367 lines
16 KiB
Python
367 lines
16 KiB
Python
import os
|
|
import time
|
|
import argparse
|
|
import wandb
|
|
import random
|
|
import numpy as np
|
|
import scipy as sp
|
|
import tensorflow as tf
|
|
from custom_utils.custom_utils import UV2Mesh, maskFacesBatch
|
|
from tensorboardX import SummaryWriter
|
|
import plotly.graph_objects as go
|
|
import plotly.express as ex
|
|
|
|
from utils import str_to_bool
|
|
from tf_utils import check_tf_version, allow_memory_growth, split_gpu_for_testing
|
|
from load_models import load_generator, load_discriminator
|
|
from losses import d_logistic, d_logistic_r1_reg, g_logistic_non_saturating, g_logistic_ns_pathreg, edge_loss, uv_mask_loss, bend_loss
|
|
# from losses import collision_penalty as collision_penalty
|
|
from losses import collision_loss as collision_penalty
|
|
from convnext.models.convnext_tf import ConvNeXt, ConvNeXtReverse, Block
|
|
from stylegan2_ae.generator import Generator
|
|
from custom_utils.graph_utils import Reconstructor_V4_Decoder as GNNDecoder, sparse_to_tensor
|
|
from custom_utils.graph_utils import Reconstructor_v2 as SimpleReconstructor
|
|
from dataset import DatasetFactory, StaticDataset
|
|
from custom_utils.custom_utils import batch_to_dict, plot_3d
|
|
from custom_utils.smpl import SMPL
|
|
|
|
name='vae_11'
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
|
|
run = wandb.init(project='vae', name=name,
|
|
notes='vae, ms_ssim with range 3. with l1 loss, larger latent_dim=1024. 4 dowscaling steps, 4 upscaling, the image resize for final resolution. upscale from 256*128')
|
|
artifact = wandb.Artifact('code', type='code')
|
|
artifact.add_file('vae.py')
|
|
artifact.add_file('dataset.py')
|
|
artifact.add_file('losses.py')
|
|
artifact.add_file('convnext/models/convnext_tf.py')
|
|
run.log_artifact(artifact)
|
|
|
|
uv2mesh = UV2Mesh(res=256)
|
|
|
|
|
|
class Bottleneck(tf.keras.Model):
|
|
|
|
def __init__(self, dim):
|
|
super(Bottleneck, self).__init__()
|
|
|
|
self.bottleneck = tf.keras.Sequential(
|
|
[Block(dim=dim, drop_path=0,
|
|
layer_scale_init_value=1e-6, prefix=f'bottleneck{j}') for j in range(5)]
|
|
)
|
|
|
|
def call(self, inputs, training=None):
|
|
x = inputs
|
|
|
|
x = self.bottleneck(x, training=training)
|
|
|
|
return x
|
|
|
|
|
|
class Encoder(tf.keras.Model):
|
|
|
|
def __init__(self, latent_dim):
|
|
super(Encoder, self).__init__()
|
|
|
|
self.encoder = ConvNeXt(depths=[3, 3, 9, 3], dims=[32, 64, 128, 128], include_top=False)
|
|
self.flatten = tf.keras.layers.Flatten()
|
|
self.mean = tf.keras.layers.Dense(latent_dim)
|
|
self.log_var = tf.keras.layers.Dense(latent_dim)
|
|
|
|
def call(self, inputs, training=None):
|
|
x = inputs
|
|
|
|
x = self.encoder(x)
|
|
x = self.flatten(x)
|
|
mean = self.mean(x)
|
|
log_var = self.log_var(x)
|
|
|
|
return mean, log_var
|
|
|
|
|
|
class Decoder(tf.keras.Model):
|
|
|
|
def __init__(self):
|
|
super(Decoder, self).__init__()
|
|
|
|
self.dense = tf.keras.layers.Dense(8 * 4 * 512)
|
|
self.reshape = tf.keras.layers.Reshape((8, 4, 512))
|
|
self.decoder = ConvNeXtReverse(depths=list(reversed([3, 3, 9, 3])), dims=list(reversed([128, 128, 64, 16])))
|
|
|
|
|
|
def call(self, inputs, training=None):
|
|
x = inputs
|
|
|
|
x = self.dense(x)
|
|
x = self.reshape(x)
|
|
x = self.decoder(x)
|
|
|
|
return x
|
|
|
|
|
|
class VAE(tf.keras.Model):
|
|
|
|
def __init__(self, latent_dim=1024):
|
|
super(VAE, self).__init__()
|
|
|
|
|
|
self.encoder = Encoder(latent_dim)
|
|
self.decoder = Decoder()
|
|
|
|
|
|
@tf.function
|
|
def sample(self, eps=None):
|
|
if eps is None:
|
|
eps = tf.random.normal(shape=(4, self.latent_dim))
|
|
return self.decode(eps, apply_sigmoid=True)
|
|
|
|
|
|
def reparameterize(self, mean, logvar):
|
|
eps = tf.random.normal(shape=mean.shape)
|
|
return eps * tf.exp(logvar * .5) + mean
|
|
|
|
|
|
def decode(self, z, apply_sigmoid=False):
|
|
logits = self.decoder(z)
|
|
if apply_sigmoid:
|
|
probs = tf.sigmoid(logits)
|
|
return probs
|
|
return logits
|
|
|
|
|
|
def call(self, inputs, training=None):
|
|
x, uv_mask = inputs
|
|
|
|
x = tf.transpose(x, [0,2,3,1])
|
|
|
|
mean, logvar = self.encoder(x)
|
|
z = self.reparameterize(mean, logvar)
|
|
x = self.decoder(z)
|
|
|
|
|
|
# x = tf.transpose(x, [0,3,1,2])
|
|
# x = x * uv_mask[:,tf.newaxis,:,:]
|
|
x = tf.image.resize(x, (512, 256), tf.image.ResizeMethod.LANCZOS5)
|
|
x = x * uv_mask[:,:,:,tf.newaxis]
|
|
|
|
return x, mean, logvar, z
|
|
|
|
|
|
def log_normal_pdf(sample, mean, logvar, raxis=1):
|
|
log2pi = tf.math.log(2. * np.pi)
|
|
res = tf.reduce_sum(-.5 * ((sample - mean) ** 2. * tf.exp(-logvar) + logvar + log2pi), axis=raxis)
|
|
return res
|
|
|
|
|
|
@tf.function
|
|
def train_step(batch):
|
|
|
|
with tf.GradientTape() as tape:
|
|
|
|
pred, mean, logvar, z = model([batch['uv_unposed'], batch['uv_mask']])
|
|
|
|
mesh_pred = uv2mesh(tf.transpose(pred, [0,3,1,2]))
|
|
mesh_orig = batch['mesh_unposed']
|
|
|
|
|
|
e_loss = 1e-3 * tf.reduce_mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
|
|
c_loss = 1e-3 * tf.reduce_mean(collision_penalty(mesh_pred, batch['mesh_body_posed'], batch['mesh_mask']))
|
|
l1_loss = tf.reduce_sum(tf.reduce_sum((tf.math.abs(tf.transpose(batch['uv_unposed'], [0,2,3,1]) - pred)), axis=-1) * batch['uv_mask'], axis=[1, 2])
|
|
# ms_ssim = 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.image.ssim_multiscale(pred + 1, tf.transpose(batch['uv_unposed'], [0,2,3,1]) + 1, max_val=3)
|
|
# ms_ssim = tf.reduce_mean(tf.image.ssim_multiscale((tf.transpose(pred, [0,2,3,1]) + 1) * batch['uv_mask'][:,:,:,tf.newaxis], (tf.transpose(batch['uv_unposed'], [0,2,3,1]) + 1) * batch['uv_mask'][:,:,:,tf.newaxis], max_val=2))
|
|
rfft_loss = 0 #tf.reduce_mean(tf.reduce_sum(tf.math.abs(tf.signal.rfft3d(pred) - tf.signal.rfft3d(batch['uv_unposed'])), axis=[1,2,3]))
|
|
fft_loss = tf.reduce_sum(tf.math.abs(tf.signal.fft3d(tf.complex(pred, tf.zeros_like(pred))) - tf.signal.fft3d(tf.complex(tf.transpose(batch['uv_unposed'], [0,2,3,1]), tf.zeros_like(tf.transpose(batch['uv_unposed'], [0,2,3,1]))))), axis=[1,2,3])
|
|
fft_loss_pc, fft_loss_pc_low, fft_loss_pc_high = 0,0,0 #fft_loss_per_channel(pred, batch['uv_unposed'])
|
|
tv_loss = tf.image.total_variation(pred)
|
|
|
|
reconstr_loss = l1_loss
|
|
logpx_z = -reconstr_loss
|
|
logpz = log_normal_pdf(z, 0., 0.)
|
|
logqz_x = log_normal_pdf(z, mean, logvar)
|
|
loss = -tf.reduce_mean(logpx_z + logpz - logqz_x)
|
|
ms_ssim = tf.reduce_mean(ms_ssim)
|
|
l1_loss = tf.reduce_mean(l1_loss)
|
|
fft_loss = tf.reduce_mean(fft_loss)
|
|
reconstr_loss = tf.reduce_mean(reconstr_loss)
|
|
tv_loss = tf.reduce_mean(tv_loss)
|
|
|
|
|
|
mesh_orig_2 = uv2mesh(batch['uv_unposed'])
|
|
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'])
|
|
metric_2 = tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig_2 - mesh_pred) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask'])
|
|
euclid_posed = 0 # tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_skinned - batch['mesh_posed']) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask'])
|
|
gradients = tape.gradient(loss, model.trainable_variables)
|
|
# gradients = [tf.clip_by_norm(g, 1.0) for g in gradients]
|
|
# gradients = [tf.clip_by_value(g, -1.0, 1.0) for g in gradients]
|
|
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
|
|
return { 'train/loss': loss, 'train/edge_loss': e_loss, 'train/collision_loss': c_loss, 'train/l1_loss': l1_loss, 'train/euclid_error': metric, 'train/ms_ssim': ms_ssim,
|
|
'train/fft_loss': fft_loss, 'train/euclid_error_reconst': metric_2, 'train/rfft_loss': rfft_loss, 'train/fft_oc_low': fft_loss_pc_low, 'train/fft_pc_high': fft_loss_pc_high,
|
|
'train/fft_loss_per_channel': fft_loss_pc, 'train/euclid_error_posed': euclid_posed, 'train/reconstr_loss': reconstr_loss, 'train/tv_loss': tv_loss }
|
|
|
|
|
|
|
|
@tf.function
|
|
def valid_step(batch):
|
|
pred, mean, logvar, z = model([batch['uv_unposed'], batch['uv_mask']])
|
|
|
|
mesh_pred = uv2mesh(tf.transpose(pred, [0,3,1,2]))
|
|
mesh_orig = batch['mesh_unposed']
|
|
e_loss = 1e-2 * tf.reduce_mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
|
|
c_loss = 1e-2 * tf.reduce_mean(collision_penalty(mesh_pred, batch['mesh_body_posed'], batch['mesh_mask']))
|
|
l1_loss = tf.reduce_sum(tf.reduce_sum((tf.math.abs(tf.transpose(batch['uv_unposed'], [0,2,3,1]) - pred)), axis=-1) * batch['uv_mask'], axis=[1, 2])
|
|
b_loss = 0 #1e-3 * bend_loss(mesh_pred, smpl_faces, laplacian_loss, batch['mesh_mask'])
|
|
# ms_ssim = 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.image.ssim_multiscale(pred + 1, tf.transpose(batch['uv_unposed'], [0,2,3,1]) + 1, max_val=3)
|
|
fft_loss = 0 #tf.reduce_sum(tf.math.abs(tf.signal.fft3d(tf.complex(pred, tf.zeros_like(pred))) - tf.signal.fft3d(tf.complex(tf.transpose(batch['uv_unposed'], [0,2,3,1]), tf.zeros_like(tf.transpose(batch['uv_unposed'], [0,2,3,1]))))), axis=[1,2,3])
|
|
# mesh_orig_2 = uv2mesh(batch['uv_unposed'])
|
|
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'])
|
|
metric_2 = 0 #tf.reduce_sum(tf.math.sqrt(tf.reduce_sum(batch['mesh_mask'][:,:,tf.newaxis] * (mesh_orig_2 - mesh_pred) ** 2, axis=-1))) / tf.reduce_sum(batch['mesh_mask'])
|
|
|
|
reconstr_loss = l1_loss
|
|
logpx_z = -reconstr_loss
|
|
logpz = log_normal_pdf(z, 0., 0.)
|
|
logqz_x = log_normal_pdf(z, mean, logvar)
|
|
loss = -tf.reduce_mean(logpx_z + logpz - logqz_x)
|
|
ms_ssim = tf.reduce_mean(ms_ssim)
|
|
l1_loss = tf.reduce_mean(l1_loss)
|
|
fft_loss = tf.reduce_mean(fft_loss)
|
|
b_loss = tf.reduce_mean(b_loss)
|
|
reconstr_loss = tf.reduce_mean(reconstr_loss)
|
|
|
|
return { 'valid/loss': loss, 'valid/edge_loss': e_loss, 'valid/collision_loss': c_loss, 'valid/l1_loss': l1_loss, 'valid/euclid_error': metric, 'valid/ms_ssim': ms_ssim,
|
|
'valid/fft_loss': fft_loss, 'valid/euclid_error_reconst': metric_2, 'train/bend_loss': b_loss, 'train/reconstr_loss': reconstr_loss
|
|
}
|
|
|
|
|
|
def plot_everything(batch: dict, type: str):
|
|
|
|
pred, mean, logvar, z = model([batch['uv_unposed'], batch['uv_mask']])
|
|
|
|
mesh_pred = uv2mesh(tf.transpose(pred, [0,3,1,2]))
|
|
mesh_orig = batch['mesh_unposed']
|
|
|
|
orig_img = tf.transpose(batch['uv_unposed'], [0,2,3,1])
|
|
pred_img = pred
|
|
img = tf.concat((orig_img, pred_img), axis=2)
|
|
imgs = [wandb.Image((img[i] + 1) / 2.0) for i in range(4)]
|
|
|
|
F = np.load("custom_utils/smpl_faces.npy")
|
|
mask = batch['mesh_mask'].numpy().astype(bool)
|
|
garment_unposed, faces = maskFacesBatch(mesh_pred.numpy(), F, mask)
|
|
gt = [mesh_orig[i].numpy()[mask[i], :] for i in range(pred.shape[0])]
|
|
|
|
body = batch['mesh_body_posed'].numpy()
|
|
mesh_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred).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] + 14475), 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))]
|
|
|
|
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) }
|
|
|
|
return { **figs, **{ type + '/images': imgs } }
|
|
|
|
|
|
|
|
|
|
|
|
def repeat_faces(faces, times):
|
|
idx_offset = np.linspace(0, 14475 * (times-1), times).astype(np.int32)
|
|
idx_offset = np.repeat(idx_offset, faces.shape[0], axis=0)
|
|
merged = np.tile(faces, [times,1])
|
|
merged += idx_offset[:, np.newaxis]
|
|
return merged
|
|
|
|
|
|
def repeat_laplacian(L, times):
|
|
idx = np.repeat(np.mat([L.row, L.col]), repeats=times, axis=1)
|
|
idx_offset = np.linspace(0, L.shape[0] * (times-1), times).astype(np.int32)
|
|
idx_offset = np.repeat(idx_offset, L.row.shape[0], axis=0)
|
|
idx += idx_offset[np.newaxis,:]
|
|
data = np.repeat(L.data, repeats=times, axis=0)
|
|
shape = (L.shape[0] * times, L.shape[1] * times)
|
|
return sp.sparse.coo_matrix((data, (np.squeeze(np.asarray(idx[0,:])), np.squeeze(np.asarray(idx[1,:])))), shape)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
allow_memory_growth()
|
|
parser = argparse.ArgumentParser(description='')
|
|
parser.add_argument('--name', type=str)
|
|
args = vars(parser.parse_args())
|
|
|
|
batch_size = 8
|
|
|
|
smpl_faces = np.load('custom_utils/smpl_faces.npy')
|
|
faces_loss_single = smpl_faces
|
|
faces_loss = repeat_faces(smpl_faces, batch_size)
|
|
laplacian_loss = sp.sparse.load_npz("custom_utils/laplacian/laplacian0.npz").todense().astype(np.float32)
|
|
# edges = np.stack(np.nonzero(laplacian_loss.todense())).transpose()
|
|
# laplacian_loss = sparse_to_tensor(laplacian_loss)
|
|
# laplacian_loss = sparse_to_tensor(repeat_laplacian(laplacian_loss, batch_size))
|
|
|
|
smpl = SMPL()
|
|
|
|
factory = DatasetFactory()
|
|
signature_dynamic = { 'id': (),
|
|
'pose_shape': (72 + 11,),
|
|
# 'uv_posed': (3, 512, 256,),
|
|
# 'uv_body_posed': (3, 512, 256,),
|
|
'mesh_body_posed': (14475, 3,),
|
|
'mesh_posed': (14475, 3,),
|
|
'mesh_unposed': (14475, 3,),
|
|
'uv_unposed': (3, 512, 256,)
|
|
}
|
|
signature_static = { #'uv_static': (3, 512, 256,),
|
|
# 'mesh_static': (14475, 3,),
|
|
'mesh_mask': (14475,),
|
|
'uv_mask': (512, 256,),
|
|
# 'uv_shaped_body': (3, 512, 256,),
|
|
}
|
|
|
|
dataset_train = factory.get_dataset(folder='/home/workspace/local-data/jumpsuit_20/train',
|
|
batch_size=batch_size,
|
|
signature=signature_dynamic)
|
|
static_dataset_train = StaticDataset(folder='/home/workspace/local-data/jumpsuit_20/train',
|
|
signature=signature_static)
|
|
dataset_valid = factory.get_dataset(folder='/home/workspace/local-data/jumpsuit_20/valid',
|
|
batch_size=batch_size,
|
|
signature=signature_dynamic)
|
|
static_dataset_valid = StaticDataset(folder='/home/workspace/local-data/jumpsuit_20/valid',
|
|
signature=signature_static)
|
|
|
|
|
|
model = VAE()
|
|
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(0.0001, decay_steps=30000, decay_rate=0.1, staircase=False)
|
|
optimizer = tf.optimizers.Adam(learning_rate=lr_schedule)
|
|
folder = 'models/ae_logs/'
|
|
valid_iterator = iter(dataset_valid)
|
|
|
|
step = 0
|
|
for batch in dataset_train:
|
|
batch = static_dataset_train.expand_batch(batch_to_dict(batch, signature_dynamic))
|
|
|
|
# _, masked_faces = maskFacesBatch(batch['mesh_unposed'].numpy(), smpl_faces, batch['mesh_mask'].numpy().astype(bool))
|
|
# # batch['mesh_masked_verts'] = masked_verts
|
|
# batch['mesh_masked_faces'] = [masked_faces[i].astype(np.int32) for i in range(batch_size)]
|
|
# batch['mesh_mask_bool'] = batch['mesh_mask'].numpy().astype(bool)
|
|
# batch['laplacian'] = [sparse_to_tensor(sp.sparse.coo_matrix(laplacian_loss[batch['mesh_mask_bool'][i], :][:, batch['mesh_mask_bool'][i]])) for i in range(batch_size)]
|
|
# batch['laplacian'] = [tf.sparse.to_dense(sparse_to_tensor(sp.sparse.coo_matrix(laplacian_loss[batch['mesh_mask_bool'][i], :][:, batch['mesh_mask_bool'][i]]))) for i in range(batch_size)]
|
|
# batch['laplacian'] = [sp.sparse.coo_matrix(laplacian_loss[batch['mesh_mask_bool'][i], :][:, batch['mesh_mask_bool'][i]]) for i in range(batch_size)]
|
|
|
|
train_loss = train_step(batch)
|
|
|
|
data_valid = valid_iterator.get_next()
|
|
data_valid = static_dataset_valid.expand_batch(batch_to_dict(data_valid, signature_dynamic))
|
|
valid_loss = valid_step(data_valid)
|
|
|
|
if step % 500 == 0:
|
|
train_figs = plot_everything(batch, 'train')
|
|
valid_figs = plot_everything(data_valid, 'valid')
|
|
wandb.log({ **train_loss, **valid_loss, **{ 'current_step': step}, **train_figs, **valid_figs }, step=step)
|
|
else:
|
|
wandb.log({ **train_loss, **valid_loss, **{ 'current_step': step} }, step=step)
|
|
|
|
|
|
print(step)
|
|
|
|
step += 1 |