def save_to_mongo(_results, _saved, model_update_time, data_source):
done = True
for key in _results.keys():
result_pair = _results[key]
if result_pair[1].return_value is None:
done = False
logging.warn("Haven't finish yet for key %s, continue..." % key)
continue
else:
if _saved[key] == False:
_saved[key] = True
to_save = (result_pair[0], result_pair[1].return_value, result_pair[2])
region = to_save[0]
for single_hour_prediction in zip(to_save[1], to_save[2]):
p = Prediction()
p.setRegion(region)
p.setModelUpdateTime(model_update_time)
p.setPredictedValues(
float(single_hour_prediction[0][1]), math.sqrt(float(single_hour_prediction[0][2]))
)
p.setTime(str(single_hour_prediction[1]))
p_json = p.toDict()
if data_source == "twitter":
save_interface = PredictionInterface()
save_interface.setDB(TwitterConfig.prediction_db)
save_interface.setCollection(TwitterConfig.prediction_collection)
save_interface.saveDocument(p_json)
elif data_source == "instagram":
save_interface = PredictionInterface()
save_interface.setDB(InstagramConfig.prediction_db)
save_interface.setCollection(InstagramConfig.prediction_collection)
save_interface.saveDocument(p_json)
return done
def save_to_mongo(_results, _saved, model_update_time):
done = True
for key in _results.keys():
result_pair = _results[key]
if result_pair[1].return_value is None:
done = False
continue
else:
if _saved[key] == False:
_saved[key] = True
to_save = (result_pair[0], result_pair[1].return_value, result_pair[2])
region = to_save[0]
for single_hour_prediction in zip(to_save[1], to_save[2]):
p = Prediction()
p.setRegion(region)
p.setModelUpdateTime(model_update_time)
p.setPredictedValues( float(single_hour_prediction[0][1]), math.sqrt(float(single_hour_prediction[0][2])))
p.setTime( str(single_hour_prediction[1]) )
p_json = p.toJSON()
save_interface = PredictionInterface()
save_interface.saveDocument( p_json )
return done
def session_length_performance(prediction: Prediction, dataset: RecSysDataset):
_, score = prediction.get_score()
score.reset_index(level=1, inplace=True)
score.set_index(['session_id'], inplace=True)
performances = {}
for index in range(len(dataset)):
indices = np.array(
dataset.rec_sys_data.groups[dataset.session_ids[index]])
session = dataset.rec_sys_data.session_df.loc[indices]
length = len(session)
session_id = session['session_id'].iloc[0]
if length not in performances:
performances[length] = []
performances[length].append(score.loc[session_id]['score'])
for length in performances.keys():
arr = np.array(performances[length])
performances[length] = (arr.mean(), len(arr))
return performances
def train(config, state=None):
random.seed(42)
config = prepare_config(config)
batch_size = config.get('batch_size')
patience = config.get('patience')
num_epochs = config.get('num_epochs')
learning_rate = config.get('learning_rate')
phases = config.get('phases')
reduce_factor = config.get('reduce_factor')
reduce_patience = config.get('reduce_patience')
weight_decay = config.get('weight_decay')
use_cosine_similarity = config.get('use_cosine_similarity')
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
data = get_rec_sys_data(size=config.get('dataset_size'))
train_dataset = RecSysDataset(
rec_sys_data=data,
split_strategy=RandomSampleStrategy(split=0.7),
include_impressions=False,
train_mode=True)
train_val_dataset = RecSysDataset(rec_sys_data=data,
split_strategy=AllSamplesIncludeStrategy(
include=train_dataset.session_ids),
include_impressions=True)
val_dataset = RecSysDataset(rec_sys_data=data,
split_strategy=AllSamplesExceptStrategy(
exclude=train_dataset.session_ids),
include_impressions=True)
print("Num Train Sessions", len(train_dataset))
print("Num Validation Sessions", len(val_dataset))
recommend_network = RecommenderNetwork(
config=config,
item_size=train_dataset.item_size,
target_item_size=train_dataset.target_item_size,
)
loss_function = nn.CosineEmbeddingLoss()
rc_optimizer = optim.Adam(recommend_network.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
rc_lr_scheduler = ReduceLROnPlateau(rc_optimizer,
mode='max',
factor=reduce_factor,
patience=reduce_patience)
start_epoch = 0
best_score_so_far = None
if state:
rc_optimizer.load_state_dict(state['rc_optimizer_state_dict'])
rc_lr_scheduler.load_state_dict(state['rc_optimizer_state_dict'])
recommend_network.load_state_dict(state['network_state_dict'])
start_epoch = state['epoch']
best_score_so_far = state['best_score_so_far']
datasets = {
"train": train_dataset,
"train_val": train_dataset,
"val": val_dataset,
}
data_loaders = {
"train":
DataLoader(train_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=6,
collate_fn=train_dataset.collator),
"train_val":
DataLoader(train_val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=6,
collate_fn=train_val_dataset.collator),
"val":
DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=6,
collate_fn=val_dataset.collator),
}
sizes = {
"train": int(len(train_dataset) / batch_size) + 1,
"train_val": int(len(train_dataset) / batch_size) + 1,
"val": int(len(val_dataset) / batch_size) + 1,
}
losses = np.zeros(int(len(train_dataset) / batch_size + 1.0))
recommend_network = recommend_network.to(device)
cur_patience = 0
print("Uses CUDA: {0}".format(use_cuda))
for epoch in range(start_epoch, num_epochs):
print('-' * 15, "Epoch: ", epoch + 1, '\t', '-' * 15)
cur_phase = phases
for phase in cur_phase: # 'train_val',
cur_dataset = datasets[phase]
cur_prediction = Prediction(
dataset=cur_dataset,
device=device,
use_cosine_similarity=use_cosine_similarity,
)
do_validation = phase != 'train'
if phase == 'train':
recommend_network.train()
else:
recommend_network.eval()
losses.fill(0)
with progressbar.ProgressBar(max_value=sizes[phase],
redirect_stdout=True) as bar:
for idx, data in enumerate(data_loaders[phase]):
if phase != 'train':
sessions, session_lengths, session_targets, item_impressions, impression_ids, target_index, prices, ids = data
else:
sessions, session_lengths, session_targets = data
impression_ids = None
item_impressions = None
ids = None
target_index = None
prices = None
sessions = sessions.to(device)
session_targets = session_targets.to(device)
if phase == 'train':
rc_optimizer.zero_grad()
with torch.set_grad_enabled(True):
item_scores = recommend_network(
sessions, session_lengths).float()
loss = loss_function(
item_scores, session_targets,
torch.ones(session_targets.size(0)).to(device))
loss.backward()
rc_optimizer.step()
losses[idx] = loss.item()
else:
with torch.set_grad_enabled(False):
item_scores = recommend_network(
sessions, session_lengths).float()
cur_prediction.add_predictions(
ids=ids,
impression_ids=impression_ids,
item_impressions=item_impressions,
item_scores=item_scores,
)
bar.update(idx)
if do_validation:
score, _ = cur_prediction.get_score()
print(phase, " Score: ", score)
rc_lr_scheduler.step(score)
if phase == 'val':
if best_score_so_far is None or score > best_score_so_far:
best_score_so_far = score
torch.save(
{
'epoch':
epoch,
'best_score_so_far':
best_score_so_far,
'rc_optimizer_state_dict':
rc_optimizer.state_dict(),
'network_state_dict':
recommend_network.state_dict(),
'config':
config,
}, MODEL_PATH)
cur_patience = 0
print("New best \\o/")
else:
cur_patience += 1
if cur_patience > patience:
print("Not patient anymore => Quit")
break
if not do_validation:
print(phase, " Loss: ", losses.mean())
if cur_patience > patience:
break
print("Final best model: ", best_score_so_far)
target_path = os.path.join(
MODEL_BASE_PATH, '{}_{}.pth'.format(MODEL_NAME,
round(best_score_so_far, 2)))
shutil.move(MODEL_PATH, target_path)
return best_score_so_far, target_path
def save_to_mongo(_results, _saved, model_update_time):
done = True
for key in _results.keys():
result_pair = _results[key]
if result_pair[1].return_value is None:
done = False
continue
else:
if _saved[key] == False:
_saved[key] = True
to_save = (result_pair[0], result_pair[1].return_value,
result_pair[2])
region = to_save[0]
for single_hour_prediction in zip(to_save[1], to_save[2]):
p = Prediction()
p.setRegion(region)
p.setModelUpdateTime(model_update_time)
p.setPredictedValues(
float(single_hour_prediction[0][1]),
math.sqrt(float(single_hour_prediction[0][2])))
p.setTime(str(single_hour_prediction[1]))
p_json = p.toJSON()
save_interface = PredictionInterface()
save_interface.saveDocument(p_json)
return done
def do_prediction(recommender_state, ranking_state, dataset):
random.seed(42)
from recommender_configs import prepare_config
from utility.prediction import Prediction
from network.recommender_network import RecommenderNetwork
from network.impression_network import ImpressionRankNetwork
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
recommender_state_config = prepare_config(recommender_state.get('config'))
rc_network_state_dict = recommender_state.get('network_state_dict')
ir_state_config = prepare_config(ranking_state.get('config'))
ir_network_state_dict = ranking_state.get('network_state_dict')
recommender_network = RecommenderNetwork(
config=recommender_state_config,
item_size=dataset.item_size,
target_item_size=dataset.target_item_size,
)
recommender_network.load_state_dict(rc_network_state_dict)
recommender_network.eval()
rank_network = ImpressionRankNetwork(
config=ir_state_config,
item_size=dataset.item_size,
device=device,
)
rank_network.load_state_dict(ir_network_state_dict)
rank_network.eval()
data_loader = DataLoader(dataset,
batch_size=SUBMISSION_BATCH_SIZE,
shuffle=True,
num_workers=6,
collate_fn=dataset.collator)
cur_prediction = Prediction(
dataset=dataset,
device=device,
)
recommender_network = recommender_network.to(device)
rank_network = rank_network.to(device)
print("Begin predicting...")
with progressbar.ProgressBar(
max_value=int(len(dataset) / SUBMISSION_BATCH_SIZE + 1),
redirect_stdout=True) as bar:
for idx, data in enumerate(data_loader):
sessions, session_lengths, _, item_impressions, impression_ids, target_index, prices, ids = data
sessions = sessions.to(device)
item_scores = recommender_network(sessions,
session_lengths).float()
selected_impression = rank_network(item_impressions, item_scores,
prices)
cur_prediction.add_predictions(
ids=ids,
impression_ids=impression_ids,
item_impressions=item_impressions,
item_scores=item_scores,
selected_impression=selected_impression)
bar.update(idx)
return cur_prediction
