the loss of unsupervised training could not decline about deepcluster HOT 2 CLOSED

Hi,
The classification loss you're observing is the one at the first iteration of the epoch ([0/32]). The final prediction layer (top_layer) is reinitialized at the beginning of each epoch so it makes sense that at the very beginning of the epoch (ie first iteration) the loss is high.
You should look the loss at the end of the epoch (which actually decreases from 2.133 to 1.752 in your example).
Hope that helps

from deepcluster.

Hi,
The classification loss you're observing is the one at the first iteration of the epoch ([0/32]). The final prediction layer (top_layer) is reinitialized at the beginning of each epoch so it makes sense that at the very beginning of the epoch (ie first iteration) the loss is high.
You should look the loss at the end of the epoch (which actually decreases from 2.133 to 1.752 in your example).
Hope that helps

thanks

from deepcluster.