diff --git a/configs/abl_all.py b/configs/abl_all.py new file mode 100644 index 0000000..972b78a --- /dev/null +++ b/configs/abl_all.py @@ -0,0 +1,194 @@ +from utils.convnext_utils import CNBlockConfig + +config = dict() + +config['gpu'] = '0' +config['name'] = 'abl_torch_all' +config['notes'] = '' +config['tags'] = ['iccv', 'abl_all', 'torch'] +config['script'] = 'scripts/cvae.py' + + +config['use_skinning'] = True # and deshaping +config['remove_global_orientation'] = True +config['use_curvature'] = True +config['use_extra_losses'] = True # edge, collision, 3d mesh +config['use_offsets'] = True + +config['use_norm_loss'] = True + +ablation_count = [config['use_skinning'], config['remove_global_orientation'], config['use_curvature'], config['use_extra_losses'], config['use_offsets']].count(False) +assert ablation_count == 1 or ablation_count == 0 + +config['avg_garment_path'] = 'custom_data/avg/train_no_shape_avg_uv.npy' #custom_data/train_no_shape_avg_uv.npy # train_no_shape_avg_uv train_no_shape_one_hot_avg_uv +if config['use_skinning']: + assert (config['use_offsets'] and config['avg_garment_path'] != '') or (not config['use_offsets'] and config['avg_garment_path'] == '') + +config['reg_loss_coeff'] = 0.1 # set to 0 if don't want to use edge+collision + +config['batch_size'] = 4 +config['resolution'] = 512 +config['is_skirt'] = False +config['dataset_in_memory'] = False +config['is_cape'] = False +config['device'] = 'cuda' + +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['static_model'] = 'checkpoints/iccv/static_torch_orig_2' +config['body_model'] = 'checkpoints/iccv/body_torch_1' +config['curvature_model'] = 'checkpoints/iccv/curv_torch_1' + + +config['smpl_res'] = 14475 +if config['is_cape']: + config['smpl_res'] = 3638 + +config['latent_dim'] = 128 + +config['pretrained_model'] = '' + +############### +config['lr'] = 0.0001# 0.0001 +############### +config['starting_balancing_coeff'] = 0.0001 + +config['signature']: set = { 'id', + 'frame_nr', + 'pose_shape', + 'uv_body_posed', + 'mesh_body_posed', + 'mesh_posed', + 'uv_curvature', + 'uv_static', + 'mesh_mask', + 'uv_mask', + } + +if config['use_skinning']: + config['signature'].add('uv_unposed') + config['signature'].add('mesh_unposed') +else: + config['signature'].add('uv_posed') + +config['upscale_enhancer'] = '' +config['freeze_encoders'] = True +config['use_one_hot'] = False +config['bottleneck_scale'] = 2 + +config['cvae'] = dict() +config['cvae']['device'] = config['device'] +config['cvae']['is_label'] = False +config['cvae']['input_features'] = 3+3+3+7 if config['use_curvature'] else 3+3+3 +config['cvae']['kernel_size'] = 7 +config['cvae']['bottleneck_size'] = 64 * config['bottleneck_scale'] +config['cvae']['latent_dim'] = config['cvae']['bottleneck_size'] + +block_setting = [ + CNBlockConfig(32, 64, 3), + CNBlockConfig(64, 128, 3), + CNBlockConfig(128, 256, 3), + CNBlockConfig(256, config['cvae']['bottleneck_size'], 9), + CNBlockConfig(config['cvae']['bottleneck_size'], None, 3), +] + +block_setting_reverse = [ + CNBlockConfig(config['cvae']['bottleneck_size'] * 2, 256, 3), + CNBlockConfig(256, 128, 9), + CNBlockConfig(128, 64, 9), + CNBlockConfig(64, 32, 3), + CNBlockConfig(32, 3, 3), +] + +config['cvae']['encoder_block_settings'] = block_setting +config['cvae']['decoder_block_settings'] = block_setting_reverse + + + +####################### cond_ae_curv ############################ +config['cond_ae_curv'] = dict() +config['cond_ae_curv']['device'] = config['device'] +config['cond_ae_curv']['is_label'] = True +config['cond_ae_curv']['latent_dim'] = config['latent_dim'] +config['cond_ae_curv']['kernel_size'] = 7 +config['cond_ae_curv']['bottleneck_size'] = 40 * config['bottleneck_scale'] + +block_setting = [ + CNBlockConfig(16, 32, 3), + CNBlockConfig(32, 64, 3), + CNBlockConfig(64, 128, 9), + CNBlockConfig(128, config['cond_ae_curv']['bottleneck_size'], 9), + CNBlockConfig(config['cond_ae_curv']['bottleneck_size'], None, 3), +] + +block_setting_reverse = [ + CNBlockConfig(config['cond_ae_curv']['bottleneck_size'], 128, 3), + CNBlockConfig(128, 64, 9), + CNBlockConfig(64, 32, 9), + CNBlockConfig(32, 16, 3), + CNBlockConfig(16, 3, 3), +] + +config['cond_ae_curv']['encoder_block_settings'] = block_setting +config['cond_ae_curv']['decoder_block_settings'] = block_setting_reverse +config['cond_ae_curv']['activation'] = 'tanh' + +####################### cond_ae_body ############################ +config['cond_ae_body'] = dict() +config['cond_ae_body']['device'] = config['device'] +config['cond_ae_body']['is_label'] = False +config['cond_ae_body']['latent_dim'] = config['latent_dim'] +config['cond_ae_body']['kernel_size'] = 7 +config['cond_ae_body']['bottleneck_size'] = 16 * config['bottleneck_scale'] + +block_setting = [ + CNBlockConfig(16, 32, 3), + CNBlockConfig(32, 64, 3), + CNBlockConfig(64, 128, 9), + CNBlockConfig(128, config['cond_ae_body']['bottleneck_size'], 9), + CNBlockConfig(config['cond_ae_body']['bottleneck_size'], None, 3), +] + +block_setting_reverse = [ + CNBlockConfig(config['cond_ae_body']['bottleneck_size'], 128, 3), + CNBlockConfig(128, 64, 9), + CNBlockConfig(64, 32, 9), + CNBlockConfig(32, 16, 3), + CNBlockConfig(16, 3, 3), +] + +config['cond_ae_body']['encoder_block_settings'] = block_setting +config['cond_ae_body']['decoder_block_settings'] = block_setting_reverse +config['cond_ae_body']['activation'] = 'tanh' + +####################### cond_ae_static ############################ +config['cond_ae_static'] = dict() +config['cond_ae_static']['device'] = config['device'] +config['cond_ae_static']['is_label'] = False +config['cond_ae_static']['latent_dim'] = config['latent_dim'] +config['cond_ae_static']['kernel_size'] = 7 +config['cond_ae_static']['bottleneck_size'] = 8 * config['bottleneck_scale'] + +block_setting = [ + CNBlockConfig(8, 16, 3), + CNBlockConfig(16, 32, 3), + CNBlockConfig(32, 64, 3), + CNBlockConfig(64, config['cond_ae_static']['bottleneck_size'], 9), + CNBlockConfig(config['cond_ae_static']['bottleneck_size'], None, 3), +] + +block_setting_reverse = [ + CNBlockConfig(config['cond_ae_static']['bottleneck_size'], 64, 3), + CNBlockConfig(64, 32, 9), + CNBlockConfig(32, 16, 3), + CNBlockConfig(16, 8, 3), + CNBlockConfig(8, 3, 3), +] + +config['cond_ae_static']['encoder_block_settings'] = block_setting +config['cond_ae_static']['decoder_block_settings'] = block_setting_reverse +config['cond_ae_static']['activation'] = 'tanh' \ No newline at end of file diff --git a/custom_data/avg/train_no_shape_avg_uv.npy b/custom_data/avg/train_no_shape_avg_uv.npy new file mode 100644 index 0000000..7868198 Binary files /dev/null and b/custom_data/avg/train_no_shape_avg_uv.npy differ diff --git a/custom_data/avg/train_no_shape_skirt_avg_uv.npy b/custom_data/avg/train_no_shape_skirt_avg_uv.npy new file mode 100644 index 0000000..ac60c6e Binary files /dev/null and b/custom_data/avg/train_no_shape_skirt_avg_uv.npy differ diff --git a/models/base.py b/models/base.py index bd4b595..406d3d7 100644 --- a/models/base.py +++ b/models/base.py @@ -6,22 +6,24 @@ from typing import Any, Callable, List, Optional, Sequence class VariationalEncoder(nn.Module): - def __init__(self, block_setting: CNBlockConfig, kernel_size: int = 7): + def __init__(self, block_setting: List[CNBlockConfig], kernel_size: int = 7, input_features: int = 16, device: str = 'cuda'): super(VariationalEncoder, self).__init__() - self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size) + self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, input_features=input_features, device=device) - self.mean_branch = nn.Sequential([CNBlock(block_setting[-1].out_channels, kernel_size=kernel_size) for _ in range(5)]) - self.var_branch = nn.Sequential([CNBlock(block_setting[-1].out_channels, kernel_size=kernel_size) for _ in range(5)]) + 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 call(self, inputs, training): + def forward(self, inputs): x = inputs - x = self.encoder(x, training=training) + x = self.encoder(x) mean = self.mean_branch(x) log_var = self.var_branch(x) diff --git a/models/convnext_orig.py b/models/convnext_orig.py index 01b0a84..b466bf5 100644 --- a/models/convnext_orig.py +++ b/models/convnext_orig.py @@ -66,6 +66,7 @@ class ConvNeXtBase(nn.Module): kernel_size: int = 7, is_label: bool = False, nr_of_labels: Optional[int] = None, + input_features: int = 3, ) -> None: super().__init__() @@ -85,7 +86,7 @@ class ConvNeXtBase(nn.Module): # Stem firstconv_output_channels = block_setting[0].input_channels self.stem = Conv2dNormActivation( - 3 if not is_label else nr_of_labels, + input_features if not is_label else nr_of_labels, firstconv_output_channels, kernel_size=4, stride=4, @@ -229,6 +230,7 @@ def _convnext( is_label: bool, nr_of_labels: Optional[int], kernel_size: int, + input_features: int, device='cuda', **kwargs: Any, ) -> ConvNeXtBase: @@ -236,17 +238,17 @@ def _convnext( if reversed: model = ConvNeXtReverseBase(block_setting, stochastic_depth_prob=stochastic_depth_prob, is_label=is_label, nr_of_labels=nr_of_labels, kernel_size=kernel_size, device=device, **kwargs) else: - model = ConvNeXtBase(block_setting, stochastic_depth_prob=stochastic_depth_prob, kernel_size=kernel_size, is_label=is_label, nr_of_labels=nr_of_labels, device=device, **kwargs) + model = ConvNeXtBase(block_setting, stochastic_depth_prob=stochastic_depth_prob, kernel_size=kernel_size, is_label=is_label, nr_of_labels=nr_of_labels, input_features=input_features, device=device, **kwargs) return model class ConvNeXt(nn.Module): - def __init__(self, block_setting, stochastic_depth_prob=0.0, reversed=False, is_label=False, nr_of_labels=None, kernel_size=7, device='cuda', **kwargs) -> None: + def __init__(self, block_setting, stochastic_depth_prob=0.0, reversed=False, is_label=False, nr_of_labels=None, kernel_size=7, input_features=3, device='cuda', **kwargs) -> None: super().__init__() - self.model = _convnext(block_setting, stochastic_depth_prob, reversed, is_label, nr_of_labels, kernel_size, device=device, **kwargs) + self.model: nn.Module = _convnext(block_setting, stochastic_depth_prob, reversed, is_label, nr_of_labels, kernel_size, input_features=input_features, device=device, **kwargs) def forward(self, x: Tensor) -> Tensor: x = self.model(x) diff --git a/models/cvae_model.py b/models/cvae_model.py index f851741..a92cd2a 100644 --- a/models/cvae_model.py +++ b/models/cvae_model.py @@ -1,63 +1,54 @@ import os import time import random -import tensorflow as tf +import torch +from torch import nn import numpy as np import scipy as sp # import plotly.graph_objects as go # import plotly.express as ex -from custom_utils.custom_utils import load_model +from utils.checkpoints import load_model from models.base import AEEncoder, VariationalEncoder, Decoder -from models.base import AEEncoderNew, VariationalEncoderNew, DecoderNew -class CVAE(tf.keras.Model): +class CVAE(nn.Module): def __init__(self, config): super(CVAE, self).__init__() - self.latent_dim = config['cvae']['latent_dim'] self.size = config['resolution'] if not('only_curvatures' in config.keys() and config['only_curvatures'] == True): - self.static_enc : AEEncoder = AEEncoder(config['cond_ae']['latent_dim'], config['cond_ae']['encoder_depths'], config['cond_ae']['encoder_dims'], config['cond_ae']['bottleneck_size'], config['cond_ae']['kernel_size']) + self.static_enc : AEEncoder = AEEncoder(config['cond_ae_static']['encoder_block_settings'], config['cond_ae_static']['kernel_size'], config['cond_ae_static']['activation'], device=config['device']) if config['static_model'] != '': load_model(self.static_enc, config['static_model'] + '_enc') - self.body_enc : AEEncoder = AEEncoder(config['cond_ae_body']['latent_dim'], config['cond_ae_body']['encoder_depths'], config['cond_ae_body']['encoder_dims'], config['cond_ae_body']['bottleneck_size'], config['cond_ae_body']['kernel_size']) + self.body_enc : AEEncoder = AEEncoder(config['cond_ae_body']['encoder_block_settings'], config['cond_ae_body']['kernel_size'], config['cond_ae_body']['activation'], device=config['device']) if config['body_model'] != '': load_model(self.body_enc, config['body_model'] + '_enc') if config['use_curvature']: - self.curv_enc : AEEncoder = AEEncoder(config['cond_ae']['latent_dim'], config['cond_ae']['encoder_depths'], config['cond_ae']['encoder_dims'], config['cond_ae']['bottleneck_size'], config['cond_ae']['kernel_size']) + self.curv_enc : AEEncoder = AEEncoder(config['cond_ae_curv']['encoder_block_settings'], config['cond_ae_curv']['kernel_size'], config['cond_ae_curv']['activation'], device=config['device'], is_label=True, nr_of_labels=7) if config['curvature_model'] != '': load_model(self.curv_enc, config['curvature_model'] + '_enc') if config['freeze_encoders']: - if not('only_curvatures' in config.keys() and config['only_curvatures'] == True): - self.static_enc.trainable = False - self.body_enc.trainable = False + for param in self.static_enc.parameters(): param.requires_grad = False + for param in self.body_enc.parameters(): param.requires_grad = False if config['use_curvature']: - self.curv_enc.trainable = False + for param in self.curv_enc.parameters(): param.requires_grad = False - self.cond_encoder = tf.keras.Sequential([tf.keras.layers.Dense(512, activation='relu'), - tf.keras.layers.Dense(256, activation='relu'), - tf.keras.layers.Dense(128, activation='relu'), - # tf.keras.layers.Lambda(lambda x: x / 1000), - tf.keras.layers.Dense(config['latent_dim'], activation='tanh')]) + self.encoder = VariationalEncoder(config['cvae']['encoder_block_settings'], config['cvae']['kernel_size'], config['cvae']['input_features'], device=config['device']) + self.decoder : Decoder = Decoder(config['cvae']['decoder_block_settings'], kernel_size=config['cvae']['kernel_size'], device=config['device']) - self.encoder = VariationalEncoder(config['cvae']['latent_dim'], config['cvae']['encoder_depths'], config['cvae']['encoder_dims'], kernel_size=config['cvae']['kernel_size']) - self.decoder = Decoder(config['cvae']['decoder_input'], config['cvae']['decoder_depths'], config['cvae']['decoder_dims'], config['cvae']['kernel_size']) + self.latent_shape = (config['cvae']['latent_dim'], 8, 4) if config['avg_garment_path'] != '': - self.avg_garment = np.load(config['avg_garment_path']) + self.avg_garment = torch.permute(torch.tensor(np.load(config['avg_garment_path']), device=config['device']), [2,0,1]) else: self.avg_garment = None - self.add_noise = config['cvae']['add_noise'] - self.noise_amount = config['cvae']['noise_amount'] self.use_curvature = config['use_curvature'] - self.only_curvatures = config['only_curvatures'] if 'only_curvatures' in config.keys() and config['only_curvatures'] == True else None self.use_offsets_over_body = not config['use_skinning'] if self.use_offsets_over_body: @@ -68,7 +59,7 @@ class CVAE(tf.keras.Model): # @tf.function def sample(self, uv_template, uv_body_posed, uv_curv, uv_mask, noise=None, add_noise=False): if noise is None: - eps = tf.random.normal(shape=(uv_template.shape[0], self.latent_dim)) + eps = tf.random.normal(shape=(uv_template.shape[0], 8, 4, self.latent_dim)) else: eps = noise @@ -91,195 +82,49 @@ class CVAE(tf.keras.Model): return x - def reparameterize(self, mean, logvar, training, fix_noise=False): + def reparameterize(self, mean: torch.Tensor, logvar: torch.Tensor, training: bool = True, fix_noise: bool = False): if training == True: if not fix_noise: - eps = tf.random.normal(shape=mean.shape) + eps = torch.randn(size=mean.shape).to(mean.device) else: - eps = tf.random.normal(shape=(1, mean.shape[1])) - eps = tf.tile(eps, [mean.shape[0],1]) - out = eps * tf.exp(logvar * .5) + mean + eps = torch.randn(size=(1, mean.shape[1])) + eps = torch.tile(eps, [mean.shape[0],1]) + out = eps * torch.exp(logvar * 0.5) + mean else: - out = tf.exp(logvar * .5) + mean + out = torch.exp(logvar * 0.5) + mean return out - def encode_conds(self, uv_template, uv_body_posed, uv_curv, training): - if self.only_curvatures: - z_cond = self.curv_enc(uv_curv, training=training) + def encode_conds(self, uv_template, uv_body_posed, uv_curv): + z_body = self.body_enc(uv_body_posed) + z_static = self.static_enc(uv_template) + if self.use_curvature: + z_curv = self.curv_enc(uv_curv) + z_cond = torch.concat([z_static, z_body, z_curv], dim=1) else: - z_body = self.body_enc(uv_body_posed, training=training) - z_static = self.static_enc(uv_template, training=training) - if self.use_curvature: - z_curv = self.curv_enc(uv_curv, training=training) - z_cond = tf.concat([z_static, z_body, z_curv], axis=-1) - else: - z_cond = tf.concat([z_static, z_body], axis=-1) - - z_cond = self.cond_encoder(z_cond, training=training) + z_cond = torch.concat([z_static, z_body], dim=1) return z_cond - def call(self, inputs, training): + def forward(self, inputs): uv_unposed, uv_template, uv_body_posed, uv_curv, uv_mask = inputs - if self.only_curvatures: - x = tf.concat([uv_unposed, uv_curv], axis=-1) - else: - x = tf.concat([uv_unposed, uv_template, uv_body_posed, uv_curv], axis=-1) + x = torch.concat([uv_unposed, uv_template, uv_body_posed, uv_curv], dim=1) - mean, logvar = self.encoder(x, training=training) - z = self.reparameterize(mean, logvar, training) - z_cond = self.encode_conds(uv_template, uv_body_posed, uv_curv, training=training) + mean, logvar = self.encoder(x) + z: torch.Tensor = self.reparameterize(mean, logvar) + z = z.view((mean.shape[0], *self.latent_shape)) + z_cond = self.encode_conds(uv_template, uv_body_posed, uv_curv) - if self.add_noise and training == True: - z_cond = z_cond + (tf.random.uniform(z_cond.shape, maxval=self.noise_amount * 2) - self.noise_amount) - - - z = tf.concat([z, z_cond], axis=-1) - x = self.decoder(z, training=training) + z = torch.concat([z, z_cond], dim=1) + x = self.decoder(z) if self.avg_garment is not None: - x = x + self.avg_garment[tf.newaxis,:,:,:] + x = x + self.avg_garment[None,:,:,:] if self.use_offsets_over_body: x = x + uv_body_posed - x = x * uv_mask[:,:,:,tf.newaxis] - - return x, mean, logvar - - - - -class CVAENew(tf.keras.Model): - - def __init__(self, config): - super(CVAENew, self).__init__() - self.size = config['resolution'] - - if not('only_curvatures' in config.keys() and config['only_curvatures'] == True): - self.static_enc : AEEncoderNew = AEEncoderNew(config['cond_ae_static']['encoder_depths'], config['cond_ae_static']['encoder_dims'], config['cond_ae_static']['kernel_size'], config['cond_ae_static']['activation']) - if config['static_model'] != '': - load_model(self.static_enc, config['static_model'] + '_enc') - - self.body_enc : AEEncoderNew = AEEncoderNew(config['cond_ae_body']['encoder_depths'], config['cond_ae_body']['encoder_dims'], config['cond_ae_body']['kernel_size']) - if config['body_model'] != '': - load_model(self.body_enc, config['body_model'] + '_enc') - - if config['use_curvature']: - self.curv_enc : AEEncoderNew = AEEncoderNew(config['cond_ae_curv']['encoder_depths'], config['cond_ae_curv']['encoder_dims'], config['cond_ae_curv']['kernel_size'], config['cond_ae_curv']['activation']) - if config['curvature_model'] != '': - load_model(self.curv_enc, config['curvature_model'] + '_enc') - - if config['freeze_encoders']: - if not('only_curvatures' in config.keys() and config['only_curvatures'] == True): - self.static_enc.trainable = False - self.body_enc.trainable = False - if config['use_curvature']: - self.curv_enc.trainable = False - - self.encoder = VariationalEncoderNew(config['cvae']['encoder_depths'], config['cvae']['encoder_dims'], kernel_size=config['cvae']['kernel_size']) - self.decoder = DecoderNew(config['cvae']['decoder_depths'], config['cvae']['decoder_dims'], config['cvae']['kernel_size']) - - self.reshape = tf.keras.layers.Reshape((8, 4, 64)) - - if config['avg_garment_path'] != '': - self.avg_garment = np.load(config['avg_garment_path']) - else: - self.avg_garment = None - - self.add_noise = config['cvae']['add_noise'] - self.noise_amount = config['cvae']['noise_amount'] - self.use_curvature = config['use_curvature'] - self.only_curvatures = config['only_curvatures'] if 'only_curvatures' in config.keys() and config['only_curvatures'] == True else None - self.use_offsets_over_body = not config['use_skinning'] - - if self.use_offsets_over_body: - assert self.avg_garment is None, 'When we are not using skinning, we can\'t use avg_garment!' - if self.avg_garment is not None: - assert not self.use_offsets_over_body, 'When avg_garment is given, we must use skinning!' - - # @tf.function - def sample(self, uv_template, uv_body_posed, uv_curv, uv_mask, noise=None, add_noise=False): - if noise is None: - eps = tf.random.normal(shape=(uv_template.shape[0], self.latent_dim)) - else: - eps = noise - - z_cond = self.encode_conds(uv_template, uv_body_posed, uv_curv, training=False) - if add_noise: - z_cond = z_cond + (tf.random.uniform(z_cond.shape, maxval=self.noise_amount * 2) - self.noise_amount) - - z = tf.concat([eps, z_cond], axis=-1) - - x = self.decoder(z, training=False) - - if self.avg_garment is not None: - x = x + self.avg_garment[tf.newaxis,:,:,:] - - if self.use_offsets_over_body: - x = x + uv_body_posed - - x = x * uv_mask[:,:,:,tf.newaxis] - - return x - - - def reparameterize(self, mean, logvar, training, fix_noise=False): - if training == True: - if not fix_noise: - eps = tf.random.normal(shape=mean.shape) - else: - eps = tf.random.normal(shape=(1, mean.shape[1])) - eps = tf.tile(eps, [mean.shape[0],1]) - out = eps * tf.exp(logvar * .5) + mean - else: - out = tf.exp(logvar * .5) + mean - return out - - - def encode_conds(self, uv_template, uv_body_posed, uv_curv, training): - if self.only_curvatures: - z_cond = self.curv_enc(uv_curv, training=training) - else: - z_body = self.body_enc(uv_body_posed, training=training) - z_static = self.static_enc(uv_template, training=training) - if self.use_curvature: - z_curv = self.curv_enc(uv_curv, training=training) - z_cond = tf.concat([z_static, z_body, z_curv], axis=-1) - else: - z_cond = tf.concat([z_static, z_body], axis=-1) - - return z_cond - - - def call(self, inputs, training): - uv_unposed, uv_template, uv_body_posed, uv_curv, uv_mask = inputs - if self.only_curvatures: - x = tf.concat([uv_unposed, uv_curv], axis=-1) - else: - x = tf.concat([uv_unposed, uv_template, uv_body_posed, uv_curv], axis=-1) - - mean, logvar = self.encoder(x, training=training) - z = self.reparameterize(mean, logvar, training) - z = self.reshape(z) - z_cond = self.encode_conds(uv_template, uv_body_posed, uv_curv, training=training) - - if self.add_noise and training == True: - z_cond = z_cond + (tf.random.uniform(z_cond.shape, maxval=self.noise_amount * 2) - self.noise_amount) - - - z = tf.concat([z, z_cond], axis=-1) - x = self.decoder(z, training=training) - - if self.avg_garment is not None: - x = x + self.avg_garment[tf.newaxis,:,:,:] - - if self.use_offsets_over_body: - x = x + uv_body_posed - - x = x * uv_mask[:,:,:,tf.newaxis] - - return x, mean, logvar + x = x * uv_mask[:,None,:,:] + return x, mean, logvar \ No newline at end of file diff --git a/scripts/cvae.py b/scripts/cvae.py index 27637a3..8e9f118 100644 --- a/scripts/cvae.py +++ b/scripts/cvae.py @@ -1,21 +1,20 @@ import os import sys -from configs.config_abl_all_post import config +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) +args = vars(parser.parse_args()) + +config_module = importlib.import_module(args['config']) +config = config_module.config + +os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = config['gpu'] -import tensorflow as tf - -gpus = tf.config.experimental.list_physical_devices('GPU') -if gpus: - try: - # Currently, memory growth needs to be the same across GPUs - for gpu in gpus: - tf.config.experimental.set_memory_growth(gpu, True) - logical_gpus = tf.config.experimental.list_logical_devices('GPU') - print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs") - except RuntimeError as e: - # Memory growth must be set before GPUs have been initialized - print(e) +# os.environ["CUDA_LAUNCH_BLOCKING"] = '1' import time import argparse @@ -26,256 +25,205 @@ import numpy as np import plotly.graph_objects as go import plotly.express as ex from PIL import Image -from utils.postprocess import fix_collision if config['is_skirt']: - from utils.losses import edge_loss_skirt as edge_loss + from losses.losses import edge_loss_skirt as edge_loss else: - from utils.losses import edge_loss + from losses.losses import edge_loss -from utils.losses import collision_ratio, kl_loss, normal_loss, collision_loss -from custom_utils.custom_utils import batch_to_dict, load_model, plot_3d, UV2Mesh, get_summary_html, maskFacesBatch, _maskFaces, plot_3d_collision -from data.dataset import DatasetFactory, StaticDataset -from custom_utils.smpl import SMPL -from models.cvae_model import CVAENew as CVAE -from utils.utils import Accumulator +import torch +import torch.nn as nn +import torch.optim as optim +from data.dataset import Dataset +from models.cond_encoder import AutoEncoder +from utils.checkpoints import CheckpointSaver, load_model +from utils.utils import InfiniteDataLoader as DataLoader +from utils.postprocess import fix_collision +from losses.losses import collision_loss, normal_loss +from pytorch_msssim import MS_SSIM +from models.cvae_model import CVAE +from smpl.smpl_torch import SMPL +from custom_utils.custom_utils import UV2Mesh, Accumulator, plot_3d, get_summary_html, maskFacesBatch, get_summary_str, plot_3d_collision def init_wandb(): - run = wandb.init(project='ablation', name=config['name'], notes=config['notes'], dir='/wandb', tags=config['tags']) + run = wandb.init(project='iccv', name=config['name'], notes=config['notes'], dir='/tmp', tags=config['tags']) artifact = wandb.Artifact('code', type='code') - config_file_names = [key for key in sys.modules.keys() if (key.startswith('configs.') or key.startswith('custom_utils.') or key.startswith('utils.') or key.startswith('data.') or key.startswith('models.') or key.startswith('scripts.'))] + artifact.add_file('scripts/cvae.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('models/cvae_model.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_adaptive=False, is_smpl_orig=config['is_cape']) +uv2mesh = UV2Mesh(res=256, is_skirt=config['is_skirt'], is_smpl_orig=config['is_cape']) -def kl_loss(mean, logvar): - kl_loss = 1 + logvar - tf.square(mean) - tf.exp(logvar) - kl_loss = tf.reduce_mean(kl_loss) +def kl_loss(mean: torch.Tensor, logvar: torch.Tensor) -> torch.Tensor: + kl_loss = 1 + logvar - torch.square(mean) - torch.exp(logvar) + kl_loss = torch.mean(kl_loss) kl_loss *= -0.5 return kl_loss -# 3d loss on posed !!!!!!!!!!!! - -# @tf.function def train_step(batch): + if config['use_skinning']: + mesh_orig, uv_orig = batch['mesh_unposed'], batch['uv_unposed'] + else: + mesh_orig, uv_orig = batch['mesh_posed'], batch['uv_posed'] - with tf.GradientTape() as tape: - if config['use_skinning']: - mesh_orig, uv_orig = batch['mesh_unposed'], batch['uv_unposed'] - else: - mesh_orig, uv_orig = batch['mesh_posed'], batch['uv_posed'] + model.train() + pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']]) + mesh_pred = uv2mesh(pred) - pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']], training=True) + if config['use_skinning']: + 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])]) + # shape_offsets = torch.stack([torch.matmul(shape[i, :10], smpl.shapedirs[gender[i]][:, smpl.to_keep]) for i in range(shape.shape[0])]) - mesh_pred = uv2mesh(pred) + 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] + else: + mesh_pred_posed = mesh_pred - if config['use_skinning']: - shape = batch['pose_shape'][:, 72:] - gender = tf.cast(tf.greater_equal(shape[:, 10] , 0.5), tf.int32) - shape_offsets = tf.stack([tf.gather(tf.tensordot(shape[i,:10][tf.newaxis, :], smpl.shapedirs[gender[i]], axes=[[1], [2]])[0,:,:], smpl.to_keep, axis=0) for i in range(shape.shape[0])]) - mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,tf.newaxis] - mesh_pred_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped) - else: - mesh_pred_posed = mesh_pred + 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(((uv_orig - 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']) + kl = kl_loss(mean, logvar) - # is the output already masked? + reconstr_loss = l2_loss + if config['use_extra_losses']: + reconstr_loss = reconstr_loss + l2_mesh_loss + (e_loss + c_loss + n_loss) * config['reg_loss_coeff'] - e_loss = tf.reduce_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 = tf.reduce_mean(c_loss), tf.reduce_mean(c_ratio) - l2_loss = tf.reduce_mean(1e3 * tf.reduce_mean(tf.reduce_sum(((uv_orig - pred) ** 2), axis=-1), axis=[1, 2])) - l2_mesh_loss = tf.reduce_mean(1e3 * tf.reduce_mean(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (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']) - # conc_loss = concept_loss(pred, batch['uv_unposed']) - kl = kl_loss(mean, logvar) + loss = reconstr_loss + g_balancing_coeff * kl - reconstr_loss = l2_loss - if config['use_extra_losses']: - if 'block_collision_loss' in config.keys() and config['block_collision_loss'] == True: - reconstr_loss = reconstr_loss + l2_mesh_loss + (e_loss) * config['reg_loss_coeff'] - else: - reconstr_loss = reconstr_loss + l2_mesh_loss + (e_loss + c_loss) * config['reg_loss_coeff'] - if config['use_norm_loss'] and step > 2000: - reconstr_loss = reconstr_loss + n_loss * config['reg_loss_coeff'] + optimizer.zero_grad() + loss.backward() + optimizer.step() - loss = reconstr_loss + g_balancing_coeff * kl - - mean = tf.reduce_mean(mean) - logvar = tf.reduce_mean(logvar) - - # ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred + 5, batch['uv_unposed'] + 5, max_val=10, power_factors=(0.5, 0.2, 0.1, 0.1, 0.1))) - 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']) - gradients = tape.gradient(loss, model.trainable_variables) - optimizer.apply_gradients(zip(gradients, model.trainable_variables)) + model.eval() + with torch.inference_mode(): + mean = torch.mean(mean) + logvar = torch.mean(logvar) + ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_static'] + 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/collision_loss': c_loss, 'train/euclid_error': metric, 'train/reconstr_loss': reconstr_loss, 'train/collision_ratio': c_ratio, 'train/mean': mean, 'train/logvar': logvar, 'train/kl_loss': kl, 'train/var': np.e ** logvar, 'train/l2_loss': l2_loss, 'train/l2_mesh_loss': l2_mesh_loss, 'train/normal_loss': n_loss, - # 'train/ms_ssim': ms_ssim, + 'train/ms_ssim': ms_ssim, } + + -@tf.function def valid_step(batch): if config['use_skinning']: mesh_orig, uv_orig = batch['mesh_unposed'], batch['uv_unposed'] else: mesh_orig, uv_orig = batch['mesh_posed'], batch['uv_posed'] - pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']], training=False) + model.eval() + with torch.inference_mode(): + pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']]) + mesh_pred = uv2mesh(pred) - mesh_pred = uv2mesh(pred) + if config['use_skinning']: + 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])]) + # shape_offsets = torch.stack([torch.matmul(shape[i, :10], smpl.shapedirs[gender[i]][:, smpl.to_keep]) for i in range(shape.shape[0])]) - if config['use_skinning']: - shape = batch['pose_shape'][:, 72:] - gender = tf.cast(tf.greater_equal(shape[:, 10] , 0.5), tf.int32) - shape_offsets = tf.stack([tf.gather(tf.tensordot(shape[i,:10][tf.newaxis, :], smpl.shapedirs[gender[i]], axes=[[1], [2]])[0,:,:], smpl.to_keep, axis=0) 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] + else: + mesh_pred_posed = mesh_pred - mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,tf.newaxis] - mesh_pred_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped) - else: - mesh_pred_posed = mesh_pred - # is the output already masked? + 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(((uv_orig - 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']) + kl = kl_loss(mean, logvar) - e_loss = tf.reduce_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 = tf.reduce_mean(c_loss), tf.reduce_mean(c_ratio) - l2_loss = tf.reduce_mean(1e3 * tf.reduce_mean(tf.reduce_sum(((uv_orig - pred) ** 2), axis=-1), axis=[1, 2])) - l2_mesh_loss = tf.reduce_mean(1e3 * tf.reduce_mean(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (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']) - # conc_loss = concept_loss(pred, batch['uv_unposed']) - kl = kl_loss(mean, logvar) + reconstr_loss = l2_loss + if config['use_extra_losses']: + reconstr_loss = reconstr_loss + l2_mesh_loss + (e_loss + c_loss + n_loss) * config['reg_loss_coeff'] - reconstr_loss = l2_loss - if config['use_extra_losses']: - reconstr_loss = reconstr_loss + l2_mesh_loss + (e_loss + c_loss) * config['reg_loss_coeff'] - if config['use_norm_loss']: - reconstr_loss = reconstr_loss + n_loss * config['reg_loss_coeff'] + loss = reconstr_loss + g_balancing_coeff * kl - loss = reconstr_loss + g_balancing_coeff * kl - - mean = tf.reduce_mean(mean) - logvar = tf.reduce_mean(logvar) - - # ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred + 5, batch['uv_unposed'] + 5, max_val=10, power_factors=(0.5, 0.2, 0.1, 0.1, 0.1))) - 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']) - # c_ratio = tf.reduce_mean(collision_ratio(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask'])) + mean = torch.mean(mean) + logvar = torch.mean(logvar) + ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_static'] + 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 { 'valid/loss': loss, 'valid/edge_loss': e_loss, 'valid/collision_loss': c_loss, 'valid/euclid_error': metric, 'valid/reconstr_loss': reconstr_loss, 'valid/l2_mesh_loss': l2_mesh_loss, 'valid/collision_ratio': c_ratio, 'valid/mean': mean, 'valid/logvar': logvar, 'valid/kl_loss': kl, 'valid/var': np.e ** logvar, - 'valid/normal_loss': n_loss,# 'valid/ms_ssim': ms_ssim, + 'valid/normal_loss': n_loss, 'valid/ms_ssim': ms_ssim, } -@tf.function -def test_step(batch): - if config['use_skinning']: - mesh_orig, uv_orig = batch['mesh_unposed'], batch['uv_unposed'] - else: - mesh_orig, uv_orig = batch['mesh_posed'], batch['uv_posed'] - - pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']], training=False) - - mesh_pred = uv2mesh(pred) - - if config['use_skinning']: - shape = batch['pose_shape'][:, 72:] - gender = tf.cast(tf.greater_equal(shape[:, 10] , 0.5), tf.int32) - shape_offsets = tf.stack([tf.gather(tf.tensordot(shape[i,:10][tf.newaxis, :], smpl.shapedirs[gender[i]], axes=[[1], [2]])[0,:,:], smpl.to_keep, axis=0) for i in range(shape.shape[0])]) - - mesh_pred_shaped = mesh_pred + shape_offsets * batch['mesh_mask'][:,:,tf.newaxis] - mesh_pred_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred_shaped) - mesh_orig_shaped = mesh_orig + shape_offsets * batch['mesh_mask'][:,:,tf.newaxis] - mesh_orig_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_orig_shaped) - else: - mesh_pred_posed = mesh_pred - mesh_orig_posed = mesh_orig - - mesh_pred_posed_fixed = fix_collision(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask']) - - - # is the output already masked? - - e_loss = tf.reduce_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 = tf.reduce_mean(c_loss), tf.reduce_mean(c_ratio) - l2_loss = tf.reduce_mean(1e3 * tf.reduce_mean(tf.reduce_sum(((uv_orig - pred) ** 2), axis=-1), axis=[1, 2])) - l2_mesh_loss = tf.reduce_mean(1e3 * tf.reduce_mean(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig - mesh_pred) ** 2, axis=-1), axis=-1)) - # conc_loss = concept_loss(pred, batch['uv_unposed']) - kl = kl_loss(mean, logvar) - - reconstr_loss = l2_loss - if config['use_extra_losses']: - reconstr_loss = reconstr_loss + l2_mesh_loss + (e_loss + c_loss) * config['reg_loss_coeff'] - - loss = reconstr_loss + g_balancing_coeff * kl - - mean = tf.reduce_mean(mean) - logvar = tf.reduce_mean(logvar) - - ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale(pred + 5, uv_orig + 5, max_val=10, power_factors=(0.5, 0.2, 0.1, 0.1, 0.1))) - 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']) - euclid_posed = tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig_posed - mesh_pred_posed) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask']) - euclid_posed_fixed = tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig_posed - mesh_pred_posed_fixed) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask']) - - return { 'test/loss': loss, 'test/edge_loss': e_loss, 'test/collision_loss': c_loss, 'test/euclid_error': metric, 'test/ms_ssim': ms_ssim, - 'test/reconstr_loss': reconstr_loss, 'test/collision_ratio': c_ratio, 'test/l2_mesh_loss': l2_mesh_loss, - 'test/mean': mean, 'test/logvar': logvar, 'test/kl_loss': kl, 'test/var': np.e ** logvar, - 'test/euclid_posed': euclid_posed, 'test/euclid_posed_fixed': euclid_posed_fixed } - - def plot_everything(batch: dict, type: str): if config['use_skinning']: - mesh_orig, uv_orig = batch['mesh_unposed'], batch['uv_unposed'] + mesh_orig, uv_orig = batch['mesh_unposed'].cpu().numpy(), batch['uv_unposed'] else: - mesh_orig, uv_orig = batch['mesh_posed'], batch['uv_posed'] + mesh_orig, uv_orig = batch['mesh_posed'].cpu().numpy(), batch['uv_posed'] - pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']], training=False) + model.eval() + with torch.inference_mode(): + pred, mean, logvar = model([uv_orig, batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature'], batch['uv_mask']]) - mesh_pred = uv2mesh(pred).numpy() + mesh_pred = uv2mesh(pred).detach().cpu().numpy() - orig_img = uv_orig - uv_static_img = batch['uv_static'] - uv_body_posed_img = batch['uv_body_posed'] - pred_img = pred - img = tf.concat((uv_static_img, uv_body_posed_img, orig_img, pred_img), axis=2) - imgs = [wandb.Image((img[i] + 1) / 2.0) for i in range(4)] + orig_img = torch.permute(uv_orig, [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("custom_utils/smpl_faces" + skirt_modifier + ("_orig" if config['is_cape'] else "") + ".npy") - mask = batch['mesh_mask'].numpy().astype(bool) + F = np.load("smpl/smpl_faces" + skirt_modifier + ("_orig" if config['is_cape'] else "") + ".npy") + mask = batch['mesh_mask'].cpu().numpy().astype(bool) if config['use_skinning']: shape = batch['pose_shape'][:, 72:] - gender = tf.cast(tf.greater_equal(shape[:, 10] , 0.5), tf.int32) - shape_offsets = tf.stack([tf.gather(tf.tensordot(shape[i,:10][tf.newaxis, :], smpl.shapedirs[gender[i]], axes=[[1], [2]])[0,:,:], smpl.to_keep, axis=0) for i in range(shape.shape[0])]) + 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.numpy()[i, mask[i], :] for i in range(mask.shape[0])] - gt = [mesh_orig[i].numpy()[mask[i], :] + shape_offsets.numpy()[i, mask[i], :] for i in range(pred.shape[0])] + 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.numpy()[i, mask[i], :] - mesh_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred).numpy() + 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() else: - mesh_posed = mesh_pred - gt = [mesh_orig[i].numpy()[mask[i], :] for i in range(pred.shape[0])] - garment_unposed, faces = maskFacesBatch(mesh_pred, F, mask) + mesh_posed = mesh_pred.cpu().numpy() + gt = [mesh_orig[i][mask[i], :] for i in range(pred.shape[0])] + garment_unposed, faces = maskFacesBatch(mesh_pred.cpu().numpy(), F, mask) - gt_posed = [batch['mesh_posed'].numpy()[i, mask[i], :] for i in range(mask.shape[0])] + gt_posed = [batch['mesh_posed'].cpu().numpy()[i, mask[i], :] for i in range(mask.shape[0])] - body = batch['mesh_body_posed'].numpy() - mesh_posed_fixed = fix_collision(mesh_posed, body, mask).numpy() + 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))] @@ -289,8 +237,8 @@ def plot_everything(batch: dict, type: str): figs = { type + "/garment_" + str(i): plot_3d(gt[i], garment_unposed[i], faces[i], collision[i], collision_faces[i], collision_color[i], flip_xy=(not config['use_skinning'] and config['remove_global_orientation'])) 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], flip_xy=(not config['use_skinning'] and config['remove_global_orientation'])) for i in range(4) } - mean_hist = wandb.Histogram(sequence=mean.numpy().mean(axis=0)) - logvar_hist = wandb.Histogram(sequence=logvar.numpy().mean(axis=0)) + mean_hist = wandb.Histogram(sequence=mean.cpu().numpy().mean(axis=0)) + logvar_hist = wandb.Histogram(sequence=logvar.cpu().numpy().mean(axis=0)) return { **figs, **figs_coll, **{ type + '/images': imgs, type + '/mean_hist': mean_hist, type + '/logvar_hist': logvar_hist } } @@ -414,6 +362,7 @@ def sample(folder): print() + def test(): str_split = config['name'].split('_') test_idx = str_split.index('test') @@ -462,108 +411,71 @@ def test(): -if __name__ == '__main__': - # init_wandb() - parser = argparse.ArgumentParser(description='') - parser.add_argument('--name', type=str) - args = vars(parser.parse_args()) +def train(): + valid_iterator = iter(valid_dataloader) - smpl = SMPL(orig_smpl=config['is_cape']) + 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)) } } + + wandb.log(to_log, step=step) + train_accumulator.reset() + valid_accumulator.reset() + print(step) + + if train_loss['train/euclid_error'] < (config['balancing_activation_limit'] if 'balancing_activation_limit' in config.keys() else 0.017): + if balancing_activated_step == 0: + balancing_activated_step = step + g_balancing_coeff = g_balancing_coeff + 1 / 10000 + + if step % 10000 == 0: + ckpt_saver(model, step, optimizer) + + step += 1 + + +if __name__ == '__main__': + init_wandb() + # parser = argparse.ArgumentParser(description='') + # parser.add_argument('--name', type=str) + # parser.add_argument('--debug', action='store_true') + # args = vars(parser.parse_args()) + + smpl = SMPL(orig_smpl=config['is_cape'], path='smpl', device=config['device']) skirt_modifier = '_skirt' if config['is_skirt'] else '' - factory = DatasetFactory() - - dataset_train = factory.get_dataset(folder=config['train_dir'], - batch_size=config['batch_size'], - signature=config['signature_dynamic'], - resolutions=[config['resolution']], - in_memory=config['dataset_in_memory'], - use_one_hot=config['use_one_hot'], - use_no_orient=config['remove_global_orientation'], - frames_file=config['frames_file'] if 'frames_file' in config.keys() else None, - is_cape=config['is_cape']) - static_dataset_train = StaticDataset(folder=config['train_dir'], - signature=config['signature_static'], - resolutions=[config['resolution']], - in_memory=True, - is_cape=config['is_cape']) - - dataset_valid = factory.get_dataset(folder=config['valid_dir'], - batch_size=config['batch_size'], - signature=config['signature_dynamic'], - resolutions=[config['resolution']], - in_memory=config['dataset_in_memory'], - use_one_hot=config['use_one_hot'], - use_no_orient=config['remove_global_orientation'], - frames_file=config['frames_file_valid'] if 'frames_file_valid' in config.keys() else None, - is_cape=config['is_cape']) - static_dataset_valid = StaticDataset(folder=config['valid_dir'], - signature=config['signature_static'], - resolutions=[config['resolution']], - in_memory=True, - is_cape=config['is_cape']) - + dataset_train = Dataset(config['train_dir'], config['signature'], device=config['device'], frames_file=config['train_frames'], use_cache=config['dataset_in_memory']) + dataset_valid = Dataset(config['valid_dir'], config['signature'], device=config['device'], frames_file=config['valid_frames'], use_cache=config['dataset_in_memory']) + 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 = CVAE(config) - optimizer = tf.optimizers.Adam(learning_rate=config['lr']) + optimizer = optim.Adam(model.parameters(), lr=config['lr']) + _MSSSIM = MS_SSIM(data_range=2, size_average=True, channel=3) balancing_activated_step = 0 g_balancing_coeff = config['starting_balancing_coeff'] - # sample('checkpoints/cvae_13') - # test() - - ckpt = tf.train.Checkpoint(model=model) - manager = tf.train.CheckpointManager(ckpt, 'checkpoints/post/' + config['name'], max_to_keep=20) + ckpt_saver = CheckpointSaver('checkpoints/iccv/' + config['name'], top_n=5) if config['pretrained_model'] != '': load_model(model, config['pretrained_model']) - # load_model(model, 'checkpoints/' + config['name']) - - valid_iterator = iter(dataset_valid) - step = 0 - for batch in dataset_train: - batch = static_dataset_train.expand_batch(batch_to_dict(batch, config['signature_dynamic'])) - - train_loss = train_step(batch) - train_accumulator.update(train_loss) - - data_valid = valid_iterator.get_next() - data_valid = static_dataset_valid.expand_batch(batch_to_dict(data_valid, config['signature_dynamic'])) - valid_loss = valid_step(data_valid) - valid_accumulator.update(valid_loss) - - - if train_accumulator.get()['train/euclid_error'] < (config['balancing_activation_limit'] if 'balancing_activation_limit' in config.keys() else 0.017): - if balancing_activated_step == 0: - balancing_activated_step = step - g_balancing_coeff = g_balancing_coeff + 1 / 10000 - - - if step % 1000 == 0 or (step < 1000 and step % 100 == 0): - to_log = { **train_accumulator.get(), **valid_accumulator.get(), **{ '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)) } } - - to_log = { **to_log, **{ 'train/balancing_coeff': g_balancing_coeff } } - - wandb.log(to_log, step = step) - train_accumulator.reset() - valid_accumulator.reset() - - if step % 5000 == 0: - manager.save(checkpoint_number=step) - - print(step) - - step += 1 - \ No newline at end of file + train() \ No newline at end of file