trying skip connection

configs/direct_1.py

@@ -0,0 +1,77 @@
+from utils.convnext_utils import CNBlockConfig
+
+config = dict()
+
+config['gpu'] = '5'
+config['name'] = 'direct_1'
+config['notes'] = ' ' 
+config['tags'] = ['torch', 'direct'] 
+config['script'] = 'scripts/direct_train.py'
+
+config['pretrained_model'] = ''
+config['use_curvature'] = False
+
+config['batch_size'] = 4
+config['train_dir'] = '/data-local/127/train'
+config['valid_dir'] = '/data-local/127/valid'
+config['test_dir'] = '/data-net/dataset/test_ready'
+config['train_frames'] = None
+config['valid_frames'] = None
+
+config['resolution'] = 512
+config['is_skirt'] = False
+config['dataset_in_memory'] = False
+config['is_cape'] = False
+config['remove_global_orientation'] = True
+config['device'] = 'cuda'
+config['input_features'] = 6
+
+config['smpl_res'] = 14475
+if config['is_cape']:
+  config['smpl_res'] = 3638
+
+config['latent_dim'] = 128
+
+config['bottleneck_scale'] = 2
+
+config['is_label'] = False
+config['kernel_size'] = 7
+config['bottleneck_size'] = 64 * config['bottleneck_scale']
+
+block_setting = [
+    CNBlockConfig(16, 32, 3),
+    CNBlockConfig(32, 128, 3),
+    CNBlockConfig(128, 256, 9),
+    CNBlockConfig(256, config['bottleneck_size'], 9),
+    CNBlockConfig(config['bottleneck_size'], None, 3),
+]
+
+block_setting_reverse = [
+    CNBlockConfig(config['bottleneck_size'], 256, 3),
+    CNBlockConfig(256, 128, 9),
+    CNBlockConfig(128, 32, 9),
+    CNBlockConfig(32, 16, 3),
+    CNBlockConfig(16, 3, 3),
+    # CNBlockConfig(3, 3, 3),
+]
+
+config['encoder_block_settings'] = block_setting
+config['decoder_block_settings'] = block_setting_reverse
+config['activation'] = 'tanh'
+
+
+config['lr'] = 0.0001
+config['reg_loss_coeff'] = 0.1
+
+config['signature'] = { 'id',
+                        'frame_nr',
+                        'pose_shape',
+                        'uv_body_posed',
+                        'mesh_mask',
+                        'mesh_body_posed',
+                        'mesh_posed',
+                        'uv_unposed',
+                        'mesh_unposed',
+                        'uv_static',
+                        'uv_mask',
+                        }

configs/direct_2.py

@@ -0,0 +1,78 @@
+from utils.convnext_utils import CNBlockConfig
+
+config = dict()
+
+config['gpu'] = '0'
+config['name'] = 'direct_2'
+config['notes'] = 'with curvature input' 
+config['tags'] = ['torch', 'direct'] 
+config['script'] = 'scripts/direct_train.py'
+
+config['pretrained_model'] = ''
+config['use_curvature'] = True
+
+config['batch_size'] = 4
+config['train_dir'] = '/data-local/127/train'
+config['valid_dir'] = '/data-local/127/valid'
+config['test_dir'] = '/data-net/dataset/test_ready'
+config['train_frames'] = None
+config['valid_frames'] = None
+
+config['resolution'] = 512
+config['is_skirt'] = False
+config['dataset_in_memory'] = False
+config['is_cape'] = False
+config['remove_global_orientation'] = True
+config['device'] = 'cuda'
+config['input_features'] = 10
+
+config['smpl_res'] = 14475
+if config['is_cape']:
+  config['smpl_res'] = 3638
+
+config['latent_dim'] = 128
+
+config['bottleneck_scale'] = 2
+
+config['is_label'] = False
+config['kernel_size'] = 7
+config['bottleneck_size'] = 64 * config['bottleneck_scale']
+
+block_setting = [
+    CNBlockConfig(16, 32, 3),
+    CNBlockConfig(32, 128, 3),
+    CNBlockConfig(128, 256, 9),
+    CNBlockConfig(256, config['bottleneck_size'], 9),
+    CNBlockConfig(config['bottleneck_size'], None, 3),
+]
+
+block_setting_reverse = [
+    CNBlockConfig(config['bottleneck_size'], 256, 3),
+    CNBlockConfig(256, 128, 9),
+    CNBlockConfig(128, 32, 9),
+    CNBlockConfig(32, 16, 3),
+    CNBlockConfig(16, 3, 3),
+    # CNBlockConfig(3, 3, 3),
+]
+
+config['encoder_block_settings'] = block_setting
+config['decoder_block_settings'] = block_setting_reverse
+config['activation'] = 'tanh'
+
+
+config['lr'] = 0.0001
+config['reg_loss_coeff'] = 0.1
+
+config['signature'] = { 'id',
+                        'frame_nr',
+                        'pose_shape',
+                        'uv_curvature',
+                        'mesh_mask',
+                        'mesh_body_posed',
+                        'mesh_posed',
+                        'uv_unposed',
+                        'mesh_unposed',
+                        'uv_static',
+                        'uv_mask',
+                        'uv_body_posed',
+                        }

configs/direct_3.py

@@ -0,0 +1,78 @@
+from utils.convnext_utils import CNBlockConfig
+
+config = dict()
+
+config['gpu'] = '4'
+config['name'] = 'direct_3'
+config['notes'] = 'with curvature input and skip connection' 
+config['tags'] = ['torch', 'direct'] 
+config['script'] = 'scripts/direct_train.py'
+
+config['pretrained_model'] = ''
+config['use_curvature'] = True
+
+config['batch_size'] = 4
+config['train_dir'] = '/data-local/127/train'
+config['valid_dir'] = '/data-local/127/valid'
+config['test_dir'] = '/data-net/dataset/test_ready'
+config['train_frames'] = None
+config['valid_frames'] = None
+
+config['resolution'] = 512
+config['is_skirt'] = False
+config['dataset_in_memory'] = False
+config['is_cape'] = False
+config['remove_global_orientation'] = True
+config['device'] = 'cuda'
+config['input_features'] = 10
+
+config['smpl_res'] = 14475
+if config['is_cape']:
+  config['smpl_res'] = 3638
+
+config['latent_dim'] = 128
+
+config['bottleneck_scale'] = 2
+
+config['is_label'] = False
+config['kernel_size'] = 7
+config['bottleneck_size'] = 64 * config['bottleneck_scale']
+
+block_setting = [
+    CNBlockConfig(16, 32, 3),
+    CNBlockConfig(32, 128, 3),
+    CNBlockConfig(128, 256, 9),
+    CNBlockConfig(256, config['bottleneck_size'], 9),
+    CNBlockConfig(config['bottleneck_size'], None, 3),
+]
+
+block_setting_reverse = [
+    CNBlockConfig(config['bottleneck_size'], 256, 3),
+    CNBlockConfig(256*2, 128, 9),
+    CNBlockConfig(128*2, 32, 9),
+    CNBlockConfig(32*2, 16, 3),
+    CNBlockConfig(16*2, 8, 3),
+    CNBlockConfig(8, 3, 3),
+]
+
+config['encoder_block_settings'] = block_setting
+config['decoder_block_settings'] = block_setting_reverse
+config['activation'] = 'tanh'
+
+
+config['lr'] = 0.0001
+config['reg_loss_coeff'] = 0.1
+
+config['signature'] = { 'id',
+                        'frame_nr',
+                        'pose_shape',
+                        'uv_curvature',
+                        'mesh_mask',
+                        'mesh_body_posed',
+                        'mesh_posed',
+                        'uv_unposed',
+                        'mesh_unposed',
+                        'uv_static',
+                        'uv_mask',
+                        'uv_body_posed',
+                        }

custom_utils/custom_utils.py

@@ -43,6 +43,7 @@ class UV2Mesh(nn.Module):
         if uv.shape[1] == 3 or uv.shape[1] == 7:
             uv = torch.permute(uv, [0,2,3,1])
         if uv.shape[1] != 512:
+            raise NotImplementedError
             uv = Resize(uv[:,:,None], size=(512, 256), interpolation=InterpolationMode.BILINEAR)
             
         # batch_size = uv.shape[0]

models/base.py

@@ -35,10 +35,10 @@ class VariationalEncoder(nn.Module):
 
 class AEEncoder(nn.Module):
     
-    def __init__(self, block_setting: CNBlockConfig, kernel_size: int = 7, activation='tanh', is_label=False, nr_of_labels=None, device='cuda'):
+    def __init__(self, block_setting: CNBlockConfig, kernel_size: int = 7, activation='tanh', is_label=False, nr_of_labels=None, device='cuda', input_features=3):
         super(AEEncoder, self).__init__()
 
-        self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, is_label=is_label, nr_of_labels=nr_of_labels, device=device)
+        self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, is_label=is_label, nr_of_labels=nr_of_labels, device=device, input_features=input_features)
         if activation == 'tanh':
             self.act = nn.Tanh()
         elif activation == 'sigmoid':
@@ -53,13 +53,13 @@ class AEEncoder(nn.Module):
         else:
             x = inputs
 
-        x = self.encoder(x)
+        x, features = self.encoder(x)
         x = self.act(x)
 
         if len(inputs) == 2:
-            return x, inputs[1]
+            return x, features, inputs[1]
         else:
-            return x
+            return x, features
 
 
 class Decoder(nn.Module):
@@ -75,12 +75,12 @@ class Decoder(nn.Module):
             self.softmax = nn.Softmax(dim=1)
 
     def forward(self, inputs):
-        if len(inputs) == 2:
-            x, mask = inputs
+        if len(inputs) == 3:
+            x, features, mask = inputs
         else:
-            x = inputs
+            x, features = inputs
 
-        x = self.decoder(x)
+        x = self.decoder([x, features])
 
         if self.is_label:
             x = x * mask[:,None,:,:]

models/cond_encoder.py

@@ -8,7 +8,7 @@ class AutoEncoder(nn.Module):
     def __init__(self, config):
         super(AutoEncoder, self).__init__()
 
-        self.encoder : AEEncoder = AEEncoder(config['encoder_block_settings'], config['kernel_size'], config['activation'], is_label=config['is_label'], device=config['device'], nr_of_labels=7)
+        self.encoder : AEEncoder = AEEncoder(config['encoder_block_settings'], config['kernel_size'], config['activation'], is_label=config['is_label'], device=config['device'], nr_of_labels=7, input_features=config['input_features'])
         self.decoder : Decoder = Decoder(config['decoder_block_settings'], kernel_size=config['kernel_size'], is_label=config['is_label'], device=config['device'], nr_of_classes=6)
 
 

models/convnext_orig.py

@@ -81,8 +81,6 @@ class ConvNeXtBase(nn.Module):
         if norm_layer is None:
             norm_layer = partial(LayerNorm2d, eps=1e-6)
 
-        layers_list: List[nn.Module] = []
-
         # Stem
         firstconv_output_channels = block_setting[0].input_channels
         self.stem = Conv2dNormActivation(
@@ -95,7 +93,8 @@ class ConvNeXtBase(nn.Module):
                 activation_layer=None,
                 bias=True,
             ).to(device)
-        layers_list.append(self.stem)
+        self.stages = []
+        self.downsamples = []
 
         total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
         stage_block_id = 0
@@ -107,17 +106,20 @@ class ConvNeXtBase(nn.Module):
                 sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
                 stage.append(block(cnf.input_channels, layer_scale, sd_prob, kernel_size))
                 stage_block_id += 1
-            layers_list.append(nn.Sequential(*stage).to(device))
+            self.stages.append(nn.Sequential(*stage).to(device))
             if cnf.out_channels is not None:
                 # Downsampling
-                layers_list.append(
+                self.downsamples.append(
                     nn.Sequential(
                         norm_layer(cnf.input_channels),
                         nn.Conv2d(cnf.input_channels, cnf.out_channels, kernel_size=2, stride=2),
                     ).to(device)
                 )
+        self.stages = nn.ModuleList(self.stages)
+        self.downsamples = nn.ModuleList(self.downsamples)
 
-        self.layers = nn.Sequential(*layers_list).to(device)
+        # self.layers = nn.Sequential(*layers_list).to(device)
+        # self.layers = nn.ModuleList(layers_list)
         self.act = nn.Tanh().to(device)
 
         # for m in self.modules():
@@ -127,9 +129,16 @@ class ConvNeXtBase(nn.Module):
         #             nn.init.zeros_(m.bias)
 
     def _forward_impl(self, x: Tensor) -> Tensor:
-        x = self.layers(x)
+        x = self.stem(x)
+        features = []
+        for i in range(len(self.stages)):
+            x = self.stages[i](x)
+            if i == len(self.stages) - 1:
+                continue
+            features.append(x)
+            x = self.downsamples[i](x)
         x = self.act(x)
-        return x
+        return x, features
 
     def forward(self, x: Tensor) -> Tensor:
         return self._forward_impl(x)
@@ -165,13 +174,14 @@ class ConvNeXtReverseBase(nn.Module):
             norm_layer = partial(LayerNorm2d, eps=1e-6)
 
         self.stem = nn.Sequential(
-                        nn.ConvTranspose2d(block_setting[0].input_channels, block_setting[0].out_channels, kernel_size=4, stride=4),
+                        nn.ConvTranspose2d(block_setting[0].input_channels, block_setting[0].out_channels, kernel_size=2, stride=2),
                         norm_layer(block_setting[0].out_channels),
                     ).to(device)
         
         total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
         stage_block_id = 0
-        layers_list: List[nn.Module] = []
+        self.stages = []
+        self.upsamples = []
 
 
         for i, cnf in enumerate(block_setting):
@@ -182,17 +192,17 @@ class ConvNeXtReverseBase(nn.Module):
                 sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
                 stage.append(block(cnf.input_channels, layer_scale, sd_prob, kernel_size))
                 stage_block_id += 1
-            layers_list.append(nn.Sequential(*stage))
+            self.stages.append(nn.Sequential(*stage).to(device))
 
             if i == 0:
-                layers_list.append(self.stem)
+                self.upsamples.append(self.stem)
             else:
                 # Upsampling
-                layers_list.append(
+                self.upsamples.append(
                     nn.Sequential(
                         norm_layer(cnf.input_channels),
                         nn.ConvTranspose2d(cnf.input_channels, cnf.out_channels, kernel_size=2, stride=2),
-                    )
+                    ).to(device)
                 )
 
         self.last = nn.Sequential(
@@ -201,7 +211,8 @@ class ConvNeXtReverseBase(nn.Module):
                         Permute([0, 3, 1, 2]),
                     ).to(device)
         
-        self.layers = nn.Sequential(*layers_list).to(device)
+        self.stages = nn.ModuleList(self.stages)
+        self.upsamples = nn.ModuleList(self.upsamples)
         self.activation = nn.Tanh().to(device) if not is_label else nn.Identity().to(device)
 
 
@@ -213,7 +224,15 @@ class ConvNeXtReverseBase(nn.Module):
 
 
     def _forward_impl(self, x) -> Tensor:
-        x = self.layers(x)
+        x, features = x
+        for i in range(len(self.stages)-1):
+            if i == 0:
+                x = self.stages[i](x)
+            else:
+                x = self.stages[i](torch.cat([x, features[-i]], dim=1))
+            x = self.upsamples[i](x)
+        x = self.stages[-1](x)
+        x = self.upsamples[-1](x)
         x = self.last(x)
         x = self.activation(x)
         return x

models/direct.py

@@ -0,0 +1,29 @@
+import torch
+from torch import nn
+from models.convnext_orig import ConvNeXt, CNBlock, CNBlockConfig
+from typing import Any, Callable, List, Optional, Sequence
+
+
+class DirectModel(nn.Module):
+    
+    def __init__(self, block_setting: List[CNBlockConfig], kernel_size: int = 7, input_features: int = 16, device: str = 'cuda'):
+        super(DirectModel, self).__init__()
+
+        self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, input_features=input_features, device=device)
+        self.decoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, device=device, reversed=True)
+
+        layers = [CNBlock(block_setting[-1].input_channels, kernel_size=kernel_size) for _ in range(5)]
+        self.mean_branch = nn.Sequential(*layers).to(device)
+        layers = [CNBlock(block_setting[-1].input_channels, kernel_size=kernel_size) for _ in range(5)]
+        self.var_branch = nn.Sequential(*layers).to(device)
+
+        self.flatten_mean = nn.Flatten()
+        self.flatten_var = nn.Flatten()
+
+
+    def forward(self, inputs):
+        body, template = inputs
+
+        x = self.encoder(x)
+
+        return mean

run.py

@@ -1,6 +1,6 @@
 import os
 import sys
-from configs.pyr_64_cape_m import config
+from configs.direct_1 import config
 import argparse
 
 # parser = argparse.ArgumentParser(description='')

scripts/direct_train.py

@@ -0,0 +1,281 @@
+import os
+import sys
+sys.path.append(os.path.dirname(os.path.abspath(__file__)) + '/..')
+
+import importlib
+import argparse
+parser = argparse.ArgumentParser(description='')
+parser.add_argument('--name', type=str)
+parser.add_argument('--config', type=str)
+parser.add_argument('--debug', action='store_true')
+args = vars(parser.parse_args())
+
+config_module = importlib.import_module(args['config'])
+config = config_module.config
+if args['debug']:
+    from configs.direct_3 import config
+    # from configs.ae_static_orig_2 import config
+os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
+os.environ["CUDA_VISIBLE_DEVICES"] = config['gpu']
+if args['debug']:
+    os.environ["CUDA_LAUNCH_BLOCKING"] = '1'
+
+import time
+import argparse
+import wandb
+import random
+import numpy as np
+import torch
+# from torchmetrics import MultiScaleStructuralSimilarityIndexMeasure as MSSSIM
+from custom_utils.custom_utils import UV2Mesh, get_summary_html, maskFacesBatch, plot_3d_collision
+import plotly.graph_objects as go
+import plotly.express as ex
+import torch.optim as optim
+
+if config['is_skirt']:
+    from losses.losses import edge_loss_skirt as edge_loss
+else:
+    from losses.losses import edge_loss
+
+from data.dataset import Dataset
+# from torch.utils.data import DataLoader
+from losses.losses import collision_loss, normal_loss
+from utils.postprocess import fix_collision
+from custom_utils.custom_utils import plot_3d
+from utils.checkpoints import CheckpointSaver, load_model
+from custom_utils.custom_utils import Accumulator
+from utils.utils import InfiniteDataLoader as DataLoader
+from models.cond_encoder import AutoEncoder as Model
+from smpl.smpl_torch import SMPL
+
+from pytorch_msssim import MS_SSIM
+
+
+def init_wandb():
+    run = wandb.init(project='iccv', name=config['name'], notes=config['notes'], dir='/data-local/127/temp', tags=config['tags'])
+    artifact = wandb.Artifact('code', type='code')
+    artifact.add_file('scripts/direct_train.py')
+    artifact.add_file('losses/losses.py')
+    artifact.add_file('models/base.py')
+    artifact.add_file('models/cond_encoder.py')
+    artifact.add_file('models/convnext.py')
+    artifact.add_file('data/dataset.py')
+    config_file_names = [key for key in sys.modules.keys()  if key.startswith('configs.')]
+    for config_file_name in config_file_names:
+        file_name = '/'.join(config_file_name.split('.')) + '.py'
+        if os.path.exists(file_name):
+            artifact.add_file(file_name)
+    run.log_artifact(artifact)
+
+uv2mesh = UV2Mesh(res=256, is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape'])
+
+
+
+def train_step(batch):
+    model.train()
+    # with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=use_amp):
+    if config['use_curvature']:
+        input = torch.cat([batch['uv_curvature'], batch['uv_static']], axis=1)
+    else:
+        input = torch.cat([batch['uv_body_posed'], batch['uv_static']], axis=1)
+    pred = model([input, batch['uv_mask']])
+
+    mesh_pred = uv2mesh(pred)
+    mesh_orig = batch['mesh_unposed']
+    shape = batch['pose_shape'][:, 72:]
+    gender = torch.greater_equal(shape[:, 10] , 0.5).to(torch.long)
+    shape_offsets = torch.stack([torch.tensordot(shape[i,:10][None, :], smpl.shapedirs[gender[i]], dims=[[1], [2]])[0,smpl.to_keep,:] for i in range(shape.shape[0])])
+    mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,None]
+    mesh_pred_posed = smpl(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped)[0]
+ 
+    e_loss = torch.mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
+    c_loss, c_ratio = collision_loss(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask'])
+    c_loss, c_ratio = torch.mean(c_loss), torch.mean(c_ratio)
+    # l2_loss = torch.mean(1e3 * torch.mean(torch.sum(((batch['uv_unposed'] - pred) ** 2), axis=-1), axis=[1, 2]))
+    l2_mesh_loss = torch.mean(1e3 * torch.mean(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_posed'] - mesh_pred_posed) ** 2, axis=-1), axis=-1))
+    n_loss = normal_loss(mesh_pred_posed, batch['mesh_posed'], batch['mesh_mask'], is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape'])
+
+    loss = l2_mesh_loss + (e_loss + c_loss + n_loss) * config['reg_loss_coeff']
+
+    optimizer.zero_grad()
+    loss.backward()
+    optimizer.step()
+    # scaler.scale(loss).backward()
+    # scaler.step(optimizer)
+    # scaler.update()
+    # optimizer.zero_grad()
+
+    model.eval()
+    with torch.inference_mode():
+        ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_unposed'] + 1))
+        metric = torch.sum(torch.sqrt(torch.sum(batch['mesh_mask'][:,:,None] * (mesh_orig - mesh_pred.detach()) ** 2, axis=-1))) / torch.sum(batch['mesh_mask'])
+    
+    return { 'train/loss': loss, 'train/edge_loss': e_loss, 'train/euclid_error': metric, 'train/collision_ratio': c_ratio,
+             'train/ms_ssim': ms_ssim, 'train/collision_loss': c_loss, 'train/l2_mesh_loss': l2_mesh_loss,
+             'train/normal_loss': n_loss }
+        
+
+def valid_step(batch):
+    model.eval()
+
+    with torch.inference_mode():
+        if config['use_curvature']:
+            input = torch.cat([batch['uv_curvature'], batch['uv_static']], axis=1)
+        else:
+            input = torch.cat([batch['uv_body_posed'], batch['uv_static']], axis=1)
+        pred = model([input, batch['uv_mask']])
+        mesh_pred = uv2mesh(pred)
+        mesh_orig = batch['mesh_unposed']
+        shape = batch['pose_shape'][:, 72:]
+        gender = torch.greater_equal(shape[:, 10] , 0.5).to(torch.long)
+        shape_offsets = torch.stack([torch.tensordot(shape[i,:10][None, :], smpl.shapedirs[gender[i]], dims=[[1], [2]])[0,smpl.to_keep,:] for i in range(shape.shape[0])])
+        mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,None]
+        mesh_pred_posed = smpl(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped)[0]
+    
+        e_loss = torch.mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
+        c_loss, c_ratio = collision_loss(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask'])
+        c_loss, c_ratio = torch.mean(c_loss), torch.mean(c_ratio)
+        # l2_loss = torch.mean(1e3 * torch.mean(torch.sum(((batch['uv_unposed'] - pred) ** 2), axis=-1), axis=[1, 2]))
+        l2_mesh_loss = torch.mean(1e3 * torch.mean(torch.sum(batch['mesh_mask'][:,:,None] * (batch['mesh_posed'] - mesh_pred_posed) ** 2, axis=-1), axis=-1))
+        n_loss = normal_loss(mesh_pred_posed, batch['mesh_posed'], batch['mesh_mask'], is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape'])
+
+        ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_unposed'] + 1))
+        loss = l2_mesh_loss + (e_loss + c_loss + n_loss) * config['reg_loss_coeff']
+
+        metric = torch.sum(torch.sqrt(torch.sum(batch['mesh_mask'][:,:,None] * (mesh_orig - mesh_pred) ** 2, axis=-1))) / torch.sum(batch['mesh_mask'])
+    
+    return { 'valid/loss': loss, 'valid/edge_loss': e_loss, 'valid/euclid_error': metric, 'valid/collision_ratio': c_ratio,
+             'valid/ms_ssim': ms_ssim, 'valid/collision_loss': c_loss, 'valid/l2_mesh_loss': l2_mesh_loss,
+             'valid/normal_loss': n_loss }
+       
+
+def plot_everything(batch: dict, type: str):
+    model.eval()
+    with torch.inference_mode():
+        if config['use_curvature']:
+            input = torch.cat([batch['uv_curvature'], batch['uv_static']], axis=1)
+        else:
+            input = torch.cat([batch['uv_body_posed'], batch['uv_static']], axis=1)
+        pred = model([input, batch['uv_mask']]).detach()
+
+    mesh_pred = uv2mesh(pred).cpu().numpy()
+    mesh_orig = batch['mesh_unposed'].cpu().numpy()
+    
+    orig_img = torch.permute(batch['uv_unposed'], [0,2,3,1])
+    uv_static_img = torch.permute(batch['uv_static'], [0,2,3,1])
+    uv_body_posed_img = torch.permute(batch['uv_body_posed'], [0,2,3,1])
+    pred_img = torch.permute(pred, [0,2,3,1])
+    img = torch.concat((uv_static_img, uv_body_posed_img, orig_img, pred_img), axis=2)
+    imgs = [wandb.Image((img[i].cpu().numpy() + 1) / 2.0) for i in range(4)]
+    
+    F = np.load("smpl/smpl_faces" + skirt_modifier + ("_orig" if config['is_cape'] else "") + ".npy")
+    mask = batch['mesh_mask'].cpu().numpy().astype(bool)
+
+    shape = batch['pose_shape'][:, 72:]
+    gender = torch.greater_equal(shape[:, 10] , 0.5).to(torch.long)
+    shape_offsets = torch.stack([torch.tensordot(shape[i,:10][None, :], smpl.shapedirs[gender[i]], dims=[[1], [2]])[0,smpl.to_keep,:] for i in range(shape.shape[0])]).cpu().numpy()
+
+    garment_unposed, faces = maskFacesBatch(mesh_pred, F, mask)
+    garment_unposed = [garment_unposed[i] + shape_offsets[i, mask[i], :] for i in range(mask.shape[0])]
+    gt = [mesh_orig[i][mask[i], :] + shape_offsets[i, mask[i], :] for i in range(pred.shape[0])]
+
+    for i in range(mesh_orig.shape[0]):
+        mesh_pred[i, mask[i], :] += shape_offsets[i, mask[i], :]
+    mesh_posed = smpl(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], torch.tensor(mesh_pred, device=config['device']))[0].cpu().numpy()
+
+    gt_posed = [batch['mesh_posed'].cpu().numpy()[i, mask[i], :] for i in range(mask.shape[0])]
+
+    mesh_posed_fixed = fix_collision(torch.tensor(mesh_posed, device=config['device']), batch['mesh_body_posed'], torch.tensor(mask, device=config['device'])).cpu().numpy()
+    body = batch['mesh_body_posed'].cpu().numpy()
+
+    garment_posed = [mesh_posed[i, mask[i], :] for i in range(body.shape[0])]
+    collision = [np.concatenate((body[i], garment_posed[i]), axis=0) for i in range(len(garment_posed))]
+    collision_faces = [np.concatenate((F, faces[i] + config['smpl_res']), axis=0) for i in range(len(garment_posed))]
+    collision_color = [np.concatenate((np.ones_like(body[i], dtype=np.int32)*100, np.ones_like(garment_posed[i], dtype=np.int32)*230), axis=0) for i in range(len(garment_posed))]
+
+    garment_posed_fixed = [mesh_posed_fixed[i, mask[i], :] for i in range(body.shape[0])]
+    collision_fixed = [np.concatenate((body[i], garment_posed_fixed[i]), axis=0) for i in range(len(garment_posed))]
+    collision_gt = [np.concatenate((body[i], gt_posed[i]), axis=0) for i in range(len(garment_posed))]
+
+    figs = { type + "/garment_" + str(i): plot_3d(gt[i], garment_unposed[i], faces[i], collision[i], collision_faces[i], collision_color[i]) for i in range(4) }
+    figs_coll = { type + "/garment_" + str(i) + '_collision': plot_3d_collision(collision_gt[i], collision[i], collision_fixed[i], collision_faces[i], collision_color[i]) for i in range(4) }
+
+    return { **figs, **figs_coll, **{ type + '/images': imgs, } }
+
+
+def train():
+    valid_iterator = iter(valid_dataloader)
+
+    step = 0
+    for batch in train_dataloader:
+        train_loss = train_step(batch)
+        train_accumulator.update(train_loss)
+
+        if step % 25 == 0:
+            data_valid = next(valid_iterator)
+            valid_loss = valid_step(data_valid)
+            valid_accumulator.update(valid_loss)
+
+        if step % 1000 == 0 or (step < 1000 and step % 100 == 0):
+            to_log = dict({ **train_loss, **valid_loss, **{ 'current_step': step } })
+            if step % 5000 == 0 or (step < 5000 and step % 1000 == 0):
+                train_figs = plot_everything(batch, 'train')
+                valid_figs = plot_everything(data_valid, 'valid')
+                to_log = { **to_log, **train_figs, **valid_figs }
+            if step == 0:
+                to_log = { **to_log, **{ 'model/summary': wandb.Html(get_summary_html(model)) } }
+
+            if not args['debug']:
+                wandb.log(to_log, step=step)
+            train_accumulator.reset()
+            valid_accumulator.reset()
+            print(step)
+
+        if step % 10000 == 0:
+            ckpt_saver_enc(model.encoder, step, optimizer, scaler)
+            ckpt_saver_dec(model.decoder, step, optimizer, scaler)
+
+        step += 1
+
+
+if __name__ == '__main__':
+    if not args['debug']:
+        init_wandb()
+    
+    use_amp = True
+    skirt_modifier = '_skirt' if config['is_skirt'] else ''
+
+    # dataset = Dataset('/data-local/127/train', config['signature'], resolutions=[512], device=device, frames_file='frames_jumpsuit.txt')
+    dataset_train = Dataset(config['train_dir'], config['signature'], device=config['device'], frames_file=config['train_frames'], use_cache=config['dataset_in_memory'], is_cape=config['is_cape'])
+    dataset_valid = Dataset(config['valid_dir'], config['signature'], device=config['device'], frames_file=config['valid_frames'], use_cache=config['dataset_in_memory'], is_cape=config['is_cape'])
+    train_dataloader = DataLoader(dataset_train, batch_size=4, drop_last=True, shuffle=True, )
+    valid_dataloader = DataLoader(dataset_valid, batch_size=4, drop_last=True, shuffle=True)
+    train_accumulator = Accumulator()
+    valid_accumulator = Accumulator()
+
+    model : Model = Model(config)
+    optimizer = optim.Adam(model.parameters(), lr=config['lr'])
+    scaler = torch.cuda.amp.GradScaler(enabled=use_amp)
+    smpl = SMPL(orig_smpl=config['is_cape'], path='smpl', device=config['device'])
+
+    # _MSSSIM = MSSSIM(data_range=2).to(config['device'])
+    _MSSSIM = MS_SSIM(data_range=2, size_average=True, channel=3)
+
+    if config['pretrained_model'] != '':
+        load_model(model.encoder, config['pretrained_model'] + '_enc')
+        load_model(model.decoder, config['pretrained_model'] + '_dec')
+
+    ckpt_saver_enc = CheckpointSaver('checkpoints/iccv/torch/' + config['name'] + '_enc', top_n=5)
+    ckpt_saver_dec = CheckpointSaver('checkpoints/iccv/torch/' + config['name'] + '_dec', top_n=5)
+
+    run_name = config['name']
+    print(f'started training: {run_name}')
+
+    if not args['debug']:
+        # model = torch.compile(model)
+        wandb.watch(model, log='all', log_freq=1000, log_graph=True)
+
+    train()
+
+
+