Open TamuraMasayuki opened 1 year ago
Batch sizeが小さいときは最初から小さいので気づきにくいが、値が大きいまま0に近づかない
latent loss(KLダイバージェンス)が大きい。つまり平均と分散が0、1に収束していっていない。
なぜ?
[x] ともりさんのデータでlossが下がるか確認する
[x] データ1つで学習
潜在変数の話
連続的に形が変わると思っている
やらずにpganだけやってもいいけど、
オンライン学習
形状1つでやって、もう一つでやった時にどう変わるか
学習のさせ方(データの数)を研究する
潜在変数の評価にもつながる
dropoutをonにする
青木さんの修論
最後の推論もdropoutをonにする
outputがばらつく→ばらつきが大きければ
あるラベルと潜在変数
平均値を持っておいて、
結局は信頼性が高いか低いかを知りたいだけ
# labelt_dim + class_num
self.model = nn.Sequential(
*block(latent_dim+CLASS_NUM, 64, normalize=False),
*block(64, 128, dropout=0.2),
*block(128, 256, dropout=0.2),
*block(256, 512, dropout=0.2),
*block(512, 1024, dropout=0.2),
nn.Linear(1024, coord_size),
nn.Tanh()
)
decoder.load_state_dict(torch.load(G_PATH, map_location=torch.device('cpu')))
RuntimeError: Error(s) in loading state_dict for Decoder:
Missing key(s) in state_dict: "model.7.weight", "model.7.bias", "model.7.running_mean", "model.7.running_var", "model.10.weight", "model.10.bias", "model.11.running_mean", "model.11.running_var", "model.15.weight", "model.15.bias", "model.15.running_mean", "model.15.running_var", "model.18.weight", "model.18.bias".
Unexpected key(s) in state_dict: "model.5.weight", "model.5.bias", "model.6.running_mean", "model.6.running_var", "model.6.num_batches_tracked", "model.8.weight", "model.8.bias", "model.9.weight", "model.9.bias", "model.9.running_mean", "model.9.running_var", "model.9.num_batches_tracked", "model.12.weight", "model.12.bias", "model.12.running_mean", "model.12.running_var", "model.12.num_batches_tracked".
size mismatch for model.6.weight: copying a param with shape torch.Size([256]) from checkpoint, the shape in current model is torch.Size([256, 128]).
size mismatch for model.11.weight: copying a param with shape torch.Size([1024, 512]) from checkpoint, the shape in current model is torch.Size([512]).
size mismatch for model.11.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]).
size mismatch for model.14.weight: copying a param with shape torch.Size([292, 1024]) from checkpoint, the shape in current model is torch.Size([1024, 512]).
size mismatch for model.14.bias: copying a param with shape torch.Size([292]) from checkpoint, the shape in current model is torch.Size([1024]).
ToDo