Generative Loss is negative

[wkgreat]

I use t2vec to train ads-b data, the dataset is from opensky, download link is opensky
the date of dataset I used is from 2022-05-02 to 2022-06-27.
The hyper-parameters are:

{
  "region": {
    "cityname": "opensky",
    "minlon": 73.49901302691623,
    "minlat": 3.83788851813507,
    "maxlon": 135.08737696307114,
    "maxlat": 53.5616571938264,
    "cellsize": 100.0,
    "minfreq": 20
  }
}

args are:

Namespace(data='E:/codes/github/t2vec/data', checkpoint='E:/codes/github/t2vec/data/checkpoint.pt', prefix='opensky', pretrained_embedding=None, num_layers=3, bidirectional=True, hidden_size=256, embedding_size=256, dropout=0.2, max_grad_norm=5.0, learning_rate=0.001, batch=128, generator_batch=32, t2vec_batch=256, start_iteration=0, epochs=15, print_freq=50, save_freq=1000, cuda=True, use_discriminative=False, discriminative_w=0.1, criterion_name='KLDIV', knearestvocabs='data/opensky-vocab-dist-cell100.h5', dist_decay_speed=0.8, max_num_line=20000000, max_length=200, mode=0, vocab_size=20000, bucketsize=[(20, 30), (30, 30), (30, 50), (50, 50), (50, 70), (70, 70), (70, 100), (100, 100)])

training logs are:

Iteration: 0    Generative Loss: 2.175  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 50   Generative Loss: 1.196  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 100  Generative Loss: 1.089  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 150  Generative Loss: 0.731  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 200  Generative Loss: 0.528  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 250  Generative Loss: 0.376  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 300  Generative Loss: 0.278  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 350  Generative Loss: 0.226  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 400  Generative Loss: 0.195  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 450  Generative Loss: 0.092  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 500  Generative Loss: 0.077  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 550  Generative Loss: 0.077  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 600  Generative Loss: 0.076  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 650  Generative Loss: 0.047  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 700  Generative Loss: 0.052  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 750  Generative Loss: -0.023 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 800  Generative Loss: -0.016 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 850  Generative Loss: -0.010 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 900  Generative Loss: -0.031 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 950  Generative Loss: -0.023 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1000 Generative Loss: -0.021 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Saving the model at iteration 1000 validation loss 76.77668204471983
Iteration: 1050 Generative Loss: -0.057 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1100 Generative Loss: -0.052 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1150 Generative Loss: -0.050 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1200 Generative Loss: -0.056 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1250 Generative Loss: -0.060 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1300 Generative Loss: -0.081 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1350 Generative Loss: -0.056 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1400 Generative Loss: 0.070  Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1450 Generative Loss: -0.063 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1500 Generative Loss: -0.068 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1550 Generative Loss: -0.095 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1600 Generative Loss: -0.076 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1650 Generative Loss: -0.073 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1700 Generative Loss: -0.002 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1750 Generative Loss: -0.066 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1800 Generative Loss: -0.069 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1850 Generative Loss: -0.073 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1900 Generative Loss: -0.069 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 1950 Generative Loss: -0.073 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2000 Generative Loss: -0.100 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Saving the model at iteration 2000 validation loss 81.64155062971444
Iteration: 2050 Generative Loss: -0.077 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2100 Generative Loss: -0.069 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2150 Generative Loss: -0.072 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2200 Generative Loss: -0.079 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2250 Generative Loss: -0.074 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2300 Generative Loss: -0.073 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2350 Generative Loss: -0.076 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2400 Generative Loss: -0.071 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2450 Generative Loss: -0.070 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2500 Generative Loss: -0.088 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2550 Generative Loss: -0.104 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2600 Generative Loss: -0.091 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2650 Generative Loss: -0.075 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2700 Generative Loss: -0.083 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2750 Generative Loss: -0.063 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2800 Generative Loss: -0.069 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2850 Generative Loss: -0.071 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2900 Generative Loss: -0.075 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 2950 Generative Loss: -0.072 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Iteration: 3000 Generative Loss: -0.100 Discriminative Cross Loss: 0.000        Discriminative Inner Loss: 0.000
Saving the model at iteration 3000 validation loss 85.59408001077587

As we can see, from iteration 750, the generative loss becomes negative, and validation loss ascendes.
How can explains this result?