tfp.sts.forcast only applies to future data. Is it possible to "predict" training data?

[HuangKY]

I am using the Structural Time Series of TensorFlow Probability. I separated my data(1 year) into training data (10 months) and test data (2 months).

After building the model and forecasting, I wondered whether it is possible to "predict" the training data so that I can compare the performance of the model in the training data interval. Since the arguments of "tfp.sts.forcast" can only specify "num_steps_forecast" which does not include that option, I would ask here.

[jeffpollock9]

@HuangKY if you run forward_filter the resulting structure has (predicted/forecasted) observation means and covariances, is that what you are after? For example:

import tensorflow as tf
import tensorflow_probability as tfp

print(tf.__version__)
# 2.18.0

print(tfp.__version__)
# 0.25.0

sts = tfp.sts
tfd = tfp.distributions

num_timesteps = 10
param_vals = [1.0, 2.0]
observations = tf.ones([num_timesteps, 1])

local_level = sts.LocalLevel()

model = sts.Sum(components=[local_level])

ssm = model.make_state_space_model(num_timesteps=num_timesteps, param_vals=param_vals)

results = ssm.forward_filter(observations)

tf.concat(
    [
        observations,
        results.observation_means,
        results.observation_covs[..., 0],
    ],
    axis=-1,
)
# <tf.Tensor: shape=(10, 3), dtype=float32, numpy=
# array([[1.        , 0.        , 1.9999999 ],
#        [1.        , 0.5       , 5.5       ],
#        [1.        , 0.9090909 , 5.818182  ],
#        [1.        , 0.984375  , 5.828125  ],
#        [1.        , 0.99731904, 5.8284183 ],
#        [1.        , 0.99954003, 5.8284264 ],
#        [1.        , 0.9999211 , 5.828428  ],
#        [1.        , 0.99998647, 5.828428  ],
#        [1.        , 0.9999977 , 5.828428  ],
#        [1.        , 0.9999996 , 5.828428  ]], dtype=float32)>