my loss looks extremely unusual with my own data,

[nuass]

my dataset categories number is 631, the program logs display that loss is even smaller than 20news ,Is my setting improper? I set config file like this : {

"training_path": "ens_cluster_datasets/xx/train_X.npy" ,
"ground_truth_path" : "ens_cluster_datasets/xx/train_Y.npy"  ,
"n_epochs_per_iter":5 ,
"n_epochs_per_iter0":5 ,
"batch_size":2048 ,
"n_iter":10  ,
"noise_std" : 0.2 ,
"mi_loss_w_init":0.3 ,
"ladder_loss_w_init":1 ,
"noise_kl_w_init":1 ,
"last_layer_bn":false ,
"n_class" : 631 ,
"fc_layer_sizes":[  1200  , 1200 , 631 ] ,
"thresh":0

}

the program log :

Epoch 1/5 2021-02-26 09:48:40.609729: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0

• 10s - loss: -7.1115e-01 Epoch 2/5
• 6s - loss: -2.6099e-01 Epoch 3/5
• 7s - loss: 0.0490 Epoch 4/5
• 7s - loss: 0.2618 Epoch 5/5
• 6s - loss: 0.4247 /root/anaconda3/envs/zcctf/lib/python3.6/site-packages/keras/engine/training.py:297: UserWarning: Discrepancy between trainable weights and collected trainable weights, did you set model.trainable without calling model.compile after ? 'Discrepancy between trainable weights and collected trainable' Epoch 1/5
• 9s - loss: -7.6037e-01 Epoch 2/5
• 7s - loss: -1.7202e-01 Epoch 3/5
• 6s - loss: 0.1432 Epoch 4/5
• 6s - loss: 0.3263 Epoch 5/5
• 6s - loss: 0.4878 /root/anaconda3/envs/zcctf/lib/python3.6/site-packages/keras/engine/training.py:297: UserWarning: Discrepancy between trainable weights and collected trainable weights, did you set model.trainable without calling model.compile after ? 'Discrepancy between trainable weights and collected trainable' Epoch 1/5
• 10s - loss: -7.5107e-01 Epoch 2/5
• 7s - loss: -1.5359e-01 Epoch 3/5
• 7s - loss: 0.1047 Epoch 4/5
• 6s - loss: 0.3391 Epoch 5/5
• 6s - loss: 0.4869