how to train neural network mixture model with tfp

[isaaccs]

Hello, I am trying to write a simple gaussian mixture model using tfp. Basically I want to implement

an MDN with tensorflow probability, but i have a mistake

`from keras.models import Sequential

from keras.layers.core import Dense, Dropout from keras.callbacks import History from keras.layers.recurrent import LSTM from keras.models import model_from_json

c = 1 #The number of outputs we want to predict m = 2 #The number of distributions we want to use in the mixture

from keras import backend as K

i=tf.keras.layers.Input(shape=((14,40))) x=tf.keras.layers.LSTM(64, return_sequences=True)(i)

x=Dropout(0.5)(x)

x=tf.keras.layers.LSTM(64, return_sequences=True)(x)

x=Dropout(0.5)(x)

x=Dropout(0.5)(x)

x=tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(64))(x) x=tf.keras.layers.AveragePooling1D()(x) x=tf.keras.layers.Flatten()(x) mu = tf.keras.layers.Dense(int(c*m), name='mdn_mus')(x)#(self.inputs) sigma = tf.keras.layers.Dense(m, activation=K.softplus, name='mdn_sigmas')(x)#(self.inputs) xi = tf.keras.layers.Dense(m, activation=K.softplus, name='mdn_xi')(x)#(self.inputs) pi = tf.keras.layers.Dense(m, activation=K.softmax, name='mdn_pi')(x)#(self.inputs)

mdn_out = tf.keras.layers.concatenate([mu, sigma,xi, pi], name='mdn_out')

mdn_out=Dense(1, activation=K.softmax)(mdn_out)

model = tf.keras.Model(inputs=i, outputs=mdn_out ) print(model.summary()) optim = tf.keras.optimizers.Adam(lr=0.01) model.compile(optimizer=optim, loss=mdn_loss)

model.fit(X_unrolled_train, y_train, epochs=5, batch_size=64, shuffle=False, validation_data=(X_unrolled_test, y_test))`

and the loss function

def mdn_loss(y_true, y_pred): out_mu, out_sigma,out_xi, out_pi = tf.split(y_pred, num_or_size_splits=[m, m, m,m],axis=-1, name='mdn_coef_split') mus = tf.split(out_mu, num_or_size_splits=m, axis=1) sigs = tf.split(out_sigma, num_or_size_splits=m, axis=1) xis = tf.split(out_xi, num_or_size_splits=m, axis=1) cat = tfd.Categorical(logits=out_pi) coll = [tfd.Normal(loc=loc, scale=scale) for loc, scale in zip(mus, sigs)] mixture = tfd.Mixture(cat=cat, components=coll) loss = mixture.log_prob(y_true) loss = tf.negative(loss) loss = tf.reduce_mean(loss) return loss

i have this error

ValueError: components[0] batch shape must be compatible with cat shape >and other component batch shapes

[isaaccs]

i change my loss and it dosent work

`def mdn_loss(y_true, y_pred):

out_mu, out_sigma,out_xi, out_pi = tf.split(y_pred, num_or_size_splits=[m, m,
m,m],axis=-1, name='mdn_coef_split')
mus = tf.split(out_mu, num_or_size_splits=m, axis=1)
sigs = tf.split(out_sigma, num_or_size_splits=m, axis=1)
xis = tf.split(out_xi, num_or_size_splits=m, axis=1)
pis = tf.split(out_pi, num_or_size_splits=m, axis=1)

pis_mean=tf.reduce_mean(pis, 1)
xis_mean=tf.reduce_mean(xis, 1)
sigs_mean=tf.reduce_mean(sigs, 1)
mus_mean=tf.reduce_mean(mus, 1)

cat = tfd.Categorical(logits=[tf.gather(pis_mean, 0),tf.gather(pis_mean, 1)])

coll = [tfd.Normal(loc=tf.gather(mus_mean, 0), scale=tf.gather(sigs_mean, 0)),
    tfd.GeneralizedPareto(loc=tf.gather(mus_mean, 0), scale=tf.gather(sigs_mean, 1), concentration=tf.gather(xis_mean, 1))]

mixture = tfd.Mixture(cat=cat, components=coll)

loss = mixture.log_prob(y_true)
loss = tf.negative(loss)
loss = tf.reduce_mean(loss)  

return loss`

[Strategy24]

I did like this p = keras.layers.Dense(2, activation='softmax')(x)

and then categorical = tensorflow_probability.distributions.Categorical(probs=tensorflow.reshape(p, (-1, 1, 2)))

for some reason Categorical needs tensor with extra dimension