convolution based upscale enhancement

models/upscaler.py

@@ -99,11 +99,11 @@ class UpscalerBasicNew(tf.keras.Model):
         self.flatten = tf.keras.layers.Flatten(name='upscale_enhancer_flattent')
 
         self.seq = tf.keras.Sequential([
-                    tf.keras.layers.Dense(1024, activation='relu', name='upscale_enhancer_dense_1'),
+                    tf.keras.layers.Dense(4096, activation='relu', name='upscale_enhancer_dense_1'),
                     tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_1'),
-                    tf.keras.layers.Dense(1024, activation='relu', name='upscale_enhancer_dense_2'),
+                    tf.keras.layers.Dense(2048, activation='relu', name='upscale_enhancer_dense_2'),
                     tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_2'),
-                    tf.keras.layers.Dense(1024, activation='relu', name='upscale_enhancer_dense_3'),
+                    tf.keras.layers.Dense(2048, activation='relu', name='upscale_enhancer_dense_3'),
                     tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_3'),
                     tf.keras.layers.Dense(low_res * (low_res//2) * 3, activation='tanh', name='upscale_enhancer_dense_final'),
         ])
@@ -123,6 +123,36 @@ class UpscalerBasicNew(tf.keras.Model):
         return x
 
 
+class UpscalerBasicConv(tf.keras.Model):
+    
+    def __init__(self, low_res=32):
+        super(UpscalerBasicConv, self).__init__()
+
+        self.low_res = low_res
+
+        self.seq = tf.keras.Sequential([
+                    tf.keras.layers.Conv2D(4096, kernel_size=3, padding='same', activation='relu', name='upscale_enhancer_conv_1'),
+                    tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_1'),
+                    tf.keras.layers.Conv2D(2048, kernel_size=3, padding='same', activation='relu', name='upscale_enhancer_conv_2'),
+                    tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_2'),
+                    tf.keras.layers.Conv2D(1024, kernel_size=3, padding='same', activation='relu', name='upscale_enhancer_conv_3'),
+                    tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_3'),
+                    tf.keras.layers.Conv2D(128, kernel_size=3, padding='same', activation='relu', name='upscale_enhancer_conv_4'),
+                    tf.keras.layers.BatchNormalization(name='upscale_enhancer_batch_norm_4'),
+                    tf.keras.layers.Dense(3, activation='tanh', name='upscale_enhancer_dense_final'),
+        ])
+
+
+    def call(self, inputs, training=None):
+        img, mask = inputs
+
+        x = self.seq(img, training=training)
+
+        x = x * (1 - mask[:,:,:,tf.newaxis])
+        x = x + img
+
+        return x
+
 # class Upscaler(tf.keras.Model):
     
 #     def __init__(self, bottleneck_size):

upscale_test.py

@@ -1,5 +1,5 @@
 import os
-os.environ["CUDA_VISIBLE_DEVICES"] = "1"
+os.environ["CUDA_VISIBLE_DEVICES"] = "3"
 import tensorflow as tf
 
 gpus = tf.config.experimental.list_physical_devices('GPU')
@@ -30,11 +30,14 @@ from dataset import DatasetFactory, StaticDataset
 from custom_utils.custom_utils import batch_to_dict
 # from custom_utils.smpl import SMPL
 # from models.upscaler import Upscaler
-from models.upscaler import UpscalerBasicNew as Upscaler
+from models.upscaler import UpscalerBasicConv as Upscaler
 
-name='up_3_d256_512_2'
+# !!!!!!!!!! CHANGE SSIM BASED ON RES
+# !!!!!!!!!! CHANGE SKIRT OR NOT
+
+name='32_512_3'
 def init_wandb():
-    run = wandb.init(project='vae', name=name,
+    run = wandb.init(project='upscale', name=name,
             notes='\
 changed ms-ssim scale weights | \
 inpainting style | \
@@ -44,6 +47,8 @@ use mask for loss | \
 using ms ssim for higher res | \
 training from low res to low res | \
 named the layers, added tanh to final layer | \
+changed l1 loss reduce axis | \
+try convolutional architecture | \
 ')
     artifact = wandb.Artifact('code', type='code')
     artifact.add_file('upscale_test.py')
@@ -53,7 +58,6 @@ named the layers, added tanh to final layer | \
     artifact.add_file('convnext/models/convnext_tf.py')
     run.log_artifact(artifact)
 
-# !!!!!!!!!! CHANGE SKIRT OR NOT
 uv2mesh = UV2Mesh(res=256, is_skirt=False)
 
 
@@ -98,17 +102,15 @@ def plot_3d(mesh_gt, mesh_pred_unposed, faces):
 @tf.function
 def train_step(batch):
     with tf.GradientTape() as tape:
-        pred = model([batch['uv_unposed' + low_resolution_modifier], batch['uv_mask' + low_resolution_modifier]])
+        pred = model([batch['uv_unposed' + low_resolution_modifier], batch['uv_mask' + low_resolution_modifier]], training=True)
         pred = tf.image.resize(pred, size=(high_res, high_res//2), method=tf.image.ResizeMethod.LANCZOS5)
 
         mesh_pred = uv2mesh(pred)
         mesh_orig = batch['mesh_unposed']
 
-        l1_loss = tf.reduce_mean(tf.reduce_sum(((tf.math.abs(batch['uv_unposed' + high_resolution_modifier] - pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis]) )), axis=1))
+        l1_loss = tf.reduce_mean(tf.reduce_sum(((tf.math.abs(batch['uv_unposed' + high_resolution_modifier] - pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis]) )), axis=-1))
-        # ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred * batch['uv_mask'][:,:,:,tf.newaxis] + 1, tf.transpose(batch['uv_unposed'], [0,2,3,1]) + 1, max_val=2, power_factors=(0.5, 0.2, 0.2, 0.1)))
+        # ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, batch['uv_unposed' + high_resolution_modifier] * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, max_val=2, power_factors=(0.5, 0.2, 0.2, 0.1)))
-        # ms_ssim = tf.reduce_mean(tf.image.ssim(pred * batch['uv_mask'][:,:,:,tf.newaxis] + 1, batch['uv_unposed'] * batch['uv_mask'][:,:,:,tf.newaxis] + 1, max_val=2))
+        ms_ssim = tf.reduce_mean(tf.image.ssim(pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, batch['uv_unposed' + high_resolution_modifier] * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, max_val=2))
-        ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, batch['uv_unposed' + high_resolution_modifier] * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, max_val=2, power_factors=(0.5, 0.2, 0.2, 0.1)))
-        # ms_ssim = tf.reduce_mean(tf.image.ssim(pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, batch['uv_unposed' + high_resolution_modifier] * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, max_val=2))
         loss = 1e3 * (1 - ms_ssim)
 
     metric = tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig - mesh_pred) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask'])
@@ -118,17 +120,17 @@ def train_step(batch):
         
 
 
-# UPDATE!
 @tf.function
 def valid_step(batch):
-    pred = model(batch['uv_unposed' + low_resolution_modifier], batch['uv_mask'])
+    pred = model([batch['uv_unposed' + low_resolution_modifier], batch['uv_mask' + low_resolution_modifier]], training=False)
+    pred = tf.image.resize(pred, size=(high_res, high_res//2), method=tf.image.ResizeMethod.LANCZOS5)
 
     mesh_pred = uv2mesh(pred)
     mesh_orig = batch['mesh_unposed']
 
-    l1_loss = tf.reduce_mean(tf.reduce_sum(((tf.math.abs(batch['uv_unposed'] - pred) )), axis=1))
+    l1_loss = tf.reduce_mean(tf.reduce_sum(((tf.math.abs(batch['uv_unposed' + high_resolution_modifier] - pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis]) )), axis=-1))
-    ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(tf.transpose(pred, [0,2,3,1]) + 1, tf.transpose(batch['uv_unposed'], [0,2,3,1]) + 1, max_val=2))
+    # ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, batch['uv_unposed' + high_resolution_modifier] * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, max_val=2, power_factors=(0.5, 0.2, 0.2, 0.1)))
-    # ms_ssim = tf.reduce_mean(tf.image.ssim(tf.transpose(pred * uv_mask_body[tf.newaxis,tf.newaxis,:,:], [0,2,3,1]) + 1, tf.transpose(batch['uv_body_posed' + resolution_modifier] * uv_mask_body[tf.newaxis,tf.newaxis,:,:], [0,2,3,1]) + 1, max_val=2))
+    ms_ssim = tf.reduce_mean(tf.image.ssim(pred * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, batch['uv_unposed' + high_resolution_modifier] * batch['uv_mask' + high_resolution_modifier][:,:,:,tf.newaxis] + 1, max_val=2))
     loss = 1e3 * (1 - ms_ssim)
 
     metric = tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig - mesh_pred) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask'])
@@ -138,7 +140,7 @@ def valid_step(batch):
 
 def plot_everything(batch: dict, type: str):
 
-    pred = model([batch['uv_unposed' + low_resolution_modifier], batch['uv_mask' + low_resolution_modifier]])
+    pred = model([batch['uv_unposed' + low_resolution_modifier], batch['uv_mask' + low_resolution_modifier]], training=False)
     pred = tf.image.resize(pred, size=(high_res, high_res//2), method=tf.image.ResizeMethod.LANCZOS5)
     
     mesh_pred = uv2mesh(pred)
@@ -162,20 +164,20 @@ def plot_everything(batch: dict, type: str):
 
 
 def train():
-    # valid_iterator = iter(dataset_valid)
+    valid_iterator = iter(dataset_valid)
 
     step = 0
     for batch in dataset_train:
         batch = static_dataset_train.expand_batch(batch_to_dict(batch, signature_dynamic))
         train_loss = train_step(batch)
 
-        # data_valid = valid_iterator.get_next()
+        data_valid = valid_iterator.get_next()
-        # data_valid = static_dataset_valid.expand_batch(batch_to_dict(data_valid, signature_dynamic))
+        data_valid = static_dataset_valid.expand_batch(batch_to_dict(data_valid, signature_dynamic))
-        valid_loss = {}#valid_step(data_valid)
+        valid_loss = valid_step(data_valid)
 
-        if step % 10000 == 0:
+        if step % 10000 == 0 or (step < 5000 and step % 500 == 0):
             train_figs = plot_everything(batch, 'train')
-            valid_figs = {}#plot_everything(data_valid, 'valid')
+            valid_figs = plot_everything(data_valid, 'valid')
             wandb.log({ **train_loss, **valid_loss, **train_figs, **valid_figs }, step=step)
         else:
             wandb.log({ **train_loss, **valid_loss }, step=step)
@@ -187,8 +189,8 @@ def train():
 
         print(step)
 
-        if step > 300000:
+        # if step > 900000:
-            break
+        #     break
 
         step += 1
 
@@ -203,7 +205,7 @@ if __name__ == '__main__':
     # uv_mask_body = np.load('custom_utils/mask_256x512.npy')
     # uv_mask_body = cv2.dilate(uv_mask_body.astype(np.float32), cv2.getStructuringElement(cv2.MORPH_RECT, (2 * 2 + 1, 2 * 2 + 1), (2, 2)), iterations=2)
 
-    low_res = 256
+    low_res = 32
     high_res = 512
     uv_mask_body = tf.ones((low_res, low_res//2))
     low_resolution_modifier = '_' + str(low_res)
@@ -222,10 +224,10 @@ if __name__ == '__main__':
                          'uv_mask' + low_resolution_modifier: (low_res, low_res//2,),
                         }
 
-    dataset_train = factory.get_dataset(folder='/home/workspace/local-data/train_100', batch_size=8, signature=signature_dynamic, use_filled=False, resolutions=[high_res, low_res])
+    dataset_train = factory.get_dataset(folder='/home/workspace/local-data/train', batch_size=8, signature=signature_dynamic, use_filled=False, resolutions=[high_res, low_res], in_memory=False)
-    static_dataset_train = StaticDataset(folder='/home/workspace/local-data/train_100', signature=signature_static, use_filled=False, resolutions=[high_res, low_res])
+    static_dataset_train = StaticDataset(folder='/home/workspace/local-data/train', signature=signature_static, use_filled=False, resolutions=[high_res, low_res], in_memory=True)
-    dataset_valid = None #factory.get_dataset(folder='/home/workspace/local-data/jumpsuit_20/valid', batch_size=32, signature=signature_dynamic)
+    dataset_valid = factory.get_dataset(folder='/home/workspace/local-data/valid', batch_size=8, signature=signature_dynamic, use_filled=False, resolutions=[high_res, low_res], in_memory=False)
-    static_dataset_valid = None #StaticDataset(folder='/home/workspace/local-data/jumpsuit_20/valid', signature=signature_static)
+    static_dataset_valid = StaticDataset(folder='/home/workspace/local-data/valid', signature=signature_static, use_filled=False, resolutions=[high_res, low_res], in_memory=True)
 
     model = Upscaler(low_res=low_res)