class TrainLoop_fusion_gen():
def __init__(self, opt, is_finetune):
self.opt=opt
self.train_dataset=dataset('data/train_data.jsonl',opt)
self.dict=self.train_dataset.word2index
self.index2word={self.dict[key]:key for key in self.dict}
self.batch_size=self.opt['batch_size']
self.epoch=self.opt['epoch']
self.use_cuda=opt['use_cuda']
if opt['load_dict']!=None:
self.load_data=True
else:
self.load_data=False
self.is_finetune=False
self.movie_ids = pkl.load(open("data/movie_ids.pkl", "rb"))
# Note: we cannot change the type of metrics ahead of time, so you
# should correctly initialize to floats or ints here
self.metrics_rec={"recall@1":0,"recall@10":0,"recall@50":0,"loss":0,"count":0}
self.metrics_gen={"dist1":0,"dist2":0,"dist3":0,"dist4":0,"bleu1":0,"bleu2":0,"bleu3":0,"bleu4":0,"count":0}
self.build_model(is_finetune=True)
if opt['load_dict'] is not None:
# load model parameters if available
print('[ Loading existing model params from {} ]'
''.format(opt['load_dict']))
states = self.model.load(opt['load_dict'])
else:
states = {}
self.init_optim(
[p for p in self.model.parameters() if p.requires_grad],
optim_states=states.get('optimizer'),
saved_optim_type=states.get('optimizer_type')
)
def build_model(self,is_finetune):
self.model = CrossModel(self.opt, self.dict, is_finetune)
if self.opt['embedding_type'] != 'random':
pass
if self.use_cuda:
self.model.cuda()
def train(self):
self.model.load_model()
losses=[]
best_val_gen=1000
gen_stop=False
for i in range(self.epoch*3):
train_set=CRSdataset(self.train_dataset.data_process(True),self.opt['n_entity'],self.opt['n_concept'])
train_dataset_loader = torch.utils.data.DataLoader(dataset=train_set,
batch_size=self.batch_size,
shuffle=False)
num=0
for context,c_lengths,response,r_length,mask_response,mask_r_length,entity,entity_vector,movie,concept_mask,dbpedia_mask,concept_vec, db_vec,rec in tqdm(train_dataset_loader):
seed_sets = []
batch_size = context.shape[0]
for b in range(batch_size):
seed_set = entity[b].nonzero().view(-1).tolist()
seed_sets.append(seed_set)
self.model.train()
self.zero_grad()
scores, preds, rec_scores, rec_loss, gen_loss, mask_loss, info_db_loss, info_con_loss=self.model(context.cuda(), response.cuda(), mask_response.cuda(), concept_mask, dbpedia_mask, seed_sets, movie, concept_vec, db_vec, entity_vector.cuda(), rec, test=False)
joint_loss=gen_loss
losses.append([gen_loss])
self.backward(joint_loss)
self.update_params()
if num%50==0:
print('gen loss is %f'%(sum([l[0] for l in losses])/len(losses)))
losses=[]
num+=1
output_metrics_gen = self.val(True)
if best_val_gen < output_metrics_gen["dist4"]:
pass
else:
best_val_gen = output_metrics_gen["dist4"]
self.model.save_model()
print("generator model saved once------------------------------------------------")
_=self.val(is_test=True)
def val(self,is_test=False):
self.metrics_gen={"ppl":0,"dist1":0,"dist2":0,"dist3":0,"dist4":0,"bleu1":0,"bleu2":0,"bleu3":0,"bleu4":0,"count":0}
self.metrics_rec={"recall@1":0,"recall@10":0,"recall@50":0,"loss":0,"gate":0,"count":0,'gate_count':0}
self.model.eval()
if is_test:
val_dataset = dataset('data/test_data.jsonl', self.opt)
else:
val_dataset = dataset('data/valid_data.jsonl', self.opt)
val_set=CRSdataset(val_dataset.data_process(True),self.opt['n_entity'],self.opt['n_concept'])
val_dataset_loader = torch.utils.data.DataLoader(dataset=val_set,
batch_size=self.batch_size,
shuffle=False)
inference_sum=[]
golden_sum=[]
context_sum=[]
losses=[]
recs=[]
for context, c_lengths, response, r_length, mask_response, mask_r_length, entity, entity_vector, movie, concept_mask, dbpedia_mask, concept_vec, db_vec, rec in tqdm(val_dataset_loader):
with torch.no_grad():
seed_sets = []
batch_size = context.shape[0]
for b in range(batch_size):
seed_set = entity[b].nonzero().view(-1).tolist()
seed_sets.append(seed_set)
_, _, _, _, gen_loss, mask_loss, info_db_loss, info_con_loss = self.model(context.cuda(), response.cuda(), mask_response.cuda(), concept_mask, dbpedia_mask, seed_sets, movie, concept_vec, db_vec, entity_vector.cuda(), rec, test=False)
scores, preds, rec_scores, rec_loss, _, mask_loss, info_db_loss, info_con_loss = self.model(context.cuda(), response.cuda(), mask_response.cuda(), concept_mask, dbpedia_mask, seed_sets, movie, concept_vec, db_vec, entity_vector.cuda(), rec, test=True, maxlen=20, bsz=batch_size)
golden_sum.extend(self.vector2sentence(response.cpu()))
inference_sum.extend(self.vector2sentence(preds.cpu()))
context_sum.extend(self.vector2sentence(context.cpu()))
recs.extend(rec.cpu())
losses.append(torch.mean(gen_loss))
#print(losses)
#exit()
self.metrics_cal_gen(losses,inference_sum,golden_sum,recs)
output_dict_gen={}
for key in self.metrics_gen:
if 'bleu' in key:
output_dict_gen[key]=self.metrics_gen[key]/self.metrics_gen['count']
else:
output_dict_gen[key]=self.metrics_gen[key]
print(output_dict_gen)
f=open('context_test.txt','w',encoding='utf-8')
f.writelines([' '.join(sen)+'\n' for sen in context_sum])
f.close()
f=open('output_test.txt','w',encoding='utf-8')
f.writelines([' '.join(sen)+'\n' for sen in inference_sum])
f.close()
return output_dict_gen
def metrics_cal_gen(self,rec_loss,preds,responses,recs):
def bleu_cal(sen1, tar1):
bleu1 = sentence_bleu([tar1], sen1, weights=(1, 0, 0, 0))
bleu2 = sentence_bleu([tar1], sen1, weights=(0, 1, 0, 0))
bleu3 = sentence_bleu([tar1], sen1, weights=(0, 0, 1, 0))
bleu4 = sentence_bleu([tar1], sen1, weights=(0, 0, 0, 1))
return bleu1, bleu2, bleu3, bleu4
def distinct_metrics(outs):
# outputs is a list which contains several sentences, each sentence contains several words
unigram_count = 0
bigram_count = 0
trigram_count=0
quagram_count=0
unigram_set = set()
bigram_set = set()
trigram_set=set()
quagram_set=set()
for sen in outs:
for word in sen:
unigram_count += 1
unigram_set.add(word)
for start in range(len(sen) - 1):
bg = str(sen[start]) + ' ' + str(sen[start + 1])
bigram_count += 1
bigram_set.add(bg)
for start in range(len(sen)-2):
trg=str(sen[start]) + ' ' + str(sen[start + 1]) + ' ' + str(sen[start + 2])
trigram_count+=1
trigram_set.add(trg)
for start in range(len(sen)-3):
quag=str(sen[start]) + ' ' + str(sen[start + 1]) + ' ' + str(sen[start + 2]) + ' ' + str(sen[start + 3])
quagram_count+=1
quagram_set.add(quag)
dis1 = len(unigram_set) / len(outs)#unigram_count
dis2 = len(bigram_set) / len(outs)#bigram_count
dis3 = len(trigram_set)/len(outs)#trigram_count
dis4 = len(quagram_set)/len(outs)#quagram_count
return dis1, dis2, dis3, dis4
predict_s=preds
golden_s=responses
#print(rec_loss[0])
#self.metrics_gen["ppl"]+=sum([exp(ppl) for ppl in rec_loss])/len(rec_loss)
generated=[]
for out, tar, rec in zip(predict_s, golden_s, recs):
bleu1, bleu2, bleu3, bleu4=bleu_cal(out, tar)
generated.append(out)
self.metrics_gen['bleu1']+=bleu1
self.metrics_gen['bleu2']+=bleu2
self.metrics_gen['bleu3']+=bleu3
self.metrics_gen['bleu4']+=bleu4
self.metrics_gen['count']+=1
dis1, dis2, dis3, dis4=distinct_metrics(generated)
self.metrics_gen['dist1']=dis1
self.metrics_gen['dist2']=dis2
self.metrics_gen['dist3']=dis3
self.metrics_gen['dist4']=dis4
def vector2sentence(self,batch_sen):
sentences=[]
for sen in batch_sen.numpy().tolist():
sentence=[]
for word in sen:
if word>3:
sentence.append(self.index2word[word])
elif word==3:
sentence.append('_UNK_')
sentences.append(sentence)
return sentences
@classmethod
def optim_opts(self):
"""
Fetch optimizer selection.
By default, collects everything in torch.optim, as well as importing:
- qhm / qhmadam if installed from github.com/facebookresearch/qhoptim
Override this (and probably call super()) to add your own optimizers.
"""
# first pull torch.optim in
optims = {k.lower(): v for k, v in optim.__dict__.items()
if not k.startswith('__') and k[0].isupper()}
try:
import apex.optimizers.fused_adam as fused_adam
optims['fused_adam'] = fused_adam.FusedAdam
except ImportError:
pass
try:
# https://openreview.net/pdf?id=S1fUpoR5FQ
from qhoptim.pyt import QHM, QHAdam
optims['qhm'] = QHM
optims['qhadam'] = QHAdam
except ImportError:
# no QHM installed
pass
return optims
def init_optim(self, params, optim_states=None, saved_optim_type=None):
"""
Initialize optimizer with model parameters.
:param params:
parameters from the model
:param optim_states:
optional argument providing states of optimizer to load
:param saved_optim_type:
type of optimizer being loaded, if changed will skip loading
optimizer states
"""
opt = self.opt
# set up optimizer args
lr = opt['learningrate']
kwargs = {'lr': lr}
kwargs['amsgrad'] = True
kwargs['betas'] = (0.9, 0.999)
optim_class = self.optim_opts()[opt['optimizer']]
self.optimizer = optim_class(params, **kwargs)
def backward(self, loss):
"""
Perform a backward pass. It is recommended you use this instead of
loss.backward(), for integration with distributed training and FP16
training.
"""
loss.backward()
def update_params(self):
"""
Perform step of optimization, clipping gradients and adjusting LR
schedule if needed. Gradient accumulation is also performed if agent
is called with --update-freq.
It is recommended (but not forced) that you call this in train_step.
"""
update_freq = 1
if update_freq > 1:
# we're doing gradient accumulation, so we don't only want to step
# every N updates instead
self._number_grad_accum = (self._number_grad_accum + 1) % update_freq
if self._number_grad_accum != 0:
return
if self.opt['gradient_clip'] > 0:
torch.nn.utils.clip_grad_norm_(
self.model.parameters(), self.opt['gradient_clip']
)
self.optimizer.step()
def zero_grad(self):
"""
Zero out optimizer.
It is recommended you call this in train_step. It automatically handles
gradient accumulation if agent is called with --update-freq.
"""
self.optimizer.zero_grad()
class IConversationalRecommender():
def __init__(self, opt):
# tensorboard
self.writer = SummaryWriter()
# word from corpus trained by gensim
self.word2index = json.load(
open('data/word2index_redial.json', encoding='utf-8'))
self.index2word = {
self.word2index[key]: key
for key in self.word2index
}
# mapping from movie ids to names
self.id2moviename = pkl.load(open('data/movie_id2name.pkl', 'rb'))
# dbpedia
self.entity2entityId = pkl.load(open('data/entity2entityId.pkl', 'rb'))
self.entityId2entity = dict([
(val, key) for key, val in self.entity2entityId.items()
])
# model
self.opt = opt
self.model = CrossModel(opt, self.word2index, is_finetune=True).cuda()
self.processor = Processor(opt)
# conversation logs
self.logs = []
def visualize_model(self):
sample_sen = 'Sample sentence for model visualization'
context, length, concept_mask, concept_bitmask, dbpedia_mask, dbpedia_bitmask, \
entities, entity_bitmask = self.to_batch_tensor(
*(self.processor.data_process([sample_sen])))
seed_sets = [123]
self.writer.add_graph(
self.model,
(context.cuda(), concept_mask, dbpedia_mask, concept_bitmask,
dbpedia_bitmask, seed_sets, entities.cuda(), 3))
def prompt(self):
self.input = input('KGSF> ').strip()
self.logs.append(self.input)
# TODO add popup of movie items for selection
def to_batch_tensor(self, *args):
args = [torch.from_numpy(arg) for arg in args]
return [torch.unsqueeze(arg, 0) for arg in args]
def vector2sentence(self, sen: List[int]):
sentence_id = []
sentence_name = []
for idx in sen:
try:
if idx >= 3:
# without replace movie id with name
word = '_UNK_' if idx == 3 else self.index2word[idx]
sentence_id.append(word)
# replace id with name
word = self.convert_id_to_name(
word) if word[0] == '@' else word
sentence_name.append(word)
except:
print("OOV", idx)
return ' '.join(w
for w in sentence_name), ' '.join(w
for w in sentence_id)
def convert_id_to_name(self, movieId: str) -> str:
try:
return self.id2moviename[movieId[1:]]
except:
return movieId
def start(self):
self.model.load_model(model_path=self.opt['model_path'])
while True:
self.prompt()
if self.input == '':
print('End of conversation ...')
self.logs = []
continue
elif self.input == 'exit()':
break
# get model input from logs
context, length, concept_mask, concept_bitmask, dbpedia_mask, dbpedia_bitmask, \
entities, entity_bitmask = self.processor.data_process(self.logs)
seed_sets = [entities.nonzero()[0].tolist()]
# inference
self.model.eval()
with torch.no_grad():
context, concept_mask, concept_bitmask, dbpedia_mask, dbpedia_bitmask, entities, entity_bitmask = \
self.to_batch_tensor(context, concept_mask, concept_bitmask,
dbpedia_mask, dbpedia_bitmask, entities, entity_bitmask)
scores, preds, rec_scores, _, _, _, _, _ = self.model(
context.cuda(),
concept_mask,
dbpedia_mask,
concept_bitmask,
dbpedia_bitmask,
seed_sets,
entities.cuda(),
TrainType.INFER,
maxlen=20,
bsz=1)
response_display, response_log = self.vector2sentence(
preds.squeeze().detach().cpu().numpy().tolist())
self.logs.append(response_log)
print("Response> ", response_display)