def att_plot(test_set, load_dir, plot_path, use_gpu, max_seq_len, beam_width): """ generate attention alignment plots Args: test_set: test dataset load_dir: model dir use_gpu: on gpu/cpu max_seq_len Returns: """ # check devide print('cuda available: {}'.format(torch.cuda.is_available())) use_gpu = use_gpu and torch.cuda.is_available() # load model # latest_checkpoint_path = Checkpoint.get_latest_checkpoint(load_dir) # latest_checkpoint_path = Checkpoint.get_thirdlast_checkpoint(load_dir) latest_checkpoint_path = load_dir resume_checkpoint = Checkpoint.load(latest_checkpoint_path) model = resume_checkpoint.model.to(device) print('Model dir: {}'.format(latest_checkpoint_path)) print('Model laoded') # reset batch_size: model.reset_max_seq_len(max_seq_len) model.reset_use_gpu(use_gpu) model.reset_batch_size(test_set.batch_size) # in plotting mode always turn off beam search model.set_beam_width(beam_width=0) model.check_var('ptr_net') print('max seq len {}'.format(model.max_seq_len)) print('ptr_net {}'.format(model.ptr_net)) # load test if type(test_set.attkey_path) == type(None): test_batches, vocab_size = test_set.construct_batches(is_train=False) else: test_batches, vocab_size = test_set.construct_batches_with_ddfd_prob(is_train=False) # start eval model.eval() match = 0 total = 0 count = 0 with torch.no_grad(): for batch in test_batches: src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] tgt_ids = batch['tgt_word_ids'] tgt_lengths = batch['tgt_sentence_lengths'] src_probs = None if 'src_ddfd_probs' in batch: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, use_gpu).unsqueeze(2) src_ids = _convert_to_tensor(src_ids, use_gpu) tgt_ids = _convert_to_tensor(tgt_ids, use_gpu) decoder_outputs, decoder_hidden, other = model(src_ids, tgt_ids, is_training=False, att_key_feats=src_probs, beam_width=0) # Evaluation # default batch_size = 1 # attention: 31 * [1 x 1 x 32] ( tgt_len(query_len) * [ batch_size x 1 x src_len(key_len)] ) attention = other['attention_score'] seqlist = other['sequence'] # traverse over time not batch bsize = test_set.batch_size max_seq = test_set.max_seq_len vocab_size = len(test_set.tgt_word2id) for idx in range(len(decoder_outputs)): # loop over max_seq step = idx step_output = decoder_outputs[idx] # 64 x vocab_size # count correct target = tgt_ids[:, step+1] non_padding = target.ne(PAD) correct = seqlist[step].view(-1).eq(target).masked_select(non_padding).sum().item() match += correct total += non_padding.sum().item() # Print sentence by sentence srcwords = _convert_to_words_batchfirst(src_ids, test_set.src_id2word) refwords = _convert_to_words_batchfirst(tgt_ids[:,1:], test_set.tgt_id2word) seqwords = _convert_to_words(seqlist, test_set.tgt_id2word) # print(type(attention)) # print(len(attention)) # print(type(attention[0])) # print(attention[0].size()) # input('...') n_q = len(attention) n_k = attention[0].size(2) b_size = attention[0].size(0) att_score = torch.empty(n_q, n_k, dtype=torch.float) # att_score = np.empty([n_q, n_k]) for i in range(len(seqwords)): # loop over sentences outline_src = ' '.join(srcwords[i]) outline_ref = ' '.join(refwords[i]) outline_gen = ' '.join(seqwords[i]) print('SRC: {}'.format(outline_src)) print('REF: {}'.format(outline_ref)) print('GEN: {}'.format(outline_gen)) for j in range(len(attention)): # i: idx of batch # j: idx of query gen = seqwords[i][j] ref = refwords[i][j] att = attention[j][i] # record att scores att_score[j] = att # print('REF:GEN - {}:{}'.format(ref,gen)) # print('{}th ATT size: {}'.format(j, attention[j][i].size())) # print(att) # print(torch.argmax(att)) # print(sum(sum(att))) # input('Press enter to continue ...') # plotting # print(att_score) loc_eos_k = srcwords[i].index('</s>') + 1 loc_eos_q = seqwords[i].index('</s>') + 1 loc_eos_ref = refwords[i].index('</s>') + 1 print('eos_k: {}, eos_q: {}'.format(loc_eos_k, loc_eos_q)) att_score_trim = att_score[:loc_eos_q, :loc_eos_k] # each row (each query) sum up to 1 print(att_score_trim) print('\n') # import pdb; pdb.set_trace() choice = input('Plot or not ? - y/n\n') if choice: if choice.lower()[0] == 'y': print('plotting ...') plot_dir = os.path.join(plot_path, '{}.png'.format(count)) src = srcwords[i][:loc_eos_k] hyp = seqwords[i][:loc_eos_q] ref = refwords[i][:loc_eos_ref] # x-axis: src; y-axis: hyp # plot_alignment(att_score_trim.numpy(), plot_dir, src=src, hyp=hyp, ref=ref) plot_alignment(att_score_trim.numpy(), plot_dir, src=src, hyp=hyp, ref=None) # no ref count += 1 input('Press enter to continue ...') if total == 0: accuracy = float('nan') else: accuracy = match / total print(saccuracy)
def translate(test_set, load_dir, test_path_out, use_gpu, max_seq_len, beam_width, seqrev=False): """ no reference tgt given - Run translation. Args: test_set: test dataset src, tgt using the same dir test_path_out: output dir load_dir: model dir use_gpu: on gpu/cpu """ # load model # latest_checkpoint_path = Checkpoint.get_latest_checkpoint(load_dir) # latest_checkpoint_path = Checkpoint.get_thirdlast_checkpoint(load_dir) latest_checkpoint_path = load_dir resume_checkpoint = Checkpoint.load(latest_checkpoint_path) model = resume_checkpoint.model.to(device) print('Model dir: {}'.format(latest_checkpoint_path)) print('Model laoded') # reset batch_size: model.reset_max_seq_len(max_seq_len) model.reset_use_gpu(use_gpu) model.reset_batch_size(test_set.batch_size) model.check_var('ptr_net') print('max seq len {}'.format(model.max_seq_len)) sys.stdout.flush() # load test if type(test_set.attkey_path) == type(None): test_batches, vocab_size = test_set.construct_batches(is_train=False) else: test_batches, vocab_size = test_set.construct_batches_with_ddfd_prob(is_train=False) # f = open(os.path.join(test_path_out, 'translate.txt'), 'w') -> use proper encoding with open(os.path.join(test_path_out, 'translate.txt'), 'w', encoding="utf8") as f: model.eval() match = 0 total = 0 with torch.no_grad(): for batch in test_batches: src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] src_probs = None if 'src_ddfd_probs' in batch: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, use_gpu).unsqueeze(2) src_ids = _convert_to_tensor(src_ids, use_gpu) decoder_outputs, decoder_hidden, other = model(src=src_ids, is_training=False, att_key_feats=src_probs, beam_width=beam_width) # memory usage mem_kb, mem_mb, mem_gb = get_memory_alloc() mem_mb = round(mem_mb, 2) print('Memory used: {0:.2f} MB'.format(mem_mb)) # write to file seqlist = other['sequence'] seqwords = _convert_to_words(seqlist, test_set.src_id2word) for i in range(len(seqwords)): # skip padding sentences in batch (num_sent % batch_size != 0) if src_lengths[i] == 0: continue words = [] for word in seqwords[i]: if word == '<pad>': continue elif word == '</s>': break else: words.append(word) if len(words) == 0: outline = '' else: if seqrev: words = words[::-1] outline = ' '.join(words) f.write('{}\n'.format(outline)) # if i == 0: # print(outline) sys.stdout.flush()
def evaluate(test_set, load_dir, test_path_out, use_gpu, max_seq_len, beam_width, seqrev=False): """ with reference tgt given - Run translation. Args: test_set: test dataset test_path_out: output dir load_dir: model dir use_gpu: on gpu/cpu Returns: accuracy (excluding PAD tokens) """ # load model # latest_checkpoint_path = Checkpoint.get_latest_checkpoint(load_dir) # latest_checkpoint_path = Checkpoint.get_thirdlast_checkpoint(load_dir) latest_checkpoint_path = load_dir resume_checkpoint = Checkpoint.load(latest_checkpoint_path) model = resume_checkpoint.model.to(device) print('Model dir: {}'.format(latest_checkpoint_path)) print('Model laoded') # reset batch_size: model.reset_max_seq_len(max_seq_len) model.reset_use_gpu(use_gpu) model.reset_batch_size(test_set.batch_size) model.set_beam_width(beam_width) model.check_var('ptr_net') print('max seq len {}'.format(model.max_seq_len)) sys.stdout.flush() # load test if type(test_set.attkey_path) == type(None): test_batches, vocab_size = test_set.construct_batches(is_train=False) else: test_batches, vocab_size = test_set.construct_batches_with_ddfd_prob(is_train=False) # f = open(os.path.join(test_path_out, 'test.txt'), 'w') with open(os.path.join(test_path_out, 'translate.txt'), 'w', encoding="utf8") as f: model.eval() match = 0 total = 0 with torch.no_grad(): for batch in test_batches: src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] tgt_ids = batch['tgt_word_ids'] tgt_lengths = batch['tgt_sentence_lengths'] src_probs = None if 'src_ddfd_probs' in batch: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, use_gpu).unsqueeze(2) src_ids = _convert_to_tensor(src_ids, use_gpu) tgt_ids = _convert_to_tensor(tgt_ids, use_gpu) decoder_outputs, decoder_hidden, other = model(src_ids, tgt_ids, is_training=False, att_key_feats=src_probs, beam_width=beam_width) # Evaluation seqlist = other['sequence'] # traverse over time not batch if beam_width > 1: full_seqlist = other['topk_sequence'] decoder_outputs = decoder_outputs[:-1] for step, step_output in enumerate(decoder_outputs): target = tgt_ids[:, step+1] non_padding = target.ne(PAD) correct = seqlist[step].view(-1).eq(target) .masked_select(non_padding).sum().item() match += correct total += non_padding.sum().item() # write to file seqwords = _convert_to_words(seqlist, test_set.tgt_id2word) for i in range(len(seqwords)): # skip padding sentences in batch (num_sent % batch_size != 0) if src_lengths[i] == 0: continue words = [] for word in seqwords[i]: if word == '<pad>': continue elif word == '</s>': break else: words.append(word) if len(words) == 0: outline = '' else: if seqrev: words = words[::-1] outline = ' '.join(words) f.write('{}\n'.format(outline)) if i == 0: print(outline) sys.stdout.flush() if total == 0: accuracy = float('nan') else: accuracy = match / total
def translate(test_set, load_dir, test_path_out, use_gpu, max_seq_len, beam_width, mode='gec', seqrev=False): """ no reference tgt given - Run translation. Args: test_set: test dataset src, tgt using the same dir test_path_out: output dir load_dir: model dir use_gpu: on gpu/cpu """ # load model # latest_checkpoint_path = Checkpoint.get_latest_checkpoint(load_dir) # latest_checkpoint_path = Checkpoint.get_thirdlast_checkpoint(load_dir) latest_checkpoint_path = load_dir resume_checkpoint = Checkpoint.load(latest_checkpoint_path) model = resume_checkpoint.model print('Model dir: {}'.format(latest_checkpoint_path)) print('Model laoded') # reset batch_size: model.reset_max_seq_len(max_seq_len) model.reset_use_gpu(use_gpu) model.reset_batch_size(test_set.batch_size) # fix compatibility model.check_classvar('shared_embed') model.check_classvar('additional_key_size') model.check_classvar('gec_num_bilstm_dec') model.check_classvar('num_unilstm_enc') model.check_classvar('residual') model.check_classvar('add_discriminator') if model.ptr_net == 'none': model.ptr_net = 'null' model.to(device) print('max seq len {}'.format(model.max_seq_len)) sys.stdout.flush() # load test print('--- constrcut gec test set ---') if type(test_set.tsv_path) == type(None): test_batches, vocab_size = test_set.construct_batches(is_train=False) else: test_batches, vocab_size = test_set.construct_batches_with_ddfd_prob(is_train=False) with open(os.path.join(test_path_out, 'translate.txt'), 'w', encoding="utf8") as f: model.eval() match = 0 total = 0 with torch.no_grad(): for batch in test_batches: src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] src_probs = None if 'src_ddfd_probs' in batch and model.dd_additional_key_size > 0: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, use_gpu).unsqueeze(2) src_ids = _convert_to_tensor(src_ids, use_gpu) tgt_ids = None if mode.lower() == 'gec' and beam_width <= 1: gec_dd_decoder_outputs, gec_dd_dec_hidden, gec_dd_ret_dict, \ decoder_outputs, dec_hidden, ret_dict = \ model.gec_eval(src_ids, tgt_ids, is_training=False, gec_dd_att_key_feats=src_probs, beam_width=beam_width) elif mode.lower() == 'gec' and beam_width > 1: decoder_outputs, dec_hidden, ret_dict = \ model.gec_eval(src_ids, tgt_ids, is_training=False, gec_dd_att_key_feats=src_probs, beam_width=beam_width) elif mode.lower() == 'dd': decoder_outputs, dec_hidden, ret_dict = \ model.dd_eval(src_ids, tgt_ids, is_training=False, dd_att_key_feats=src_probs, beam_width=beam_width) else: assert False, 'Unrecognised eval mode - choose from gec/dd' # memory usage mem_kb, mem_mb, mem_gb = get_memory_alloc() mem_mb = round(mem_mb, 2) print('Memory used: {0:.2f} MB'.format(mem_mb)) # gec output write to file # import pdb; pdb.set_trace() seqlist = ret_dict['sequence'] seqwords = _convert_to_words(seqlist, test_set.src_id2word) for i in range(len(seqwords)): # skip padding sentences in batch (num_sent % batch_size != 0) if src_lengths[i] == 0: continue words = [] for word in seqwords[i]: if word == '<pad>': continue elif word == '</s>': break else: words.append(word) if len(words) == 0: outline = '' else: if seqrev: words = words[::-1] outline = ' '.join(words) f.write('{}\n'.format(outline)) # if i == 0: # print(outline) sys.stdout.flush()
def debug_beam_search(test_set, load_dir, use_gpu, max_seq_len, beam_width): """ with reference tgt given - debug beam search. Args: test_set: test dataset load_dir: model dir use_gpu: on gpu/cpu Returns: accuracy (excluding PAD tokens) """ # load model # latest_checkpoint_path = Checkpoint.get_latest_checkpoint(load_dir) # latest_checkpoint_path = Checkpoint.get_thirdlast_checkpoint(load_dir) latest_checkpoint_path = load_dir resume_checkpoint = Checkpoint.load(latest_checkpoint_path) model = resume_checkpoint.model print('Model dir: {}'.format(latest_checkpoint_path)) print('Model laoded') # reset batch_size: model.reset_max_seq_len(max_seq_len) model.reset_use_gpu(use_gpu) model.reset_batch_size(test_set.batch_size) print('max seq len {}'.format(model.max_seq_len)) sys.stdout.flush() # load test if type(test_set.attkey_path) == type(None): test_batches, vocab_size = test_set.construct_batches(is_train=False) else: test_batches, vocab_size = test_set.construct_batches_with_ddfd_prob( is_train=False) model.eval() match = 0 total = 0 with torch.no_grad(): for batch in test_batches: src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] tgt_ids = batch['tgt_word_ids'] tgt_lengths = batch['tgt_sentence_lengths'] src_probs = None if 'src_ddfd_probs' in batch: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, use_gpu).unsqueeze(2) src_ids = _convert_to_tensor(src_ids, use_gpu) tgt_ids = _convert_to_tensor(tgt_ids, use_gpu) decoder_outputs, decoder_hidden, other = model( src_ids, tgt_ids, is_training=False, att_key_feats=src_probs, beam_width=beam_width) # Evaluation seqlist = other['sequence'] # traverse over time not batch if beam_width > 1: # print('dict:sequence') # print(len(seqlist)) # print(seqlist[0].size()) full_seqlist = other['topk_sequence'] # print('dict:topk_sequence') # print(len(full_seqlist)) # print((full_seqlist[0]).size()) # input('...') seqlists = [] for i in range(beam_width): seqlists.append([seq[:, i] for seq in full_seqlist]) # print(decoder_outputs[0].size()) # print('tgt id size {}'.format(tgt_ids.size())) # input('...') decoder_outputs = decoder_outputs[:-1] # print(len(decoder_outputs)) for step, step_output in enumerate( decoder_outputs): # loop over time steps target = tgt_ids[:, step + 1] non_padding = target.ne(PAD) # print('step', step) # print('target', target) # print('hyp', seqlist[step]) # if beam_width > 1: # print('full_seqlist', full_seqlist[step]) # input('...') correct = seqlist[step].view(-1).eq(target).masked_select( non_padding).sum().item() match += correct total += non_padding.sum().item() # write to file refwords = _convert_to_words_batchfirst(tgt_ids[:, 1:], test_set.tgt_id2word) seqwords = _convert_to_words(seqlist, test_set.tgt_id2word) seqwords_list = [] for i in range(beam_width): seqwords_list.append( _convert_to_words(seqlists[i], test_set.tgt_id2word)) for i in range(len(seqwords)): outline_ref = ' '.join(refwords[i]) print('REF', outline_ref) outline_hyp = ' '.join(seqwords[i]) # print(outline_hyp) outline_hyps = [] for j in range(beam_width): outline_hyps.append(' '.join(seqwords_list[j][i])) print('{}th'.format(j), outline_hyps[-1]) # skip padding sentences in batch (num_sent % batch_size != 0) # if src_lengths[i] == 0: # continue # words = [] # for word in seqwords[i]: # if word == '<pad>': # continue # elif word == '</s>': # break # else: # words.append(word) # if len(words) == 0: # outline = '' # else: # outline = ' '.join(words) input('...') sys.stdout.flush() if total == 0: accuracy = float('nan') else: accuracy = match / total return accuracy
def _train_epoches(self, train_set, model, n_epochs, start_epoch, start_step, dev_set=None): log = self.logger print_loss_total = 0 # Reset every print_every epoch_loss_total = 0 # Reset every epoch att_print_loss_total = 0 # Reset every print_every att_epoch_loss_total = 0 # Reset every epoch attcls_print_loss_total = 0 # Reset every print_every attcls_epoch_loss_total = 0 # Reset every epoch step = start_step step_elapsed = 0 prev_acc = 0.0 count_no_improve = 0 count_num_rollback = 0 ckpt = None # ******************** [loop over epochs] ******************** for epoch in range(start_epoch, n_epochs + 1): for param_group in self.optimizer.optimizer.param_groups: print('epoch:{} lr: {}'.format(epoch, param_group['lr'])) lr_curr = param_group['lr'] # ----------construct batches----------- # allow re-shuffling of data if type(train_set.attkey_path) == type(None): print('--- construct train set ---') train_batches, vocab_size = train_set.construct_batches( is_train=True) if dev_set is not None: print('--- construct dev set ---') dev_batches, vocab_size = dev_set.construct_batches( is_train=False) else: print('--- construct train set ---') train_batches, vocab_size = train_set.construct_batches_with_ddfd_prob( is_train=True) if dev_set is not None: print('--- construct dev set ---') assert type(dev_set.attkey_path) != type( None), 'Dev set missing ddfd probabilities' dev_batches, vocab_size = dev_set.construct_batches_with_ddfd_prob( is_train=False) # --------print info for each epoch---------- steps_per_epoch = len(train_batches) total_steps = steps_per_epoch * n_epochs log.info("steps_per_epoch {}".format(steps_per_epoch)) log.info("total_steps {}".format(total_steps)) log.debug( " ----------------- Epoch: %d, Step: %d -----------------" % (epoch, step)) mem_kb, mem_mb, mem_gb = get_memory_alloc() mem_mb = round(mem_mb, 2) print('Memory used: {0:.2f} MB'.format(mem_mb)) self.writer.add_scalar('Memory_MB', mem_mb, global_step=step) sys.stdout.flush() # ******************** [loop over batches] ******************** model.train(True) for batch in train_batches: # update macro count step += 1 step_elapsed += 1 # load data src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] tgt_ids = batch['tgt_word_ids'] tgt_lengths = batch['tgt_sentence_lengths'] src_probs = None src_labs = None if 'src_ddfd_probs' in batch and model.additional_key_size > 0: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, self.use_gpu).unsqueeze(2) if 'src_ddfd_labs' in batch: src_labs = batch['src_ddfd_labs'] src_labs = _convert_to_tensor(src_labs, self.use_gpu).unsqueeze(2) # sanity check src-tgt pair if step == 1: print('--- Check src tgt pair ---') log_msgs = check_srctgt(src_ids, tgt_ids, train_set.src_id2word, train_set.tgt_id2word) for log_msg in log_msgs: sys.stdout.buffer.write(log_msg) # convert variable to tensor src_ids = _convert_to_tensor(src_ids, self.use_gpu) tgt_ids = _convert_to_tensor(tgt_ids, self.use_gpu) # Get loss loss, att_loss, attcls_loss = self._train_batch( src_ids, tgt_ids, model, step, total_steps, src_probs=src_probs, src_labs=src_labs) print_loss_total += loss epoch_loss_total += loss att_print_loss_total += att_loss att_epoch_loss_total += att_loss attcls_print_loss_total += attcls_loss attcls_epoch_loss_total += attcls_loss if step % self.print_every == 0 and step_elapsed > self.print_every: print_loss_avg = print_loss_total / self.print_every att_print_loss_avg = att_print_loss_total / self.print_every attcls_print_loss_avg = attcls_print_loss_total / self.print_every print_loss_total = 0 att_print_loss_total = 0 attcls_print_loss_total = 0 log_msg = 'Progress: %d%%, Train nlll: %.4f, att: %.4f, attcls: %.4f' % ( step / total_steps * 100, print_loss_avg, att_print_loss_avg, attcls_print_loss_avg) # print(log_msg) log.info(log_msg) self.writer.add_scalar('train_loss', print_loss_avg, global_step=step) self.writer.add_scalar('att_train_loss', att_print_loss_avg, global_step=step) self.writer.add_scalar('attcls_train_loss', attcls_print_loss_avg, global_step=step) # Checkpoint if step % self.checkpoint_every == 0 or step == total_steps: # save criteria if dev_set is not None: dev_loss, accuracy, dev_attlosses = \ self._evaluate_batches(model, dev_batches, dev_set) dev_attloss = dev_attlosses['att_loss'] dev_attclsloss = dev_attlosses['attcls_loss'] log_msg = 'Progress: %d%%, Dev loss: %.4f, accuracy: %.4f, att: %.4f, attcls: %.4f' % ( step / total_steps * 100, dev_loss, accuracy, dev_attloss, dev_attclsloss) log.info(log_msg) self.writer.add_scalar('dev_loss', dev_loss, global_step=step) self.writer.add_scalar('dev_acc', accuracy, global_step=step) self.writer.add_scalar('att_dev_loss', dev_attloss, global_step=step) self.writer.add_scalar('attcls_dev_loss', dev_attclsloss, global_step=step) # save if prev_acc < accuracy: # save best model ckpt = Checkpoint(model=model, optimizer=self.optimizer, epoch=epoch, step=step, input_vocab=train_set.vocab_src, output_vocab=train_set.vocab_tgt) saved_path = ckpt.save(self.expt_dir) print('saving at {} ... '.format(saved_path)) # reset prev_acc = accuracy count_no_improve = 0 count_num_rollback = 0 else: count_no_improve += 1 # roll back if count_no_improve > MAX_COUNT_NO_IMPROVE: # resuming latest_checkpoint_path = Checkpoint.get_latest_checkpoint( self.expt_dir) if type(latest_checkpoint_path) != type(None): resume_checkpoint = Checkpoint.load( latest_checkpoint_path) print( 'epoch:{} step: {} - rolling back {} ...'. format(epoch, step, latest_checkpoint_path)) model = resume_checkpoint.model self.optimizer = resume_checkpoint.optimizer # A walk around to set optimizing parameters properly resume_optim = self.optimizer.optimizer defaults = resume_optim.param_groups[0] defaults.pop('params', None) defaults.pop('initial_lr', None) self.optimizer.optimizer = resume_optim\ .__class__(model.parameters(), **defaults) # start_epoch = resume_checkpoint.epoch # step = resume_checkpoint.step # reset count_no_improve = 0 count_num_rollback += 1 # update learning rate if count_num_rollback > MAX_COUNT_NUM_ROLLBACK: # roll back latest_checkpoint_path = Checkpoint.get_latest_checkpoint( self.expt_dir) if type(latest_checkpoint_path) != type(None): resume_checkpoint = Checkpoint.load( latest_checkpoint_path) print( 'epoch:{} step: {} - rolling back {} ...'. format(epoch, step, latest_checkpoint_path)) model = resume_checkpoint.model self.optimizer = resume_checkpoint.optimizer # A walk around to set optimizing parameters properly resume_optim = self.optimizer.optimizer defaults = resume_optim.param_groups[0] defaults.pop('params', None) defaults.pop('initial_lr', None) self.optimizer.optimizer = resume_optim\ .__class__(model.parameters(), **defaults) start_epoch = resume_checkpoint.epoch step = resume_checkpoint.step # decrease lr for param_group in self.optimizer.optimizer.param_groups: param_group['lr'] *= 0.5 lr_curr = param_group['lr'] print('reducing lr ...') print('step:{} - lr: {}'.format( step, param_group['lr'])) # check early stop if lr_curr < 0.000125: print('early stop ...') break # reset count_no_improve = 0 count_num_rollback = 0 model.train(mode=True) if ckpt is None: ckpt = Checkpoint(model=model, optimizer=self.optimizer, epoch=epoch, step=step, input_vocab=train_set.vocab_src, output_vocab=train_set.vocab_tgt) saved_path = ckpt.save(self.expt_dir) ckpt.rm_old(self.expt_dir, keep_num=KEEP_NUM) print('n_no_improve {}, num_rollback {}'.format( count_no_improve, count_num_rollback)) sys.stdout.flush() else: if dev_set is None: # save every epoch if no dev_set ckpt = Checkpoint(model=model, optimizer=self.optimizer, epoch=epoch, step=step, input_vocab=train_set.vocab_src, output_vocab=train_set.vocab_tgt) # saved_path = ckpt.save(self.expt_dir) saved_path = ckpt.save_epoch(self.expt_dir, epoch) print('saving at {} ... '.format(saved_path)) continue else: continue # break nested for loop break if step_elapsed == 0: continue epoch_loss_avg = epoch_loss_total / min(steps_per_epoch, step - start_step) epoch_loss_total = 0 log_msg = "Finished epoch %d: Train %s: %.4f" % ( epoch, self.loss.name, epoch_loss_avg) log.info('\n') log.info(log_msg)
def _evaluate_batches(self, model, batches, dataset): model.eval() loss = NLLLoss() loss.reset() match = 0 total = 0 out_count = 0 with torch.no_grad(): for batch in batches: src_ids = batch['src_word_ids'] src_lengths = batch['src_sentence_lengths'] tgt_ids = batch['tgt_word_ids'] tgt_lengths = batch['tgt_sentence_lengths'] src_probs = None if 'src_ddfd_probs' in batch and model.additional_key_size > 0: src_probs = batch['src_ddfd_probs'] src_probs = _convert_to_tensor(src_probs, self.use_gpu).unsqueeze(2) src_labs = None if 'src_ddfd_labs' in batch: src_labs = batch['src_ddfd_labs'] src_labs = _convert_to_tensor(src_labs, self.use_gpu).unsqueeze(2) src_ids = _convert_to_tensor(src_ids, self.use_gpu) tgt_ids = _convert_to_tensor(tgt_ids, self.use_gpu) non_padding_mask_tgt = tgt_ids.data.ne(PAD) non_padding_mask_src = src_ids.data.ne(PAD) decoder_outputs, decoder_hidden, other = model( src_ids, tgt_ids, is_training=False, att_key_feats=src_probs) # Evaluation logps = torch.stack(decoder_outputs, dim=1).to(device=device) if not self.eval_with_mask: loss.eval_batch(logps.reshape(-1, logps.size(-1)), tgt_ids[:, 1:].reshape(-1)) else: loss.eval_batch_with_mask( logps.reshape(-1, logps.size(-1)), tgt_ids[:, 1:].reshape(-1), non_padding_mask_tgt[:, 1:].reshape(-1)) seqlist = other['sequence'] seqres = torch.stack(seqlist, dim=1).to(device=device) correct = seqres.view(-1).eq(tgt_ids[:,1:].reshape(-1))\ .masked_select(non_padding_mask_tgt[:,1:].reshape(-1)).sum().item() match += correct total += non_padding_mask_tgt[:, 1:].sum().item() if not self.eval_with_mask: loss.norm_term = 1.0 * tgt_ids.size( 0) * tgt_ids[:, 1:].size(1) else: loss.norm_term = 1.0 * torch.sum(non_padding_mask_tgt[:, 1:]) loss.normalise() if out_count < 3: srcwords = _convert_to_words_batchfirst( src_ids, dataset.tgt_id2word) refwords = _convert_to_words_batchfirst( tgt_ids[:, 1:], dataset.tgt_id2word) seqwords = _convert_to_words(seqlist, dataset.tgt_id2word) outsrc = 'SRC: {}\n'.format(' '.join( srcwords[0])).encode('utf-8') outref = 'REF: {}\n'.format(' '.join( refwords[0])).encode('utf-8') outline = 'GEN: {}\n'.format(' '.join( seqwords[0])).encode('utf-8') sys.stdout.buffer.write(outsrc) sys.stdout.buffer.write(outref) sys.stdout.buffer.write(outline) out_count += 1 att_resloss = 0 attcls_resloss = 0 resloss = loss.get_loss() if total == 0: accuracy = float('nan') else: accuracy = match / total torch.cuda.empty_cache() losses = {} losses['att_loss'] = att_resloss losses['attcls_loss'] = attcls_resloss return resloss, accuracy, losses