Saving Custom Estimators In Tensorflow

I am trying to save a custom estimator after training, but always receive an error. I am using TensorFlow v.1.4, and have tried various solutions I could search on the web and in t

Solution 1:

Make sure to include export_outputs in your model_fn function when the mode is Predict.

defsimple_rnn(features, labels, mode, params):

    # 0. Reformat input shape to become a sequence
    x = tf.split(features[TIMESERIES_COL], N_INPUTS, 1)
    #print 'x={}'.format(x)# 1. configure the RNN
    lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(LSTM_SIZE, forget_bias=1.0)
    outputs, _ = tf.nn.static_rnn(lstm_cell, x, dtype=tf.float32)

    # slice to keep only the last cell of the RNN
    outputs = outputs[-1]
    #print 'last outputs={}'.format(outputs)# output is result of linear activation of last layer of RNN
    weight = tf.Variable(tf.random_normal([LSTM_SIZE, N_OUTPUTS]))
    bias = tf.Variable(tf.random_normal([N_OUTPUTS]))
    predictions = tf.matmul(outputs, weight) + bias

    # 2. loss function, training/eval opsif mode == tf.contrib.learn.ModeKeys.TRAIN or mode == tf.contrib.learn.ModeKeys.EVAL:
        loss = tf.losses.mean_squared_error(labels, predictions)
        optimizer = tf.train.GradientDescentOptimizer(learning_rate=params["l_rate"])
        train_op = optimizer.minimize(loss=loss, global_step=tf.train.get_global_step())
        eval_metric_ops = {"rmse": tf.metrics.root_mean_squared_error(labels, predictions)}
        return  tf.estimator.EstimatorSpec(
            mode=mode,
            loss=loss,
            train_op=train_op,
            eval_metric_ops=eval_metric_ops)

    else:
        loss = None
        train_op = None
        eval_metric_ops = None# 3. Create predictions

        export_outputs = {'predict_output': tf.estimator.export.PredictOutput({"pred_output_classes": predictions, 'probabilities': #your probabilities})}
        predictions_dict = {"predicted": predictions}
        # 4. return ModelFnOpsreturn tf.estimator.EstimatorSpec(
            mode=mode,
            predictions=predictions_dict,
            loss=loss,
            train_op=train_op,
            eval_metric_ops=eval_metric_ops,export_outputs=export_outputs )

Solution 2:

When exporting a graph, there is a required export_outputs field in the EstimatorSpec. See the model_fn documentation for details.

I'd also note that tf.contrib.timeseries has some of this boilerplate written for you (including an RNN example).