How can I set RNNBlock's number of units?

[yurimarimo]

Sorry for the second question.

I would like to specify (or tune automatically) the number of units in the RNNBlock.
However, in the RNNBlock, there doesn't seem to be any attribute related to the number of units.

I have tried the following code, but the number of units is 1.

input_node = ak.Input()
gru_node = ak.RNNBlock(bidirectional=False, layer_type="gru")(input_node)
output_node = ak.RegressionHead(output_dim=1, dropout=0.0)(gru_node)
auto_model = ak.AutoModel(
    inputs=input_node, 
    outputs=output_node,
    overwrite=True,
    max_trials=100
)

auto_model.fit(X_train, y_train, batch_size=batch_size, epochs=100, validation_split=0.1)

model = auto_model.export_model()
model.summary()

↓output

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 200, 1)]          0         
_________________________________________________________________
cast_to_float32 (CastToFloat (None, 200, 1)            0         
_________________________________________________________________
gru (GRU)                    (None, 1)                 12        
_________________________________________________________________
dropout (Dropout)            (None, 1)                 0         
_________________________________________________________________
regression_head_1 (Dense)    (None, 1)                 2         
=================================================================
Total params: 14
Trainable params: 14
Non-trainable params: 0
_________________________________________________________________

How can I set the number of units to be specific value (or automatically tuned value), as shown below? (This is a keras description)

model = Sequential()
model.add(Input(shape=(200, 1)))
model.add(GRU(6, return_sequences=True))
model.add(GRU(6))
model.add(Dense(1, activation='linear'))
model.compile(optimizer='adam', loss='mean_squared_error')
model.summary()

↓output

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
gru (GRU)                    (None, 200, 6)            144       
_________________________________________________________________
gru_1 (GRU)                  (None, 6)                 234       
_________________________________________________________________
dense (Dense)                (None, 1)                 7         
=================================================================
Total params: 385
Trainable params: 385
Non-trainable params: 0
_________________________________________________________________

Thank you.

[haifeng-jin]

You are right.
Feel free to submit an issue about this feature.
If more people want it, we will add it.

[yurimarimo]

Thank you for your reply.
After this question, I could solve the problem like below.
However, I'm going to submit an issue in case anyone else is thinking the same thing.

import autokeras as ak
import tensorflow as tf
from tensorflow.keras.layers import Dense, TimeDistributed

class TimeDistributedDenseBlock(ak.Block):
    def build(self, hp, inputs=None):
        input_node = tf.nest.flatten(inputs)[0]
        layer = TimeDistributed(Dense(hp.Int('units', min_value=2, max_value=10)))
        output_node = layer(input_node)
        return output_node

input_node = ak.Input()
dense_node = TimeDistributedDenseBlock()(input_node)
gru_node = ak.RNNBlock(bidirectional=False, layer_type="gru")(dense_node)
output_node = ak.RegressionHead(output_dim=1, dropout=0.0)(gru_node)

auto_model = ak.AutoModel(
    inputs=input_node, 
    outputs=output_node,
    overwrite=True
)

auto_model.fit(x, y)

model = auto_model.export_model()
model.summary()

↓output

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 200, 1)]          0         
_________________________________________________________________
cast_to_float32 (CastToFloat (None, 200, 1)            0         
_________________________________________________________________
time_distributed (TimeDistri (None, 200, 10)           20        
_________________________________________________________________
gru (GRU)                    (None, 200, 10)           660       
_________________________________________________________________
gru_1 (GRU)                  (None, 200, 10)           660       
_________________________________________________________________
gru_2 (GRU)                  (None, 10)                660       
_________________________________________________________________
regression_head_1 (Dense)    (None, 1)                 11        
=================================================================
Total params: 2,011
Trainable params: 2,011
Non-trainable params: 0
_________________________________________________________________