Like half year ago, I trained this model in a A100. Now, it seems easy to train less params. So my question is as I have checked generator code, there no change at all. But when I load the old model to new, it seems different.
I list the difference of two models.
Is there any suggestion to transfer to a new model efficiently?
old one:
0 | discriminator_loss | NowcastingLoss | 0
1 | grid_regularizer | GridCellLoss | 0
2 | conditioning_stack | ContextConditioningStack | 4.2 M
3 | latent_stack | LatentConditioningStack | 7.2 M
4 | sampler | Sampler | 42.1 M
5 | generator | Generator | 53.6 M
6 | discriminator | Discriminator | 44.7 M
98.3 M Trainable params
0 Non-trainable params
98.3 M Total params
393.086 Total estimated model params size (MB)
Like half year ago, I trained this model in a A100. Now, it seems easy to train less params. So my question is as I have checked generator code, there no change at all. But when I load the old model to new, it seems different. I list the difference of two models.
Is there any suggestion to transfer to a new model efficiently?
old one: 0 | discriminator_loss | NowcastingLoss | 0
1 | grid_regularizer | GridCellLoss | 0
2 | conditioning_stack | ContextConditioningStack | 4.2 M 3 | latent_stack | LatentConditioningStack | 7.2 M 4 | sampler | Sampler | 42.1 M 5 | generator | Generator | 53.6 M 6 | discriminator | Discriminator | 44.7 M
98.3 M Trainable params 0 Non-trainable params 98.3 M Total params 393.086 Total estimated model params size (MB)
new one:
| Name | Type | Params
0 | discriminator_loss | NowcastingLoss | 0
1 | grid_regularizer | GridCellLoss | 0
2 | conditioning_stack | ContextConditioningStack | 1.1 M 3 | latent_stack | LatentConditioningStack | 1.8 M 4 | sampler | Sampler | 10.5 M 5 | generator | Generator | 13.4 M 6 | discriminator | Discriminator | 44.7 M
58.1 M Trainable params 0 Non-trainable params 58.1 M Total params 232.417 Total estimated model params size (MB)
Possible Implementation