drop old models

Author: GreenWizard2015

NN/__init__.py

@@ -49,9 +49,8 @@ def _nerf_from_config(config):
       trainingLoss=_makeTrainingLoss(config.get('training loss', None)),
       residual=config.get('residual', False),
       extraLatents=config.get('extra latents', None),
+      format=config['format']
     )
-    # If format is not specified, use BGR, because old models were trained to predict BGR
-    nerfParams['format'] = config.get('format', 'bgr')
 
     return lambda encoder, renderer: CNerf2D(
       encoder=encoder,

NN/encoding/CCoordsEncodingLayer.py

@@ -1,22 +1,8 @@
 import tensorflow as tf
-from .CCoordsEncodingLayer import CCoordsEncodingLayerV1
 from .CCoordsEncodingLayerV2 import CCoordsEncodingLayerV2 as CCoordsEncodingLayer
 from .CCoordsGridLayer import CCoordsGridLayer
 from .CFixedSinCosEncoding import CFixedSinCosEncoding
 
-# Old and incorrect implementation of the encoding layer
-class CFlatCoordsEncodingLayer_OLD(tf.keras.layers.Layer):
-  def __init__(self, N=32, **kwargs):
-    super().__init__(**kwargs)
-    self._enc = CCoordsEncodingLayerV1(N)
-    return
-
-  def call(self, x):
-    B = tf.shape(x)[0]
-    tf.assert_equal(tf.shape(x)[:-1], (B, ))
-    x = tf.cast(x, tf.float32)[..., None]
-    return self._enc(x)[:, 0]
-  
 # Correct implementation of the encoding layer
 class CFlatCoordsEncodingLayer(tf.keras.layers.Layer):
   def __init__(self, encoder, **kwargs):
@@ -38,12 +24,8 @@ def encoding_from_config(config):
   if isinstance(config, dict):
     name = config['name']
     params = { k: v for k, v in config.items() if k != 'name' }
-    if 'learned' == name: return CFlatCoordsEncodingLayer_OLD(**params)
     if 'fixed' == name: return CFixedSinCosEncoding(**params)
 
-    if 'learned v2' == name: return CFlatCoordsEncodingLayer(
-      encoder=CCoordsEncodingLayerV1(**params)
-    )
     if 'learned v3' == name: return CFlatCoordsEncodingLayer(
       encoder=CCoordsEncodingLayer(**params)
     )    

NN/encoding/__init__.py

@@ -1,14 +1,28 @@
-from Utils.utils import setupGPU, load_config, setGPUMemoryLimit
+from Utils.utils import setupGPU, load_config, setGPUMemoryLimit, upgrade_configs_structure
 setupGPU() # call it on startup to prevent OOM errors on my machine
 
 import argparse, os, shutil, json
 import tensorflow as tf
 from NN import model_from_config, model_to_architecture
 from Utils import dataset_from_config
 
+def validateLayersNames(model):
+  not_unique_layers = []
+  layers_names = set()
+  for layer in model.trainable_variables:
+    if layer.name in layers_names:
+      not_unique_layers.append(layer.name)
+    layers_names.add(layer.name)
+    continue
+  for layer in not_unique_layers:
+    print(f"Layer name '{layer}' is not unique")
+  assert not not_unique_layers, "Model contains not unique layers names"
+  return
+
 def main(args):
   folder = os.path.dirname(__file__)
   config = load_config(args.config, folder=folder)
+
   assert "experiment" in config, "Config must contain 'experiment' key"
   # store args as part of config
   config['experiment']['command line arguments'] = vars(args)
@@ -37,17 +51,7 @@ def main(args):
   # Create model
   model = model_from_config(config["model"], compile=True)
   model.summary(expand_nested=True)
-  # check if model is contain only unique layers names
-  not_unique_layers = []
-  layers_names = set()
-  for layer in model.trainable_variables:
-    if layer.name in layers_names:
-      not_unique_layers.append(layer.name)
-    layers_names.add(layer.name)
-    continue
-  for layer in not_unique_layers:
-    print(f"Layer name '{layer}' is not unique")
-  assert not not_unique_layers, "Model contains not unique layers names"
+  validateLayersNames(model)
   # save to config model architecture and number of parameters
   config['architecture'] = model_to_architecture(model)
 
@@ -82,10 +86,13 @@ def main(args):
     ),
     tf.keras.callbacks.TerminateOnNaN(),
   ]
-
-  if args.wandb: # init wandb
+  
+  if args.wandb:
     import wandb
+
     wandb.init(project=args.wandb, entity=args.wandb_entity, config=config)
+    # assign run name if specified
+    if args.wandb_name: wandb.run.name = args.wandb_name
     # track model metrics only
     callbacks.append(wandb.keras.WandbCallback(
       save_model=False, # save model to wandb manually
@@ -130,6 +137,7 @@ def main(args):
 
   parser.add_argument('--wandb', type=str, help='Wandb project name (optional)')
   parser.add_argument('--wandb-entity', type=str, help='Wandb entity name (optional)')
+  parser.add_argument('--wandb-name', type=str, help='Wandb run name (optional)')
 
   args = parser.parse_args()
   if args.gpu_memory_mb: setGPUMemoryLimit(args.gpu_memory_mb)