batch_size=len(idx2song),
pl_length=dp_train.out_size - 1,
feat_dim=model.FEAT_DIM,
)
# initialize network
my_net = Network(net)
# load previously fit parameters
params_file = os.path.join(out_dir, '{}_params.pkl'.format(model.name))
my_net.load(file_path=params_file)
# shape withheld continuations for evaluation
_, Y_cont = shape_data(playlists_idx,
songs_idx,
idx2song,
features,
subset=cont_idx)
# songs in the "query" playlists need to be masked to make sure that
# they are not recommended as continuations
_, Y_query = shape_data(playlists_idx,
songs_idx,
idx2song,
features,
subset=query_idx)
# get num of song occurrences when model was fit for cold-start analysis
# Y_fit = Y_query
train_occ = np.asarray(Y_query.sum(axis=1)).flatten()
# provide data information
show_data_splits(playlists_idx, songs_idx, idx2song, song2artist,
train_idx, valid_idx, fit_idx, query_idx, cont_idx)
# provide model information
print('\nNetwork:')
show_design(model)
if args.train:
#
# train the hybrid model while validating on withheld playlists
#
# prepare input song features and playlist targets at training
X_train, Y_train = shape_data(
playlists_idx, songs_idx, idx2song, features,
mode='train', subset=train_idx
)
# prepare input song features and playlist targets at validation
X_valid, Y_valid = shape_data(
playlists_idx, songs_idx, idx2song, features,
mode='test', subset=valid_idx
)
# preprocess input features if required
# use the training song features to standardize the validation data
if model.standardize:
scaler = prep.RobustScaler()
X_train = scaler.fit_transform(X_train)
X_valid = scaler.transform(X_valid)
query_idx = np.intersect1d(test_idx_dsj, train_idx_cnt)
cont_idx = np.intersect1d(test_idx_dsj, test_idx_cnt)
# provide data information
show_data_splits(playlists_idx, songs_idx, idx2song, song2artist,
train_idx, valid_idx, fit_idx, query_idx, cont_idx)
#
# extend the playlists in the query split and evaluate the
# continuations by comparing them to actual withheld continuations
#
# prepare song-playlist matrix in test continuations
_, Y_cont = shape_data(playlists_idx,
songs_idx,
idx2song=None,
features=None,
subset=cont_idx)
# prepare song-playlist matrix in test queries
# used to mask songs from queries
_, Y_query = shape_data(playlists_idx,
songs_idx,
idx2song=None,
features=None,
subset=query_idx)
# prepare song-playlist matrix in fit playlists
_, Y_fit = shape_data(playlists_idx,
songs_idx,
idx2song=None,
print('\nMatrix Factorization:')
show_design(model)
if args.train:
#
# train the factorization while validating on withheld playlists
# for tuning purposes, one fold is enough
#
if fold > 0:
print('\nI won\'t train anymore. Just did it for one fold.')
break
# prepare song-playlist matrix in training
_, Y_train = shape_data(
playlists_idx, songs_idx, idx2song=None, features=None,
subset=train_idx
)
# prepare song-playlist matrix in validation
_, Y_valid = shape_data(
playlists_idx, songs_idx, idx2song=None, features=None,
subset=valid_idx
)
# train the model
train(
model=model,
train_target=Y_train,
valid_target=Y_valid,
out_dir=out_dir,
use_gpu=args.use_gpu