Val Real and Fake loss are much lower might imply it already learn what is real and what is fake. what is the disadvantage of this?
without sgd dx goes to 0
never stop the training
batch norm shoudlnt be in the gen out and critic in -> after applying this still very bad and LeakyRelu on critic coldnt make it better
0.0000, 70.5655, 70.3995, 111.1400, 61.1991, 77.6169, 113.1387, 46.8771, 39.3540, 52.4166, 65.7305, 46.6447, 71.7455, 94.6454, 45.0927, 72.9331, 93.1769
45.55371856689453+ 64.40406799316406+ 54.48160171508789+ 60.053070068359375+ 59.9969367980957+ 47.60749435424805+ 158.171630859375+ 83.24748229980469+ 121.00409698486328+ 59.45954132080078+ 67.34489440917969+ 62.718833923339844+ 45.82394027709961+ 61.519256591796875+ 43.1134033203125
Per action results
Start validation epoch 2d_2_3d Validation: Loss: 0.0708 ReCon: 0.0042 KLD: 1.7719 critic_loss: 1.2320 gen_loss: 7.1504 D_x: 0.6343 D_G_z1: 0.0769 D_G_z2: 0.0769 2d_2_3d - * MPJPE * : 74.5449 per joint tensor([ 0.0000, 77.1236, 81.1871, 121.5773, 68.6131, 84.4415, 126.7168, 50.5929, 43.1049, 54.2732, 73.1803, 49.2222, 80.5940, 109.1315, 49.7943, 83.8382, 113.8718])
[52.382102966308594, 70.56153869628906, 68.28710174560547, 64.94527435302734, 69.23880004882812, 57.96500778198242, 146.91574096679688, 93.15471649169922, 132.41453552246094, 68.19825744628906, 76.4754409790039, 71.19129943847656, 52.16413116455078, 69.61522674560547, 51.443939208984375]