def predict_loss(prev, paths, split_df):
# prev: true value
# paths: paths to the model weights
t_preds = []
for tm in range(3):
tf.reset_default_graph()
t_preds.append(
predict(paths[-1:],
build_hparams(hparams.params_s32),
back_offset=0,
predict_window=288,
n_models=3,
target_model=tm,
seed=2,
batch_size=50,
asgd=True,
split_df=split_df))
preds = sum(t_preds) / 3
preds.index = [idx.decode('ascii') for idx in preds.index]
# mean mae
res = 0
for idx in preds.index:
res += np.abs(preds.loc[idx, :] - prev.loc[idx, -288:]).sum()
res /= len(preds.index) * 288
return preds, res
def predict_loss(paths):
# paths: paths to the model weights
t_preds = []
for tm in range(3):
tf.reset_default_graph()
t_preds.append(predict(paths[-1:], build_hparams(hparams.params_s32),
n_models=3, target_model=tm, seed=5, batch_size=50, asgd=True))
preds=sum(t_preds) /3
return preds
def predict_loss(prev, paths):
# prev: true value
# paths: paths to the model weights
t_preds = []
for tm in range(3):
tf.reset_default_graph()
t_preds.append(
predict(paths[-1:],
build_hparams(hparams.params_s32),
n_models=3,
target_model=tm,
seed=5,
batch_size=50,
asgd=True))
preds = sum(t_preds) / 3
# mean mae
res = 0
for idx in preds.index:
res += np.abs(preds[idx] - prev[idx]) / prev[idx]
res /= 72
return preds, res
type=int,
help=
"Save model on each evaluation (10 evals per epoch), starting from this step"
)
parser.add_argument('--predict_window',
default=63,
type=int,
help="Number of days to predict")
args = parser.parse_args()
#param_dict = dict(vars(args))
#param_dict['hparams'] = build_from_set(args.hparam_set)
#del param_dict['hparam_set']
#train(**param_dict)
hparams = build_hparams()
result = train("definc_attn",
hparams,
n_models=1,
train_sampling=1.0,
eval_sampling=1.0,
patience=5,
multi_gpu=False,
save_best_model=False,
gpu=0,
eval_memsize=15,
seed=5,
verbose=True,
forward_split=False,
write_summaries=True,
side_split=True,
raw_smape = smape(true, pred)
masked_smape = np.ma.array(raw_smape, mask=np.isnan(raw_smape))
return masked_smape.mean()
from make_features import read_all
df_all = read_all()
df_all.columns
prev = df_all
paths = [p for p in tf.train.get_checkpoint_state('data/cpt/s32').all_model_checkpoint_paths]
t_preds = []
for tm in range(3):
tf.reset_default_graph()
t_preds.append(predict(paths, build_hparams(hparams.params_s32), back_offset=0, predict_window=63,
n_models=3, target_model=tm, seed=2, batch_size=2048, asgd=True))
preds=sum(t_preds) /3
missing_pages = prev.index.difference(preds.index)
# Use zeros for missing pages
rmdf = pd.DataFrame(index=missing_pages,
data=np.tile(0, (len(preds.columns),len(missing_pages))).T, columns=preds.columns)
f_preds = preds.append(rmdf).sort_index()
# Use zero for negative predictions
f_preds[f_preds < 0.5] = 0
# Rouns predictions to nearest int
f_preds = np.round(f_preds).astype(np.int64)