A question about the LOSS function in CRLI models

[PBemmm]

In the tensorflow source code of paper, there are five LOSS functions (loss_d, loss_G, loss_pre, loss_re, loss_km).

#loss_D
loss_D = tf.nn.sigmoid_cross_entropy_with_logits(logits = disc_reshaped,labels = mask_inputs)
loss_D = tf.multiply(loss_D,length_mark)
loss_D = tf.reduce_sum(tf.reduce_mean(loss_D,0))
#loss_G
loss_G = tf.nn.sigmoid_cross_entropy_with_logits(logits = disc_reshaped,labels = 1-mask_inputs)
loss_G = tf.multiply(loss_G,length_mark)
loss_G = tf.multiply(loss_G,1-mask_inputs)
loss_G = tf.reduce_sum(tf.reduce_mean(loss_G,0))
#loss_pre
pre_tmp = tf.multiply(impute_vector,masks)
targ_pre = tf.multiply(masks,corpt_inputs)
loss_pre = mse_error(pre_tmp,targ_pre)
#loss_re
output_trim = decoder_outputs[:,:-1,:]
out_tmp = tf.multiply(output_trim,masks)
targ_re = tf.multiply(corpt_inputs,masks)
loss_re = mse_error(out_tmp,targ_re)
#loss_km
HTH = tf.matmul(latent_rep,tf.transpose(latent_rep))
F_copy = tf.get_variable('F', shape=[config.batch_size, config.k_cluster],
initializer=tf.orthogonal_initializer(gain=1.0, seed=None, dtype=tf.float32),
dtype=tf.float32,
trainable=False
)

But I can't seem to find anything about loss_km in pypots.crli

[WenjieDu]

PyPOTS/pypots/clustering/crli/modules/core.py

Line 106 in eb02440

l_kmeans = torch.trace(HTH) - torch.trace(FTHTHF) # k-means loss

[PBemmm]

I got it, thank you

[WenjieDu]

@PBemmm No problem. You can follow me on GitHub to receive our latest open-source research. We have interesting work about clustering to be published this year ;-)