data = _extend_attention_weight_format * _extend_input_data
# shape of weighted_output: <n_batch, n_seq, hidden_size>
weighted_output = torch.sum(_weighted_input_data, 2)
return weighted_output
[docs]class GeneralAttention(nn.Module):
def __init__(self, hidden_size, device):
super(GeneralAttention, self).__init__()
self.hidden_size = hidden_size
self.correlated_value_ori_func = nn.Linear(self.hidden_size, self.hidden_size)
self.device = device
[docs] def forward(self, input_data):
# shape of input_data: <n_batch, n_seq, hidden_size>
n_batch, n_seq, hidden_size = input_data.shape
# shape of reshape_feat: <n_batch*n_seq, hidden_size>
reshape_feat = input_data.reshape(n_batch*n_seq, hidden_size)
# shape of correlated_value_ori: <n_batch, n_seq, hidden_size>
correlated_value_ori = self.correlated_value_ori_func(reshape_feat).reshape(n_batch, n_seq, hidden_size)
# shape of _extend_correlated_value_ori: <n_batch, n_seq, 1, hidden_size>
_extend_correlated_value_ori = correlated_value_ori.unsqueeze(-2)
# shape of _extend_input_data: <n_batch, 1, n_seq, hidden_size>
_extend_input_data = input_data.unsqueeze(1)
# shape of _extend_input_data: <n_batch, n_seq, n_seq, hidden_size>
_correlat_value = _extend_correlated_value_ori * _extend_input_data
# shape of attention_value_format: <n_batch, n_seq, n_seq>
attention_value_format = torch.exp(torch.sum(_correlat_value, dim = -1))
# shape of ensemble flag format: <1, n_seq, n_seq>
# if n_seq = 3, ensemble_flag_format can get below flag data
# [[[ 0 0 0
# 1 0 0
# 1 1 0 ]]]
ensemble_flag_format = torch.triu(torch.ones([n_seq, n_seq]), diagonal = 1).permute(1, 0).unsqueeze(0).to(self.device)
# shape of accumulate_attention_value: <n_batch, n_seq, 1>
accumulate_attention_value = torch.sum(attention_value_format * ensemble_flag_format, -1).unsqueeze(-1) + 1e-10
# shape of each_attention_value: <n_batch, n_seq, n_seq>
each_attention_value = attention_value_format * ensemble_flag_format
# shape of attention_weight_format: <n_batch, n_seq, n_seq>
attention_weight_format = each_attention_value/accumulate_attention_value
# shape of _extend_attention_weight_format: <n_batch, n_seq, n_seq, 1>
_extend_attention_weight_format = attention_weight_format.unsqueeze(-1)
# shape of _extend_input_data: <n_batch, 1, n_seq, hidden_size>
_extend_input_data = input_data.unsqueeze(1)
# shape of _weighted_input_data: <n_batch, n_seq, n_seq, hidden_size>
_weighted_input_data = _extend_attention_weight_format * _extend_input_data
# shape of weighted_output: <n_batch, n_seq, hidden_size>
weighted_output = torch.sum(_weighted_input_data, 2)
return weighted_output
[docs]class ConcatenationAttention(nn.Module):
def __init__(self, hidden_size, attention_dim = 16, device = None):
super(ConcatenationAttention, self).__init__()
self.hidden_size = hidden_size
self.attention_dim = attention_dim
self.attention_map_func = nn.Linear(2 * self.hidden_size, self.attention_dim)
self.activate_func = nn.Tanh()
self.correlated_value_ori_func = nn.Linear(self.attention_dim, 1)
self.device = device
[docs] def forward(self, input_data):
# shape of input_data: <n_batch, n_seq, hidden_size>
n_batch, n_seq, hidden_size = input_data.shape
# shape of _extend_input_data: <n_batch, n_seq, 1, hidden_size>
_extend_input_data_f = input_data.unsqueeze(-2)
# shape of _repeat_extend_correlated_value_ori: <n_batch, n_seq, n_seq, hidden_size>
_repeat_extend_input_data_f = _extend_input_data_f.repeat(1,1,n_seq,1)
# shape of _extend_input_data: <n_batch, 1, n_seq, hidden_size>
_extend_input_data_b = input_data.unsqueeze(1)
# shape of _repeat_extend_input_data: <n_batch, n_seq, n_seq, hidden_size>
_repeat_extend_input_data_b = _extend_input_data_b.repeat(1,n_seq,1,1)
# shape of _concate_value: <n_batch, n_seq, n_seq, 2 * hidden_size>
_concate_value = torch.cat([_repeat_extend_input_data_f, _repeat_extend_input_data_b], dim = -1)
# shape of _correlat_value: <n_batch, n_seq, n_seq>
_correlat_value = self.activate_func(self.attention_map_func(_concate_value.reshape(-1, 2 * hidden_size)))
_correlat_value = self.correlated_value_ori_func(_correlat_value).reshape(n_batch, n_seq, n_seq)
# shape of attention_value_format: <n_batch, n_seq, n_seq>
attention_value_format = torch.exp(_correlat_value)
# shape of ensemble flag format: <1, n_seq, n_seq>
# if n_seq = 3, ensemble_flag_format can get below flag data
# [[[ 0 0 0
# 1 0 0
# 1 1 0 ]]]
ensemble_flag_format = torch.triu(torch.ones([n_seq, n_seq]), diagonal = 1).permute(1, 0).unsqueeze(0).to(self.device)
# shape of accumulate_attention_value: <n_batch, n_seq, 1>
accumulate_attention_value = torch.sum(attention_value_format * ensemble_flag_format, -1).unsqueeze(-1) + 1e-10
# shape of each_attention_value: <n_batch, n_seq, n_seq>
each_attention_value = attention_value_format * ensemble_flag_format
# shape of attention_weight_format: <n_batch, n_seq, n_seq>
attention_weight_format = each_attention_value/accumulate_attention_value
# shape of _extend_attention_weight_format: <n_batch, n_seq, n_seq, 1>
_extend_attention_weight_format = attention_weight_format.unsqueeze(-1)
# shape of _extend_input_data: <n_batch, 1, n_seq, hidden_size>
_extend_input_data = input_data.unsqueeze(1)
# shape of _weighted_input_data: <n_batch, n_seq, n_seq, hidden_size>
_weighted_input_data = _extend_attention_weight_format * _extend_input_data
# shape of weighted_output: <n_batch, n_seq, hidden_size>
weighted_output = torch.sum(_weighted_input_data, 2)
return weighted_output
[docs]class callPredictor(nn.Module):
def __init__(self,
input_size = None,
embed_size = 16,
hidden_size = 8,
output_size = 10,
bias = True,
dropout = 0.5,
batch_first = True,
label_size = 1,
attention_type = 'location_based',
attention_dim = 8,
device = None):
super(callPredictor, self).__init__()
assert input_size != None and isinstance(input_size, int), 'fill in correct input_size'
self.input_size = input_size
self.embed_size = embed_size
self.hidden_size = hidden_size
self.output_size = output_size
self.label_size = label_size
self.embed_func = nn.Linear(self.input_size, self.embed_size)
self.rnn_model = nn.GRU(input_size = embed_size,
hidden_size = hidden_size,
bias = bias,
dropout = dropout,
bidirectional = True,
batch_first = batch_first)
if attention_type == 'location_based':
self.attention_func = LocationAttention(2*hidden_size, device)
elif attention_type == 'general':
self.attention_func = GeneralAttention(2*hidden_size, device)
elif attention_type == 'concatenation_based':
self.attention_func = ConcatenationAttention(2*hidden_size, attention_dim = attention_dim, device = device)
else:
raise Exception('fill in correct attention_type, [location_based, general, concatenation_based]')
self.output_func = nn.Linear(4 * hidden_size, self.output_size)
self.output_activate = nn.Tanh()
self.predict_func = nn.Linear(self.output_size, self.label_size)
[docs] def forward(self, input_data):
"""
Parameters
----------
input_data = {
'X': shape (batchsize, n_timestep, n_featdim)
'M': shape (batchsize, n_timestep)
'cur_M': shape (batchsize, n_timestep)
'T': shape (batchsize, n_timestep)
}
Return
----------
all_output, shape (batchsize, n_timestep, n_labels)
predict output of each time step
cur_output, shape (batchsize, n_labels)
predict output of last time step
"""
X = input_data['X']
M = input_data['M']
cur_M = input_data['cur_M']
batchsize, n_timestep, n_orifeatdim = X.shape
_ori_X = X.view(-1, n_orifeatdim)
_embed_X = self.embed_func(_ori_X)
_embed_X = _embed_X.reshape(batchsize, n_timestep, self.embed_size)
_embed_F, _ = self.rnn_model(_embed_X)
_embed_F_w = self.attention_func(_embed_F)
_mix_F = torch.cat([_embed_F, _embed_F_w], dim = -1)
_mix_F_reshape = _mix_F.view(-1, 4 * self.hidden_size)
outputs = self.output_activate(self.output_func(_mix_F_reshape)).reshape(batchsize, n_timestep, self.output_size)
n_batchsize, n_timestep, output_size = outputs.shape
all_output = self.predict_func(outputs.reshape(n_batchsize*n_timestep, output_size)).\
reshape(n_batchsize, n_timestep, self.label_size) * M.unsqueeze(-1)
cur_output = (all_output * cur_M.unsqueeze(-1)).sum(dim=1)
return all_output, cur_output
[docs]class Dipole(BaseControler):
def __init__(self,
expmodel_id = 'test.new',
n_epoch = 100,
n_batchsize = 5,
learn_ratio = 1e-4,
weight_decay = 1e-4,
n_epoch_saved = 1,
attention_type = 'location_based',
attention_dim = 8,
embed_size = 16,
hidden_size = 8,
output_size = 8,
bias = True,
dropout = 0.5,
batch_first = True,
loss_name = 'L1LossSigmoid',
target_repl = False,
target_repl_coef = 0.,
aggregate = 'sum',
optimizer_name = 'adam',
use_gpu = False,
gpu_ids = '0'
):
"""
Applies an Attention-based Bidirectional Recurrent Neural Networks for an healthcare data sequence
Parameters
----------
exp_id : str, optional (default='init.test')
name of current experiment
n_epoch : int, optional (default = 100)
number of epochs with the initial learning rate
n_batchsize : int, optional (default = 5)
batch size for model training
learn_ratio : float, optional (default = 1e-4)
initial learning rate for adam
weight_decay : float, optional (default = 1e-4)
weight decay (L2 penalty)
n_epoch_saved : int, optional (default = 1)
frequency of saving checkpoints at the end of epochs
attention_type : str, optional (default = 'location_based')
Apply attention mechnism to derive a context vector that captures relevant information to
help predict target.
Current support attention methods in [location_based, general, concatenation_based] proposed in KDD2017
'location_based' ---> Location-based Attention. Alocation-based attention function is to
calculate the weights solely from hidden state
'general' ---> General Attention. An easy way to capture the relationship between two hidden states
'concatenation_based' ---> Concatenation-based Attention. Via concatenating two hidden states, then use multi-layer
perceptron(MLP) to calculatethe contextvector
attention_dim : int, optional (default = 8)
It is the latent dimensionality used for attention weight computing just for for concatenation_based attention mechnism
embed_size: int, optional (default = 16)
The number of the embeded features of original input
hidden_size : int, optional (default = 8)
The number of features of the hidden state h
output_size : int, optional (default = 8)
The number of mix features
bias : bool, optional (default = True)
If False, then the layer does not use bias weights b_ih and b_hh.
dropout : float, optional (default = 0.5)
If non-zero, introduces a Dropout layer on the outputs of each GRU layer except the last layer,
with dropout probability equal to dropout.
batch_first : bool, optional (default = False)
If True, then the input and output tensors are provided as (batch, seq, feature).
loss_name : str, optional (default='SigmoidCELoss')
Name or objective function.
use_gpu : bool, optional (default=False)
If yes, use GPU recources; else use CPU recources
gpu_ids : str, optional (default='')
If yes, assign concrete used gpu ids such as '0,2,6'; else use '0'
"""
super(Dipole, self).__init__(expmodel_id)
self.n_batchsize = n_batchsize
self.n_epoch = n_epoch
self.learn_ratio = learn_ratio
self.weight_decay = weight_decay
self.n_epoch_saved = n_epoch_saved
self.attention_type = attention_type
self.attention_dim = attention_dim
self.embed_size = embed_size
self.hidden_size = hidden_size
self.output_size = output_size
self.bias = bias
self.dropout = dropout
self.batch_first = batch_first
self.loss_name = loss_name
self.target_repl = target_repl
self.target_repl_coef = target_repl_coef
self.aggregate = aggregate
self.optimizer_name = optimizer_name
self.use_gpu = use_gpu
self.gpu_ids = gpu_ids
self._args_check()
def _build_model(self):
"""
Build the crucial components for model training
"""
if self.is_loadmodel is False:
_config = {
'input_size': self.input_size,
'embed_size': self.embed_size,
'hidden_size': self.hidden_size,
'output_size': self.output_size,
'bias': self.bias,
'dropout': self.dropout,
'batch_first': self.batch_first,
'label_size': self.label_size,
'attention_type': self.attention_type,
'attention_dim': self.attention_dim,
'device': self.device
}
self.predictor = callPredictor(**_config)
self.predictor.to(self.device)
self._save_predictor_config({key: value for key, value in _config.items() if key != 'device'})
if self.dataparallal:
self.predictor= torch.nn.DataParallel(self.predictor)
self.criterion = callLoss(task = self.task_type,
loss_name = self.loss_name,
target_repl = self.target_repl,
target_repl_coef = self.target_repl_coef,
aggregate = self.aggregate)
self.optimizer = self._get_optimizer(self.optimizer_name)
[docs] def fit(self, train_data, valid_data, assign_task_type = None):
"""
Parameters
----------
train_data : {
'x':list[episode_file_path],
'y':list[label],
'l':list[seq_len],
'feat_n': n of feature space,
'label_n': n of label space
}
The input train samples dict.
valid_data : {
'x':list[episode_file_path],
'y':list[label],
'l':list[seq_len],
'feat_n': n of feature space,
'label_n': n of label space
}
The input valid samples dict.
assign_task_type: str (default = None)
predifine task type to model mapping <feature, label>
current support ['binary','multiclass','multilabel','regression']
Returns
-------
self : object
Fitted estimator.
"""
self.task_type = assign_task_type
self._data_check([train_data, valid_data])
self._build_model()
train_reader = self._get_reader(train_data, 'train')
valid_reader = self._get_reader(valid_data, 'valid')
self._fit_model(train_reader, valid_reader)
[docs] def load_model(self,
loaded_epoch = '',
config_file_path = '',
model_file_path = ''):
"""
Parameters
----------
loaded_epoch : str, loaded model name
we save the model by <epoch_count>.epoch, latest.epoch, best.epoch
Returns
-------
self : object
loaded estimator.
"""
predictor_config = self._load_predictor_config(config_file_path)
predictor_config['device'] = self.device
self.predictor = callPredictor(**predictor_config).to(self.device)
self._load_model(loaded_epoch, model_file_path)
def _args_check(self):
"""
Check args whether valid/not and give tips
"""
assert isinstance(self.n_batchsize,int) and self.n_batchsize>0, \
'fill in correct n_batchsize (int, >0)'
assert isinstance(self.n_epoch,int) and self.n_epoch>0, \
'fill in correct n_epoch (int, >0)'
assert isinstance(self.learn_ratio,float) and self.learn_ratio>0., \
'fill in correct learn_ratio (float, >0.)'
assert isinstance(self.weight_decay,float) and self.weight_decay>=0., \
'fill in correct weight_decay (float, >=0.)'
assert isinstance(self.n_epoch_saved,int) and self.n_epoch_saved>0 and self.n_epoch_saved < self.n_epoch, \
'fill in correct n_epoch (int, >0 and <{0}).format(self.n_epoch)'
assert isinstance(self.attention_type,str) and self.attention_type in ['location_based', 'general', 'concatenation_based'], \
'fill in correct attention_type (str, [\'location_based\', \'general\', \'concatenation_based\'])'
assert isinstance(self.attention_dim,int) and self.attention_dim>0, \
'fill in correct attention_dim (int, >0)'
assert isinstance(self.embed_size,int) and self.embed_size>0, \
'fill in correct embed_size (int, >0)'
assert isinstance(self.hidden_size,int) and self.hidden_size>0, \
'fill in correct hidden_size (int, 8)'
assert isinstance(self.output_size,int) and self.output_size>0, \
'fill in correct output_size (int, 8)'
assert isinstance(self.bias,bool), \
'fill in correct bias (bool)'
assert isinstance(self.dropout,float) and self.dropout>0. and self.dropout<1., \
'fill in correct learn_ratio (float, >0 and <1.)'
assert isinstance(self.batch_first,bool), \
'fill in correct batch_first (bool)'
assert isinstance(self.target_repl,bool), \
'fill in correct target_repl (bool)'
assert isinstance(self.target_repl_coef,float) and self.target_repl_coef>=0. and self.target_repl_coef<=1., \
'fill in correct target_repl_coef (float, >=0 and <=1.)'
assert isinstance(self.aggregate,str) and self.aggregate in ['sum','avg'], \
'fill in correct aggregate (str, [\'sum\',\'avg\'])'
assert isinstance(self.optimizer_name,str) and self.optimizer_name in ['adam'], \
'fill in correct optimizer_name (str, [\'adam\'])'
assert isinstance(self.use_gpu,bool), \
'fill in correct use_gpu (bool)'
assert isinstance(self.loss_name,str), \
'fill in correct optimizer_name (str)'
assert isinstance(self.gpu_ids,str), \
'fill in correct use_gpu (str, \'0,2,7\')'
self.device = self._get_device()