# Modified from https://github.com/quark0/darts import os, gc, sys, time, math import numpy as np from copy import deepcopy import torch import torch.nn as nn from utils import print_log, obtain_accuracy, AverageMeter from utils import time_string, convert_secs2time from utils import count_parameters_in_MB from datasets import Corpus from nas_rnn import batchify, get_batch, repackage_hidden from nas_rnn import DARTSCell, RNNModel def obtain_best(accuracies): if len(accuracies) == 0: return (0, 0) tops = [value for key, value in accuracies.items()] s2b = sorted( tops ) return s2b[-1] def main_procedure(config, genotype, save_dir, print_freq, log): print_log('-'*90, log) print_log('save-dir : {:}'.format(save_dir), log) print_log('genotype : {:}'.format(genotype), log) print_log('config : {:}'.format(config), log) corpus = Corpus(config.data_path) train_data = batchify(corpus.train, config.train_batch, True) valid_data = batchify(corpus.valid, config.eval_batch , True) test_data = batchify(corpus.test, config.test_batch , True) ntokens = len(corpus.dictionary) print_log("Train--Data Size : {:}".format(train_data.size()), log) print_log("Valid--Data Size : {:}".format(valid_data.size()), log) print_log("Test---Data Size : {:}".format( test_data.size()), log) print_log("ntokens = {:}".format(ntokens), log) model = RNNModel(ntokens, config.emsize, config.nhid, config.nhidlast, config.dropout, config.dropouth, config.dropoutx, config.dropouti, config.dropoute, cell_cls=DARTSCell, genotype=genotype) model = model.cuda() print_log('Network =>\n{:}'.format(model), log) print_log('Genotype : {:}'.format(genotype), log) print_log('Parameters : {:.3f} MB'.format(count_parameters_in_MB(model)), log) checkpoint_path = os.path.join(save_dir, 'checkpoint-{:}.pth'.format(config.data_name)) Soptimizer = torch.optim.SGD (model.parameters(), lr=config.LR, weight_decay=config.wdecay) Aoptimizer = torch.optim.ASGD(model.parameters(), lr=config.LR, t0=0, lambd=0., weight_decay=config.wdecay) if os.path.isfile(checkpoint_path): checkpoint = torch.load(checkpoint_path) model.load_state_dict( checkpoint['state_dict'] ) Soptimizer.load_state_dict( checkpoint['SGD_optimizer'] ) Aoptimizer.load_state_dict( checkpoint['ASGD_optimizer'] ) epoch = checkpoint['epoch'] use_asgd = checkpoint['use_asgd'] print_log('load checkpoint from {:} and start train from {:}'.format(checkpoint_path, epoch), log) else: epoch, use_asgd = 0, False start_time, epoch_time = time.time(), AverageMeter() valid_loss_from_sgd, losses = [], {-1 : 1e9} while epoch < config.epochs: need_time = convert_secs2time(epoch_time.val * (config.epochs-epoch), True) print_log("\n==>>{:s} [Epoch={:04d}/{:04d}] {:}".format(time_string(), epoch, config.epochs, need_time), log) if use_asgd : optimizer = Aoptimizer else : optimizer = Soptimizer try: Dtime, Btime = train(model, optimizer, corpus, train_data, config, epoch, print_freq, log) except: torch.cuda.empty_cache() checkpoint = torch.load(checkpoint_path) model.load_state_dict( checkpoint['state_dict'] ) Soptimizer.load_state_dict( checkpoint['SGD_optimizer'] ) Aoptimizer.load_state_dict( checkpoint['ASGD_optimizer'] ) epoch = checkpoint['epoch'] use_asgd = checkpoint['use_asgd'] valid_loss_from_sgd = checkpoint['valid_loss_from_sgd'] continue if use_asgd: tmp = {} for prm in model.parameters(): tmp[prm] = prm.data.clone() prm.data = Aoptimizer.state[prm]['ax'].clone() val_loss = evaluate(model, corpus, valid_data, config.eval_batch, config.bptt) for prm in model.parameters(): prm.data = tmp[prm].clone() else: val_loss = evaluate(model, corpus, valid_data, config.eval_batch, config.bptt) if len(valid_loss_from_sgd) > config.nonmono and val_loss > min(valid_loss_from_sgd): use_asgd = True valid_loss_from_sgd.append( val_loss ) print_log('{:} end of epoch {:3d} with {:} | valid loss {:5.2f} | valid ppl {:8.2f}'.format(time_string(), epoch, 'ASGD' if use_asgd else 'SGD', val_loss, math.exp(val_loss)), log) if val_loss < min(losses.values()): if use_asgd: tmp = {} for prm in model.parameters(): tmp[prm] = prm.data.clone() prm.data = Aoptimizer.state[prm]['ax'].clone() torch.save({'epoch' : epoch, 'use_asgd' : use_asgd, 'valid_loss_from_sgd': valid_loss_from_sgd, 'state_dict': model.state_dict(), 'SGD_optimizer' : Soptimizer.state_dict(), 'ASGD_optimizer': Aoptimizer.state_dict()}, checkpoint_path) if use_asgd: for prm in model.parameters(): prm.data = tmp[prm].clone() print_log('save into {:}'.format(checkpoint_path), log) if use_asgd: tmp = {} for prm in model.parameters(): tmp[prm] = prm.data.clone() prm.data = Aoptimizer.state[prm]['ax'].clone() test_loss = evaluate(model, corpus, test_data, config.test_batch, config.bptt) if use_asgd: for prm in model.parameters(): prm.data = tmp[prm].clone() print_log('| epoch={:03d} | test loss {:5.2f} | test ppl {:8.2f}'.format(epoch, test_loss, math.exp(test_loss)), log) losses[epoch] = val_loss epoch = epoch + 1 # measure elapsed time epoch_time.update(time.time() - start_time) start_time = time.time() print_log('--------------------- Finish Training ----------------', log) checkpoint = torch.load(checkpoint_path) model.load_state_dict( checkpoint['state_dict'] ) test_loss = evaluate(model, corpus, test_data , config.test_batch, config.bptt) print_log('| End of training | test loss {:5.2f} | test ppl {:8.2f}'.format(test_loss, math.exp(test_loss)), log) vali_loss = evaluate(model, corpus, valid_data, config.eval_batch, config.bptt) print_log('| End of training | valid loss {:5.2f} | valid ppl {:8.2f}'.format(vali_loss, math.exp(vali_loss)), log) def evaluate(model, corpus, data_source, batch_size, bptt): # Turn on evaluation mode which disables dropout. model.eval() total_loss, total_length = 0.0, 0.0 with torch.no_grad(): ntokens = len(corpus.dictionary) hidden = model.init_hidden(batch_size) for i in range(0, data_source.size(0) - 1, bptt): data, targets = get_batch(data_source, i, bptt) targets = targets.view(-1) log_prob, hidden = model(data, hidden) loss = nn.functional.nll_loss(log_prob.view(-1, log_prob.size(2)), targets) total_loss += loss.item() * len(data) total_length += len(data) hidden = repackage_hidden(hidden) return total_loss / total_length def train(model, optimizer, corpus, train_data, config, epoch, print_freq, log): # Turn on training mode which enables dropout. total_loss, data_time, batch_time = 0, AverageMeter(), AverageMeter() start_time = time.time() ntokens = len(corpus.dictionary) hidden_train = model.init_hidden(config.train_batch) batch, i = 0, 0 while i < train_data.size(0) - 1 - 1: bptt = config.bptt if np.random.random() < 0.95 else config.bptt / 2. # Prevent excessively small or negative sequence lengths seq_len = max(5, int(np.random.normal(bptt, 5))) # There's a very small chance that it could select a very long sequence length resulting in OOM seq_len = min(seq_len, config.bptt + config.max_seq_len_delta) lr2 = optimizer.param_groups[0]['lr'] optimizer.param_groups[0]['lr'] = lr2 * seq_len / config.bptt model.train() data, targets = get_batch(train_data, i, seq_len) targets = targets.contiguous().view(-1) # count data preparation time data_time.update(time.time() - start_time) optimizer.zero_grad() hidden_train = repackage_hidden(hidden_train) log_prob, hidden_train, rnn_hs, dropped_rnn_hs = model(data, hidden_train, return_h=True) raw_loss = nn.functional.nll_loss(log_prob.view(-1, log_prob.size(2)), targets) loss = raw_loss # Activiation Regularization if config.alpha > 0: loss = loss + sum(config.alpha * dropped_rnn_h.pow(2).mean() for dropped_rnn_h in dropped_rnn_hs[-1:]) # Temporal Activation Regularization (slowness) loss = loss + sum(config.beta * (rnn_h[1:] - rnn_h[:-1]).pow(2).mean() for rnn_h in rnn_hs[-1:]) loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), config.clip) optimizer.step() gc.collect() optimizer.param_groups[0]['lr'] = lr2 total_loss += raw_loss.item() assert torch.isnan(loss) == False, '--- Epoch={:04d} :: {:03d}/{:03d} Get Loss = Nan'.format(epoch, batch, len(train_data)//config.bptt) batch_time.update(time.time() - start_time) start_time = time.time() batch, i = batch + 1, i + seq_len if batch % print_freq == 0: cur_loss = total_loss / print_freq print_log(' >> Epoch: {:04d} :: {:03d}/{:03d} || loss = {:5.2f}, ppl = {:8.2f}'.format(epoch, batch, len(train_data) // config.bptt, cur_loss, math.exp(cur_loss)), log) total_loss = 0 return data_time.sum, batch_time.sum