static encoder runs, but doesn't work yet
This commit is contained in:
@@ -0,0 +1,71 @@
|
||||
from utils.convnext_utils import CNBlockConfig
|
||||
|
||||
config = dict()
|
||||
|
||||
config['gpu'] = '5'
|
||||
config['name'] = 'static_torch_1'
|
||||
config['notes'] = ' '
|
||||
config['tags'] = ['torch', 'static']
|
||||
|
||||
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'] = True
|
||||
config['is_cape'] = False
|
||||
config['remove_global_orientation'] = True
|
||||
config['device'] = 'cuda'
|
||||
|
||||
config['smpl_res'] = 14475
|
||||
if config['is_cape']:
|
||||
config['smpl_res'] = 3638
|
||||
|
||||
config['latent_dim'] = 128
|
||||
|
||||
config['bottleneck_scale'] = 2
|
||||
|
||||
config['cond_ae'] = dict()
|
||||
config['cond_ae']['device'] = config['device']
|
||||
config['cond_ae']['is_label'] = False
|
||||
config['cond_ae']['latent_dim'] = config['latent_dim']
|
||||
config['cond_ae']['kernel_size'] = 7
|
||||
config['cond_ae']['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']['bottleneck_size'], 9),
|
||||
CNBlockConfig(config['cond_ae']['bottleneck_size'], None, 3),
|
||||
]
|
||||
|
||||
block_setting_reverse = [
|
||||
CNBlockConfig(config['cond_ae']['bottleneck_size'], 64, 3),
|
||||
CNBlockConfig(64, 32, 9),
|
||||
CNBlockConfig(32, 16, 3),
|
||||
CNBlockConfig(16, 8, 3),
|
||||
CNBlockConfig(8, 3, 3),
|
||||
CNBlockConfig(3, 3, 3),
|
||||
]
|
||||
|
||||
config['cond_ae']['encoder_block_settings'] = block_setting
|
||||
config['cond_ae']['decoder_block_settings'] = block_setting_reverse
|
||||
config['cond_ae']['activation'] = 'tanh'
|
||||
|
||||
|
||||
config['lr'] = 0.0005
|
||||
config['cond_ae']['upscale_output'] = False
|
||||
config['pretrained_model'] = ''
|
||||
|
||||
config['signature'] = { 'id',
|
||||
'frame_nr',
|
||||
'uv_static',
|
||||
'mesh_static',
|
||||
'mesh_mask',
|
||||
'uv_mask',
|
||||
}
|
||||
@@ -1,12 +1,16 @@
|
||||
import json
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
# import tensorflow as tf
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import scipy as sp
|
||||
import plotly.graph_objects as go
|
||||
from plotly.subplots import make_subplots
|
||||
from torchvision.transforms import Resize
|
||||
from torchvision.transforms.functional import InterpolationMode
|
||||
|
||||
from tensorflow_graphics.geometry.convolution.graph_pooling import pool, unpool
|
||||
from tensorflow_graphics.geometry.convolution.graph_convolution import edge_convolution_template
|
||||
# from tensorflow_graphics.geometry.convolution.graph_pooling import pool, unpool
|
||||
# from tensorflow_graphics.geometry.convolution.graph_convolution import edge_convolution_template
|
||||
|
||||
|
||||
def m2uv_array(m2uv, is_smpl_orig=False):
|
||||
@@ -23,8 +27,8 @@ def m2uv_array(m2uv, is_smpl_orig=False):
|
||||
return result
|
||||
|
||||
|
||||
class UV2Mesh(tf.keras.Model):
|
||||
def __init__(self, res=256, is_skirt=False, is_smpl_orig=False, is_adaptive=False):
|
||||
class UV2Mesh(nn.Module):
|
||||
def __init__(self, res=256, is_skirt=False, is_smpl_orig=False):
|
||||
super(UV2Mesh, self).__init__()
|
||||
self.res = res
|
||||
uv_map = np.load('custom_utils/stretched_uv_map_' + ('skirt' if is_skirt else '3') + ('_orig' if is_smpl_orig else '') + '.npy')
|
||||
@@ -33,48 +37,40 @@ class UV2Mesh(tf.keras.Model):
|
||||
m2uv = m2uv_array(m2uv, is_smpl_orig=is_smpl_orig)
|
||||
self.uv_pixels = self.uv_map_discreet[m2uv]
|
||||
self.smpl_res = 14475 if not is_smpl_orig else 3638
|
||||
self.is_adaptive = is_adaptive
|
||||
|
||||
if self.is_adaptive:
|
||||
laplacian = sp.sparse.load_npz("custom_utils/laplacian/laplacian0" + ("_skirt" if is_skirt else "") + ".npz")
|
||||
self.gc = GraphConvolution((3, 3), L=laplacian, batchSize=4)
|
||||
|
||||
|
||||
@tf.function
|
||||
def uv_to_mesh(self, uv):
|
||||
if uv.shape[1] == 3:
|
||||
uv = tf.transpose(uv, [0,2,3,1])
|
||||
uv = torch.permute(uv, [0,2,3,1])
|
||||
if uv.shape[1] != 512:
|
||||
uv = tf.image.resize(uv, (512, 256), tf.image.ResizeMethod.LANCZOS5)
|
||||
uv = Resize(uv[:,:,None], size=(512, 256), interpolation=InterpolationMode.BILINEAR)
|
||||
|
||||
batch_size = uv.shape[0]
|
||||
# batch_size = uv.shape[0]
|
||||
|
||||
x1_ind = tf.concat((tf.tile(self.uv_pixels[:, 0, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.zeros((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
y1_ind = tf.concat((tf.tile(self.uv_pixels[:, 0, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
z1_ind = tf.concat((tf.tile(self.uv_pixels[:, 0, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32) * 2), axis=-1)
|
||||
x2_ind = tf.concat((tf.tile(self.uv_pixels[:, 1, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.zeros((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
y2_ind = tf.concat((tf.tile(self.uv_pixels[:, 1, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
z2_ind = tf.concat((tf.tile(self.uv_pixels[:, 1, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32) * 2), axis=-1)
|
||||
# x1_ind = tf.concat((tf.tile(self.uv_pixels[:, 0, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.zeros((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
# y1_ind = tf.concat((tf.tile(self.uv_pixels[:, 0, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
# z1_ind = tf.concat((tf.tile(self.uv_pixels[:, 0, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32) * 2), axis=-1)
|
||||
# x2_ind = tf.concat((tf.tile(self.uv_pixels[:, 1, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.zeros((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
# y2_ind = tf.concat((tf.tile(self.uv_pixels[:, 1, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32)), axis=-1)
|
||||
# z2_ind = tf.concat((tf.tile(self.uv_pixels[:, 1, :][tf.newaxis, :, :], [batch_size, 1, 1]), tf.ones((batch_size, self.smpl_res, 1), dtype=tf.int32) * 2), axis=-1)
|
||||
|
||||
|
||||
x1 = tf.gather_nd(uv, x1_ind, batch_dims=1)
|
||||
y1 = tf.gather_nd(uv, y1_ind, batch_dims=1)
|
||||
z1 = tf.gather_nd(uv, z1_ind, batch_dims=1)
|
||||
x2 = tf.gather_nd(uv, x2_ind, batch_dims=1)
|
||||
y2 = tf.gather_nd(uv, y2_ind, batch_dims=1)
|
||||
z2 = tf.gather_nd(uv, z2_ind, batch_dims=1)
|
||||
# x1 = tf.gather_nd(uv, x1_ind, batch_dims=1)
|
||||
# y1 = tf.gather_nd(uv, y1_ind, batch_dims=1)
|
||||
# z1 = tf.gather_nd(uv, z1_ind, batch_dims=1)
|
||||
# x2 = tf.gather_nd(uv, x2_ind, batch_dims=1)
|
||||
# y2 = tf.gather_nd(uv, y2_ind, batch_dims=1)
|
||||
# z2 = tf.gather_nd(uv, z2_ind, batch_dims=1)
|
||||
|
||||
p1 = tf.stack((x1, y1, z1), axis=-1)
|
||||
p2 = tf.stack((x2, y2, z2), axis=-1)
|
||||
# p1 = tf.stack((x1, y1, z1), axis=-1)
|
||||
# p2 = tf.stack((x2, y2, z2), axis=-1)
|
||||
|
||||
mesh = (p1 + p2) / 2
|
||||
mesh = torch.mean(uv[:, self.uv_pixels[:,:,0], self.uv_pixels[:,:,1]], dim=2)
|
||||
return mesh
|
||||
|
||||
|
||||
def __call__(self, uv):
|
||||
x = self.uv_to_mesh(uv)
|
||||
if self.is_adaptive:
|
||||
x = self.gc(x)
|
||||
return x
|
||||
|
||||
|
||||
@@ -313,10 +309,16 @@ def load_model(model, checkpoint, disable_assert=False, assert_consumed_on=False
|
||||
|
||||
|
||||
def get_summary_str(model):
|
||||
lines = []
|
||||
model.summary(print_fn=lines.append, line_length=150,expand_nested=True,show_trainable=True)
|
||||
# from torchview import draw_graph
|
||||
# draw_graph(model, input_size=((4,3,512,256), (4,512,256)), expand_nested=True)
|
||||
|
||||
# lines = []
|
||||
# model.summary(print_fn=lines.append, line_length=150,expand_nested=True,show_trainable=True)
|
||||
# Add initial spaces to avoid markdown formatting in TensorBoard
|
||||
return lines
|
||||
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
||||
total_params = sum(param.numel() for param in model.parameters())
|
||||
|
||||
return [f'all_params: {total_params}, trainable: {trainable_params}']
|
||||
|
||||
def get_summary_html(model):
|
||||
lines = get_summary_str(model)
|
||||
@@ -341,7 +343,7 @@ def merge_verts(Vs, batchDim):
|
||||
return V
|
||||
|
||||
|
||||
class GraphConvolution(tf.keras.layers.Layer):
|
||||
class GraphConvolution(nn.Module):
|
||||
def __init__(self, shape, activation=None, name='gcn', L=None, batchSize=None, normalization=None):
|
||||
super(GraphConvolution, self).__init__()
|
||||
self.shape = shape
|
||||
@@ -406,3 +408,45 @@ def group_inputs_for_levels(batch, config):
|
||||
levels_input.append([uv_orig, batch['uv_static' + res_modifier], batch['uv_body_posed' + res_modifier], batch['uv_curvature' + res_modifier], batch['uv_mask' + res_modifier], batch['uv_mask']])
|
||||
conds = [batch['uv_static'], batch['uv_body_posed'], batch['uv_curvature']]
|
||||
return levels_input, conds
|
||||
|
||||
|
||||
class Accumulator():
|
||||
def __init__(self):
|
||||
self.dict = dict()
|
||||
|
||||
def update(self, dict):
|
||||
for key in dict.keys():
|
||||
if key in self.dict.keys():
|
||||
self.dict[key].append(dict[key])
|
||||
else:
|
||||
self.dict[key] = [dict[key]]
|
||||
|
||||
def get(self):
|
||||
result = dict()
|
||||
for key in self.dict.keys():
|
||||
result[key] = np.mean(self.dict[key])
|
||||
return result
|
||||
|
||||
def reset(self):
|
||||
self.dict = dict()
|
||||
|
||||
|
||||
from stl import mesh as mesh_stl
|
||||
|
||||
def save_stl(file, mesh, faces):
|
||||
obj = mesh_stl.Mesh(np.zeros(faces.shape[0], dtype=mesh_stl.Mesh.dtype))
|
||||
for i, f in enumerate(faces):
|
||||
for j in range(3):
|
||||
obj.vectors[i][j] = mesh[f[j],:]
|
||||
|
||||
obj.save(file)
|
||||
|
||||
|
||||
def write_obj(file, vertices, faces, colors):
|
||||
with open(file, 'w') as file:
|
||||
for v, c in zip(vertices, colors):
|
||||
line = 'v ' + ' '.join([str(_) for _ in v]) + ' ' + ' '.join([str(_) for _ in c]) + '\n'
|
||||
file.write(line)
|
||||
for f in faces:
|
||||
line = 'f ' + ' '.join([str(_) for _ in f]) + '\n'
|
||||
file.write(line)
|
||||
+7
-1
@@ -37,6 +37,8 @@ def read_frames(folder, shuffle=True, is_cape=False, testing=False, frames_file=
|
||||
frames = sorted(os.listdir(os.path.join(folder, 'mesh_static') ))
|
||||
frames = [frame.split('.')[0] + '_4' for frame in frames]
|
||||
|
||||
frames = [Frame(frame.split('_')[0], frame.split('_')[1]) for frame in frames]
|
||||
|
||||
if frames_file is not None:
|
||||
frames = []
|
||||
with open(os.path.join(folder, frames_file), 'r') as f:
|
||||
@@ -122,7 +124,7 @@ class StaticCache():
|
||||
|
||||
|
||||
class Dataset(Dataset):
|
||||
def __init__(self, folder, signature, resolutions, frames_file='frames_all.txt', device='cuda', use_cache=False, is_cape=False):
|
||||
def __init__(self, folder, signature, resolutions=[512], frames_file='frames_all.txt', device='cuda', use_cache=False, is_cape=False):
|
||||
self.folder = folder
|
||||
self.device = device
|
||||
self.is_cape = is_cape
|
||||
@@ -210,6 +212,10 @@ class Dataset(Dataset):
|
||||
|
||||
data = { key: (torch.from_numpy(value).to(self.device) if not (type(value) == str or type(value) == torch.Tensor) else value) for key, value in data.items()}
|
||||
|
||||
for key, value in data.items():
|
||||
if key.startswith('uv_') and value.shape[-1] == 3:
|
||||
data[key] = torch.permute(value, [2,0,1])
|
||||
|
||||
if self.cache is not None:
|
||||
self.cache[frame] = data
|
||||
return data
|
||||
|
||||
@@ -0,0 +1,244 @@
|
||||
import torch
|
||||
|
||||
import numpy as np
|
||||
# import tensorflow_graphics as tfg
|
||||
import scipy as sp
|
||||
import scipy.io as sio
|
||||
|
||||
from pytorch3d.structures import Meshes
|
||||
from scipy.spatial import cKDTree
|
||||
# from tensorflow_graphics.geometry.representation.mesh.normals import vertex_normals
|
||||
# from tensorflow_graphics.geometry.convolution.graph_convolution import edge_convolution_template
|
||||
|
||||
|
||||
|
||||
# from hbertiche/PBNS
|
||||
""" COLLISION FUNCTIONS START """
|
||||
F_coll = np.load("smpl/smpl_faces.npy")
|
||||
F_coll_skirt = np.load("smpl/smpl_faces_skirt.npy")
|
||||
F_coll_orig = np.load("smpl/smpl_faces_orig.npy")
|
||||
F_norm = np.tile(F_coll[np.newaxis,:,:], [4,1,1])
|
||||
F_norm_skirt = np.tile(F_coll_skirt[np.newaxis,:,:], [4,1,1])
|
||||
F_norm_orig = np.tile(F_coll_orig[np.newaxis,:,:], [4,1,1])
|
||||
|
||||
def _nearest_neighbour(V, B):
|
||||
tree = cKDTree(B.numpy())
|
||||
return tree.query(V.numpy(), workers=-1, distance_upper_bound=0.2)[1]
|
||||
|
||||
# @tf.function
|
||||
# def _tf_nn_parallel(V, B):
|
||||
# return tf.map_fn(fn=lambda elem: tf.py_function(func=_nearest_neighbour, inp=elem, Tout=tf.int32), elems=(V,B), fn_output_signature=tf.int32, parallel_iterations=V.shape[0])
|
||||
|
||||
# @tf.function
|
||||
def collision_loss(V, B, mask, thr=.004):
|
||||
mesh = Meshes(B, torch.tile(F_coll[None], [B.shape[0], 1, 1]))
|
||||
B_vn = torch.stack(mesh.verts_normals_list())
|
||||
|
||||
# compute correspondences
|
||||
idx = torch.stack([_nearest_neighbour(V[i], B[i]) for i in range(V.shape[0])])
|
||||
N, M = V.shape[:2]
|
||||
# _idx = tf.tile(np.array(list(range(N)), np.int32)[:,None], [1, M])
|
||||
_idx = torch.ones((N, M), dtype=torch.int32, device=V.get_device()) * np.array(list(range(N)), np.int32)[:,None]
|
||||
idx = torch.stack((_idx, idx), axis=-1)
|
||||
# compute loss
|
||||
D = V - B[idx]
|
||||
b_vn = B_vn[idx]
|
||||
dot = torch.einsum('abc,abc->ab', D, b_vn)
|
||||
thresholded = torch.minimum(dot - thr, 0) * mask
|
||||
loss = torch.sum(thresholded ** 2, axis=-1)
|
||||
ratio = torch.sum((thresholded < 0).to(torch.float32)) / torch.sum(mask)
|
||||
|
||||
return loss, ratio
|
||||
|
||||
|
||||
# def collision_ratio(V, B, mask, thr=.004):
|
||||
# B_vn = vertex_normals(B, tf.tile(F[None], [B.shape[0], 1, 1]))
|
||||
|
||||
# # compute correspondences
|
||||
# idx = _tf_nn_parallel(V, B)
|
||||
# N, M = V.shape[:2]
|
||||
# # _idx = tf.tile(np.array(list(range(N)), np.int32)[:,None], [1, M])
|
||||
# _idx = tf.ones((N, M), dtype=tf.int32) * np.array(list(range(N)), np.int32)[:,None]
|
||||
# idx = tf.stack((_idx, idx), axis=-1)
|
||||
# # compute loss
|
||||
# D = V - tf.gather_nd(B, idx)
|
||||
# b_vn = tf.gather_nd(B_vn, idx)
|
||||
# dot = tf.einsum('abc,abc->ab', D, b_vn)
|
||||
# loss = tf.minimum(dot - thr, 0) * mask # / tf.reduce_sum(mask, axis=-1)
|
||||
# ratio = tf.reduce_sum(tf.cast(loss < 0, dtype=tf.float32)) / tf.reduce_sum(mask)
|
||||
|
||||
# return ratio
|
||||
|
||||
""" COLLISION FUNCTIONS END """
|
||||
|
||||
|
||||
laplacian_loss = sio.loadmat("custom_utils/ssmpl0.mat")['L'].tocoo()
|
||||
edges: torch.Tensor = torch.tensor(np.stack(np.nonzero(laplacian_loss.todense())).transpose(), device='cuda')
|
||||
|
||||
laplacian_loss_skirt = sio.loadmat("custom_utils/ssmpl_skirt0.mat")['L'].tocoo()
|
||||
edges_skirt: torch.Tensor = torch.tensor(np.stack(np.nonzero(laplacian_loss.todense())).transpose(), device='cuda')
|
||||
|
||||
|
||||
def get_edge_mask(masks: torch.Tensor):
|
||||
if edges.get_device() != masks.get_device():
|
||||
edges.to(masks.get_device())
|
||||
return torch.min(masks[:,edges], dim=-1)[0]
|
||||
|
||||
def edge_loss(x, y, mask):
|
||||
e_mask = get_edge_mask(mask)
|
||||
x_e = x[:,edges[:,0],:] - x[:,edges[:,1],:]
|
||||
x_e = torch.sum(x_e ** 2, -1)
|
||||
x_e = torch.sqrt(x_e + 1e-9)
|
||||
y_e = y[:,edges[:,0],:] - y[:,edges[:,1],:]
|
||||
y_e = torch.sum(y_e ** 2, -1)
|
||||
y_e = torch.sqrt(y_e + 1e-9)
|
||||
d_e = y_e - x_e
|
||||
d_e = d_e * e_mask
|
||||
loss = torch.sum(d_e ** 2, dim=-1)# / tf.reduce_sum(e_mask, axis=-1)
|
||||
return loss
|
||||
|
||||
def get_edge_mask_skirt(masks: torch.Tensor):
|
||||
if edges.get_device() != masks.get_device():
|
||||
edges.to(masks.get_device())
|
||||
return torch.min(masks[:,edges_skirt], dim=-1)[0]
|
||||
|
||||
def edge_loss_skirt(x, y, mask):
|
||||
e_mask = get_edge_mask_skirt(mask)
|
||||
x_e = x[:,edges_skirt[:,0],:] - x[:,edges_skirt[:,1],:]
|
||||
x_e = torch.sum(x_e ** 2, -1)
|
||||
x_e = torch.sqrt(x_e + 1e-9)
|
||||
y_e = y[:,edges_skirt[:,0],:] - y[:,edges_skirt[:,1],:]
|
||||
y_e = torch.sum(y_e ** 2, -1)
|
||||
y_e = torch.sqrt(y_e + 1e-9)
|
||||
d_e = y_e - x_e
|
||||
d_e = d_e * e_mask
|
||||
loss = torch.sum(d_e ** 2, dim=-1)# / tf.reduce_sum(e_mask, axis=-1)
|
||||
return loss
|
||||
|
||||
|
||||
# class PairwiseDistance(tf.keras.layers.Layer):
|
||||
# def __init__(self, **kwargs):
|
||||
# super(PairwiseDistance, self).__init__(**kwargs)
|
||||
|
||||
# def call(self, A, B):
|
||||
# rA = tf.reduce_sum(tf.square(A), axis=-1)
|
||||
# rB = tf.reduce_sum(tf.square(B), axis=-1)
|
||||
# transpose_axes = [0, 2, 1]
|
||||
# distances = - 2*tf.matmul(A, tf.transpose(B, transpose_axes)) + rA[:, :, tf.newaxis] + rB[:, tf.newaxis, :]
|
||||
# return distances
|
||||
|
||||
|
||||
# class NearestNeighbour(tf.keras.layers.Layer):
|
||||
# def __init__(self, **kwargs):
|
||||
# super(NearestNeighbour, self).__init__(**kwargs)
|
||||
|
||||
# def call(self, A, B):
|
||||
# distances = PairwiseDistance(dtype=self.dtype)(A, B)
|
||||
# nearest_neighbour = tf.argmin(distances, axis=-1)
|
||||
# return tf.cast(nearest_neighbour, dtype=tf.int32)
|
||||
|
||||
#
|
||||
# @tf.function
|
||||
def collision_penalty(va, vb, mask, eps=1e-3):
|
||||
nb = vertex_normals(vb, tf.tile(F[None], [vb.shape[0], 1, 1]))
|
||||
batch_size = tf.cast(tf.shape(va)[0], va.dtype)
|
||||
|
||||
closest_vertices = NearestNeighbour(dtype=va.dtype)(va, vb)
|
||||
vb = tf.gather(vb, closest_vertices, batch_dims=1)
|
||||
nb = tf.gather(nb, closest_vertices, batch_dims=1)
|
||||
|
||||
distance = tf.reduce_sum(nb*(va - vb), axis=-1)
|
||||
interpenetration = tf.maximum(eps - distance * mask, 0)
|
||||
|
||||
return tf.reduce_sum(interpenetration**3) / batch_size
|
||||
|
||||
whole_garment_uv_mask = np.load('custom_utils/mask_256x512.npy')
|
||||
|
||||
def uv_mask_loss(pred, uv_mask):
|
||||
inv_mask = (1 - uv_mask) * whole_garment_uv_mask
|
||||
pred = tf.reduce_sum(pred, axis=1)
|
||||
invalid = pred * inv_mask
|
||||
invalid = invalid > 0.01
|
||||
invalid_count = tf.math.count_nonzero(invalid, dtype=tf.float32)
|
||||
total_count = tf.math.count_nonzero(inv_mask, dtype=tf.float32)
|
||||
return invalid_count / total_count
|
||||
|
||||
|
||||
|
||||
|
||||
def merge_verts(Vs, batchDim):
|
||||
V = tf.concat(tf.unstack(Vs, num=batchDim), axis=0)
|
||||
return V
|
||||
|
||||
|
||||
def _bend_loss_single(x, F, L, mask):
|
||||
verts = tf.boolean_mask(x, mask, axis=0)
|
||||
VN = vertex_normals(verts, F)
|
||||
bend = edge_convolution_template(
|
||||
VN, L, sizes=None,
|
||||
edge_function=lambda x,y: (x - y)**2,
|
||||
reduction='weighted',
|
||||
edge_function_kwargs={}
|
||||
)
|
||||
loss = tf.reduce_sum(bend)
|
||||
return loss
|
||||
|
||||
def bend_loss(x, F, L, mask):
|
||||
# x: predicted outfits
|
||||
# F: faces
|
||||
# L: laplacian
|
||||
loss = tf.stack([_bend_loss_single(x[i], F[i], L[i], mask[i]) for i in range(x.shape[0])])
|
||||
return loss
|
||||
|
||||
|
||||
# def bend_loss(x, F, L, mask):
|
||||
# # x: predicted outfits
|
||||
# # F: faces
|
||||
# # L: laplacian
|
||||
# # x = x * mask[:,:,tf.newaxis]
|
||||
# batch_size = x.shape[0]
|
||||
# verts, faces = maskFacesBatch(x, F, mask)
|
||||
# nr_valid_verts = [tf.reduce_sum(mask[i]) for i in range(batch_size)]
|
||||
# x = tf.concat(verts, axis=0)
|
||||
# F = tf.concat([faces[i] + tf.reduce_sum(nr_valid_verts[:i]) for i in range(1, batch_size)])
|
||||
# F = tf.concat((faces[0], F), axis=0)
|
||||
# VN = vertex_normals(x, F)
|
||||
# bend = edge_convolution_template(
|
||||
# VN, L, sizes=None,
|
||||
# edge_function=lambda x,y: (x - y)**2,
|
||||
# reduction='weighted',
|
||||
# edge_function_kwargs={}
|
||||
# )
|
||||
# loss = tf.reduce_sum(bend, axis=[1])
|
||||
# loss = tf.reshape(loss, (batch_size, 14475))
|
||||
# loss = tf.reduce_mean(loss, axis=1)
|
||||
# return loss
|
||||
|
||||
|
||||
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
|
||||
|
||||
|
||||
def kl_loss(mean, logvar):
|
||||
kl_loss = 1 + logvar - tf.square(mean) - tf.exp(logvar)
|
||||
kl_loss = tf.reduce_mean(kl_loss)
|
||||
kl_loss *= -0.5
|
||||
return kl_loss
|
||||
|
||||
|
||||
def normal_loss(pred, gt, mask, is_skirt=False, is_smpl_orig=False):
|
||||
f_norm = None
|
||||
if not is_skirt and not is_smpl_orig:
|
||||
f_norm = F_norm
|
||||
elif is_skirt:
|
||||
f_norm = F_norm_skirt
|
||||
elif is_smpl_orig:
|
||||
f_norm = F_norm_orig
|
||||
else:
|
||||
raise NotImplementedError
|
||||
|
||||
pred_norm = vertex_normals(pred, f_norm)
|
||||
gt_norm = vertex_normals(gt, f_norm)
|
||||
return tf.reduce_mean(tf.reduce_mean((tf.reduce_sum(tf.math.abs(pred_norm - gt_norm), axis=-1) * mask), axis=-1))
|
||||
+36
-148
@@ -1,134 +1,21 @@
|
||||
import tensorflow as tf
|
||||
from convnext.models.convnext_tf import ConvNeXt, ConvNeXtReverse, Block, ConvNeXtReverseLabels
|
||||
|
||||
class Bottleneck(tf.keras.Model):
|
||||
|
||||
def __init__(self, dim, latent_dim, kernel_size=7):
|
||||
super(Bottleneck, self).__init__()
|
||||
|
||||
self.bottleneck = tf.keras.Sequential(
|
||||
[Block(dim=dim, drop_path=0, kernel_size=kernel_size,
|
||||
layer_scale_init_value=1e-6, prefix=f'bottleneck{j}') for j in range(5)]
|
||||
)
|
||||
|
||||
self.flatten = tf.keras.layers.Flatten()
|
||||
self.dense = tf.keras.layers.Dense(latent_dim)
|
||||
|
||||
def call(self, inputs, training):
|
||||
x = inputs
|
||||
|
||||
x = self.bottleneck(x, training=training)
|
||||
x = self.flatten(x)
|
||||
x = self.dense(x, training=training)
|
||||
|
||||
return x
|
||||
import torch
|
||||
from torch import nn
|
||||
from models.convnext import ConvNeXt, CNBlock, CNBlockConfig
|
||||
from typing import Any, Callable, List, Optional, Sequence
|
||||
|
||||
|
||||
class VariationalEncoder(tf.keras.Model):
|
||||
class VariationalEncoder(nn.Module):
|
||||
|
||||
def __init__(self, latent_dim, depths, dims, kernel_size=7):
|
||||
def __init__(self, block_setting: CNBlockConfig, kernel_size: int = 7):
|
||||
super(VariationalEncoder, self).__init__()
|
||||
|
||||
self.encoder = ConvNeXt(depths=depths, dims=dims, include_top=False, kernel_size=kernel_size)
|
||||
self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size)
|
||||
|
||||
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):
|
||||
x = inputs
|
||||
|
||||
x = self.encoder(x, training=training)
|
||||
x = self.flatten(x)
|
||||
mean = self.mean(x)
|
||||
log_var = self.log_var(x)
|
||||
|
||||
return mean, log_var
|
||||
|
||||
|
||||
class AEEncoder(tf.keras.Model):
|
||||
|
||||
def __init__(self, latent_dim, depths, dims, bottleneck_size=32, kernel_size=7):
|
||||
super(AEEncoder, self).__init__()
|
||||
|
||||
self.encoder = ConvNeXt(depths=depths, dims=dims, include_top=False, kernel_size=kernel_size)
|
||||
self.bottleneck = Bottleneck(bottleneck_size, latent_dim, kernel_size)
|
||||
|
||||
|
||||
def call(self, inputs, training):
|
||||
if len(inputs) == 2:
|
||||
x, _ = inputs
|
||||
else:
|
||||
x = inputs
|
||||
|
||||
x = self.encoder(x, training=training)
|
||||
x = self.bottleneck(x, training=training)
|
||||
|
||||
if len(inputs) == 2:
|
||||
return x, inputs[1]
|
||||
else:
|
||||
return x
|
||||
|
||||
|
||||
class Decoder(tf.keras.Model):
|
||||
|
||||
def __init__(self, input_size, depths, dims, kernel_size=7, is_label=False, upscale=False):
|
||||
super(Decoder, self).__init__()
|
||||
|
||||
self.dense = tf.keras.layers.Dense(tf.reduce_prod(input_size))
|
||||
self.reshape = tf.keras.layers.Reshape(input_size)
|
||||
self.upscale = upscale
|
||||
|
||||
if not is_label:
|
||||
self.decoder = ConvNeXtReverse(depths=list(reversed(depths)), dims=list(reversed(dims)), kernel_size=kernel_size)
|
||||
else:
|
||||
self.decoder = ConvNeXtReverseLabels(depths=list(reversed(depths)), dims=list(reversed(dims)), nr_of_classes=7)
|
||||
|
||||
|
||||
def call(self, inputs, training):
|
||||
if len(inputs) == 2:
|
||||
x, mask = inputs
|
||||
else:
|
||||
x = inputs
|
||||
|
||||
x = self.dense(x)
|
||||
x = self.reshape(x)
|
||||
x = self.decoder(x, training=training)
|
||||
|
||||
if self.upscale:
|
||||
x = tf.image.resize(x, size=(512, 256), method=tf.image.ResizeMethod.LANCZOS5)
|
||||
|
||||
|
||||
if len(inputs) == 2:
|
||||
if len(mask.shape) == 2:
|
||||
x = x * mask[tf.newaxis,:,:,tf.newaxis]
|
||||
else:
|
||||
x = x * mask[:,:,:,tf.newaxis]
|
||||
|
||||
return x
|
||||
|
||||
|
||||
|
||||
class VariationalEncoderNew(tf.keras.Model):
|
||||
|
||||
def __init__(self, depths, dims, kernel_size=7):
|
||||
super(VariationalEncoderNew, self).__init__()
|
||||
|
||||
self.encoder = ConvNeXt(depths=depths, dims=dims, include_top=False, kernel_size=kernel_size)
|
||||
|
||||
self.mean_branch = tf.keras.Sequential(
|
||||
[Block(dim=64, drop_path=0, kernel_size=kernel_size,
|
||||
layer_scale_init_value=1e-6, prefix=f'bottleneck{j}') for j in range(5)]
|
||||
)
|
||||
|
||||
self.var_branch = tf.keras.Sequential(
|
||||
[Block(dim=64, drop_path=0, kernel_size=kernel_size,
|
||||
layer_scale_init_value=1e-6, prefix=f'bottleneck{j}') for j in range(5)]
|
||||
)
|
||||
|
||||
self.flatten_mean = tf.keras.layers.Flatten()
|
||||
self.flatten_var = tf.keras.layers.Flatten()
|
||||
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)])
|
||||
|
||||
self.flatten_mean = nn.Flatten()
|
||||
self.flatten_var = nn.Flatten()
|
||||
|
||||
|
||||
def call(self, inputs, training):
|
||||
@@ -144,23 +31,27 @@ class VariationalEncoderNew(tf.keras.Model):
|
||||
return mean, log_var
|
||||
|
||||
|
||||
class AEEncoder(nn.Module):
|
||||
|
||||
class AEEncoderNew(tf.keras.Model):
|
||||
def __init__(self, block_setting: CNBlockConfig, kernel_size: int = 7, activation='tanh', device='cuda'):
|
||||
super(AEEncoder, self).__init__()
|
||||
|
||||
def __init__(self, depths, dims, kernel_size=7, activation='tanh'):
|
||||
super(AEEncoderNew, self).__init__()
|
||||
|
||||
self.encoder = ConvNeXt(depths=depths, dims=dims, include_top=False, kernel_size=kernel_size)
|
||||
self.act = tf.keras.layers.Activation(activation)
|
||||
self.encoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, device=device)
|
||||
if activation == 'tanh':
|
||||
self.act = nn.Tanh()
|
||||
elif activation == 'sigmoid':
|
||||
self.act = nn.Sigmoid()
|
||||
else:
|
||||
raise ValueError('Activation function not supported')
|
||||
|
||||
|
||||
def call(self, inputs, training):
|
||||
def forward(self, inputs):
|
||||
if len(inputs) == 2:
|
||||
x, _ = inputs
|
||||
else:
|
||||
x = inputs
|
||||
|
||||
x = self.encoder(x, training=training)
|
||||
x = self.encoder(x)
|
||||
x = self.act(x)
|
||||
|
||||
if len(inputs) == 2:
|
||||
@@ -169,39 +60,36 @@ class AEEncoderNew(tf.keras.Model):
|
||||
return x
|
||||
|
||||
|
||||
class DecoderNew(tf.keras.Model):
|
||||
class Decoder(nn.Module):
|
||||
|
||||
def __init__(self, depths, dims, kernel_size=7, is_label=False):
|
||||
super(DecoderNew, self).__init__()
|
||||
def __init__(self, block_setting: CNBlockConfig, kernel_size=7, is_label=False, nr_of_classes=None, device='cuda'):
|
||||
super(Decoder, self).__init__()
|
||||
|
||||
self.is_label = is_label
|
||||
|
||||
if not is_label:
|
||||
self.decoder = ConvNeXtReverse(depths=depths, dims=dims, kernel_size=kernel_size)
|
||||
else:
|
||||
self.decoder = ConvNeXtReverseLabels(depths=depths, dims=dims, nr_of_classes=7)
|
||||
self.decoder = ConvNeXt(block_setting=block_setting, kernel_size=kernel_size, is_label=is_label, nr_of_labels=nr_of_classes, device=device, reversed=True)
|
||||
|
||||
self.softmax = nn.Softmax(dim=1)
|
||||
|
||||
def call(self, inputs, training):
|
||||
def forward(self, inputs):
|
||||
if len(inputs) == 2:
|
||||
x, mask = inputs
|
||||
else:
|
||||
x = inputs
|
||||
|
||||
x = self.decoder(x, training=training)
|
||||
x = self.decoder(x)
|
||||
|
||||
if self.is_label:
|
||||
x = x * mask[:,:,:,tf.newaxis]
|
||||
x = x * mask[:,:,:,None]
|
||||
|
||||
background = tf.ones((mask.shape[0],512,256,1)) * (1 - mask[:,:,:,tf.newaxis])
|
||||
x = tf.concat((background, x), axis=-1)
|
||||
|
||||
x = tf.math.softmax(x)
|
||||
background = torch.ones((mask.shape[0],512,256,1), device=mask.get_device()) * (1 - mask[:,:,:,None])
|
||||
x = torch.concat((background, x), axis=-1)
|
||||
x = self.softmax(x)
|
||||
|
||||
if len(inputs) == 2:
|
||||
if len(mask.shape) == 2:
|
||||
x = x * mask[tf.newaxis,:,:,tf.newaxis]
|
||||
x = x * mask[None,None,:,:]
|
||||
else:
|
||||
x = x * mask[:,:,:,tf.newaxis]
|
||||
x = x * mask[:,None,:,:]
|
||||
|
||||
return x
|
||||
+10
-27
@@ -1,38 +1,21 @@
|
||||
import tensorflow as tf
|
||||
from models.base import AEEncoder, Decoder, AEEncoderNew, DecoderNew
|
||||
import torch
|
||||
from torch import nn
|
||||
from models.base import AEEncoder, Decoder
|
||||
|
||||
|
||||
class AutoEncoder(tf.keras.Model):
|
||||
class AutoEncoder(nn.Module):
|
||||
|
||||
def __init__(self, config):
|
||||
super(AutoEncoder, self).__init__()
|
||||
|
||||
self.encoder : AEEncoder = AEEncoder(config['latent_dim'], config['encoder_depths'], config['encoder_dims'], config['bottleneck_size'], config['kernel_size'])
|
||||
self.decoder : Decoder = Decoder(config['decoder_input'], config['decoder_depths'], config['decoder_dims'], kernel_size=config['kernel_size'], is_label=config['is_label'], upscale=config['upscale_output'])
|
||||
self.encoder : AEEncoder = AEEncoder(config['encoder_block_settings'], config['kernel_size'], config['activation'], device=config['device'])
|
||||
self.decoder : Decoder = Decoder(config['decoder_block_settings'], kernel_size=config['kernel_size'], is_label=config['is_label'], device=config['device'])
|
||||
|
||||
|
||||
def call(self, inputs, training):
|
||||
x = inputs
|
||||
def forward(self, input):
|
||||
x = input
|
||||
|
||||
x = self.encoder(x, training=training)
|
||||
x = self.decoder(x, training=training)
|
||||
|
||||
return x
|
||||
|
||||
|
||||
class AutoEncoderNew(tf.keras.Model):
|
||||
|
||||
def __init__(self, config):
|
||||
super(AutoEncoderNew, self).__init__()
|
||||
|
||||
self.encoder : AEEncoderNew = AEEncoderNew(config['encoder_depths'], config['encoder_dims'], config['kernel_size'], config['activation'])
|
||||
self.decoder : DecoderNew = DecoderNew(config['decoder_depths'], config['decoder_dims'], kernel_size=config['kernel_size'], is_label=config['is_label'])
|
||||
|
||||
|
||||
def call(self, inputs, training):
|
||||
x = inputs
|
||||
|
||||
x = self.encoder(x, training=training)
|
||||
x = self.decoder(x, training=training)
|
||||
x = self.encoder(x)
|
||||
x = self.decoder(x)
|
||||
|
||||
return x
|
||||
+378
-260
@@ -1,306 +1,424 @@
|
||||
from unittest import skip
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.layers import (Conv2D, Conv2DTranspose, Dense, DepthwiseConv2D,
|
||||
GlobalAveragePooling2D, Layer,
|
||||
LayerNormalization, Flatten, Reshape, UpSampling2D)
|
||||
# taken from torchvision.models.convnext
|
||||
|
||||
def drop_path(inputs, drop_prob, is_training):
|
||||
# borrowed from https://github.com/rishigami/Swin-Transformer-TF/blob/main/swintransformer/model.py
|
||||
if (not is_training) or (drop_prob == 0.):
|
||||
return inputs
|
||||
from functools import partial
|
||||
from typing import Any, Callable, List, Optional, Sequence
|
||||
|
||||
# Compute keep_prob
|
||||
keep_prob = 1.0 - drop_prob
|
||||
import torch
|
||||
from torch import nn, Tensor
|
||||
from torch.nn import functional as F
|
||||
|
||||
# Compute drop_connect tensor
|
||||
random_tensor = keep_prob
|
||||
shape = (tf.shape(inputs)[0],) + (1,) * \
|
||||
(len(tf.shape(inputs)) - 1)
|
||||
random_tensor += tf.random.uniform(shape, dtype=inputs.dtype)
|
||||
binary_tensor = tf.floor(random_tensor)
|
||||
output = tf.math.divide(inputs, keep_prob) * binary_tensor
|
||||
return output
|
||||
from torchvision.ops.misc import Conv2dNormActivation, Permute
|
||||
# from torchvision.ops.misc import ConvNormActivation as Conv2dNormActivation
|
||||
# from torchvision.ops import Permute
|
||||
from torchvision.ops.stochastic_depth import StochasticDepth
|
||||
from utils.convnext_utils import CNBlockConfig
|
||||
|
||||
|
||||
class DropPath(tf.keras.layers.Layer):
|
||||
# borrowed from https://github.com/rishigami/Swin-Transformer-TF/blob/main/swintransformer/model.py
|
||||
def __init__(self, drop_prob=None):
|
||||
__all__ = [
|
||||
"ConvNeXt",
|
||||
"ConvNeXt_Tiny_Weights",
|
||||
"ConvNeXt_Small_Weights",
|
||||
"ConvNeXt_Base_Weights",
|
||||
"ConvNeXt_Large_Weights",
|
||||
"convnext_tiny",
|
||||
"convnext_small",
|
||||
"convnext_base",
|
||||
"convnext_large",
|
||||
]
|
||||
|
||||
|
||||
class LayerNorm2d(nn.LayerNorm):
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
x = x.permute(0, 2, 3, 1)
|
||||
x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
|
||||
x = x.permute(0, 3, 1, 2)
|
||||
return x
|
||||
|
||||
|
||||
class CNBlock(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
dim,
|
||||
layer_scale: float = 1e-6,
|
||||
stochastic_depth_prob: float = 0.0,
|
||||
kernel_size: int = 7,
|
||||
norm_layer: Optional[Callable[..., nn.Module]] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.drop_prob = drop_prob
|
||||
if norm_layer is None:
|
||||
norm_layer = partial(nn.LayerNorm, eps=1e-6)
|
||||
|
||||
def call(self, x, training):
|
||||
return drop_path(x, self.drop_prob, training)
|
||||
self.block = nn.Sequential(
|
||||
nn.Conv2d(dim, dim, kernel_size=kernel_size, padding='same', groups=dim, bias=True),
|
||||
Permute([0, 2, 3, 1]),
|
||||
norm_layer(dim),
|
||||
nn.Linear(in_features=dim, out_features=4 * dim, bias=True),
|
||||
nn.GELU(),
|
||||
nn.Linear(in_features=4 * dim, out_features=dim, bias=True),
|
||||
Permute([0, 3, 1, 2]),
|
||||
)
|
||||
self.layer_scale = nn.Parameter(torch.ones(dim, 1, 1) * layer_scale)
|
||||
self.stochastic_depth = StochasticDepth(stochastic_depth_prob, "row")
|
||||
|
||||
def forward(self, input: Tensor) -> Tensor:
|
||||
result = self.layer_scale * self.block(input)
|
||||
result = self.stochastic_depth(result)
|
||||
result += input
|
||||
return result
|
||||
|
||||
|
||||
class Block(Layer):
|
||||
r""" ConvNeXt Block. There are two equivalent implementations:
|
||||
(1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
|
||||
(2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
|
||||
We use (2) as we find it slightly faster in PyTorch
|
||||
|
||||
Args:
|
||||
dim (int): Number of input channels.
|
||||
drop_path (float): Stochastic depth rate. Default: 0.0
|
||||
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
||||
"""
|
||||
|
||||
def __init__(self, dim, kernel_size=7, drop_path=0., layer_scale_init_value=1e-6, prefix=''):
|
||||
class ConvNeXtBase(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
block_setting: List[CNBlockConfig],
|
||||
stochastic_depth_prob: float = 0.0,
|
||||
layer_scale: float = 1e-6,
|
||||
block: Optional[Callable[..., nn.Module]] = None,
|
||||
norm_layer: Optional[Callable[..., nn.Module]] = None,
|
||||
device: str = 'cuda',
|
||||
kernel_size: int = 7,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.dwconv = DepthwiseConv2D(
|
||||
kernel_size=kernel_size, padding='same') # depthwise conv
|
||||
self.norm = LayerNormalization(epsilon=1e-6)
|
||||
# pointwise/1x1 convs, implemented with linear layers
|
||||
self.pwconv1 = Dense(4 * dim)
|
||||
self.act = tf.keras.activations.gelu
|
||||
self.pwconv2 = Dense(dim)
|
||||
self.drop_path = DropPath(drop_path)
|
||||
self.dim = dim
|
||||
self.layer_scale_init_value = layer_scale_init_value
|
||||
self.prefix = prefix
|
||||
|
||||
def build(self, input_shape):
|
||||
self.gamma = tf.Variable(
|
||||
initial_value=self.layer_scale_init_value * tf.ones((self.dim)),
|
||||
trainable=True,
|
||||
name=f'{self.prefix}/gamma')
|
||||
self.built = True
|
||||
if not block_setting:
|
||||
raise ValueError("The block_setting should not be empty")
|
||||
elif not (isinstance(block_setting, Sequence) and all([isinstance(s, CNBlockConfig) for s in block_setting])):
|
||||
raise TypeError("The block_setting should be List[CNBlockConfig]")
|
||||
|
||||
def call(self, x, training):
|
||||
input = x
|
||||
x = self.dwconv(x, training=training)
|
||||
# x = x.permute(0, 2, 3, 1) # (N, C, H, W) -> (N, H, W, C)
|
||||
x = self.norm(x, training=training)
|
||||
x = self.pwconv1(x, training=training)
|
||||
if block is None:
|
||||
block = CNBlock
|
||||
|
||||
if norm_layer is None:
|
||||
norm_layer = partial(LayerNorm2d, eps=1e-6)
|
||||
|
||||
# self.permute_cf = Permute([0, 3, 1, 2])
|
||||
# self.permute_cl = Permute([0, 2, 3, 1])
|
||||
|
||||
self.stages: List[nn.Module] = []
|
||||
self.downsamples: List[nn.Module] = []
|
||||
|
||||
# Stem
|
||||
firstconv_output_channels = block_setting[0].input_channels
|
||||
self.stem = Conv2dNormActivation(
|
||||
3,
|
||||
firstconv_output_channels,
|
||||
kernel_size=4,
|
||||
stride=4,
|
||||
padding=0,
|
||||
norm_layer=norm_layer,
|
||||
activation_layer=None,
|
||||
bias=True,
|
||||
).to(device)
|
||||
|
||||
|
||||
total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
|
||||
stage_block_id = 0
|
||||
for cnf in block_setting:
|
||||
# Bottlenecks
|
||||
stage: List[nn.Module] = []
|
||||
for _ in range(cnf.num_layers):
|
||||
# adjust stochastic depth probability based on the depth of the stage block
|
||||
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
|
||||
self.stages.append(nn.Sequential(*stage).to(device))
|
||||
if cnf.out_channels is not None:
|
||||
# Downsampling
|
||||
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.act = nn.Tanh()
|
||||
|
||||
for m in self.modules():
|
||||
if isinstance(m, (nn.Conv2d, nn.Linear)):
|
||||
nn.init.trunc_normal_(m.weight, std=0.02)
|
||||
if m.bias is not None:
|
||||
nn.init.zeros_(m.bias)
|
||||
|
||||
def _forward_impl(self, x: Tensor) -> Tensor:
|
||||
# x = self.permute_cf(x)
|
||||
x = self.stem(x)
|
||||
# features: List[Tensor] = []
|
||||
for i in range(len(self.stages) - 1):
|
||||
x = self.stages[i](x)
|
||||
# features.append(x)
|
||||
x = self.downsamples[i](x)
|
||||
x = self.stages[-1](x)
|
||||
x = self.act(x)
|
||||
x = self.pwconv2(x, training=training)
|
||||
if self.gamma is not None:
|
||||
x = self.gamma * x
|
||||
# x = x.permute(0, 3, 1, 2) # (N, H, W, C) -> (N, C, H, W)
|
||||
|
||||
x = input + self.drop_path(x, training=training)
|
||||
# x = self.permute_cl(x)
|
||||
return x
|
||||
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
return self._forward_impl(x)
|
||||
|
||||
class ConvNeXt(tf.keras.Model):
|
||||
r""" ConvNeXt
|
||||
A Tensorflow keras impl of : `A ConvNet for the 2020s` -
|
||||
https://arxiv.org/pdf/2201.03545.pdf
|
||||
|
||||
Args:
|
||||
num_classes (int): Number of classes for classification head. Default: 1000
|
||||
depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
|
||||
dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
|
||||
include_top (bool): whether to add head or just use it as feature extractor. Default: True
|
||||
drop_path_rate (float): Stochastic depth rate. Default: 0.
|
||||
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
||||
head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
|
||||
"""
|
||||
|
||||
def __init__(self, num_classes=1000, kernel_size=7,
|
||||
depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], include_top=True,
|
||||
drop_path_rate=0., layer_scale_init_value=1e-6, head_init_scale=1.,
|
||||
):
|
||||
class ConvNeXtReverseBase2(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
block_setting: List[CNBlockConfig],
|
||||
stochastic_depth_prob: float = 0.0,
|
||||
layer_scale: float = 1e-6,
|
||||
block: Optional[Callable[..., nn.Module]] = None,
|
||||
norm_layer: Optional[Callable[..., nn.Module]] = None,
|
||||
device: str = 'cuda',
|
||||
use_skip: bool = False,
|
||||
is_label: bool = False,
|
||||
nr_of_labels: Optional[int] = None,
|
||||
kernel_size: int = 7,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.include_top = include_top
|
||||
self.downsample_layers = [] # stem and 3 intermediate downsampling conv layers
|
||||
# stem = tf.keras.Sequential([
|
||||
# Conv2D(dims[0], kernel_size=4, strides=4, padding='same'),
|
||||
# LayerNormalization(epsilon=1e-6)]
|
||||
# )
|
||||
stem = [Conv2D(dims[0], kernel_size=4, strides=4, padding='same'), LayerNormalization(epsilon=1e-6)]
|
||||
self.downsample_layers.append(stem)
|
||||
for i in range(len(depths) - 1):
|
||||
# downsample_layer = tf.keras.Sequential([
|
||||
# LayerNormalization(epsilon=1e-6),
|
||||
# Conv2D(dims[i+1], kernel_size=2, strides=2, padding='same')]
|
||||
# )
|
||||
downsample_layer = [LayerNormalization(epsilon=1e-6), Conv2D(dims[i+1], kernel_size=2, strides=2, padding='same')]
|
||||
self.downsample_layers.append(downsample_layer)
|
||||
|
||||
self.stages = [] # 4 feature resolution stages, each consisting of multiple residual blocks
|
||||
dp_rates = [x for x in np.linspace(0, drop_path_rate, sum(depths))]
|
||||
cur = 0
|
||||
for i in range(len(depths)):
|
||||
stage = tf.keras.Sequential(
|
||||
[Block(dim=dims[i], drop_path=dp_rates[cur + j], kernel_size=kernel_size,
|
||||
layer_scale_init_value=layer_scale_init_value, prefix=f'block{i}') for j in range(depths[i])]
|
||||
)
|
||||
self.stages.append(stage)
|
||||
cur += depths[i]
|
||||
self.use_skip = use_skip
|
||||
|
||||
if self.include_top:
|
||||
self.avg = GlobalAveragePooling2D()
|
||||
self.norm = LayerNormalization(epsilon=1e-6) # final norm layer
|
||||
self.head = Dense(num_classes)
|
||||
if not block_setting:
|
||||
raise ValueError("The block_setting should not be empty")
|
||||
elif not (isinstance(block_setting, Sequence) and all([isinstance(s, CNBlockConfig) for s in block_setting])):
|
||||
raise TypeError("The block_setting should be List[CNBlockConfig]")
|
||||
|
||||
if block is None:
|
||||
block = CNBlock
|
||||
|
||||
if norm_layer is None:
|
||||
norm_layer = partial(LayerNorm2d, eps=1e-6)
|
||||
|
||||
# self.permute_cf = Permute([0, 3, 1, 2])
|
||||
# self.permute_cl = Permute([0, 2, 3, 1])
|
||||
|
||||
# Stem
|
||||
# firstconv_output_channels = block_setting[0].input_channels
|
||||
# self.first = nn.Sequential(
|
||||
# nn.ConvTranspose2d(
|
||||
# firstconv_output_channels,
|
||||
# firstconv_output_channels,
|
||||
# kernel_size=2,
|
||||
# stride=2,
|
||||
# padding=0,
|
||||
# ),
|
||||
# norm_layer(firstconv_output_channels),
|
||||
# ).to(device)
|
||||
self.first = nn.Sequential(
|
||||
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
|
||||
self.stages: List[nn.Module] = []
|
||||
self.upsamples: List[nn.Module] = []
|
||||
for cnf in block_setting[1:]:
|
||||
if cnf.input_channels is not None:
|
||||
# Bottlenecks
|
||||
stage: List[nn.Module] = []
|
||||
for _ in range(cnf.num_layers):
|
||||
# adjust stochastic depth probability based on the depth of the stage block
|
||||
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
|
||||
self.stages.append(nn.Sequential(*stage).to(device))
|
||||
else:
|
||||
self.avg = None
|
||||
self.norm = None
|
||||
self.head = None
|
||||
self.stages.append(nn.Identity().to(device))
|
||||
|
||||
if cnf.out_channels is not None:
|
||||
# Upsampling
|
||||
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.Conv2d(
|
||||
block_setting[-1].input_channels,
|
||||
3 if not is_label else nr_of_labels,
|
||||
kernel_size=3,
|
||||
padding='same',
|
||||
).to(device)
|
||||
self.activation = nn.Tanh() if not is_label else nn.Softmax(dim=1)
|
||||
|
||||
# self.last = nn.Sequential(
|
||||
# Permute([0, 2, 3, 1]),
|
||||
# nn.Linear(in_features=block_setting[-1].input_channels, out_features=1, bias=True),
|
||||
# Permute([0, 3, 1, 2]),
|
||||
# ).to(device)
|
||||
|
||||
|
||||
for m in self.modules():
|
||||
if isinstance(m, (nn.Conv2d, nn.Linear)):
|
||||
nn.init.trunc_normal_(m.weight, std=0.02)
|
||||
if m.bias is not None:
|
||||
nn.init.zeros_(m.bias)
|
||||
|
||||
|
||||
def _forward_impl(self, x) -> Tensor:
|
||||
# input, features = x
|
||||
# x = self.permute_cf(x)
|
||||
|
||||
x = self.first(x)
|
||||
for i in range(len(self.stages)):
|
||||
# if self.use_skip:
|
||||
# x = self.stages[i](torch.concat((x, features[2-i]), dim=1))
|
||||
# else:
|
||||
x = self.stages[i](x)
|
||||
x = self.upsamples[i](x)
|
||||
|
||||
# x = self.stages[-2](x)
|
||||
# x = self.upsamples[-1](x)
|
||||
|
||||
# x = self.stages[-1](x)
|
||||
x = self.last(x)
|
||||
x = self.activation(x)
|
||||
# x = self.permute_cl(x)
|
||||
|
||||
def forward_features(self, x, training):
|
||||
for i in range(len(self.downsample_layers)):
|
||||
# x = self.downsample_layers[i](x)
|
||||
x = self.downsample_layers[i][0](x, training=training)
|
||||
x = self.downsample_layers[i][1](x, training=training)
|
||||
x = self.stages[i](x, training=training)
|
||||
return x
|
||||
|
||||
def call(self, x, training):
|
||||
x = self.forward_features(x, training=training)
|
||||
return x
|
||||
def forward(self, x) -> Tensor:
|
||||
return self._forward_impl(x)
|
||||
|
||||
|
||||
class ConvNeXtReverse(tf.keras.Model):
|
||||
r""" ConvNeXt
|
||||
A Tensorflow keras impl of : `A ConvNet for the 2020s` -
|
||||
https://arxiv.org/pdf/2201.03545.pdf
|
||||
|
||||
Args:
|
||||
num_classes (int): Number of classes for classification head. Default: 1000
|
||||
depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
|
||||
dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
|
||||
include_top (bool): whether to add head or just use it as feature extractor. Default: True
|
||||
drop_path_rate (float): Stochastic depth rate. Default: 0.
|
||||
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
||||
head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], kernel_size=7,
|
||||
drop_path_rate=0., layer_scale_init_value=1e-6, head_init_scale=1.,
|
||||
):
|
||||
class ConvNeXtReverseBase(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
block_setting: List[CNBlockConfig],
|
||||
stochastic_depth_prob: float = 0.0,
|
||||
layer_scale: float = 1e-6,
|
||||
block: Optional[Callable[..., nn.Module]] = None,
|
||||
norm_layer: Optional[Callable[..., nn.Module]] = None
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.upsample_layers = [] # stem and 3 intermediate downsampling conv layers
|
||||
stem = tf.keras.Sequential([
|
||||
Conv2DTranspose(dims[0], kernel_size=4, strides=4, padding='same'),
|
||||
# UpSampling2D((4, 4), interpolation='nearest'),
|
||||
# Conv2D(dims[0], kernel_size=4, strides=1, padding='same'),
|
||||
# Dense(dims[0]),
|
||||
LayerNormalization(epsilon=1e-6)]
|
||||
|
||||
if not block_setting:
|
||||
raise ValueError("The block_setting should not be empty")
|
||||
elif not (isinstance(block_setting, Sequence) and all([isinstance(s, CNBlockConfig) for s in block_setting])):
|
||||
raise TypeError("The block_setting should be List[CNBlockConfig]")
|
||||
|
||||
if block is None:
|
||||
block = CNBlock
|
||||
|
||||
if norm_layer is None:
|
||||
norm_layer = partial(LayerNorm2d, eps=1e-6)
|
||||
|
||||
layers: List[nn.Module] = []
|
||||
|
||||
|
||||
# add bottleneck which will be executed at the end, to process after the final larger upsampling
|
||||
|
||||
|
||||
|
||||
# Stem
|
||||
firstconv_output_channels = block_setting[0].input_channels
|
||||
|
||||
layers.append(
|
||||
nn.Sequential(
|
||||
nn.Conv2d(16, 1, kernel_size=7, padding=3, bias=True),
|
||||
|
||||
)
|
||||
self.upsample_layers.append(stem)
|
||||
for i in range(len(depths) - 1):
|
||||
downsample_layer = tf.keras.Sequential([
|
||||
LayerNormalization(epsilon=1e-6),
|
||||
Conv2DTranspose(dims[i+1], kernel_size=2, strides=2, padding='same')
|
||||
# UpSampling2D((2,2), interpolation='nearest'),
|
||||
# Conv2D(dims[i+1], kernel_size=2, strides=1, padding='same'),
|
||||
# Dense(dims[i+1]),
|
||||
])
|
||||
self.upsample_layers.append(downsample_layer)
|
||||
|
||||
self.stages = [] # 4 feature resolution stages, each consisting of multiple residual blocks
|
||||
dp_rates = [x for x in np.linspace(0, drop_path_rate, sum(depths))]
|
||||
cur = 0
|
||||
for i in range(len(depths)):
|
||||
stage = tf.keras.Sequential(
|
||||
[Block(dim=dims[i], drop_path=dp_rates[cur + j], kernel_size=kernel_size,
|
||||
layer_scale_init_value=layer_scale_init_value, prefix=f'block{i}') for j in range(depths[i])]
|
||||
)
|
||||
self.stages.append(stage)
|
||||
cur += depths[i]
|
||||
|
||||
self.last = Dense(3, activation='tanh')
|
||||
stage: List[nn.Module] = []
|
||||
for _ in range(3):
|
||||
stage.append(block(16, layer_scale, 0.0))
|
||||
layers.append(nn.Sequential(*stage))
|
||||
|
||||
def call(self, x, training):
|
||||
for i in reversed(range(len(self.upsample_layers))):
|
||||
x = self.upsample_layers[i](x, training=training)
|
||||
x = self.stages[i](x, training=training)
|
||||
x = self.last(x, training=training)
|
||||
layers.append(
|
||||
# Convtr NormActivation(
|
||||
# firstconv_output_channels,
|
||||
# 3,
|
||||
# kernel_size=4,
|
||||
# stride=4,
|
||||
# padding=0,
|
||||
# norm_layer=None,
|
||||
# activation_layer=None,
|
||||
# bias=True,
|
||||
# )
|
||||
nn.Sequential(
|
||||
nn.ConvTranspose2d(
|
||||
firstconv_output_channels,
|
||||
16,
|
||||
kernel_size=4,
|
||||
stride=4,
|
||||
padding=0,
|
||||
),
|
||||
)
|
||||
)
|
||||
|
||||
total_stage_blocks = sum(cnf.num_layers for cnf in block_setting)
|
||||
stage_block_id = 0
|
||||
for cnf in block_setting:
|
||||
# Bottlenecks
|
||||
stage: List[nn.Module] = []
|
||||
for _ in range(cnf.num_layers):
|
||||
# adjust stochastic depth probability based on the depth of the stage block
|
||||
sd_prob = stochastic_depth_prob * stage_block_id / (total_stage_blocks - 1.0)
|
||||
stage.append(block(cnf.input_channels, layer_scale, sd_prob))
|
||||
stage_block_id += 1
|
||||
layers.append(nn.Sequential(*stage))
|
||||
if cnf.out_channels is not None:
|
||||
# Upsampling
|
||||
layers.append(
|
||||
nn.Sequential(
|
||||
norm_layer(cnf.out_channels),
|
||||
nn.ConvTranspose2d(cnf.out_channels, cnf.input_channels, kernel_size=2, stride=2),
|
||||
)
|
||||
)
|
||||
|
||||
self.features = nn.Sequential(*reversed(layers))
|
||||
|
||||
for m in self.modules():
|
||||
if isinstance(m, (nn.Conv2d, nn.Linear)):
|
||||
nn.init.trunc_normal_(m.weight, std=0.02)
|
||||
if m.bias is not None:
|
||||
nn.init.zeros_(m.bias)
|
||||
|
||||
def _forward_impl(self, x: Tensor) -> Tensor:
|
||||
x = self.features(x)
|
||||
# x = torch.tanh(x)
|
||||
return x
|
||||
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
return self._forward_impl(x)
|
||||
|
||||
class ConvNeXtReverseLabels(tf.keras.Model):
|
||||
r""" ConvNeXt
|
||||
A Tensorflow keras impl of : `A ConvNet for the 2020s` -
|
||||
https://arxiv.org/pdf/2201.03545.pdf
|
||||
|
||||
Args:
|
||||
num_classes (int): Number of classes for classification head. Default: 1000
|
||||
depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
|
||||
dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
|
||||
include_top (bool): whether to add head or just use it as feature extractor. Default: True
|
||||
drop_path_rate (float): Stochastic depth rate. Default: 0.
|
||||
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
||||
head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
|
||||
"""
|
||||
def _convnext(
|
||||
block_setting: List[CNBlockConfig],
|
||||
stochastic_depth_prob: float,
|
||||
reversed: bool,
|
||||
is_label: bool,
|
||||
nr_of_labels: Optional[int],
|
||||
kernel_size: int,
|
||||
device='cuda',
|
||||
**kwargs: Any,
|
||||
) -> ConvNeXtBase:
|
||||
|
||||
def __init__(self,
|
||||
depths=[3, 3, 9, 3], dims=[96, 192, 384, 768],
|
||||
drop_path_rate=0., layer_scale_init_value=1e-6, head_init_scale=1.,
|
||||
nr_of_classes=7):
|
||||
if reversed:
|
||||
model = ConvNeXtReverseBase2(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, 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:
|
||||
super().__init__()
|
||||
self.downsample_layers = [] # stem and 3 intermediate downsampling conv layers
|
||||
stem = tf.keras.Sequential([
|
||||
Conv2DTranspose(dims[0], kernel_size=4, strides=4, padding='same'),
|
||||
LayerNormalization(epsilon=1e-6)]
|
||||
)
|
||||
self.downsample_layers.append(stem)
|
||||
for i in range(len(depths) - 1):
|
||||
downsample_layer = tf.keras.Sequential([
|
||||
LayerNormalization(epsilon=1e-6),
|
||||
Conv2DTranspose(dims[i+1], kernel_size=2, strides=2, padding='same')]
|
||||
)
|
||||
self.downsample_layers.append(downsample_layer)
|
||||
|
||||
self.stages = [] # 4 feature resolution stages, each consisting of multiple residual blocks
|
||||
dp_rates = [x for x in np.linspace(0, drop_path_rate, sum(depths))]
|
||||
cur = 0
|
||||
for i in range(len(depths)):
|
||||
stage = tf.keras.Sequential(
|
||||
[Block(dim=dims[i], drop_path=dp_rates[cur + j],
|
||||
layer_scale_init_value=layer_scale_init_value, prefix=f'block{i}') for j in range(depths[i])]
|
||||
)
|
||||
self.stages.append(stage)
|
||||
cur += depths[i]
|
||||
self.model = _convnext(block_setting, stochastic_depth_prob, reversed, is_label, nr_of_labels, kernel_size, device=device, **kwargs)
|
||||
|
||||
# self.last = Dense(nr_of_classes, activation='softmax')
|
||||
self.last = Dense(nr_of_classes - 1)
|
||||
|
||||
def call(self, x, training):
|
||||
for i in reversed(range(len(self.downsample_layers))):
|
||||
x = self.downsample_layers[i](x, training=training)
|
||||
x = self.stages[i](x, training=training)
|
||||
x = self.last(x, training=training)
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
x = self.model(x)
|
||||
return x
|
||||
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
model = ConvNeXt().to('cuda')
|
||||
test = model(torch.zeros((4, 3, 512, 256)).to('cuda'))
|
||||
|
||||
|
||||
model_urls = {
|
||||
"convnext_tiny_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_tiny_1k_224_ema.h5",
|
||||
"convnext_small_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_small_1k_224_ema.h5",
|
||||
"convnext_base_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_base_22k_1k_224.h5",
|
||||
"convnext_base_384": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_base_22k_1k_384.h5",
|
||||
"convnext_large_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_large_22k_1k_224.h5",
|
||||
"convnext_large_384": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_large_22k_1k_384.h5",
|
||||
"convnext_xlarge_224": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_xlarge_22k_1k_224_ema.h5",
|
||||
"convnext_xlarge_384": "https://github.com/bamps53/convnext-tf/releases/download/v0.1/convnext_xlarge_22k_1k_384_ema.h5",
|
||||
}
|
||||
|
||||
|
||||
model_configs = dict(
|
||||
convnext_tiny=dict(
|
||||
depths=[3, 3, 9, 3],
|
||||
dims=[8, 16, 32, 512]
|
||||
),
|
||||
convnext_small=dict(
|
||||
depths=[3, 3, 27, 3],
|
||||
dims=[96, 192, 384, 768]
|
||||
),
|
||||
convnext_base=dict(
|
||||
depths=[3, 3, 27, 3],
|
||||
dims=[128, 256, 512, 1024]
|
||||
),
|
||||
convnext_large=dict(
|
||||
depths=[3, 3, 27, 3],
|
||||
dims=[192, 384, 768, 1536]
|
||||
),
|
||||
convnext_xlarge=dict(
|
||||
depths=[3, 3, 27, 3],
|
||||
dims=[256, 512, 1024, 2048]
|
||||
),
|
||||
)
|
||||
|
||||
pass
|
||||
@@ -0,0 +1,9 @@
|
||||
import os
|
||||
from configs.ae_static import config
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
|
||||
name = config['name']
|
||||
current_file = '/'.join(__file__.split('/')[-2:])
|
||||
os.system(f'nohup python scripts/ae_static_train.py --name lhunor_{name} > nohup/torch/{name}.out 2>&1 &')
|
||||
@@ -0,0 +1,197 @@
|
||||
import os
|
||||
import sys
|
||||
sys.path.append(os.path.dirname(os.path.abspath(__file__)) + '/..')
|
||||
|
||||
from configs.ae_static import config
|
||||
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
|
||||
os.environ["CUDA_VISIBLE_DEVICES"] = config['gpu']
|
||||
|
||||
|
||||
import time
|
||||
import argparse
|
||||
import wandb
|
||||
import random
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch import nn
|
||||
from torchmetrics import MultiScaleStructuralSimilarityIndexMeasure as MSSSIM
|
||||
from custom_utils.custom_utils import UV2Mesh, get_summary_html, load_model, maskFacesBatch, get_summary_str
|
||||
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 custom_utils.custom_utils import batch_to_dict, plot_3d
|
||||
from models.cond_encoder import AutoEncoder
|
||||
from utils.checkpoints import CheckpointSaver
|
||||
from custom_utils.custom_utils import Accumulator
|
||||
|
||||
|
||||
def init_wandb():
|
||||
run = wandb.init(project='ae', name=config['name'], notes=config['notes'], dir='/data-local/127/temp/wandb', tags=config['tags'])
|
||||
artifact = wandb.Artifact('code', type='code')
|
||||
artifact.add_file('scripts/ae_static_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()
|
||||
pred = model([batch['uv_static'], batch['uv_mask']])
|
||||
|
||||
mesh_pred = uv2mesh(pred)
|
||||
mesh_orig = batch['mesh_static']
|
||||
|
||||
e_loss = 1e-2 * torch.mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
|
||||
l1_loss = torch.mean(torch.sum((torch.abs(batch['uv_static'] - pred)), axis=1))
|
||||
ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_static'] + 1))
|
||||
loss = 1e3 * (1 - ms_ssim)
|
||||
|
||||
optimizer.zero_grad()
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
|
||||
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/l1_loss': l1_loss, 'train/euclid_error': metric, 'train/ms_ssim': ms_ssim, }
|
||||
|
||||
|
||||
def valid_step(batch):
|
||||
model.eval()
|
||||
|
||||
with torch.inference_mode():
|
||||
pred = model([batch['uv_static'], batch['uv_mask']])
|
||||
|
||||
mesh_pred = uv2mesh(pred)
|
||||
mesh_orig = batch['mesh_static']
|
||||
|
||||
e_loss = 1e-2 * torch.mean(edge_loss(mesh_pred, mesh_orig, batch['mesh_mask']))
|
||||
l1_loss = torch.mean(torch.sum((torch.abs(batch['uv_static'] - pred)), axis=1))
|
||||
ms_ssim = torch.mean(_MSSSIM(pred + 1, batch['uv_static'] + 1))
|
||||
loss = 1e3 * (1 - ms_ssim)
|
||||
|
||||
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/l1_loss': l1_loss, 'valid/euclid_error': metric, 'valid/ms_ssim': ms_ssim, }
|
||||
|
||||
|
||||
def plot_everything(batch: dict, type: str):
|
||||
pred = model([batch['uv_static'], batch['uv_mask']]).detach()
|
||||
|
||||
mesh_pred = uv2mesh(pred)
|
||||
mesh_orig = batch['mesh_static']
|
||||
|
||||
orig_img = torch.permute(batch['uv_static'], [0,2,3,1])
|
||||
pred_img = torch.permute(pred, [0,2,3,1])
|
||||
img = torch.concat((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)
|
||||
garment_static, faces = maskFacesBatch(mesh_pred.detach().cpu().numpy(), F, mask)
|
||||
gt = [mesh_orig[i].cpu().numpy()[mask[i], :] for i in range(pred.shape[0])]
|
||||
|
||||
body = gt
|
||||
garment_posed = garment_static
|
||||
collision = [np.concatenate((body[i], garment_posed[i]), axis=0) for i in range(len(garment_posed))]
|
||||
collision_faces = [np.concatenate((faces[i], faces[i] + np.sum(mask[i])), 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_static[i], faces[i], collision[i], collision_faces[i], collision_color[i]) for i in range(4) }
|
||||
|
||||
return { **figs, **{ 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)) } }
|
||||
|
||||
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)
|
||||
ckpt_saver_dec(model.decoder, step, optimizer)
|
||||
|
||||
print(step)
|
||||
|
||||
step += 1
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
init_wandb()
|
||||
# device = "cuda" if torch.cuda.is_available() else "cpu"
|
||||
|
||||
parser = argparse.ArgumentParser(description='')
|
||||
parser.add_argument('--name', type=str)
|
||||
args = vars(parser.parse_args())
|
||||
|
||||
resolution_modifier = ''
|
||||
if config['resolution'] != 512:
|
||||
resolution_modifier = '_' + str(config['resolution'])
|
||||
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'])
|
||||
dataset_valid = Dataset(config['valid_dir'], config['signature'], device=config['device'], frames_file=config['valid_frames'])
|
||||
train_dataloader = DataLoader(dataset_train, batch_size=4, drop_last=True, shuffle=False)
|
||||
valid_dataloader = DataLoader(dataset_valid, batch_size=4, drop_last=True, shuffle=False)
|
||||
train_accumulator = Accumulator()
|
||||
valid_accumulator = Accumulator()
|
||||
|
||||
model : AutoEncoder = AutoEncoder(config['cond_ae'])
|
||||
optimizer = optim.Adam(model.parameters(), lr=config['lr'])
|
||||
|
||||
_MSSSIM = MSSSIM(data_range=2).to(config['device'])
|
||||
|
||||
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/' + config['name'] + '_enc', top_n=20)
|
||||
ckpt_saver_dec = CheckpointSaver('checkpoints/iccv/' + config['name'] + '_dec', top_n=20)
|
||||
|
||||
run_name = config['name']
|
||||
print(f'started training: {run_name}')
|
||||
train()
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,21 @@
|
||||
from typing import Optional
|
||||
|
||||
class CNBlockConfig:
|
||||
# Stores information listed at Section 3 of the ConvNeXt paper
|
||||
def __init__(
|
||||
self,
|
||||
input_channels: int,
|
||||
out_channels: Optional[int],
|
||||
num_layers: int,
|
||||
) -> None:
|
||||
self.input_channels = input_channels
|
||||
self.out_channels = out_channels
|
||||
self.num_layers = num_layers
|
||||
|
||||
def __repr__(self) -> str:
|
||||
s = self.__class__.__name__ + "("
|
||||
s += "input_channels={input_channels}"
|
||||
s += ", out_channels={out_channels}"
|
||||
s += ", num_layers={num_layers}"
|
||||
s += ")"
|
||||
return s.format(**self.__dict__)
|
||||
Reference in New Issue
Block a user