# Copyright (c) Microsoft Corporation. # Licensed under the MIT License. from __future__ import division from __future__ import print_function import numpy as np import pandas as pd from typing import Text, Union import copy from ...utils import get_or_create_path from ...log import get_module_logger import torch import torch.nn as nn import torch.optim as optim from .pytorch_utils import count_parameters from ...model.base import Model from ...data.dataset import DatasetH from ...data.dataset.handler import DataHandlerLP from ...contrib.model.pytorch_lstm import LSTMModel from ...contrib.model.pytorch_gru import GRUModel class IGMTF(Model): """IGMTF Model Parameters ---------- d_feat : int input dimension for each time step metric: str the evaluation metric used in early stop 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=0.0, n_epochs=200, lr=0.001, metric="", early_stop=20, loss="mse", base_model="GRU", model_path=None, optimizer="adam", GPU=0, seed=None, **kwargs ): # Set logger. self.logger = get_module_logger("IGMTF") self.logger.info("IMGTF 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.early_stop = early_stop self.optimizer = optimizer.lower() self.loss = loss self.base_model = base_model self.model_path = model_path self.device = torch.device("cuda:%d" % (GPU) if torch.cuda.is_available() and GPU >= 0 else "cpu") self.seed = seed self.logger.info( "IGMTF parameters setting:" "\nd_feat : {}" "\nhidden_size : {}" "\nnum_layers : {}" "\ndropout : {}" "\nn_epochs : {}" "\nlr : {}" "\nmetric : {}" "\nearly_stop : {}" "\noptimizer : {}" "\nloss_type : {}" "\nbase_model : {}" "\nmodel_path : {}" "\nvisible_GPU : {}" "\nuse_GPU : {}" "\nseed : {}".format( d_feat, hidden_size, num_layers, dropout, n_epochs, lr, metric, early_stop, optimizer.lower(), loss, base_model, model_path, GPU, self.use_gpu, seed, ) ) if self.seed is not None: np.random.seed(self.seed) torch.manual_seed(self.seed) self.igmtf_model = IGMTFModel( d_feat=self.d_feat, hidden_size=self.hidden_size, num_layers=self.num_layers, dropout=self.dropout, base_model=self.base_model, ) self.logger.info("model:\n{:}".format(self.igmtf_model)) self.logger.info("model size: {:.4f} MB".format(count_parameters(self.igmtf_model))) if optimizer.lower() == "adam": self.train_optimizer = optim.Adam(self.igmtf_model.parameters(), lr=self.lr) elif optimizer.lower() == "gd": self.train_optimizer = optim.SGD(self.igmtf_model.parameters(), lr=self.lr) else: raise NotImplementedError("optimizer {} is not supported!".format(optimizer)) self.fitted = False self.igmtf_model.to(self.device) @property def use_gpu(self): return self.device != torch.device("cpu") 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": x = pred[mask] y = label[mask] vx = x - torch.mean(x) vy = y - torch.mean(y) return torch.sum(vx * vy) / (torch.sqrt(torch.sum(vx**2)) * torch.sqrt(torch.sum(vy**2))) if self.metric == ("", "loss"): return -self.loss_fn(pred[mask], label[mask]) raise ValueError("unknown metric `%s`" % self.metric) def get_daily_inter(self, df, shuffle=False): # organize the train data into daily batches daily_count = df.groupby(level=0).size().values daily_index = np.roll(np.cumsum(daily_count), 1) daily_index[0] = 0 if shuffle: # shuffle data daily_shuffle = list(zip(daily_index, daily_count)) np.random.shuffle(daily_shuffle) daily_index, daily_count = zip(*daily_shuffle) return daily_index, daily_count def get_train_hidden(self, x_train): x_train_values = x_train.values daily_index, daily_count = self.get_daily_inter(x_train, shuffle=True) self.igmtf_model.eval() train_hidden = [] train_hidden_day = [] for idx, count in zip(daily_index, daily_count): batch = slice(idx, idx + count) feature = torch.from_numpy(x_train_values[batch]).float().to(self.device) out = self.igmtf_model(feature, get_hidden=True) train_hidden.append(out.detach().cpu()) train_hidden_day.append(out.detach().cpu().mean(dim=0).unsqueeze(dim=0)) train_hidden = np.asarray(train_hidden, dtype=object) train_hidden_day = torch.cat(train_hidden_day) return train_hidden, train_hidden_day def train_epoch(self, x_train, y_train, train_hidden, train_hidden_day): x_train_values = x_train.values y_train_values = np.squeeze(y_train.values) self.igmtf_model.train() daily_index, daily_count = self.get_daily_inter(x_train, shuffle=True) for idx, count in zip(daily_index, daily_count): batch = slice(idx, idx + count) feature = torch.from_numpy(x_train_values[batch]).float().to(self.device) label = torch.from_numpy(y_train_values[batch]).float().to(self.device) pred = self.igmtf_model(feature, train_hidden=train_hidden, train_hidden_day=train_hidden_day) loss = self.loss_fn(pred, label) self.train_optimizer.zero_grad() loss.backward() torch.nn.utils.clip_grad_value_(self.igmtf_model.parameters(), 3.0) self.train_optimizer.step() def test_epoch(self, data_x, data_y, train_hidden, train_hidden_day): # prepare training data x_values = data_x.values y_values = np.squeeze(data_y.values) self.igmtf_model.eval() scores = [] losses = [] daily_index, daily_count = self.get_daily_inter(data_x, shuffle=False) for idx, count in zip(daily_index, daily_count): batch = slice(idx, idx + count) feature = torch.from_numpy(x_values[batch]).float().to(self.device) label = torch.from_numpy(y_values[batch]).float().to(self.device) pred = self.igmtf_model(feature, train_hidden=train_hidden, train_hidden_day=train_hidden_day) 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(), save_path=None, ): df_train, df_valid = dataset.prepare( ["train", "valid"], col_set=["feature", "label"], data_key=DataHandlerLP.DK_L, ) if df_train.empty or df_valid.empty: raise ValueError("Empty data from dataset, please check your dataset config.") x_train, y_train = df_train["feature"], df_train["label"] x_valid, y_valid = df_valid["feature"], df_valid["label"] save_path = get_or_create_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.base_model == "LSTM": pretrained_model = LSTMModel() elif self.base_model == "GRU": pretrained_model = GRUModel() else: raise ValueError("unknown base model name `%s`" % self.base_model) if self.model_path is not None: self.logger.info("Loading pretrained model...") pretrained_model.load_state_dict(torch.load(self.model_path, map_location=self.device)) model_dict = self.igmtf_model.state_dict() pretrained_dict = { k: v for k, v in pretrained_model.state_dict().items() if k in model_dict # pylint: disable=E1135 } model_dict.update(pretrained_dict) self.igmtf_model.load_state_dict(model_dict) self.logger.info("Loading pretrained model Done...") # train self.logger.info("training...") self.fitted = True for step in range(self.n_epochs): self.logger.info("Epoch%d:", step) self.logger.info("training...") train_hidden, train_hidden_day = self.get_train_hidden(x_train) self.train_epoch(x_train, y_train, train_hidden, train_hidden_day) self.logger.info("evaluating...") train_loss, train_score = self.test_epoch(x_train, y_train, train_hidden, train_hidden_day) val_loss, val_score = self.test_epoch(x_valid, y_valid, train_hidden, train_hidden_day) 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.igmtf_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.igmtf_model.load_state_dict(best_param) torch.save(best_param, save_path) if self.use_gpu: torch.cuda.empty_cache() def predict(self, dataset: DatasetH, segment: Union[Text, slice] = "test"): if not self.fitted: raise ValueError("model is not fitted yet!") x_train = dataset.prepare("train", col_set="feature", data_key=DataHandlerLP.DK_L) train_hidden, train_hidden_day = self.get_train_hidden(x_train) x_test = dataset.prepare(segment, col_set="feature", data_key=DataHandlerLP.DK_I) index = x_test.index self.igmtf_model.eval() x_values = x_test.values preds = [] daily_index, daily_count = self.get_daily_inter(x_test, shuffle=False) for idx, count in zip(daily_index, daily_count): batch = slice(idx, idx + count) x_batch = torch.from_numpy(x_values[batch]).float().to(self.device) with torch.no_grad(): pred = ( self.igmtf_model(x_batch, train_hidden=train_hidden, train_hidden_day=train_hidden_day) .detach() .cpu() .numpy() ) preds.append(pred) return pd.Series(np.concatenate(preds), index=index) class IGMTFModel(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.rnn = nn.GRU( input_size=d_feat, hidden_size=hidden_size, num_layers=num_layers, batch_first=True, dropout=dropout, ) elif base_model == "LSTM": self.rnn = 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.lins = nn.Sequential() for i in range(2): self.lins.add_module("linear" + str(i), nn.Linear(hidden_size, hidden_size)) self.lins.add_module("leakyrelu" + str(i), nn.LeakyReLU()) self.fc_output = nn.Linear(hidden_size * 2, hidden_size * 2) self.project1 = nn.Linear(hidden_size, hidden_size, bias=False) self.project2 = nn.Linear(hidden_size, hidden_size, bias=False) self.fc_out_pred = nn.Linear(hidden_size * 2, 1) self.leaky_relu = nn.LeakyReLU() self.d_feat = d_feat def cal_cos_similarity(self, x, y): # the 2nd dimension of x and y are the same xy = x.mm(torch.t(y)) x_norm = torch.sqrt(torch.sum(x * x, dim=1)).reshape(-1, 1) y_norm = torch.sqrt(torch.sum(y * y, dim=1)).reshape(-1, 1) cos_similarity = xy / (x_norm.mm(torch.t(y_norm)) + 1e-6) return cos_similarity def sparse_dense_mul(self, s, d): i = s._indices() v = s._values() dv = d[i[0, :], i[1, :]] # get values from relevant entries of dense matrix return torch.sparse.FloatTensor(i, v * dv, s.size()) def forward(self, x, get_hidden=False, train_hidden=None, train_hidden_day=None, k_day=10, n_neighbor=10): # x: [N, F*T] device = x.device x = x.reshape(len(x), self.d_feat, -1) # [N, F, T] x = x.permute(0, 2, 1) # [N, T, F] out, _ = self.rnn(x) out = out[:, -1, :] out = self.lins(out) mini_batch_out = out if get_hidden is True: return mini_batch_out mini_batch_out_day = torch.mean(mini_batch_out, dim=0).unsqueeze(0) day_similarity = self.cal_cos_similarity(mini_batch_out_day, train_hidden_day.to(device)) day_index = torch.topk(day_similarity, k_day, dim=1)[1] sample_train_hidden = train_hidden[day_index.long().cpu()].squeeze() sample_train_hidden = torch.cat(list(sample_train_hidden)).to(device) sample_train_hidden = self.lins(sample_train_hidden) cos_similarity = self.cal_cos_similarity(self.project1(mini_batch_out), self.project2(sample_train_hidden)) row = ( torch.linspace(0, x.shape[0] - 1, x.shape[0]) .reshape([-1, 1]) .repeat(1, n_neighbor) .reshape(1, -1) .to(device) ) column = torch.topk(cos_similarity, n_neighbor, dim=1)[1].reshape(1, -1) mask = torch.sparse_coo_tensor( torch.cat([row, column]), torch.ones([row.shape[1]]).to(device) / n_neighbor, (x.shape[0], sample_train_hidden.shape[0]), ) cos_similarity = self.sparse_dense_mul(mask, cos_similarity) agg_out = torch.sparse.mm(cos_similarity, self.project2(sample_train_hidden)) # out = self.fc_out(out).squeeze() out = self.fc_out_pred(torch.cat([mini_batch_out, agg_out], axis=1)).squeeze() return out