Add Pretrained models.

Add pretrained models.

Add configs.

Restructure and adapt configs.

Add paths to aws.

Fix bug in checkpointer.

Add epoch_idx.

Work on pretrained models.

Work on pretrained models.

Use amazon file.

.gitignore

@@ -0,0 +1,5 @@
+output
+data
+*_lmdb
+__pycache__
+*.pyc

README.md

@@ -22,7 +22,7 @@ python train.py PATH_TO_CONFIG
 
 To compute the inception score for your model and generate samples, use
 ```
-python test.py PATH_TO_CONIFG
+python test.py PATH_TO_CONFIG
 ```
 
 Finally, you can create nice latent space interpolations using
@@ -34,8 +34,30 @@ or
 python interpolate_class.py PATH_TO_CONFIG
 ```
 
+# Pretrained models
+We also provide several pretrained models.
+
+You can use the models for sampling by entering
+```
+python test.py PATH_TO_CONFIG
+```
+where `PATH_TO_CONFIG` is one of the config files
+```
+configs/pretrained/celebA_pretrained.yaml
+configs/pretrained/celebAHQ_pretrained.yaml
+configs/pretrained/imagenet_pretrained.yaml
+configs/pretrained/lsun_bedroom_pretrained.yaml
+configs/pretrained/lsun_bridge_pretrained.yaml
+configs/pretrained/lsun_church_pretrained.yaml
+configs/pretrained/lsun_tower_pretrained.yaml
+```
+Our script will automatically download the model checkpoints and run the generation.
+You can find the outputs in the `output/pretrained` folders.
+Similarly, you can use the scripts `interpolate.py` and `interpolate_class.py` for generating interpolations for the pretrained models.
+
+Please note that the config files  `*_pretrained.yaml` are only for generation, not for training new models: when these configs are used for training, the model will be trained from scratch, but during inference our code will still use the pretrained model.
+
 # Notes
-* For the results presented in the paper, we did not use a moving average over the weights. However, using a moving average helps to reduce noise and we therefore recommend its usage. Indeed, we found that using a moving average leads to much better inception scores on Imagenet.
 * Batch normalization is currently *not* supported when using an exponential running average, as the running average is only computed over the parameters of the models and not the other buffers of the model.
 
 # Results

configs/celebAHQ.yaml

@@ -2,10 +2,7 @@ data:
   type: npy
   train_dir: data/celebA-HQ
   test_dir: data/celebA-HQ
-  lsun_categories_train: [bedroom_train] 
-  lsun_categories_test: [bedroom_test]
   img_size: 1024
-  nlabels: 1
 generator:
   name: resnet
   kwargs:
@@ -23,35 +20,11 @@ z_dist:
   dim: 256
 training:
   out_dir: output/celebAHQ
-  gan_type: standard
-  reg_type: real
-  reg_param: 10.
   batch_size: 24
-  nworkers: 16
-  take_model_average: true
-  model_average_beta: 0.999
-  model_average_reinit: false
-  monitoring: tensorboard
-  sample_every: 1000
-  sample_nlabels: 20
-  inception_every: -1 
-  save_every: 900
-  backup_every: 100000
-  restart_every: -1
-  optimizer: rmsprop
-  lr_g: 0.0001
-  lr_d: 0.0001
-  lr_anneal: 1.
-  lr_anneal_every: 150000
-  d_steps: 1
-  equalize_lr: false
 test:
   batch_size: 4
   sample_size: 6
   sample_nrow: 3
-  use_model_average: true
-  compute_inception: false
-  conditional_samples: false
 interpolations:
   nzs: 10
   nsubsteps: 75

configs/default.yaml

@@ -0,0 +1,53 @@
+data:
+  type: lsun
+  train_dir: data/LSUN
+  test_dir: data/LSUN
+  lsun_categories_train: [bedroom_train]
+  lsun_categories_test: [bedroom_test]
+  img_size: 256
+  nlabels: 1
+generator:
+  name: resnet
+  kwargs:
+discriminator:
+  name: resnet
+  kwargs:
+z_dist:
+  type: gauss
+  dim: 256
+training:
+  out_dir: output/default
+  gan_type: standard
+  reg_type: real
+  reg_param: 10.
+  batch_size: 64
+  nworkers: 16
+  take_model_average: true
+  model_average_beta: 0.999
+  model_average_reinit: false
+  monitoring: tensorboard
+  sample_every: 1000
+  sample_nlabels: 20
+  inception_every: -1 
+  save_every: 900
+  backup_every: 100000
+  restart_every: -1
+  optimizer: rmsprop
+  lr_g: 0.0001
+  lr_d: 0.0001
+  lr_anneal: 1.
+  lr_anneal_every: 150000
+  d_steps: 1
+  equalize_lr: false
+  model_file: model.pt
+test:
+  batch_size: 32
+  sample_size: 64
+  sample_nrow: 8
+  use_model_average: true
+  compute_inception: false
+  conditional_samples: false
+  model_file: model.pt
+interpolations:
+  nzs: 10
+  nsubsteps: 75

configs/imagenet.yaml

@@ -2,8 +2,6 @@ data:
   type: image
   train_dir: data/Imagenet
   test_dir: data/Imagenet
-  lsun_categories_train: bedroom_train
-  lsun_categories_test: bedroom_test
   img_size: 128
   nlabels: 1000
 generator:
@@ -24,32 +22,13 @@ z_dist:
 training:
   out_dir: output/imagenet
   gan_type: standard
-  reg_type: real
-  reg_param: 10.
-  batch_size: 128
-  nworkers: 32
-  take_model_average: true
-  model_average_beta: 0.999
-  model_average_reinit: false
-  monitoring: tensorboard
-  sample_every: 1000
   sample_nlabels: 20
   inception_every: 10000
-  save_every: 900
-  backup_every: 100000
-  restart_every: -1
-  optimizer: adam
-  lr_g: 0.0001
-  lr_d: 0.0001
-  lr_anneal: 1.
-  lr_anneal_every: 150000
-  d_steps: 1
-  equalize_lr: false
+  batch_size: 128
 test:
   batch_size: 32
-  sample_size: 32
+  sample_size: 64
   sample_nrow: 8
-  use_model_average: true
   compute_inception: true
   conditional_samples: true
 interpolations:

configs/lsun_bedroom.yaml

@@ -5,7 +5,6 @@ data:
   lsun_categories_train: [bedroom_train]
   lsun_categories_test: [bedroom_test]
   img_size: 256
-  nlabels: 1
 generator:
   name: resnet
   kwargs:
@@ -23,35 +22,10 @@ z_dist:
   dim: 256
 training:
   out_dir: output/lsun_bedroom
-  gan_type: standard
-  reg_type: real
-  reg_param: 10.
-  batch_size: 64
-  nworkers: 32
-  take_model_average: true
-  model_average_beta: 0.999
-  model_average_reinit: false
-  monitoring: tensorboard
-  sample_every: 1000
-  sample_nlabels: 20
-  inception_every: -1 
-  save_every: 900
-  backup_every: 100000
-  restart_every: -1
-  optimizer: rmsprop
-  lr_g: 0.0001
-  lr_d: 0.0001
-  lr_anneal: 1.
-  lr_anneal_every: 150000
-  d_steps: 1
-  equalize_lr: false
 test:
   batch_size: 32
-  sample_size: 15
-  sample_nrow: 5
-  use_model_average: true
-  compute_inception: false
-  conditional_samples: false
+  sample_size: 64
+  sample_nrow: 8
 interpolations:
   nzs: 10
   nsubsteps: 75

configs/lsun_bridge.yaml

@@ -5,7 +5,6 @@ data:
   lsun_categories_train: [bridge_train]
   lsun_categories_test: [bridge_train]
   img_size: 256
-  nlabels: 1
 generator:
   name: resnet
   kwargs:
@@ -23,35 +22,10 @@ z_dist:
   dim: 256
 training:
   out_dir: output/lsun_bridge
-  gan_type: standard
-  reg_type: real
-  reg_param: 10.
-  batch_size: 64
-  nworkers: 32
-  take_model_average: true
-  model_average_beta: 0.999
-  model_average_reinit: false
-  monitoring: tensorboard
-  sample_every: 1000
-  sample_nlabels: 20
-  inception_every: -1 
-  save_every: 900
-  backup_every: 100000
-  restart_every: -1
-  optimizer: rmsprop
-  lr_g: 0.0001
-  lr_d: 0.0001
-  lr_anneal: 1.
-  lr_anneal_every: 150000
-  d_steps: 1
-  equalize_lr: false
 test:
   batch_size: 32
-  sample_size: 15
-  sample_nrow: 5
-  use_model_average: true
-  compute_inception: false
-  conditional_samples: false
+  sample_size: 64
+  sample_nrow: 8
 interpolations:
   nzs: 10
   nsubsteps: 75

configs/lsun_church.yaml

@@ -5,7 +5,6 @@ data:
   lsun_categories_train: [church_outdoor_train]
   lsun_categories_test: [church_outdoor_test]
   img_size: 256
-  nlabels: 1
 generator:
   name: resnet
   kwargs:
@@ -23,35 +22,10 @@ z_dist:
   dim: 256
 training:
   out_dir: output/lsun_church
-  gan_type: standard
-  reg_type: real
-  reg_param: 10.
-  batch_size: 64
-  nworkers: 32
-  take_model_average: true
-  model_average_beta: 0.999
-  model_average_reinit: false
-  monitoring: tensorboard
-  sample_every: 1000
-  sample_nlabels: 20
-  inception_every: -1 
-  save_every: 900
-  backup_every: 100000
-  restart_every: -1
-  optimizer: rmsprop
-  lr_g: 0.0001
-  lr_d: 0.0001
-  lr_anneal: 1.
-  lr_anneal_every: 150000
-  d_steps: 1
-  equalize_lr: false
 test:
   batch_size: 32
-  sample_size: 15
-  sample_nrow: 5
-  use_model_average: true
-  compute_inception: false
-  conditional_samples: false
+  sample_size: 64
+  sample_nrow: 8
 interpolations:
   nzs: 10
   nsubsteps: 75

configs/lsun_tower.yaml

@@ -5,7 +5,6 @@ data:
   lsun_categories_train: [tower_train]
   lsun_categories_test: [tower_test]
   img_size: 256
-  nlabels: 1
 generator:
   name: resnet
   kwargs:
@@ -23,35 +22,10 @@ z_dist:
   dim: 256
 training:
   out_dir: output/lsun_tower
-  gan_type: standard
-  reg_type: real
-  reg_param: 10.
-  batch_size: 64
-  nworkers: 32
-  take_model_average: true
-  model_average_beta: 0.999
-  model_average_reinit: false
-  monitoring: tensorboard
-  sample_every: 1000
-  sample_nlabels: 20
-  inception_every: -1 
-  save_every: 900
-  backup_every: 100000
-  restart_every: -1
-  optimizer: rmsprop
-  lr_g: 0.0001
-  lr_d: 0.0001
-  lr_anneal: 1.
-  lr_anneal_every: 150000
-  d_steps: 1
-  equalize_lr: false
 test:
   batch_size: 32
-  sample_size: 15
-  sample_nrow: 5
-  use_model_average: true
-  compute_inception: false
-  conditional_samples: false
+  sample_size: 64
+  sample_nrow: 8
 interpolations:
   nzs: 10
   nsubsteps: 75