def f_eval_epoch(self, eval_data, network, optimizer, logger_obj, local_rank):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
self.m_eval_iteration = self.m_train_iteration
logger_obj.f_add_output2IO(" "*10+" eval the user and item encoder"+" "*10)
network.eval()
with torch.no_grad():
for pos_tag_batch, mask_batch, user_batch, item_batch in eval_data:
user_batch_gpu = user_batch.to(self.m_device)
item_batch_gpu = item_batch.to(self.m_device)
preds = self.f_get_pred(network, user_batch_gpu, item_batch_gpu, local_rank)
precision, recall = get_precision_recall(preds.cpu(), pos_tag_batch, mask_batch, k=3)
if precision != 0 and recall != 0:
# loss_list.append(loss.item())
precision_list.append(precision)
recall_list.append(recall)
logger_obj.f_add_output2IO("%d, precision:%.4f, recall:%.4f"%(self.m_eval_iteration, np.mean(precision_list), np.mean(recall_list)))
# logger_obj.f_add_scalar2tensorboard("eval/precision", np.mean(precision_list), self.m_eval_iteration)
# logger_obj.f_add_scalar2tensorboard("eval/recall", np.mean(recall_list), self.m_eval_iteration)
self.m_mean_eval_precision = np.mean(precision_list)
self.m_mean_eval_recall = np.mean(recall_list)
network.train()
def f_eval_epoch(self, eval_data, network, optimizer, logger_obj):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
# self.m_eval_iteration = 0
self.m_eval_iteration = self.m_train_iteration
# logger_obj.f_add_output2IO("--"*20)
logger_obj.f_add_output2IO(" "*10+" eval the user and item encoder"+" "*10)
# logger_obj.f_add_output2IO("--"*20)
network.eval()
with torch.no_grad():
for attr_item_batch, attr_tf_item_batch, attr_length_item_batch, item_batch, attr_user_batch, attr_tf_user_batch, attr_length_user_batch, user_batch, pos_target_batch, pos_length_batch, neg_target_batch, neg_length_batch in eval_data:
attr_item_gpu = attr_item_batch.to(self.m_device)
attr_tf_item_gpu = attr_tf_item_batch.to(self.m_device)
attr_length_item_gpu = attr_length_item_batch.to(self.m_device)
item_gpu = item_batch.to(self.m_device)
attr_user_gpu = attr_user_batch.to(self.m_device)
attr_tf_user_gpu = attr_tf_user_batch.to(self.m_device)
attr_length_user_gpu = attr_length_user_batch.to(self.m_device)
user_gpu = user_batch.to(self.m_device)
pos_target_gpu = pos_target_batch.to(self.m_device)
pos_length_gpu = pos_length_batch.to(self.m_device)
neg_target_gpu = neg_target_batch.to(self.m_device)
neg_length_gpu = neg_length_batch.to(self.m_device)
output = network(attr_item_gpu, attr_tf_item_gpu, attr_length_item_gpu, item_gpu, attr_user_gpu, attr_tf_user_gpu, attr_length_user_gpu, user_gpu)
logits, mask, targets = network.f_pred_forward(output, pos_target_gpu, pos_length_gpu, neg_target_gpu, neg_length_gpu)
NLL_loss = self.m_rec_loss(logits, targets, mask)
loss = NLL_loss
precision, recall = get_precision_recall(logits.cpu(), targets.cpu(), mask.cpu(), k=3)
if precision != 0 and recall != 0:
loss_list.append(loss.item())
precision_list.append(precision)
recall_list.append(recall)
logger_obj.f_add_output2IO("%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f"%(self.m_eval_iteration, np.mean(loss_list), np.mean(precision_list), np.mean(recall_list)))
logger_obj.f_add_scalar2tensorboard("eval/loss", np.mean(loss_list), self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/precision", np.mean(precision_list), self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/recall", np.mean(recall_list), self.m_eval_iteration)
self.m_mean_eval_loss = np.mean(loss_list)
self.m_mean_eval_precision = np.mean(precision_list)
self.m_mean_eval_recall = np.mean(recall_list)
network.train()
def f_eval_new(self, train_data, eval_data):
self.f_get_user_item(train_data, eval_data)
batch_index = 0
precision_list = []
recall_list = []
mrr_list = []
print('--'*10)
# print("user output weight", self.m_network.m_user_output.weight)
# print("item output weight", self.m_network.m_item_output.weight)
self.m_network.eval()
with torch.no_grad():
pop_correct_num_total = 0
non_pop_correct_num_total = 0
pred_num_total = 0
topk = 3
for pos_tag_batch, user_batch, item_batch in eval_data:
user_batch_gpu = user_batch.to(self.m_device)
item_batch_gpu = item_batch.to(self.m_device)
user_batch_gpu = user_batch.to(self.m_device)
item_batch_gpu = item_batch.to(self.m_device)
preds = self.f_get_pred(self.m_network, user_batch_gpu, item_batch_gpu)
precision, recall = get_precision_recall(preds.cpu(), pos_tag_batch, k=topk)
# precision_batch, recall_batch = self.f_eval_bow(user_item_attr_logits, target_logits)
if precision == 0 and recall == 0:
continue
batch_index += 1
precision_list.append(precision)
recall_list.append(recall)
print("precision: ", np.mean(precision_list))
print("recall: ", np.mean(recall_list))
def f_train_epoch(self, train_data, network, optimizer, logger_obj):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
# logger_obj.f_add_output2IO("--"*20)
logger_obj.f_add_output2IO(" "*10+"training the user and item encoder"+" "*10)
# logger_obj.f_add_output2IO("--"*20)
tmp_loss_list = []
tmp_precision_list = []
tmp_recall_list = []
network.train()
for attr_item_batch, attr_tf_item_batch, attr_length_item_batch, item_batch, attr_user_batch, attr_tf_user_batch, attr_length_user_batch, user_batch, pos_target_batch, pos_length_batch, neg_target_batch, neg_length_batch in train_data:
attr_item_gpu = attr_item_batch.to(self.m_device)
attr_tf_item_gpu = attr_tf_item_batch.to(self.m_device)
attr_length_item_gpu = attr_length_item_batch.to(self.m_device)
item_gpu = item_batch.to(self.m_device)
attr_user_gpu = attr_user_batch.to(self.m_device)
attr_tf_user_gpu = attr_tf_user_batch.to(self.m_device)
attr_length_user_gpu = attr_length_user_batch.to(self.m_device)
user_gpu = user_batch.to(self.m_device)
pos_target_gpu = pos_target_batch.to(self.m_device)
pos_length_gpu = pos_length_batch.to(self.m_device)
neg_target_gpu = neg_target_batch.to(self.m_device)
neg_length_gpu = neg_length_batch.to(self.m_device)
output = network(attr_item_gpu, attr_tf_item_gpu, attr_length_item_gpu, item_gpu, attr_user_gpu, attr_tf_user_gpu, attr_length_user_gpu, user_gpu)
logits, mask, targets = network.f_pred_forward(output, pos_target_gpu, pos_length_gpu, neg_target_gpu, neg_length_gpu)
NLL_loss = self.m_rec_loss(logits, targets, mask)
loss = NLL_loss
precision, recall = get_precision_recall(logits.cpu(), targets.cpu(), mask.cpu(), k=3)
# NLL_loss = self.m_rec_loss(user_attr_item_logits, target_gpu, attr_item_mask)
# loss = NLL_loss
# precision, recall = get_precision_recall(user_attr_item_logits.cpu(), target_batch, k=3)
if precision != 0 and recall != 0:
loss_list.append(loss.item())
precision_list.append(precision)
recall_list.append(recall)
tmp_loss_list.append(loss.item())
tmp_precision_list.append(precision)
tmp_recall_list.append(recall)
optimizer.zero_grad()
loss.backward()
optimizer.step()
self.m_train_iteration += 1
iteration += 1
if iteration % self.m_print_interval == 0:
logger_obj.f_add_output2IO("%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f"%(iteration, np.mean(tmp_loss_list), np.mean(tmp_precision_list), np.mean(tmp_recall_list)))
tmp_loss_list = []
tmp_precision_list = []
tmp_recall_list = []
logger_obj.f_add_output2IO("%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f"%(self.m_train_iteration, np.mean(loss_list), np.mean(precision_list), np.mean(recall_list)))
logger_obj.f_add_scalar2tensorboard("train/loss", np.mean(loss_list), self.m_train_iteration)
logger_obj.f_add_scalar2tensorboard("train/precision", np.mean(precision_list), self.m_train_iteration)
logger_obj.f_add_scalar2tensorboard("train/recall", np.mean(recall_list), self.m_train_iteration)
self.m_mean_train_loss = np.mean(loss_list)
self.m_mean_train_precision = np.mean(precision_list)
self.m_mean_train_recall = np.mean(recall_list)
def f_eval_epoch(self, eval_data, network, optimizer, logger_obj):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
# self.m_eval_iteration = 0
self.m_eval_iteration = self.m_train_iteration
# logger_obj.f_add_output2IO("--"*20)
logger_obj.f_add_output2IO(" "*10+" eval the user and item encoder"+" "*10)
# logger_obj.f_add_output2IO("--"*20)
beta = 0.1
network.eval()
with torch.no_grad():
for input_batch, input_freq_batch, input_length_batch, user_batch, item_batch, target_batch in eval_data:
# eval_flag = random.randint(1,5)
# if eval_flag != 2:
# continue
input_batch_gpu = input_batch.to(self.m_device)
input_freq_batch_gpu = input_freq_batch.to(self.m_device)
input_length_batch_gpu = input_length_batch.to(self.m_device)
user_batch_gpu = user_batch.to(self.m_device)
item_batch_gpu = item_batch.to(self.m_device)
target_batch_gpu = target_batch.to(self.m_device)
batch_size = input_batch.size(0)
user_item_attr_logits, mask = network(input_batch_gpu, input_freq_batch_gpu, input_length_batch_gpu, user_batch_gpu, item_batch_gpu)
# target_batch_gpu = torch.gather(target_batch_gpu, 1, input_batch_gpu)
NLL_loss = self.m_rec_loss(user_item_attr_logits, target_batch_gpu, mask)
# NLL_loss = NLL_loss/batch_size
loss = NLL_loss
precision, recall = get_precision_recall(user_item_attr_logits.cpu(), target_batch, k=3)
# print("precision", precision, "recall", recall)
if precision != 0 and recall != 0:
loss_list.append(loss.item())
precision_list.append(precision)
recall_list.append(recall)
logger_obj.f_add_output2IO("%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f"%(self.m_eval_iteration, np.mean(loss_list), np.mean(precision_list), np.mean(recall_list)))
logger_obj.f_add_scalar2tensorboard("eval/loss", np.mean(loss_list), self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/precision", np.mean(precision_list), self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/recall", np.mean(recall_list), self.m_eval_iteration)
self.m_mean_eval_loss = np.mean(loss_list)
self.m_mean_eval_precision = np.mean(precision_list)
self.m_mean_eval_recall = np.mean(recall_list)
network.train()
def f_train_epoch(self, train_data, network, optimizer, logger_obj):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
# logger_obj.f_add_output2IO("--"*20)
logger_obj.f_add_output2IO(" "*10+"training the user and item encoder"+" "*10)
# logger_obj.f_add_output2IO("--"*20)
beta = 0.1
tmp_loss_list = []
tmp_precision_list = []
tmp_recall_list = []
network.train()
for input_batch, input_freq_batch, input_length_batch, user_batch, item_batch, target_batch in train_data:
input_batch_gpu = input_batch.to(self.m_device)
input_freq_batch_gpu = input_freq_batch.to(self.m_device)
input_length_batch_gpu = input_length_batch.to(self.m_device)
user_batch_gpu = user_batch.to(self.m_device)
item_batch_gpu = item_batch.to(self.m_device)
target_batch_gpu = target_batch.to(self.m_device)
batch_size = input_batch.size(0)
# print("+++"*20)
# logits, z, z_mean, z_logvar = network(input_batch_gpu)
user_item_attr_logits, mask = network(input_batch_gpu, input_freq_batch_gpu, input_length_batch_gpu, user_batch_gpu, item_batch_gpu)
NLL_loss = self.m_rec_loss(user_item_attr_logits, target_batch_gpu, mask)
loss = NLL_loss
precision, recall = get_precision_recall(user_item_attr_logits.cpu(), target_batch, k=3)
if precision != 0 and recall != 0:
loss_list.append(loss.item())
precision_list.append(precision)
recall_list.append(recall)
tmp_loss_list.append(loss.item())
tmp_precision_list.append(precision)
tmp_recall_list.append(recall)
# loss = NLL_loss+beta*KL_loss
# print("loss", loss)
optimizer.zero_grad()
loss.backward()
optimizer.step()
self.m_train_iteration += 1
iteration += 1
if iteration % self.m_print_interval == 0:
logger_obj.f_add_output2IO("%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f"%(iteration, np.mean(tmp_loss_list), np.mean(tmp_precision_list), np.mean(tmp_recall_list)))
tmp_loss_list = []
tmp_precision_list = []
tmp_recall_list = []
# logger_obj.f_add_scalar2tensorboard("train/loss", np.mean(loss_list), self.m_train_iteration)
# logger_obj.f_add_scalar2tensorboard("train/precision", np.mean(precision_list), self.m_train_iteration)
# logger_obj.f_add_scalar2tensorboard("train/recall", np.mean(recall_list), self.m_train_iteration)
# loss_list = []
# precision_list = []
# recall_list = []
logger_obj.f_add_output2IO("%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f"%(self.m_train_iteration, np.mean(loss_list), np.mean(precision_list), np.mean(recall_list)))
logger_obj.f_add_scalar2tensorboard("train/loss", np.mean(loss_list), self.m_train_iteration)
logger_obj.f_add_scalar2tensorboard("train/precision", np.mean(precision_list), self.m_train_iteration)
logger_obj.f_add_scalar2tensorboard("train/recall", np.mean(recall_list), self.m_train_iteration)
self.m_mean_train_loss = np.mean(loss_list)
self.m_mean_train_precision = np.mean(precision_list)
self.m_mean_train_recall = np.mean(recall_list)
def f_eval_epoch(self, eval_data, network, optimizer, logger_obj):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
# self.m_eval_iteration = 0
self.m_eval_iteration = self.m_train_iteration
# logger_obj.f_add_output2IO("--"*20)
logger_obj.f_add_output2IO(" " * 10 +
" eval the user and item encoder" +
" " * 10)
# logger_obj.f_add_output2IO("--"*20)
network.eval()
with torch.no_grad():
for attr_item_batch, attr_tf_item_batch, attr_length_item_batch, attr_index_item_batch, attr_index_item_iter_debug_batch, item_batch, attr_user_batch, attr_tf_user_batch, attr_length_user_batch, attr_index_user_batch, user_batch, attr_input_batch, attr_length_batch, target_batch in eval_data:
attr_item_gpu = attr_item_batch.to(self.m_device)
attr_tf_item_gpu = attr_tf_item_batch.to(self.m_device)
attr_length_item_gpu = attr_length_item_batch.to(self.m_device)
attr_index_item_gpu = attr_index_item_batch.to(self.m_device)
item_gpu = item_batch.to(self.m_device)
attr_index_item_iter_debug_gpu = attr_index_item_iter_debug_batch.to(
self.m_device)
attr_user_gpu = attr_user_batch.to(self.m_device)
attr_tf_user_gpu = attr_tf_user_batch.to(self.m_device)
attr_length_user_gpu = attr_length_user_batch.to(self.m_device)
attr_index_user_gpu = attr_index_user_batch.to(self.m_device)
user_gpu = user_batch.to(self.m_device)
attr_input_gpu = attr_input_batch.to(self.m_device)
attr_length_gpu = attr_length_batch.to(self.m_device)
target_gpu = target_batch.to(self.m_device)
logits = torch.zeros_like(target_gpu).float()
# print(attr_input_gpu.masked_select(attr_index_item_iter_debug_gpu.bool()))
# print("logits size", logits.type())
# exit()
# print("=="*10)
# print("target_batch", target_batch)
# print("attr_input_batch", attr_input_batch)
# print("attr_item_batch", attr_item_batch)
# print("attr_user_batch", attr_user_batch)
# print("=="*10)
user_attr_item_logits, attr_item_mask, item_attr_user_logits, attr_user_mask, logits, mask = network(
attr_item_gpu, attr_tf_item_gpu, attr_length_item_gpu,
attr_index_item_gpu, item_gpu, attr_user_gpu,
attr_tf_user_gpu, attr_length_user_gpu,
attr_index_user_gpu, user_gpu, attr_length_gpu, logits)
# NLL_loss = self.m_rec_loss(logits, target_gpu, mask)
# loss = NLL_loss
# precision, recall = get_precision_recall(logits.cpu(), target_batch, mask.cpu(), k=3)
NLL_loss = self.m_rec_loss(logits, target_gpu, mask,
attr_index_item_iter_debug_gpu)
loss = NLL_loss
precision, recall = get_precision_recall(logits.cpu(),
target_batch,
mask.cpu(),
k=3)
if precision != 0 and recall != 0:
loss_list.append(loss.item())
precision_list.append(precision)
recall_list.append(recall)
logger_obj.f_add_output2IO(
"%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f" %
(self.m_eval_iteration, np.mean(loss_list),
np.mean(precision_list), np.mean(recall_list)))
logger_obj.f_add_scalar2tensorboard("eval/loss",
np.mean(loss_list),
self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/precision",
np.mean(precision_list),
self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/recall",
np.mean(recall_list),
self.m_eval_iteration)
self.m_mean_eval_loss = np.mean(loss_list)
self.m_mean_eval_precision = np.mean(precision_list)
self.m_mean_eval_recall = np.mean(recall_list)
network.train()
def f_eval_epoch(self, eval_data, network, optimizer, logger_obj):
loss_list = []
precision_list = []
recall_list = []
iteration = 0
self.m_eval_iteration = self.m_train_iteration
logger_obj.f_add_output2IO(" " * 10 +
" eval the user and item encoder" +
" " * 10)
network.eval()
with torch.no_grad():
for pos_tag_batch, mask_batch, user_batch, item_batch in eval_data:
# eval_flag = random.randint(1,5)
# if eval_flag != 2:
# continue
user_batch_gpu = user_batch.to(self.m_device)
item_batch_gpu = item_batch.to(self.m_device)
# pos_tag_batch_gpu = pos_tag_batch.to(self.m_device)
# logits = network(pos_tag_batch_gpu, , user_batch_gpu, item_batch_gpu)
# NLL_loss = self.m_rec_loss(logits)
# loss = NLL_loss
loss = 0.0
preds = self.f_get_pred(network, user_batch_gpu,
item_batch_gpu)
precision, recall = get_precision_recall(preds.cpu(),
pos_tag_batch,
mask_batch,
k=3)
if precision != 0 and recall != 0:
# loss_list.append(loss.item())
loss_list.append(loss)
precision_list.append(precision)
recall_list.append(recall)
logger_obj.f_add_output2IO(
"%d, NLL_loss:%.4f, precision:%.4f, recall:%.4f" %
(self.m_eval_iteration, np.mean(loss_list),
np.mean(precision_list), np.mean(recall_list)))
logger_obj.f_add_scalar2tensorboard("eval/loss",
np.mean(loss_list),
self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/precision",
np.mean(precision_list),
self.m_eval_iteration)
logger_obj.f_add_scalar2tensorboard("eval/recall",
np.mean(recall_list),
self.m_eval_iteration)
self.m_mean_eval_loss = np.mean(loss_list)
self.m_mean_eval_precision = np.mean(precision_list)
self.m_mean_eval_recall = np.mean(recall_list)
network.train()