luni/cvpr
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

large refactor

Browse Source
This commit is contained in:
luni
2022-09-25 08:31:13 +00:00
parent 84ec5a056e
commit 92891f5b26
33 changed files with 1661 additions and 1635 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cvae_low.py
+93 -66
View File
@@ -21,24 +21,22 @@ import random
import numpy as np
import scipy as sp
from custom_utils.custom_utils import UV2Mesh, maskFacesBatch, _maskFaces
from tensorboardX import SummaryWriter
import plotly.graph_objects as go
import plotly.express as ex
from PIL import Image
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 custom_utils.custom_utils import batch_to_dict, load_model
from dataset import DatasetFactory, StaticDataset
from custom_utils.smpl import SMPL
from models.cvae_model import CVAE, CVAE128, CVAE32
from models.cvae_model import CVAE#, CVAE128, CVAE32
from models.upscaler import UpscalerBasic as Upscaler
name='cvae_low_1_d128'
name='cvae_low_4_2_d32'
def init_wandb():
run = wandb.init(project='vae', name=name,
notes='learn offsets. smaller flattened latent vectors. with 25-85k pretraining. smaller batchsize. fixed balancing of reconst and kl loss for baseline. \
@@ -49,7 +47,8 @@ use mean l2 loss. small batch. l2 * 1000, balancing to 1 from 0.01\
fixed gender issue and retrained body encoder\
remove shape from unposed garment\
training on 100 garments , absolute, not offsets\
using filled versions')
using custom enhancer for upscaling, using not filled versions\
calculating loss at full res')
artifact = wandb.Artifact('code', type='code')
artifact.add_file('cvae.py')
artifact.add_file('models/cvae_model.py')
@@ -130,62 +129,57 @@ def train_step(batch):
with tf.GradientTape() as tape:
uv_curvature = tf.one_hot(batch['uv_curvature'], 7)
uv_curvature_low = tf.one_hot(batch['uv_curvature' + res_modifier], 7)
pred, mean, logvar, z = model([batch['uv_unposed' + res_modifier], batch['uv_static' + res_modifier], batch['uv_body_posed' + res_modifier], uv_curvature_low, batch['uv_static'], batch['uv_body_posed'], uv_curvature])
pred, mean, logvar, z = model([batch['uv_unposed' + res_modifier], batch['uv_static' + res_modifier], batch['uv_body_posed' + res_modifier], uv_curvature_low,
batch['uv_mask' + res_modifier], batch['uv_static'], batch['uv_body_posed'], uv_curvature, batch['uv_mask']], use_mask=True)
# pred = pred * batch['uv_mask' + res_modifier][:,:,:,tf.newaxis]
mesh_pred = uv2mesh(tf.transpose(pred, [0,3,1,2]))
mesh_pred_posed = smpl.run_garment(batch['pose_shape'][:, :72], batch['pose_shape'][:, 72:], mesh_pred)
pred_enhanced = enhance(upscale_enhancer, pred, batch['uv_mask' + res_modifier], training=False)
pred_full_res = tf.image.resize(pred_enhanced, size=(512, 256), method=tf.image.ResizeMethod.LANCZOS5)
mesh_pred = uv2mesh(pred_enhanced)
mesh_orig = batch['mesh_unposed']
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_posed = mesh_pred_posed + shape_offsets * batch['mesh_mask'][:,:,tf.newaxis]
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)
e_loss = tf.reduce_mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
c_loss = tf.reduce_mean(collision_penalty(mesh_pred_posed, batch['mesh_body_posed'], batch['mesh_mask']))
# l1_loss = tf.reduce_sum(tf.reduce_sum((tf.math.abs(tf.transpose(batch['uv_unposed' + res_modifier], [0,2,3,1]) - pred)), axis=-1) * batch['uv_mask' + res_modifier], axis=[1, 2])
l1_loss = tf.reduce_sum(tf.reduce_sum((tf.math.abs(tf.transpose(batch['uv_unposed' + res_modifier], [0,2,3,1]) - pred)), axis=-1), axis=[1, 2])
# l2_loss = 1e3 * tf.reduce_mean(tf.reduce_sum(((tf.transpose(batch['uv_unposed' + res_modifier], [0,2,3,1]) - pred) ** 2), axis=-1) * batch['uv_mask' + res_modifier], axis=[1, 2])
l2_loss = 1e3 * tf.reduce_mean(tf.reduce_sum(((tf.transpose(batch['uv_unposed' + res_modifier], [0,2,3,1]) - pred) ** 2), axis=-1), 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' + res_modifier], [0,2,3,1]) + 1, max_val=3, power_factors=(0.6, 0.4)) #(0.0448, 0.2856, 0.3001, 0.2363, 0.1333)
ms_ssim = tf.image.ssim(pred + 1, tf.transpose(batch['uv_unposed' + res_modifier], [0,2,3,1]) + 1, max_val=3) #(0.0448, 0.2856, 0.3001, 0.2363, 0.1333)
# 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' + res_modifier], [0,2,3,1]), tf.zeros_like(tf.transpose(batch['uv_unposed' + res_modifier], [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)
l1_loss = tf.reduce_sum(tf.reduce_sum((tf.math.abs(batch['uv_unposed' + res_modifier] - pred)), axis=-1), axis=[1, 2])
l2_loss = 1e3 * tf.reduce_mean(tf.reduce_sum(((batch['uv_unposed' + res_modifier] - pred) ** 2), axis=-1), axis=[1, 2])
l2_loss_full = 1e3 * tf.reduce_mean(tf.reduce_sum(((batch['uv_unposed'] - pred_full_res) ** 2), axis=-1), axis=[1, 2])
if resolution == 32:
ms_ssim = tf.image.ssim(pred + 1, batch['uv_unposed' + res_modifier] + 1, max_val=3)
ms_ssim_full = tf.image.ssim(pred_full_res + 1, batch['uv_unposed'] + 1, max_val=3)
else:
ms_ssim = tf.image.ssim_multiscale(pred + 1, batch['uv_unposed' + res_modifier] + 1, max_val=3, power_factors=(0.5, 0.2, 0.1, 0.1, 0.1))
reconstr_loss = l2_loss
reconstr_loss = l2_loss_full
reconstr_loss = tf.reduce_mean(reconstr_loss)
# logpx_z = -reconstr_loss
# logpz = log_normal_pdf(z[:,:model.latent_dim], 0., 0.)
# logqz_x = log_normal_pdf(z[:,:model.latent_dim], mean, logvar)
# loss = -tf.reduce_mean(logpx_z + balancing_coeff * (logpz - logqz_x))
kl = kl_loss(mean, logvar)
loss = reconstr_loss + balancing_coeff * kl
# loss = reconstr_loss + balancing_coeff * kl
ms_ssim = tf.reduce_mean(ms_ssim)
ms_ssim_full = tf.reduce_mean(ms_ssim_full)
l1_loss = tf.reduce_mean(l1_loss)
l2_loss = tf.reduce_mean(l2_loss)
fft_loss = tf.reduce_mean(fft_loss)
tv_loss = tf.reduce_mean(tv_loss)
l2_loss_full = tf.reduce_mean(l2_loss_full)
mean = tf.reduce_mean(mean)
logvar = tf.reduce_mean(logvar)
mesh_orig_2 = uv2mesh(batch['uv_unposed' + res_modifier])
# mesh_orig_2 = uv2mesh(batch['uv_unposed' + res_modifier])
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'])
# 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,
'train/mean': mean, 'train/logvar': logvar, 'train/kl_loss': kl, 'train/reconst_loss_norm': reconstr_loss / (2 * (gamma ** 2)), 'train/var': np.e ** logvar,
#'train/euclid_error_reconst': metric_2,
'train/l2_loss_full_res': l2_loss_full, 'train/ms_ssim_full_res': ms_ssim_full,
'train/euclid_error_posed': euclid_posed, 'train/reconstr_loss': reconstr_loss,
'train/mean': mean, 'train/logvar': logvar, 'train/kl_loss': kl, 'train/var': np.e ** logvar,
'train/reg_coeff': reg_coeff, 'train/l2_loss': l2_loss }
@@ -235,29 +229,36 @@ def plot_everything(batch: dict, type: str):
pred, mean, logvar, z = model([batch['uv_unposed' + res_modifier], batch['uv_static' + res_modifier], batch['uv_body_posed' + res_modifier], uv_curvature_low, batch['uv_static'], batch['uv_body_posed'], uv_curvature])
# pred = pred * batch['uv_mask' + res_modifier][:,:,:,tf.newaxis]
mesh_pred = uv2mesh(pred)
pred = enhance(upscale_enhancer, pred, batch['uv_mask' + res_modifier], training=False)
mesh_pred = uv2mesh(pred).numpy()
mesh_orig = batch['mesh_unposed']
orig_img = tf.transpose(batch['uv_unposed' + res_modifier], [0,2,3,1])
uv_static_img = tf.transpose(batch['uv_static' + res_modifier], [0,2,3,1])
uv_body_posed_img = tf.transpose(batch['uv_body_posed' + res_modifier], [0,2,3,1])
pred_img = pred
orig_img = batch['uv_unposed' + res_modifier]
uv_static_img = batch['uv_static' + res_modifier]
uv_body_posed_img = batch['uv_body_posed' + res_modifier]
pred_img = pred.numpy()
pred_img[pred_img > 1] = 0
pred_img[pred_img < -1] = 0
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)]
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()
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])])
garment_posed = [mesh_posed[i, mask[i], :] + shape_offsets.numpy()[i, mask[i], :] for i in range(body.shape[0])]
F = np.load("custom_utils/smpl_faces.npy")
mask = batch['mesh_mask'].numpy().astype(bool)
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])]
for i in range(pred.shape[0]):
mesh_pred[i, mask[i], :] += shape_offsets.numpy()[i, mask[i], :]
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))]
@@ -393,6 +394,28 @@ def sample(folder):
print()
def get_upscale_enhancer(checkpoint, res):
upscale_enhancer = Upscaler(512, res)
load_model(upscale_enhancer, checkpoint)
upscale_enhancer = tf.keras.Sequential([
upscale_enhancer.flatten,
upscale_enhancer.seq,
upscale_enhancer.reshape
])
upscale_enhancer.trainable = False
return upscale_enhancer
def enhance(upscale_enhancer, input, mask, training=None):
x = input * mask[:,:,:,tf.newaxis]
x = upscale_enhancer(x, training=training)
x = x * (1 - mask[:,:,:,tf.newaxis])
x = x + input
return x
if __name__ == '__main__':
init_wandb()
allow_memory_growth()
@@ -404,7 +427,7 @@ if __name__ == '__main__':
smpl = SMPL()
resolution = 128
resolution = 32
res_modifier = ''
if resolution != 512:
res_modifier = '_' + str(resolution)
@@ -412,37 +435,41 @@ if __name__ == '__main__':
factory = DatasetFactory()
signature_dynamic = { 'id': (),
'pose_shape': (72 + 11,),
# 'uv_posed': (3, 512, 256,),
'uv_body_posed': (3, 512, 256,),
'uv_body_posed' + res_modifier: (3, resolution, resolution//2,),
# 'uv_posed': (512, 256, 3),
'uv_body_posed': (512, 256, 3),
'uv_body_posed' + res_modifier: (resolution, resolution//2, 3),
'mesh_body_posed': (14475, 3,),
'mesh_posed': (14475, 3,),
'mesh_unposed': (14475, 3,),
'uv_unposed': (3, 512, 256,),
'uv_unposed' + res_modifier: (3, resolution, resolution//2,),
'uv_unposed': (512, 256, 3),
'uv_unposed' + res_modifier: (resolution, resolution//2, 3),
'uv_curvature': (512, 256,),
'uv_curvature' + res_modifier: (resolution, resolution//2),
}
signature_static = { 'uv_static': (3, 512, 256,),
'uv_static' + res_modifier: (3, resolution, resolution//2,),
signature_static = { 'uv_static': (512, 256, 3),
'uv_static' + res_modifier: (resolution, resolution//2, 3),
# 'mesh_static': (14475, 3,),
'mesh_mask': (14475,),
'uv_mask' + res_modifier: (resolution, resolution//2,),
# 'uv_shaped_body': (3, 512, 256,),
'uv_mask': (512, 256,),
}
dataset_train = factory.get_dataset(folder='/home/workspace/local-data/train_100', batch_size=batch_size, signature=signature_dynamic, use_filled=True, resolutions=[512, resolution])
static_dataset_train = StaticDataset(folder='/home/workspace/local-data/train_100', signature=signature_static, use_filled=True, resolutions=[512, resolution])
dataset_train = factory.get_dataset(folder='/home/workspace/local-data/train_100', batch_size=batch_size, signature=signature_dynamic, use_filled=False, resolutions=[512, resolution])
static_dataset_train = StaticDataset(folder='/home/workspace/local-data/train_100', signature=signature_static, use_filled=False, resolutions=[512, resolution])
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 = CVAE128(#avg_garment_path='custom_data/train_100_no_shape_avg_uv.npy',
latent_dim=128, static_model='checkpoints/ae_static_4',
body_model='checpoints/ae_body_8',
curvature_model='checpoints/ae_curvature_7',
model = CVAE(#avg_garment_path='custom_data/train_100_no_shape_avg_uv.npy',
size=resolution,
latent_dim=128, static_model='checkpoints/ae_static_6_copy',
body_model='checkpoints/ae_body_8_copy',
curvature_model='checkpoints/ae_curvature_7_copy',
freeze_encoders=True,
use_full_res_encoders=True)
encoder_res=512)
upscale_enhancer = get_upscale_enhancer('checkpoints/up_2_3', res=resolution)
# lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(0.0001, decay_steps=30000, decay_rate=0.1, staircase=False)
optimizer = tf.optimizers.Adam(learning_rate=0.0001)