# 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 __future__ import division from __future__ import print_function import os import numpy as np import pandas as pd import copy from sklearn.metrics import roc_auc_score, mean_squared_error import logging from ...utils import unpack_archive_with_buffer, save_multiple_parts_file, create_save_path, drop_nan_by_y_index from ...log import get_module_logger, TimeInspector import torch import torch.nn as nn import torch.optim as optim from ...model.base import Model from ...data.dataset import DatasetH from ...data.dataset.handler import DataHandlerLP class HATS(Model): """HATS Model Parameters ---------- input_dim : int input dimension output_dim : int output dimension layers : tuple layer sizes lr : float learning rate optimizer : str optimizer name GPU : str the GPU ID(s) used for training """ def __init__( self, d_feat=6, hidden_size=64, num_layers=2, dropout=0.5, n_epochs=200, lr=0.01, metric="IC", batch_size=800, early_stop=20, loss="mse", base_model="GRU", with_pretrain=True, optimizer="adam", GPU="0", seed=0, **kwargs ): # Set logger. self.logger = get_module_logger("HATS") self.logger.info("HATS pytorch version...") # set hyper-parameters. self.d_feat = d_feat self.hidden_size = hidden_size self.num_layers = num_layers self.dropout = dropout self.n_epochs = n_epochs self.lr = lr self.metric = metric self.batch_size = batch_size self.early_stop = early_stop self.optimizer = optimizer.lower() self.loss = loss self.base_model = base_model self.with_pretrain = with_pretrain #### True if train HATS with pretrained base model self.visible_GPU = GPU self.use_gpu = torch.cuda.is_available() self.seed = seed self.logger.info( "HATS parameters setting:" "\nd_feat : {}" "\nhidden_size : {}" "\nnum_layers : {}" "\ndropout : {}" "\nn_epochs : {}" "\nlr : {}" "\nmetric : {}" "\nbatch_size : {}" "\nearly_stop : {}" "\noptimizer : {}" "\nloss_type : {}" "\nbase_model : {}" "\nwith_pretrain : {}" ##### debug "\nvisible_GPU : {}" "\nuse_GPU : {}" "\nseed : {}".format( d_feat, hidden_size, num_layers, dropout, n_epochs, lr, metric, batch_size, early_stop, optimizer.lower(), loss, base_model, with_pretrain, ### debug GPU, self.use_gpu, seed, ) ) if loss not in {"mse", "binary"}: raise NotImplementedError("loss {} is not supported!".format(loss)) self._scorer = mean_squared_error if loss == "mse" else roc_auc_score self.HATS_model = HATSModel( d_feat=self.d_feat, hidden_size=self.hidden_size, num_layers=self.num_layers, dropout=self.dropout, base_model=self.base_model, ) if optimizer.lower() == "adam": self.train_optimizer = optim.Adam(self.HATS_model.parameters(), lr=self.lr) elif optimizer.lower() == "gd": self.train_optimizer = optim.SGD(self.HATS_model.parameters(), lr=self.lr) else: raise NotImplementedError("optimizer {} is not supported!".format(optimizer)) self._fitted = False if self.use_gpu: self.HATS_model.cuda() # set the visible GPU if self.visible_GPU: os.environ["CUDA_VISIBLE_DEVICES"] = self.visible_GPU def mse(self, pred, label): loss = (pred - label) ** 2 return torch.mean(loss) def loss_fn(self, pred, label): mask = ~torch.isnan(label) if self.loss == "mse": return self.mse(pred[mask], label[mask]) raise ValueError("unknown loss `%s`" % self.loss) def metric_fn(self, pred, label): mask = torch.isfinite(label) if self.metric == "IC": return self.cal_ic(pred[mask], label[mask]) if self.metric == "" or self.metric == "loss": # use loss return -self.loss_fn(pred[mask], label[mask]) raise ValueError("unknown metric `%s`" % self.metric) def cal_ic(self, pred, label): return torch.mean(pred * label) def train_epoch(self, x_train, y_train): x_train_values = x_train.values y_train_values = np.squeeze(y_train.values) * 100 self.HATS_model.train() indices = np.arange(len(x_train_values)) np.random.shuffle(indices) for i in range(len(indices))[:: self.batch_size]: if len(indices) - i < self.batch_size: break feature = torch.from_numpy(x_train_values[indices[i : i + self.batch_size]]).float() label = torch.from_numpy(y_train_values[indices[i : i + self.batch_size]]).float() if self.use_gpu: feature = feature.cuda() label = label.cuda() pred = self.HATS_model(feature) loss = self.loss_fn(pred, label) self.train_optimizer.zero_grad() loss.backward() torch.nn.utils.clip_grad_value_(self.HATS_model.parameters(), 3.0) self.train_optimizer.step() def test_epoch(self, data_x, data_y): # prepare training data x_values = data_x.values y_values = np.squeeze(data_y.values) self.HATS_model.eval() scores = [] losses = [] indices = np.arange(len(x_values)) np.random.shuffle(indices) for i in range(len(indices))[:: self.batch_size]: if len(indices) - i < self.batch_size: break feature = torch.from_numpy(x_values[indices[i : i + self.batch_size]]).float() label = torch.from_numpy(y_values[indices[i : i + self.batch_size]]).float() if self.use_gpu: feature = feature.cuda() label = label.cuda() pred = self.HATS_model(feature) loss = self.loss_fn(pred, label) losses.append(loss.item()) score = self.metric_fn(pred, label) scores.append(score.item()) return np.mean(losses), np.mean(scores) def fit( self, dataset: DatasetH, evals_result=dict(), verbose=True, save_path=None, ): df_train, df_valid, df_test = dataset.prepare( ["train", "valid", "test"], col_set=["feature", "label"], data_key=DataHandlerLP.DK_L ) x_train, y_train = df_train["feature"], df_train["label"] x_valid, y_valid = df_valid["feature"], df_valid["label"] if save_path == None: save_path = create_save_path(save_path) stop_steps = 0 train_loss = 0 best_score = -np.inf best_epoch = 0 evals_result["train"] = [] evals_result["valid"] = [] # load pretrained base_model if self.with_pretrain: self.logger.info("loading pretrained model...") if self.base_model == "LSTM": from ...contrib.model.pytorch_lstm import LSTMModel pretrained_model = LSTMModel() pretrained_model.load_state_dict(torch.load("benchmarks/LSTM/model_lstm_csi300.pkl")) elif self.base_model == "GRU": from ...contrib.model.pytorch_gru import GRUModel pretrained_model = GRUModel() pretrained_model.load_state_dict(torch.load("benchmarks/GRU/model_gru_csi300.pkl")) model_dict = self.HATS_model.state_dict() # filter unnecessary parameters pretrained_dict = {k: v for k, v in pretrained_model.state_dict().items() if k in model_dict} # overwrite entries in the existing state dict model_dict.update(pretrained_dict) # load the new state dict self.HATS_model.load_state_dict(model_dict) self.logger.info("loading pretrained model Done...") # train self.logger.info("training...") self._fitted = True # return for step in range(self.n_epochs): self.logger.info("Epoch%d:", step) self.logger.info("training...") self.train_epoch(x_train, y_train) self.logger.info("evaluating...") train_loss, train_score = self.test_epoch(x_train, y_train) val_loss, val_score = self.test_epoch(x_valid, y_valid) self.logger.info("train %.6f, valid %.6f" % (train_score, val_score)) evals_result["train"].append(train_score) evals_result["valid"].append(val_score) if val_score > best_score: best_score = val_score stop_steps = 0 best_epoch = step best_param = copy.deepcopy(self.HATS_model.state_dict()) else: stop_steps += 1 if stop_steps >= self.early_stop: self.logger.info("early stop") break self.logger.info("best score: %.6lf @ %d" % (best_score, best_epoch)) self.HATS_model.load_state_dict(best_param) torch.save(best_param, save_path) if self.use_gpu: torch.cuda.empty_cache() def predict(self, dataset): if not self._fitted: raise ValueError("model is not fitted yet!") x_test = dataset.prepare("test", col_set="feature") index = x_test.index self.HATS_model.eval() x_values = x_test.values sample_num = x_values.shape[0] preds = [] for begin in range(sample_num)[:: self.batch_size]: if sample_num - begin < self.batch_size: end = sample_num else: end = begin + self.batch_size x_batch = torch.from_numpy(x_values[begin:end]).float() if self.use_gpu: x_batch = x_batch.cuda() with torch.no_grad(): if self.use_gpu: pred = self.HATS_model(x_batch).detach().cpu().numpy() else: pred = self.HATS_model(x_batch).detach().numpy() preds.append(pred) return pd.Series(np.concatenate(preds), index=index) class HATSModel(nn.Module): def __init__(self, d_feat=6, hidden_size=64, num_layers=2, dropout=0.0, base_model="GRU"): super().__init__() if base_model == "GRU": self.model = nn.GRU( input_size=d_feat, hidden_size=hidden_size, num_layers=num_layers, batch_first=True, dropout=dropout, ) elif base_model == "LSTM": self.model = nn.LSTM( input_size=d_feat, hidden_size=hidden_size, num_layers=num_layers, batch_first=True, dropout=dropout, ) else: raise ValueError("unknown base model name `%s`" % base_model) self.hidden_size = hidden_size self.bn1 = nn.BatchNorm1d(num_features=hidden_size, track_running_stats=False) self.fc = nn.Linear(hidden_size, hidden_size) self.bn2 = nn.BatchNorm1d(num_features=hidden_size, track_running_stats=False) self.fc_out = nn.Linear(hidden_size, 1) self.leaky_relu = nn.LeakyReLU() self.softmax = nn.Softmax(dim=1) self.d_feat = d_feat num_head_att = [1] * num_layers hidden_dim = [hidden_size] * num_layers dims = [d_feat] + [d * nh for (d, nh) in zip(hidden_dim, num_head_att[:-1])] + [num_head_att[-1]] in_dims = dims[:-1] out_dims = [d // nh for (d, nh) in zip(dims[1:], num_head_att)] self.attn = nn.ModuleList( [GraphAttention(i, o, nh, dropout) for (i, o, nh) in zip(in_dims, out_dims, num_head_att)] ) self.bns = nn.ModuleList([nn.BatchNorm1d(dim) for dim in dims[1:-1]]) self.dropout = nn.Dropout(dropout) self.elu = nn.ELU() def forward(self, x): x = x.reshape(len(x), self.d_feat, -1) # [N, F, T] x = x.permute(0, 2, 1) # [N, T, F] out, _ = self.model(x) hidden = out[:, -1, :] hidden = self.bn1(hidden) attention = GraphAttention.cal_attention(hidden, hidden) output = attention.mm(hidden) output = self.fc(output) output = self.bn2(output) output = self.leaky_relu(output) return self.fc_out(output).squeeze() class GraphAttention(nn.Module): def __init__(self, input_dim, output_dim, num_heads, dropout=0.5): super().__init__() """ Parameters ---------- input_dim : int Dimension of input node features. output_dim : int Dimension of output node features. num_heads : list of ints Number of attention heads in each hidden layer and output layer. Must be non empty. Note that len(num_heads) = len(hidden_dims)+1. dropout : float Dropout rate. Default: 0.5. """ self.input_dim = input_dim self.output_dim = output_dim self.num_heads = num_heads self.fcs = nn.ModuleList([nn.Linear(input_dim, output_dim) for _ in range(num_heads)]) self.a = nn.ModuleList([nn.Linear(2 * output_dim, 1) for _ in range(num_heads)]) self.dropout = nn.Dropout(dropout) self.softmax = nn.Softmax(dim=0) self.leakyrelu = nn.LeakyReLU() def forward(self, features, nodes, mapping, rows): """ Parameters ---------- features : torch.Tensor An (n' x input_dim) tensor of input node features. node_layers : list of numpy array node_layers[i] is an array of the nodes in the ith layer of the computation graph. mappings : list of dictionary mappings[i] is a dictionary mapping node v (labelled 0 to |V|-1) in node_layers[i] to its position in node_layers[i]. For example, if node_layers[i] = [2,5], then mappings[i][2] = 0 and mappings[i][5] = 1. rows : numpy array rows[i] is an array of neighbors of node i. Returns ------- out : torch.Tensor An (len(node_layers[-1]) x output_dim) tensor of output node features. """ nprime = features.shape[0] rows = [np.array([mapping[v] for v in row], dtype=np.int64) for row in rows] sum_degs = np.hstack(([0], np.cumsum([len(row) for row in rows]))) mapped_nodes = [mapping[v] for v in nodes] indices = torch.LongTensor([[v, c] for (v, row) in zip(mapped_nodes, rows) for c in row]).t() out = [] for k in range(self.num_heads): h = self.fcs[k](features) nbr_h = torch.cat(tuple([h[row] for row in rows]), dim=0) self_h = torch.cat(tuple([h[mapping[nodes[i]]].repeat(len(row), 1) for (i, row) in enumerate(rows)]), dim=0) cat_h = torch.cat((self_h, nbr_h), dim=1) e = self.leakyrelu(self.a[k](cat_h)) alpha = [self.softmax(e[lo:hi]) for (lo, hi) in zip(sum_degs, sum_degs[1:])] alpha = torch.cat(tuple(alpha), dim=0) alpha = alpha.squeeze(1) alpha = self.dropout(alpha) adj = torch.sparse.FloatTensor(indices, alpha, torch.Size([nprime, nprime])) out.append(torch.sparse.mm(adj, h)[mapped_nodes]) return out def cal_attention(x, y): att_x = torch.mean(x, dim=1).reshape(-1, 1) att_y = torch.mean(y, dim=1).reshape(-1, 1) att = att_x.mm(torch.t(att_y)) x_att = x.reshape(x.shape[0], 1, x.shape[1]).repeat(1, y.shape[0], 1) y_att = y.reshape(1, y.shape[0], y.shape[1]).repeat(x.shape[0], 1, 1) return ( torch.mean( x.reshape(x.shape[0], 1, x.shape[1]).repeat(1, y.shape[0], 1) * y.reshape(1, y.shape[0], y.shape[1]).repeat(x.shape[0], 1, 1), dim=2, ) - att )