CTM-notebook / model.py
BasavarajMurali's picture
CTM notebook hosting on Hugging Face
b11690a
Raw
History Blame Contribute Delete
149 kB
import torch as th
import numpy as np
from scipy.stats import norm
from abc import ABC, abstractmethod
from torch.nn.utils import spectral_norm
import torch.nn as nn
import numpy as np
import math
import torch.nn.functional as F
import timm
from distillation_utils import *
from torch.optim import RAdam
from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
import cv2
import scipy
_src_to_module_dict = dict()
_module_to_src_dict = dict()
_import_hooks = []
_version = 6
_decorators = set()
import copy
import uuid
import types
import sys
from typing import Any, List, Tuple, Union, Optional
import io
import pickle
import inspect
from torch.nn.functional import silu
import shutil
if th.cuda.is_available():
dev = th.device("cuda")
print(dev)
else:
dev = th.device("cpu")
class EasyDict(dict):
"""Convenience class that behaves like a dict but allows access with the attribute syntax."""
def __getattr__(self, name: str) -> Any:
try:
return self[name]
except KeyError:
raise AttributeError(name)
def __setattr__(self, name: str, value: Any) -> None:
self[name] = value
def __delattr__(self, name: str) -> None:
del self[name]
### Creating model and diffusion
class ScheduleSampler(ABC):
"""
A distribution over timesteps in the diffusion process, intended to reduce
variance of the objective.
By default, samplers perform unbiased importance sampling, in which the
objective's mean is unchanged.
However, subclasses may override sample() to change how the resampled
terms are reweighted, allowing for actual changes in the objective.
"""
@abstractmethod
def weights(self, num_heun_step):
"""
Get a numpy array of weights, one per diffusion step.
The weights needn't be normalized, but must be positive.
"""
def sample_t(self, batch_size, device, num_heun_step=1, time_continuous=False):
"""
Importance-sample timesteps for a batch.
:param batch_size: the number of timesteps.
:param device: the torch device to save to.
:return: a tuple (timesteps, weights):
- timesteps: a tensor of timestep indices.
- weights: a tensor of weights to scale the resulting losses.
"""
if time_continuous:
indices_np = np.random.rand(batch_size)
indices = th.from_numpy(indices_np).to(device) * (1. - num_heun_step)
weights = th.ones_like(indices).float().to(device)
else:
w = self.weights(num_heun_step)
p = w / np.sum(w)
indices_np = np.random.choice(len(p), size=(batch_size,), p=p)
indices = th.from_numpy(indices_np).long().to(device)
weights_np = 1 / (len(p) * p[indices_np])
weights = th.from_numpy(weights_np).float().to(device)
return indices, weights
def sample_s(self, args, batch_size, device, indices, num_heun_step=1, time_continuous=False, N=40):
if time_continuous:
new_indices = th.from_numpy(np.random.rand(indices.shape[0])).to(indices.device) * (1. - indices - num_heun_step)\
+ indices + num_heun_step
else:
if args.sample_s_strategy == 'smallest':
new_indices = th.ones(indices.shape[0], device=indices.device) * (N - 1)
elif args.sample_s_strategy == 'uniform':
new_indices = th.from_numpy(np.random.randint(
low=(indices + num_heun_step).cpu().detach().numpy(), high=N,
size=(indices.shape[0],),
dtype=int)).to(indices.device)
return new_indices
class UniformSampler(ScheduleSampler):
def __init__(self, num_timesteps):
self._weights = np.ones([num_timesteps])
self.num_timesteps = num_timesteps
def weights(self, num_heun_step=1):
#if num_heun_step == 1:
# return self._weights
#else:
return np.ones([self.num_timesteps - num_heun_step])
class HalfLogNormalHalfUniformSampler:
def __init__(self, args, p_mean=-1.2, p_std=1.2, even=False):
self.args = args
self.p_mean = p_mean
self.p_std = p_std
self.even = even
if self.even:
self.inv_cdf = lambda x: norm.ppf(x, loc=p_mean, scale=p_std)
# self.rank, self.size = dist.get_rank(), dist.get_world_size()
def get_t(self, t):
t = self.args.sigma_max ** (1 / self.args.rho) + t * (
self.args.sigma_min ** (1 / self.args.rho) - self.args.sigma_max ** (1 / self.args.rho)
)
t = t ** self.args.rho
return t
def sample(self, bs, device, num_heun_step=1):
log_sigmas = self.p_mean + self.p_std * th.randn(bs//2, device=device)
sigmas = th.exp(log_sigmas)
t = th.rand(bs - bs//2, device=device) * self.args.diffusion_mult # it determines how large sigma to sample from
sigmas = th.cat((sigmas, self.get_t(t))).view(-1)
weights = th.ones_like(sigmas)
return sigmas, weights
def create_named_schedule_sampler(args, name, num_timesteps):
"""
Create a ScheduleSampler from a library of pre-defined samplers.
:param name: the name of the sampler.
:param diffusion: the diffusion object to sample for.
"""
if name == "uniform":
return UniformSampler(num_timesteps)
elif name == 'halflognormal':
return HalfLogNormalHalfUniformSampler(args)
else:
raise NotImplementedError(f"unknown schedule sampler: {name}")
def _src_to_module(src):
r"""Get or create a Python module for the given source code.
"""
module = _src_to_module_dict.get(src, None)
if module is None:
module_name = "_imported_module_" + uuid.uuid4().hex
module = types.ModuleType(module_name)
sys.modules[module_name] = module
_module_to_src_dict[module] = src
_src_to_module_dict[src] = module
exec(src, module.__dict__) # pylint: disable=exec-used
return module
def _reconstruct_persistent_obj(meta):
r"""Hook that is called internally by the `pickle` module to unpickle
a persistent object.
"""
meta = EasyDict(meta)
meta.state = EasyDict(meta.state)
for hook in _import_hooks:
meta = hook(meta)
assert meta is not None
assert meta.version == _version
module = _src_to_module(meta.module_src)
assert meta.type == 'class'
orig_class = module.__dict__[meta.class_name]
decorator_class = persistent_class(orig_class)
obj = decorator_class.__new__(decorator_class)
setstate = getattr(obj, '__setstate__', None)
if callable(setstate):
setstate(meta.state) # pylint: disable=not-callable
else:
obj.__dict__.update(meta.state)
return obj
def _check_pickleable(obj):
r"""Check that the given object is pickleable, raising an exception if
it is not. This function is expected to be considerably more efficient
than actually pickling the object.
"""
def recurse(obj):
if isinstance(obj, (list, tuple, set)):
return [recurse(x) for x in obj]
if isinstance(obj, dict):
return [[recurse(x), recurse(y)] for x, y in obj.items()]
if isinstance(obj, (str, int, float, bool, bytes, bytearray)):
return None # Python primitive types are pickleable.
if f'{type(obj).__module__}.{type(obj).__name__}' in ['numpy.ndarray', 'torch.Tensor', 'torch.nn.parameter.Parameter']:
return None # NumPy arrays and PyTorch tensors are pickleable.
if is_persistent(obj):
return None # Persistent objects are pickleable, by virtue of the constructor check.
return obj
with io.BytesIO() as f:
pickle.dump(recurse(obj), f)
def _module_to_src(module):
r"""Query the source code of a given Python module.
"""
src = _module_to_src_dict.get(module, None)
if src is None:
src = inspect.getsource(module)
_module_to_src_dict[module] = src
_src_to_module_dict[src] = module
return src
def is_persistent(obj):
r"""Test whether the given object or class is persistent, i.e.,
whether it will save its source code when pickled.
"""
try:
if obj in _decorators:
return True
except TypeError:
pass
return type(obj) in _decorators # pylint: disable=unidiomatic-typecheck
def persistent_class(orig_class):
r"""Class decorator that extends a given class to save its source code
when pickled.
Example:
from torch_utils import persistence
@persistence.persistent_class
class MyNetwork(torch.nn.Module):
def __init__(self, num_inputs, num_outputs):
super().__init__()
self.fc = MyLayer(num_inputs, num_outputs)
...
@persistence.persistent_class
class MyLayer(torch.nn.Module):
...
When pickled, any instance of `MyNetwork` and `MyLayer` will save its
source code alongside other internal state (e.g., parameters, buffers,
and submodules). This way, any previously exported pickle will remain
usable even if the class definitions have been modified or are no
longer available.
The decorator saves the source code of the entire Python module
containing the decorated class. It does *not* save the source code of
any imported modules. Thus, the imported modules must be available
during unpickling, also including `torch_utils.persistence` itself.
It is ok to call functions defined in the same module from the
decorated class. However, if the decorated class depends on other
classes defined in the same module, they must be decorated as well.
This is illustrated in the above example in the case of `MyLayer`.
It is also possible to employ the decorator just-in-time before
calling the constructor. For example:
cls = MyLayer
if want_to_make_it_persistent:
cls = persistence.persistent_class(cls)
layer = cls(num_inputs, num_outputs)
As an additional feature, the decorator also keeps track of the
arguments that were used to construct each instance of the decorated
class. The arguments can be queried via `obj.init_args` and
`obj.init_kwargs`, and they are automatically pickled alongside other
object state. This feature can be disabled on a per-instance basis
by setting `self._record_init_args = False` in the constructor.
A typical use case is to first unpickle a previous instance of a
persistent class, and then upgrade it to use the latest version of
the source code:
with open('old_pickle.pkl', 'rb') as f:
old_net = pickle.load(f)
new_net = MyNetwork(*old_obj.init_args, **old_obj.init_kwargs)
misc.copy_params_and_buffers(old_net, new_net, require_all=True)
"""
assert isinstance(orig_class, type)
if is_persistent(orig_class):
return orig_class
assert orig_class.__module__ in sys.modules
orig_module = sys.modules[orig_class.__module__]
orig_module_src = _module_to_src(orig_module)
class Decorator(orig_class):
_orig_module_src = orig_module_src
_orig_class_name = orig_class.__name__
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
record_init_args = getattr(self, '_record_init_args', True)
self._init_args = copy.deepcopy(args) if record_init_args else None
self._init_kwargs = copy.deepcopy(kwargs) if record_init_args else None
assert orig_class.__name__ in orig_module.__dict__
_check_pickleable(self.__reduce__())
@property
def init_args(self):
assert self._init_args is not None
return copy.deepcopy(self._init_args)
@property
def init_kwargs(self):
assert self._init_kwargs is not None
return EasyDict(copy.deepcopy(self._init_kwargs))
def __reduce__(self):
fields = list(super().__reduce__())
fields += [None] * max(3 - len(fields), 0)
if fields[0] is not _reconstruct_persistent_obj:
meta = dict(type='class', version=_version, module_src=self._orig_module_src, class_name=self._orig_class_name, state=fields[2])
fields[0] = _reconstruct_persistent_obj # reconstruct func
fields[1] = (meta,) # reconstruct args
fields[2] = None # state dict
return tuple(fields)
Decorator.__name__ = orig_class.__name__
Decorator.__module__ = orig_class.__module__
_decorators.add(Decorator)
return Decorator
def weight_init(shape, mode, fan_in, fan_out):
if mode == 'xavier_uniform': return np.sqrt(6 / (fan_in + fan_out)) * (th.rand(*shape) * 2 - 1)
if mode == 'xavier_normal': return np.sqrt(2 / (fan_in + fan_out)) * th.randn(*shape)
if mode == 'kaiming_uniform': return np.sqrt(3 / fan_in) * (th.rand(*shape) * 2 - 1)
if mode == 'kaiming_normal': return np.sqrt(1 / fan_in) * th.randn(*shape)
raise ValueError(f'Invalid init mode "{mode}"')
class AttentionOp(th.autograd.Function):
@staticmethod
def forward(ctx, q, k):
w = th.einsum('ncq,nck->nqk', q.to(th.float32), (k / np.sqrt(k.shape[1])).to(th.float32)).softmax(dim=2).to(q.dtype)
ctx.save_for_backward(q, k, w)
return w
@staticmethod
def backward(ctx, dw):
q, k, w = ctx.saved_tensors
db = th._softmax_backward_data(grad_output=dw.to(th.float32), output=w.to(th.float32), dim=2, input_dtype=th.float32)
dq = th.einsum('nck,nqk->ncq', k.to(th.float32), db).to(q.dtype) / np.sqrt(k.shape[1])
dk = th.einsum('ncq,nqk->nck', q.to(th.float32), db).to(k.dtype) / np.sqrt(k.shape[1])
return dq, dk
@persistent_class
class Linear(th.nn.Module):
def __init__(self, in_features, out_features, bias=True, init_mode='kaiming_normal', init_weight=1, init_bias=0):
super().__init__()
self.in_features = in_features
self.out_features = out_features
init_kwargs = dict(mode=init_mode, fan_in=in_features, fan_out=out_features)
self.weight = th.nn.Parameter(weight_init([out_features, in_features], **init_kwargs) * init_weight)
self.bias = th.nn.Parameter(weight_init([out_features], **init_kwargs) * init_bias) if bias else None
def forward(self, x):
x = x @ self.weight.to(x.dtype).t()
if self.bias is not None:
x = x.add_(self.bias.to(x.dtype))
return x
@persistent_class
class Conv2d(th.nn.Module):
def __init__(self,
in_channels, out_channels, kernel, bias=True, up=False, down=False,
resample_filter=[1,1], fused_resample=False, init_mode='kaiming_normal', init_weight=1, init_bias=0,
):
assert not (up and down)
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.up = up
self.down = down
self.fused_resample = fused_resample
init_kwargs = dict(mode=init_mode, fan_in=in_channels*kernel*kernel, fan_out=out_channels*kernel*kernel)
self.weight = th.nn.Parameter(weight_init([out_channels, in_channels, kernel, kernel], **init_kwargs) * init_weight) if kernel else None
self.bias = th.nn.Parameter(weight_init([out_channels], **init_kwargs) * init_bias) if kernel and bias else None
f = th.as_tensor(resample_filter, dtype=th.float32)
f = f.ger(f).unsqueeze(0).unsqueeze(1) / f.sum().square()
self.register_buffer('resample_filter', f if up or down else None)
def forward(self, x):
w = self.weight.to(x.dtype) if self.weight is not None else None
b = self.bias.to(x.dtype) if self.bias is not None else None
f = self.resample_filter.to(x.dtype) if self.resample_filter is not None else None
w_pad = w.shape[-1] // 2 if w is not None else 0
f_pad = (f.shape[-1] - 1) // 2 if f is not None else 0
if self.fused_resample and self.up and w is not None:
x = th.nn.functional.conv_transpose2d(x, f.mul(4).tile([self.in_channels, 1, 1, 1]), groups=self.in_channels, stride=2, padding=max(f_pad - w_pad, 0))
x = th.nn.functional.conv2d(x, w, padding=max(w_pad - f_pad, 0))
elif self.fused_resample and self.down and w is not None:
x = th.nn.functional.conv2d(x, w, padding=w_pad+f_pad)
x = th.nn.functional.conv2d(x, f.tile([self.out_channels, 1, 1, 1]), groups=self.out_channels, stride=2)
else:
if self.up:
x = th.nn.functional.conv_transpose2d(x, f.mul(4).tile([self.in_channels, 1, 1, 1]), groups=self.in_channels, stride=2, padding=f_pad)
if self.down:
x = th.nn.functional.conv2d(x, f.tile([self.in_channels, 1, 1, 1]), groups=self.in_channels, stride=2, padding=f_pad)
if w is not None:
x = th.nn.functional.conv2d(x, w, padding=w_pad)
if b is not None:
x = x.add_(b.reshape(1, -1, 1, 1))
return x
@persistent_class
class GroupNorm(th.nn.Module):
def __init__(self, num_channels, num_groups=32, min_channels_per_group=4, eps=1e-5):
super().__init__()
self.num_groups = min(num_groups, num_channels // min_channels_per_group)
self.eps = eps
self.weight = th.nn.Parameter(th.ones(num_channels))
self.bias = th.nn.Parameter(th.zeros(num_channels))
def forward(self, x):
x = th.nn.functional.group_norm(x, num_groups=self.num_groups, weight=self.weight.to(x.dtype), bias=self.bias.to(x.dtype), eps=self.eps)
return x
@persistent_class
class PositionalEmbedding(th.nn.Module):
def __init__(self, num_channels, max_positions=10000, endpoint=False):
super().__init__()
self.num_channels = num_channels
self.max_positions = max_positions
self.endpoint = endpoint
def forward(self, x):
freqs = th.arange(start=0, end=self.num_channels//2, dtype=th.float32, device=x.device)
freqs = freqs / (self.num_channels // 2 - (1 if self.endpoint else 0))
freqs = (1 / self.max_positions) ** freqs
x = x.ger(freqs.to(x.dtype))
x = th.cat([x.cos(), x.sin()], dim=1)
return x
@persistent_class
class FourierEmbedding(th.nn.Module):
def __init__(self, num_channels, scale=16):
super().__init__()
self.register_buffer('freqs', th.randn(num_channels // 2) * scale)
def forward(self, x):
x = x.ger((2 * np.pi * self.freqs).to(x.dtype))
x = th.cat([x.cos(), x.sin()], dim=1)
return x
@persistent_class
class UNetBlock(th.nn.Module):
def __init__(self,
in_channels, out_channels, emb_channels, up=False, down=False, attention=False,
num_heads=None, channels_per_head=64, dropout=0, skip_scale=1, eps=1e-5,
resample_filter=[1,1], resample_proj=False, adaptive_scale=True,
init=dict(), init_zero=dict(init_weight=0), init_attn=None,
training_mode='', linear_probing=False,
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.emb_channels = emb_channels
self.num_heads = 0 if not attention else num_heads if num_heads is not None else out_channels // channels_per_head
self.dropout = dropout
self.skip_scale = skip_scale
self.adaptive_scale = adaptive_scale
self.training_mode = training_mode
self.linear_probing = linear_probing
self.norm0 = GroupNorm(num_channels=in_channels, eps=eps)
self.conv0 = Conv2d(in_channels=in_channels, out_channels=out_channels, kernel=3, up=up, down=down, resample_filter=resample_filter, **init)
self.affine = Linear(in_features=emb_channels, out_features=out_channels*(2 if adaptive_scale else 1), **init)
if self.training_mode == 'ctm':
self.affine_s = Linear(in_features=emb_channels, out_features=out_channels*(2 if adaptive_scale else 1), **init)
self.norm1 = GroupNorm(num_channels=out_channels, eps=eps)
self.conv1 = Conv2d(in_channels=out_channels, out_channels=out_channels, kernel=3, **init_zero)
self.skip = None
if out_channels != in_channels or up or down:
kernel = 1 if resample_proj or out_channels!= in_channels else 0
self.skip = Conv2d(in_channels=in_channels, out_channels=out_channels, kernel=kernel, up=up, down=down, resample_filter=resample_filter, **init)
if self.num_heads:
self.norm2 = GroupNorm(num_channels=out_channels, eps=eps)
self.qkv = Conv2d(in_channels=out_channels, out_channels=out_channels*3, kernel=1, **(init_attn if init_attn is not None else init))
self.proj = Conv2d(in_channels=out_channels, out_channels=out_channels, kernel=1, **init_zero)
if linear_probing:
self.norm0_train = GroupNorm(num_channels=in_channels, eps=eps)
self.conv0_train = Conv2d(in_channels=in_channels, out_channels=out_channels, kernel=3, up=up, down=down,
resample_filter=resample_filter, **init)
self.norm1_train = GroupNorm(num_channels=out_channels, eps=eps)
self.conv1_train = Conv2d(in_channels=out_channels, out_channels=out_channels, kernel=3, **init_zero)
self.skip_train = None
if out_channels != in_channels or up or down:
kernel = 1 if resample_proj or out_channels != in_channels else 0
self.skip_train = Conv2d(in_channels=in_channels, out_channels=out_channels, kernel=kernel, up=up, down=down,
resample_filter=resample_filter, **init)
if self.num_heads:
self.norm2_train = GroupNorm(num_channels=out_channels, eps=eps)
self.qkv_train = Conv2d(in_channels=out_channels, out_channels=out_channels * 3, kernel=1,
**(init_attn if init_attn is not None else init))
self.proj_train = Conv2d(in_channels=out_channels, out_channels=out_channels, kernel=1, **init_zero)
def forward(self, x, emb, emb_s=None, emb_t=None):
orig = x
x = self.conv0(silu(self.norm0(x)))
params = self.affine(emb).unsqueeze(2).unsqueeze(3).to(x.dtype)
if self.training_mode == 'ctm':
params_s = self.affine_s(emb_s).unsqueeze(2).unsqueeze(3).to(x.dtype)
if not self.linear_probing:
params = params + params_s
if self.adaptive_scale:
scale, shift = params.chunk(chunks=2, dim=1)
x = silu(th.addcmul(shift, self.norm1(x), scale + 1))
else:
x = silu(self.norm1(x.add_(params)))
x = self.conv1(th.nn.functional.dropout(x, p=self.dropout, training=self.training))
x = x.add_(self.skip(orig) if self.skip is not None else orig)
x = x * self.skip_scale
if self.num_heads:
q, k, v = self.qkv(self.norm2(x)).reshape(x.shape[0] * self.num_heads, x.shape[1] // self.num_heads, 3, -1).unbind(2)
w = AttentionOp.apply(q, k)
a = th.einsum('nqk,nck->ncq', w, v)
x = self.proj(a.reshape(*x.shape)).add_(x)
x = x * self.skip_scale
if self.linear_probing:
y = self.conv0_train(silu(self.norm0_train(orig)))
assert emb_t != None and self.training_mode == 'ctm'
params_t = self.affine_s(emb_t).unsqueeze(2).unsqueeze(3).to(y.dtype)
params = params_t - params_s
params_ = params + params_s
if self.adaptive_scale:
scale, shift = params_.chunk(chunks=2, dim=1)
y = silu(th.addcmul(shift, self.norm1_train(y), scale + 1))
else:
y = silu(self.norm1_train(y.add_(params_)))
y = self.conv1_train(th.nn.functional.dropout(y, p=self.dropout, training=self.training))
y = y.add_(self.skip_train(orig) if self.skip_train is not None else orig)
y = y * self.skip_scale
if self.num_heads:
q, k, v = self.qkv_train(self.norm2_train(y)).reshape(y.shape[0] * self.num_heads, y.shape[1] // self.num_heads, 3,
-1).unbind(2)
w = AttentionOp.apply(q, k)
a = th.einsum('nqk,nck->ncq', w, v)
y = self.proj_train(a.reshape(*y.shape)).add_(y)
y = y * self.skip_scale
y = y.mul(params)
return x + y
return x
@persistent_class
class SongUNet(th.nn.Module):
def __init__(self,
img_resolution, # Image resolution at input/output.
in_channels, # Number of color channels at input.
out_channels, # Number of color channels at output.
label_dim = 0, # Number of class labels, 0 = unconditional.
augment_dim = 0, # Augmentation label dimensionality, 0 = no augmentation.
model_channels = 128, # Base multiplier for the number of channels.
channel_mult = [2,2,2], # Per-resolution multipliers for the number of channels.
channel_mult_emb = 4, # Multiplier for the dimensionality of the embedding vector.
num_blocks = 4, # Number of residual blocks per resolution.
attn_resolutions = [16], # List of resolutions with self-attention.
dropout = 0.13, # Dropout probability of intermediate activations.
label_dropout = 0, # Dropout probability of class labels for classifier-free guidance.
embedding_type = 'fourier', # Timestep embedding type: 'positional' for DDPM++, 'fourier' for NCSN++.
channel_mult_noise = 2, # Timestep embedding size: 1 for DDPM++, 2 for NCSN++.
encoder_type = 'residual', # Encoder architecture: 'standard' for DDPM++, 'residual' for NCSN++.
decoder_type = 'standard', # Decoder architecture: 'standard' for both DDPM++ and NCSN++.
resample_filter = [1,3,3,1], # Resampling filter: [1,1] for DDPM++, [1,3,3,1] for NCSN++.
training_mode = '',
linear_probing=False,
):
assert embedding_type in ['fourier', 'positional']
assert encoder_type in ['standard', 'skip', 'residual']
assert decoder_type in ['standard', 'skip']
self.training_mode = training_mode
self.linear_probing = linear_probing
self.img_resolution = img_resolution
super().__init__()
self.label_dropout = label_dropout
emb_channels = model_channels * channel_mult_emb
noise_channels = model_channels * channel_mult_noise
init = dict(init_mode='xavier_uniform')
init_zero = dict(init_mode='xavier_uniform', init_weight=1e-5)
init_attn = dict(init_mode='xavier_uniform', init_weight=np.sqrt(0.2))
block_kwargs = dict(
emb_channels=emb_channels, num_heads=1, dropout=dropout, skip_scale=np.sqrt(0.5), eps=1e-6,
resample_filter=resample_filter, resample_proj=True, adaptive_scale=False,
init=init, init_zero=init_zero, init_attn=init_attn,
)
# Mapping.
self.map_noise = PositionalEmbedding(num_channels=noise_channels, endpoint=True) if embedding_type == 'positional' else FourierEmbedding(num_channels=noise_channels)
self.map_label = Linear(in_features=label_dim, out_features=noise_channels, **init) if label_dim else None
self.map_augment = Linear(in_features=augment_dim, out_features=noise_channels, bias=False, **init) if augment_dim else None
self.map_layer0 = Linear(in_features=noise_channels, out_features=emb_channels, **init)
self.map_layer1 = Linear(in_features=emb_channels, out_features=emb_channels, **init)
if self.training_mode.lower() == 'ctm':
self.map_layer0_s = Linear(in_features=noise_channels, out_features=emb_channels, **init)
self.map_layer1_s = Linear(in_features=emb_channels, out_features=emb_channels, **init)
# Encoder.
self.enc = th.nn.ModuleDict()
cout = in_channels
caux = in_channels
for level, mult in enumerate(channel_mult):
res = img_resolution >> level
if level == 0:
cin = cout
cout = model_channels
self.enc[f'{res}x{res}_conv'] = Conv2d(in_channels=cin, out_channels=cout, kernel=3, **init)
else:
self.enc[f'{res}x{res}_down'] = UNetBlock(in_channels=cout, out_channels=cout, down=True,
training_mode=training_mode, linear_probing=linear_probing, **block_kwargs)
if encoder_type == 'skip':
self.enc[f'{res}x{res}_aux_down'] = Conv2d(in_channels=caux, out_channels=caux, kernel=0, down=True, resample_filter=resample_filter)
self.enc[f'{res}x{res}_aux_skip'] = Conv2d(in_channels=caux, out_channels=cout, kernel=1, **init)
if encoder_type == 'residual':
self.enc[f'{res}x{res}_aux_residual'] = Conv2d(in_channels=caux, out_channels=cout, kernel=3, down=True, resample_filter=resample_filter, fused_resample=True, **init)
caux = cout
for idx in range(num_blocks):
cin = cout
cout = model_channels * mult
attn = (res in attn_resolutions)
self.enc[f'{res}x{res}_block{idx}'] = UNetBlock(in_channels=cin, out_channels=cout, attention=attn,
training_mode=training_mode, linear_probing=linear_probing, **block_kwargs)
skips = [block.out_channels for name, block in self.enc.items() if 'aux' not in name]
# Decoder.
self.dec = th.nn.ModuleDict()
for level, mult in reversed(list(enumerate(channel_mult))):
res = img_resolution >> level
if level == len(channel_mult) - 1:
self.dec[f'{res}x{res}_in0'] = UNetBlock(in_channels=cout, out_channels=cout, attention=True,
training_mode=training_mode, linear_probing=linear_probing, **block_kwargs)
self.dec[f'{res}x{res}_in1'] = UNetBlock(in_channels=cout, out_channels=cout,
training_mode=training_mode, linear_probing=linear_probing, **block_kwargs)
else:
self.dec[f'{res}x{res}_up'] = UNetBlock(in_channels=cout, out_channels=cout, up=True,
training_mode=training_mode, linear_probing=linear_probing, **block_kwargs)
for idx in range(num_blocks + 1):
cin = cout + skips.pop()
cout = model_channels * mult
attn = (idx == num_blocks and res in attn_resolutions)
self.dec[f'{res}x{res}_block{idx}'] = UNetBlock(in_channels=cin, out_channels=cout, attention=attn,
training_mode=training_mode, linear_probing=linear_probing, **block_kwargs)
if decoder_type == 'skip' or level == 0:
if decoder_type == 'skip' and level < len(channel_mult) - 1:
self.dec[f'{res}x{res}_aux_up'] = Conv2d(in_channels=out_channels, out_channels=out_channels, kernel=0, up=True, resample_filter=resample_filter)
self.dec[f'{res}x{res}_aux_norm'] = GroupNorm(num_channels=cout, eps=1e-6)
self.dec[f'{res}x{res}_aux_conv'] = Conv2d(in_channels=cout, out_channels=out_channels, kernel=3, **init_zero)
if self.linear_probing:
self.dec[f'{res}x{res}_aux_norm_train'] = GroupNorm(num_channels=cout, eps=1e-6)
self.dec[f'{res}x{res}_aux_lin_train'] = Linear(in_features=emb_channels, out_features=out_channels*img_resolution*img_resolution, **init)
self.dec[f'{res}x{res}_aux_conv_train'] = Conv2d(in_channels=cout, out_channels=out_channels, kernel=3,
**init_zero)
def forward(self, x, noise_labels, noise_labels_s, class_labels):
# Mapping.
emb = self.map_noise(noise_labels)
emb = emb.reshape(emb.shape[0], 2, -1).flip(1).reshape(*emb.shape) # swap sin/cos
if self.map_label is not None:
tmp = class_labels
if self.training and self.label_dropout:
tmp = tmp * (th.rand([x.shape[0], 1], device=x.device) >= self.label_dropout).to(tmp.dtype)
emb = emb + self.map_label(tmp * np.sqrt(self.map_label.in_features))
emb = silu(self.map_layer0(emb))
emb = silu(self.map_layer1(emb))
if noise_labels_s != None:
emb_s = self.map_noise(noise_labels_s)
emb_s = emb_s.reshape(emb_s.shape[0], 2, -1).flip(1).reshape(*emb_s.shape) # swap sin/cos
if self.map_label is not None:
tmp = class_labels
if self.training and self.label_dropout:
tmp = tmp * (th.rand([x.shape[0], 1], device=x.device) >= self.label_dropout).to(tmp.dtype)
emb_s = emb_s + self.map_label(tmp * np.sqrt(self.map_label.in_features))
emb_s = silu(self.map_layer0_s(emb_s))
emb_s = silu(self.map_layer1_s(emb_s))
if self.linear_probing:
emb_t = self.map_noise(noise_labels)
emb_t = emb_t.reshape(emb_t.shape[0], 2, -1).flip(1).reshape(*emb_t.shape) # swap sin/cos
emb_t = silu(self.map_layer0_s(emb_t))
emb_t = silu(self.map_layer1_s(emb_t))
# Encoder.
skips = []
aux = x
for name, block in self.enc.items():
if 'aux_down' in name:
aux = block(aux)
elif 'aux_skip' in name:
x = skips[-1] = x + block(aux)
elif 'aux_residual' in name:
x = skips[-1] = aux = (x + block(aux)) / np.sqrt(2)
else:
x = block(x, emb, emb_s=None if noise_labels_s == None else emb_s,
emb_t=emb_t if self.linear_probing else None) if isinstance(block, UNetBlock) else block(x)
skips.append(x)
# Decoder.
aux = None
tmp = None
for name, block in self.dec.items():
if 'aux_up' in name:
aux = block(aux)
elif 'aux_norm' in name:
tmp = block(x)
elif 'aux_lin' in name:
emb_mult = (block(emb_t) - block(emb_s)).reshape(-1, 3, self.img_resolution, self.img_resolution)
elif 'aux_conv' in name:
tmp = block(silu(tmp))
aux = tmp if aux is None else tmp * emb_mult + aux
else:
if x.shape[1] != block.in_channels:
x = th.cat([x, skips.pop()], dim=1)
x = block(x, emb, emb_s=None if noise_labels_s == None else emb_s,
emb_t=emb_t if self.linear_probing else None)
return aux
@persistent_class
class EDMPrecond_CTM(th.nn.Module):
def __init__(self,
img_resolution, # Image resolution.
img_channels, # Number of color channels.
label_dim = 0, # Number of class labels, 0 = unconditional.
use_fp16 = False, # Execute the underlying model at FP16 precision?
sigma_min = 0, # Minimum supported noise level.
sigma_max = float('inf'), # Maximum supported noise level.
sigma_data = 0.5, # Expected standard deviation of the training data.
model_type = 'SongUNet', # Class name of the underlying model.
teacher = False,
teacher_model_path = '',
training_mode = '',
arch='ncsn',
linear_probing=False,
**model_kwargs, # Keyword arguments for the underlying model.
):
super().__init__()
self.teacher = teacher
self.img_resolution = img_resolution
self.img_channels = img_channels
self.label_dim = label_dim
self.use_fp16 = use_fp16
self.sigma_min = sigma_min
self.sigma_max = sigma_max
self.sigma_data = sigma_data
self.eye = th.eye(self.label_dim, device=dev)
if teacher:
# print("I am the teacher model -----------------")
import pickle
# print(f'Loading network from "{teacher_model_path}"...')
with open(teacher_model_path, 'rb') as f:
self.model = pickle.load(f)['ema']
else:
# print("I am not teacher *********************")
if arch in ['ddpmpp', 'ncsnpp']:
resample_filter = [1,1] if arch == 'ddpmpp' else [1,3,3,1]
channel_mult_noise = 1 if arch == 'ddpmpp' else 2
encoder_type = 'standard' if arch == 'ddpmpp' else 'residual'
embedding_type = 'positional' if arch == 'ddpmpp' else 'fourier'
self.model = globals()[model_type](img_resolution=img_resolution, in_channels=img_channels,
out_channels=img_channels, label_dim=label_dim,
training_mode=training_mode, resample_filter=resample_filter,
channel_mult_noise=channel_mult_noise, encoder_type=encoder_type,
embedding_type=embedding_type, linear_probing=linear_probing,
**model_kwargs)
else:
self.model = globals()[model_type](img_resolution=img_resolution, in_channels=img_channels,
out_channels=img_channels, label_dim=label_dim,
training_mode=training_mode, linear_probing=linear_probing,
**model_kwargs)
def get_c_in(self, sigma):
return 1 / (sigma**2 + self.sigma_data**2) ** 0.5
def unrescaling_t(self, rescaled_t):
return th.exp(rescaled_t / 250.) - 1e-44
def forward(self, rescaled_x, rescaled_t, s=None, teacher=False, **model_kwargs):
class_labels = None if self.label_dim == 0 else th.zeros([1, self.label_dim], device=rescaled_x.device) \
if model_kwargs == {} else self.eye[model_kwargs['y']].reshape(-1, self.label_dim)
dtype = th.float16 if self.use_fp16 and rescaled_x.device.type == 'cuda' else th.float32
if self.teacher:
#with torch.no_grad():
sigma = self.unrescaling_t(rescaled_t)
c_in = append_dims(self.get_c_in(sigma), rescaled_x.ndim)
x = rescaled_x / c_in
D_x = self.model(x.to(dtype), sigma.flatten(), class_labels=class_labels)
c_skip = append_dims(self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2), rescaled_x.ndim)
c_out = append_dims(sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt(), rescaled_x.ndim)
F_x = (D_x - c_skip * x) / c_out
else:
t = self.unrescaling_t(rescaled_t)
t = t.log() / 4
if s != None:
s = self.unrescaling_t(s)
s = s.log() / 4
F_x = self.model(rescaled_x.to(dtype), t.flatten(), None if s == None else s.flatten(), class_labels=class_labels)
#assert F_x.dtype == dtype
return F_x
def round_sigma(self, sigma):
return th.as_tensor(sigma)
def convert_to_fp16(self):
pass
def convert_to_fp32(self):
pass
class FeatureFusionBlock(nn.Module):
def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True, lowest=False):
super().__init__()
self.deconv = deconv
self.align_corners = align_corners
self.expand = expand
out_features = features
if self.expand==True:
out_features = features//2
self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
self.skip_add = nn.quantized.FloatFunctional()
def forward(self, *xs):
output = xs[0]
if len(xs) == 2:
output = self.skip_add.add(output, xs[1])
output = nn.functional.interpolate(
output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
)
output = self.out_conv(output)
return output
def ctm_data_defaults(data_name):
return dict(
train_classes=-1,
type='png',
sigma_data=0.5,
deterministic=False,
num_classes=10,
)
def ctm_loss_defaults(data_name):
return dict(
# CTM hyperparams
ctm_training=True,
consistency_weight=1.0,
ctm_estimate_outer_type='target_model_sg',
ctm_estimate_inner_type='model',
ctm_target_inner_type='model_sg',
ctm_target_matching=False,
sample_s_strategy='uniform',
heun_step_strategy='weighted',
heun_step_multiplier=1.0,
outer_parametrization='euler',
inner_parametrization='edm',
time_continuous=False,
self_learn=False,
self_learn_iterative=False,
target_matching=False,
# DSM hyperparams
diffusion_training=True,
apply_adaptive_weight=True,
denoising_weight=1.,
diffusion_mult = 0.7,
diffusion_schedule_sampler='halflognormal',
diffusion_training_frequency=1.,
# GAN hyperparams
d_lr=0.002,
gan_training=False,
gan_specific_batch=False,
gan_micro_batch=32,
gan_real_free=True,
discriminator_weight=1.0,
discriminator_start_itr=0,
use_d_fp16=False,
d_architecture='StyleGAN-XL',
g_learning_period=1,
gan_fake_outer_type='no',
gan_fake_inner_type='',
gan_real_inner_type='',
gan_target_matching=False,
data_augment=True,
d_backbone=['deit_base_distilled_patch16_224', 'tf_efficientnet_lite0'],
d_apply_adaptive_weight=True,
shift_ratio=0.125,
cutout_ratio=0.2,
gan_training_frequency=1.,
gaussian_filter=False,
blur_fade_itr=1000,
blur_init_sigma=2,
prob_aug=1.0,
gan_different_augment=False,
gan_num_heun_step=17 if data_name == 'cifar10' else 39,
gan_heun_step_strategy='uniform',
gan_specific_time=False,
gan_low_res_train=False,
d_opt_load=True,
)
def ctm_train_defaults(data_name):
return dict(
beta_min=0.1,
beta_max=20.,
multiplier=1.,
num_heun_step=17 if data_name == 'cifar10' else 39,
num_heun_step_random=True,
# Network architecture
edm_nn_ncsn=False,
edm_nn_ddpm=True if data_name == 'cifar10' else False,
in_channels=3,
linear_probing=False,
target_subtract=False,
)
def ctm_eval_defaults(data_name):
return dict(
intermediate_samples=False,
sampling_batch=64,
sample_interval=1000 if data_name == 'cifar10' else 10,
sampling_steps=18 if data_name == 'cifar10' else 40,
eval_interval=1000,
eval_num_samples=5000,
eval_batch=500,
#ref_path='/home/dongjun/EighthArticleExperimentalResults/CIFAR10/author_ckpt/cifar10-32x32.npz' if data_name == 'cifar10' else "",
ref_path='/home/acf15618av/EighthArticleExperimentalResults/CIFAR10/author_ckpt/cifar10-32x32.npz' if data_name == 'cifar10' \
else "/home/fp084243/EighthArticleExperimentalResults/ImageNet64/author_ckpt/VIRTUAL_imagenet64_labeled.npz",
ref_feat_path='',
large_log=False,
compute_ema_fids=False,
#dm_sample_path_seed_42='/data2/dongjun/EighthArticleExperimentalResults/CIFAR10/DM/EDM-VP/fp16-seed-42/edm_heun_sampler_18_steps_ond-vp_itrs_model_ema' if data_name == 'cifar10' else "",
dm_sample_path_seed_42='/home/acf15618av/EighthArticleExperimentalResults/CIFAR10/DM/heun_18_seed_42_ver2' if data_name == 'cifar10' else "",
ae_image_path_seed_42='',
eval_seed=42,
eval_fid=False,
eval_similarity=True,
save_png=False,
check_ctm_denoising_ability=False,
check_dm_performance=True,
sanity_check=False,
save_period=1000 if data_name == 'cifar10' else 10,
clip_denoised=False,
clip_output=True,
gpu_usage=False,
eval_large_nfe=True,
)
def cm_train_defaults(data_name):
return dict(
#teacher_model_path="/home/dongjun/EighthArticleExperimentalResults/CIFAR10/author_ckpt/edm-cifar10-32x32-uncond-vp.pkl" if data_name == 'cifar10' else "",
teacher_model_path="/home/acf15618av/EighthArticleExperimentalResults/CIFAR10/author_ckpt/edm-cifar10-32x32-uncond-vp.pkl" if data_name == 'cifar10' else "",
teacher_dropout=0.0 if data_name == 'cifar10' else 0.1,
training_mode="ctm",
target_ema_mode="fixed",
scale_mode="fixed",
total_training_steps=600000,
start_ema=0.999,
start_scales=18 if data_name == 'cifar10' else 40,
end_scales=18 if data_name == 'cifar10' else 40,
distill_steps_per_iter=50000,
loss_norm="lpips",
port=6,
)
def model_and_diffusion_defaults(data_name):
"""
Defaults for image training.
"""
res = dict(
sigma_min=0.002,
sigma_max=80.0,
rho=7,
image_size=32 if data_name == 'cifar10' else 64,
num_channels=192,
num_res_blocks=3,
num_heads=4,
num_heads_upsample=-1,
num_head_channels=64,
attention_resolutions="32,16,8",
channel_mult="",
dropout=0.0,
class_cond=False if data_name == 'cifar10' else True,
use_checkpoint=False,
use_scale_shift_norm=True,
resblock_updown=True,
use_new_attention_order=False,
learn_sigma=False,
weight_schedule="uniform",
weight_schedule_multiplier=1.,
diffusion_weight_schedule="karras_weight",
rescaling=False,
)
return res
def train_defaults(data_name):
"""
Defaults for model training.
"""
res = dict(
out_dir="",
#data_dir="/home/dongjun/EighthArticleExperimentalResults/CIFAR10/train" if data_name == 'cifar10' else "",
data_dir="./" if data_name == 'cifar10' else "",
schedule_sampler="uniform",
lr=0.0004 if data_name == 'cifar10' else 0.000008,
weight_decay=0.0,
lr_anneal_steps=0,
global_batch_size=128 if data_name == 'cifar10' else 2048,
batch_size=-1,
microbatch=64 if data_name.lower() == 'cifar10' else -1, # -1 disables microbatches
ema_rate="0.999,0.9999" if data_name == 'cifar10' else "0.999,0.9999,0.9999432189950708",
# comma-separated list of EMA values
log_interval=1000,
save_interval=1000000,
save_check_period=1000000,
resume_checkpoint="",
use_fp16=True,
fp16_scale_growth=1e-3,
device_id=0,
num_workers=4,
use_MPI=False,
map_location='cuda',
)
return res
def load_state_dict(path, **kwargs):
"""
Load a PyTorch file for single GPU.
"""
# Directly load the state dictionary from the file
state_dict = th.load(path, **kwargs)
return state_dict
def create_ema_and_scales_fn(
target_ema_mode,
start_ema,
scale_mode,
start_scales,
end_scales,
total_steps,
distill_steps_per_iter,
):
def ema_and_scales_fn(step):
if target_ema_mode == "fixed" and scale_mode == "fixed":
target_ema = start_ema
scales = start_scales
else:
print("check if other options are commented out")
print("printing the target_ema_mode:: ", target_ema_mode)
print("scale_mode:: ", scale_mode)
raise NotImplementedError
return float(target_ema), int(scales)
return ema_and_scales_fn
def load_feature_extractor(args, eval=True):
feature_extractor = None
if args.loss_norm == 'lpips':
from piq import LPIPS
feature_extractor = LPIPS(replace_pooling=True, reduction="none")
return feature_extractor
# i am cutting the GAN projector related function from here
# for GAN projector
class Slice(nn.Module):
def __init__(self, start_index=1):
super(Slice, self).__init__()
self.start_index = start_index
def forward(self, x):
return x[:, self.start_index :]
class AddReadout(nn.Module):
def __init__(self, start_index=1):
super(AddReadout, self).__init__()
self.start_index = start_index
def forward(self, x):
if self.start_index == 2:
readout = (x[:, 0] + x[:, 1]) / 2
else:
readout = x[:, 0]
return x[:, self.start_index :] + readout.unsqueeze(1)
class ProjectReadout(nn.Module):
def __init__(self, in_features, start_index=1):
super(ProjectReadout, self).__init__()
self.start_index = start_index
self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU())
def forward(self, x):
readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :])
features = th.cat((x[:, self.start_index :], readout), -1)
return self.project(features)
class Transpose(nn.Module):
def __init__(self, dim0, dim1):
super(Transpose, self).__init__()
self.dim0 = dim0
self.dim1 = dim1
def forward(self, x):
x = x.transpose(self.dim0, self.dim1)
return x.contiguous()
def _resize_pos_embed(self, posemb, gs_h, gs_w):
posemb_tok, posemb_grid = (
posemb[:, : self.start_index],
posemb[0, self.start_index :],
)
gs_old = int(math.sqrt(len(posemb_grid)))
posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear", align_corners=False)
posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1)
posemb = th.cat([posemb_tok, posemb_grid], dim=1)
return posemb
def forward_flex(self, x):
b, c, h, w = x.shape
pos_embed = self._resize_pos_embed(
self.pos_embed, h // self.patch_size[1], w // self.patch_size[0]
)
B = x.shape[0]
if hasattr(self.patch_embed, "backbone"):
x = self.patch_embed.backbone(x)
if isinstance(x, (list, tuple)):
x = x[-1] # last feature if backbone outputs list/tuple of features
x = self.patch_embed.proj(x).flatten(2).transpose(1, 2)
if hasattr(self, "dist_token") and self.dist_token is not None:
cls_tokens = self.cls_token.expand(
B, -1, -1
) # stole cls_tokens impl from Phil Wang, thanks
dist_token = self.dist_token.expand(B, -1, -1)
x = th.cat((cls_tokens, dist_token, x), dim=1)
else:
cls_tokens = self.cls_token.expand(
B, -1, -1
) # stole cls_tokens impl from Phil Wang, thanks
x = th.cat((cls_tokens, x), dim=1)
x = x + pos_embed
x = self.pos_drop(x)
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x
def get_readout_oper(vit_features, features, use_readout, start_index=1):
if use_readout == "ignore":
readout_oper = [Slice(start_index)] * len(features)
elif use_readout == "add":
readout_oper = [AddReadout(start_index)] * len(features)
elif use_readout == "project":
readout_oper = [
ProjectReadout(vit_features, start_index) for out_feat in features
]
else:
assert (
False
), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'"
return readout_oper
activations = {}
def get_activation(name):
def hook(model, input, output):
activations[name] = output
return hook
def _make_vit_b16_backbone(
model,
features=[96, 192, 384, 768],
size=[384, 384],
hooks=[2, 5, 8, 11],
vit_features=768,
use_readout="ignore",
start_index=1,
):
pretrained = nn.Module()
pretrained.model = model
pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
pretrained.activations = activations
readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
# 32, 48, 136, 384
pretrained.layer1 = nn.Sequential(
readout_oper[0],
Transpose(1, 2),
nn.Unflatten(2, th.Size([size[0] // 16, size[1] // 16])),
nn.Conv2d(
in_channels=vit_features,
out_channels=features[0],
kernel_size=1,
stride=1,
padding=0,
),
nn.ConvTranspose2d(
in_channels=features[0],
out_channels=features[0],
kernel_size=4,
stride=4,
padding=0,
bias=True,
dilation=1,
groups=1,
),
)
pretrained.layer2 = nn.Sequential(
readout_oper[1],
Transpose(1, 2),
nn.Unflatten(2, th.Size([size[0] // 16, size[1] // 16])),
nn.Conv2d(
in_channels=vit_features,
out_channels=features[1],
kernel_size=1,
stride=1,
padding=0,
),
nn.ConvTranspose2d(
in_channels=features[1],
out_channels=features[1],
kernel_size=2,
stride=2,
padding=0,
bias=True,
dilation=1,
groups=1,
),
)
pretrained.layer3 = nn.Sequential(
readout_oper[2],
Transpose(1, 2),
nn.Unflatten(2, th.Size([size[0] // 16, size[1] // 16])),
nn.Conv2d(
in_channels=vit_features,
out_channels=features[2],
kernel_size=1,
stride=1,
padding=0,
),
)
pretrained.layer4 = nn.Sequential(
readout_oper[3],
Transpose(1, 2),
nn.Unflatten(2, th.Size([size[0] // 16, size[1] // 16])),
nn.Conv2d(
in_channels=vit_features,
out_channels=features[3],
kernel_size=1,
stride=1,
padding=0,
),
nn.Conv2d(
in_channels=features[3],
out_channels=features[3],
kernel_size=3,
stride=2,
padding=1,
),
)
pretrained.model.start_index = start_index
pretrained.model.patch_size = [16, 16]
# We inject this function into the VisionTransformer instances so that
# we can use it with interpolated position embeddings without modifying the library source.
pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
pretrained.model._resize_pos_embed = types.MethodType(
_resize_pos_embed, pretrained.model
)
return pretrained
def calc_dims(pretrained, is_vit=False):
dims = []
inp_res = 256
tmp = th.zeros(1, 3, inp_res, inp_res)
if not is_vit:
tmp = pretrained.layer0(tmp)
dims.append(tmp.shape[1:3])
tmp = pretrained.layer1(tmp)
dims.append(tmp.shape[1:3])
tmp = pretrained.layer2(tmp)
dims.append(tmp.shape[1:3])
tmp = pretrained.layer3(tmp)
dims.append(tmp.shape[1:3])
else:
tmp = forward_vit(pretrained, tmp)
dims = [out.shape[1:3] for out in tmp]
# split to channels and resolution multiplier
dims = np.array(dims)
channels = dims[:, 0]
res_mult = dims[:, 1] / inp_res
return channels, res_mult
def _make_vit(model, name):
# print(name)
if 'base' in name:
features = [96, 192, 384, 768]
hooks = [2, 5, 8, 11]
vit_features = 768
else:
print("NOPR")
return _make_vit_b16_backbone(
model,
features=features,
size=[224, 224],
hooks=hooks,
vit_features=vit_features,
start_index=2 if 'deit' in name else 1,
)
def _make_efficientnet(model):
pretrained = nn.Module()
pretrained.layer0 = nn.Sequential(
model.conv_stem, model.bn1, model.act1, *model.blocks[0:2]
)
pretrained.layer1 = nn.Sequential(*model.blocks[2:3])
pretrained.layer2 = nn.Sequential(*model.blocks[3:5])
pretrained.layer3 = nn.Sequential(*model.blocks[5:9])
return pretrained
def _make_pretrained(backbone, verbose=False):
print("printing the name of backbone::::: ", backbone)
if backbone == "deit_base_distilled_patch16_224":
model = timm.create_model(backbone, pretrained=True)
pretrained = _make_vit(model, backbone)
is_vit = True
elif backbone == "tf_efficientnet_lite0":
model = timm.create_model(backbone, pretrained=True)
pretrained = _make_efficientnet(model)
is_vit = False
pretrained.CHANNELS, pretrained.RES_MULT = calc_dims(pretrained, is_vit)
return pretrained
def get_backbone_normstats(backbone):
if backbone :
return {
'mean': [0.5, 0.5, 0.5],
'std': [0.5, 0.5, 0.5],
}
else:
print("check if the necesarry blocks are delted. it is in pg_modules.projector.py")
raise NotImplementedError
def _make_scratch_ccm(scratch, in_channels, cout, expand=False):
# shapes
out_channels = [cout, cout*2, cout*4, cout*8] if expand else [cout]*4
scratch.layer0_ccm = nn.Conv2d(in_channels[0], out_channels[0], kernel_size=1, stride=1, padding=0, bias=True)
scratch.layer1_ccm = nn.Conv2d(in_channels[1], out_channels[1], kernel_size=1, stride=1, padding=0, bias=True)
scratch.layer2_ccm = nn.Conv2d(in_channels[2], out_channels[2], kernel_size=1, stride=1, padding=0, bias=True)
scratch.layer3_ccm = nn.Conv2d(in_channels[3], out_channels[3], kernel_size=1, stride=1, padding=0, bias=True)
scratch.CHANNELS = out_channels
return scratch
def _make_scratch_csm(scratch, in_channels, cout, expand):
scratch.layer3_csm = FeatureFusionBlock(in_channels[3], nn.ReLU(False), expand=expand, lowest=True)
scratch.layer2_csm = FeatureFusionBlock(in_channels[2], nn.ReLU(False), expand=expand)
scratch.layer1_csm = FeatureFusionBlock(in_channels[1], nn.ReLU(False), expand=expand)
scratch.layer0_csm = FeatureFusionBlock(in_channels[0], nn.ReLU(False))
# last refinenet does not expand to save channels in higher dimensions
scratch.CHANNELS = [cout, cout, cout*2, cout*4] if expand else [cout]*4
return scratch
def _make_projector(im_res, backbone, cout, proj_type, expand=False):
assert proj_type in [0, 1, 2], "Invalid projection type"
### Build pretrained feature network
pretrained = _make_pretrained(backbone)
# Following Projected GAN
im_res = 256
pretrained.RESOLUTIONS = [im_res//4, im_res//8, im_res//16, im_res//32]
if proj_type == 0: return pretrained, None
# print(pretrained.CHANNELS)
# print(pretrained)
# exit()
### Build CCM
scratch = nn.Module()
scratch = _make_scratch_ccm(scratch, in_channels=pretrained.CHANNELS, cout=cout, expand=expand)
pretrained.CHANNELS = scratch.CHANNELS
if proj_type == 1: return pretrained, scratch
### build CSM
scratch = _make_scratch_csm(scratch, in_channels=scratch.CHANNELS, cout=cout, expand=expand)
# CSM upsamples x2 so the feature map resolution doubles
pretrained.RESOLUTIONS = [res*2 for res in pretrained.RESOLUTIONS]
pretrained.CHANNELS = scratch.CHANNELS
return pretrained, scratch
def forward_vit(pretrained, x):
b, c, h, w = x.shape
_ = pretrained.model.forward_flex(x)
layer_1 = pretrained.activations["1"]
layer_2 = pretrained.activations["2"]
layer_3 = pretrained.activations["3"]
layer_4 = pretrained.activations["4"]
layer_1 = pretrained.layer1[0:2](layer_1)
layer_2 = pretrained.layer2[0:2](layer_2)
layer_3 = pretrained.layer3[0:2](layer_3)
layer_4 = pretrained.layer4[0:2](layer_4)
unflatten = nn.Sequential(
nn.Unflatten(
2,
th.Size(
[
h // pretrained.model.patch_size[1],
w // pretrained.model.patch_size[0],
]
),
)
)
if layer_1.ndim == 3:
layer_1 = unflatten(layer_1)
if layer_2.ndim == 3:
layer_2 = unflatten(layer_2)
if layer_3.ndim == 3:
layer_3 = unflatten(layer_3)
if layer_4.ndim == 3:
layer_4 = unflatten(layer_4)
layer_1 = pretrained.layer1[3 : len(pretrained.layer1)](layer_1)
layer_2 = pretrained.layer2[3 : len(pretrained.layer2)](layer_2)
layer_3 = pretrained.layer3[3 : len(pretrained.layer3)](layer_3)
layer_4 = pretrained.layer4[3 : len(pretrained.layer4)](layer_4)
return layer_1, layer_2, layer_3, layer_4
class F_Identity(nn.Module):
def forward(self, x):
return x
class F_RandomProj(nn.Module):
def __init__(
self,
backbone="tf_efficientnet_lite3",
im_res=256,
cout=64,
expand=True,
proj_type=2, # 0 = no projection, 1 = cross channel mixing, 2 = cross scale mixing
**kwargs,
):
super().__init__()
self.proj_type = proj_type
self.backbone = backbone
self.cout = cout
self.expand = expand
self.normstats = get_backbone_normstats(backbone)
# build pretrained feature network and random decoder (scratch)
self.pretrained, self.scratch = _make_projector(im_res=im_res, backbone=self.backbone, cout=self.cout,
proj_type=self.proj_type, expand=self.expand)
self.CHANNELS = self.pretrained.CHANNELS
self.RESOLUTIONS = self.pretrained.RESOLUTIONS
def forward(self, x):
# predict feature maps
if self.backbone == "deit_base_distilled_patch16_224":
# if self.backbone:
out0, out1, out2, out3 = forward_vit(self.pretrained, x)
# else:
elif self.backbone == "tf_efficientnet_lite0":
out0 = self.pretrained.layer0(x)
out1 = self.pretrained.layer1(out0)
out2 = self.pretrained.layer2(out1)
out3 = self.pretrained.layer3(out2)
# start enumerating at the lowest layer (this is where we put the first discriminator)
out = {
'0': out0,
'1': out1,
'2': out2,
'3': out3,
}
if self.proj_type == 0: return out
out0_channel_mixed = self.scratch.layer0_ccm(out['0'])
out1_channel_mixed = self.scratch.layer1_ccm(out['1'])
out2_channel_mixed = self.scratch.layer2_ccm(out['2'])
out3_channel_mixed = self.scratch.layer3_ccm(out['3'])
out = {
'0': out0_channel_mixed,
'1': out1_channel_mixed,
'2': out2_channel_mixed,
'3': out3_channel_mixed,
}
if self.proj_type == 1: return out
# from bottom to top
out3_scale_mixed = self.scratch.layer3_csm(out3_channel_mixed)
out2_scale_mixed = self.scratch.layer2_csm(out3_scale_mixed, out2_channel_mixed)
out1_scale_mixed = self.scratch.layer1_csm(out2_scale_mixed, out1_channel_mixed)
out0_scale_mixed = self.scratch.layer0_csm(out1_scale_mixed, out0_channel_mixed)
out = {
'0': out0_scale_mixed,
'1': out1_scale_mixed,
'2': out2_scale_mixed,
'3': out3_scale_mixed,
}
return out
def conv2d(*args, **kwargs):
return spectral_norm(nn.Conv2d(*args, **kwargs))
def NormLayer(c, mode='batch'):
if mode == 'group':
return nn.GroupNorm(c//2, c)
elif mode == 'batch':
return nn.BatchNorm2d(c)
class DownBlock(nn.Module):
def __init__(self, in_planes, out_planes, width=1):
super().__init__()
self.main = nn.Sequential(
conv2d(in_planes, out_planes*width, 4, 2, 1, bias=True),
NormLayer(out_planes*width),
nn.LeakyReLU(0.2, inplace=True),
)
def forward(self, feat):
return self.main(feat)
class DownBlockPatch(nn.Module):
def __init__(self, in_planes, out_planes):
super().__init__()
self.main = nn.Sequential(
DownBlock(in_planes, out_planes),
conv2d(out_planes, out_planes, 1, 1, 0, bias=False),
NormLayer(out_planes),
nn.LeakyReLU(0.2, inplace=True),
)
def forward(self, feat):
return self.main(feat)
class SingleDisc(nn.Module):
def __init__(self, nc=None, ndf=None, start_sz=256, end_sz=8, head=None, patch=False):
super().__init__()
# midas channels
nfc_midas = {4: 512, 8: 512, 16: 256, 32: 128, 64: 64, 128: 64,
256: 32, 512: 16, 1024: 8}
# interpolate for start sz that are not powers of two
if start_sz not in nfc_midas.keys():
sizes = np.array(list(nfc_midas.keys()))
start_sz = sizes[np.argmin(abs(sizes - start_sz))]
self.start_sz = start_sz
# if given ndf, allocate all layers with the same ndf
if ndf is None:
nfc = nfc_midas
else:
nfc = {k: ndf for k, v in nfc_midas.items()}
# for feature map discriminators with nfc not in nfc_midas
# this is the case for the pretrained backbone (midas.pretrained)
if nc is not None and head is None:
nfc[start_sz] = nc
layers = []
# Head if the initial input is the full modality
if head:
layers += [conv2d(nc, nfc[256], 3, 1, 1, bias=False),
nn.LeakyReLU(0.2, inplace=True)]
# Down Blocks
DB = DownBlockPatch if patch else DownBlock
while start_sz > end_sz:
layers.append(DB(nfc[start_sz], nfc[start_sz//2]))
start_sz = start_sz // 2
layers.append(conv2d(nfc[end_sz], 1, 4, 1, 0, bias=False))
self.main = nn.Sequential(*layers)
def forward(self, x, c):
return self.main(x)
class MultiScaleD(nn.Module):
def __init__(
self,
channels,
resolutions,
num_discs=4,
proj_type=2, # 0 = no projection, 1 = cross channel mixing, 2 = cross scale mixing
cond=0,
patch=False,
**kwargs,
):
super().__init__()
assert num_discs in [1, 2, 3, 4, 5]
# the first disc is on the lowest level of the backbone
self.disc_in_channels = channels[:num_discs]
self.disc_in_res = resolutions[:num_discs]
Disc = SingleDisc
mini_discs = []
for i, (cin, res) in enumerate(zip(self.disc_in_channels, self.disc_in_res)):
start_sz = res if not patch else 16
mini_discs += [str(i), Disc(nc=cin, start_sz=start_sz, end_sz=8, patch=patch)],
self.mini_discs = nn.ModuleDict(mini_discs)
def forward(self, features, c, rec=False):
all_logits = []
for k, disc in self.mini_discs.items():
all_logits.append(disc(features[k], c).view(features[k].size(0), -1))
all_logits = th.cat(all_logits, dim=1)
return all_logits
def load_discriminator_and_d_feature_extractor(args):
#assert (args.gan_training == True) == (args.d_architecture == 'StyleGAN-XL')
if args.gan_training:
# from pg_modules.projector import F_RandomProj
# from pg_modules.discriminator import MultiScaleD
backbones = ['deit_base_distilled_patch16_224', 'tf_efficientnet_lite0']
discriminator, discriminator_feature_extractor = [], []
if args.gan_low_res_train:
discriminator2 = []
backbone_kwargs = {'im_res': args.image_size}
# print("printing the backbones: ", backbones)
# exit()
for i, bb_name in enumerate(backbones):
feat = F_RandomProj(bb_name, **backbone_kwargs)
disc = MultiScaleD(
channels=feat.CHANNELS,
resolutions=feat.RESOLUTIONS,
**backbone_kwargs,
)
discriminator_feature_extractor.append([bb_name, feat])
discriminator.append([bb_name, disc])
if args.gan_low_res_train:
discriminator2.append([bb_name + '_low', disc])
discriminator_feature_extractor = nn.ModuleDict(discriminator_feature_extractor)
discriminator_feature_extractor = discriminator_feature_extractor.train(False).to(dev)
discriminator_feature_extractor.requires_grad_(False)
if args.gan_low_res_train:
discriminator = discriminator + discriminator2
discriminator = nn.ModuleDict(discriminator)
discriminator.to(dev)
discriminator.train()
if args.use_d_fp16:
discriminator.convert_to_fp16()
else:
discriminator, discriminator_feature_extractor = None, None
return discriminator, discriminator_feature_extractor
### GAN part ends here
INITIAL_LOG_LOSS_SCALE = 20.0
def find_resume_checkpoint():
# On your infrastructure, you may want to override this to automatically
# discover the latest checkpoint on your blob storage, etc.
return None
def get_param_groups_and_shapes(named_model_params):
named_model_params = list(named_model_params)
scalar = []
matrix = []
for n, p in named_model_params:
if p.ndim <= 1 and p.requires_grad:
scalar.append((n, p))
if p.ndim > 1 and p.requires_grad:
matrix.append((n, p))
scalar_vector_named_params = (
scalar,
(-1),
)
matrix_named_params = (
matrix,
(1, -1),
)
return [scalar_vector_named_params, matrix_named_params]
def make_master_params(param_groups_and_shapes):
"""
Copy model parameters into a (differently-shaped) list of full-precision
parameters.
"""
master_params = []
for param_group, shape in param_groups_and_shapes:
#a = []
#for _, param in param_group:
# if param.requires_grad == True:
# a.append([param.detach().float()])
#master_param = nn.Parameter(
# _flatten_dense_tensors(a).view(shape)
#)
master_param = nn.Parameter(
_flatten_dense_tensors(
[param.detach().float() for (_, param) in param_group]
).view(shape)
)
master_param.requires_grad = True
master_params.append(master_param)
return master_params
def zero_master_grads(master_params):
for param in master_params:
param.grad = None
def zero_grad(model_params):
for param in model_params:
# Taken from https://pytorch.org/docs/stable/_modules/torch/optim/optimizer.html#Optimizer.add_param_group
if param.grad is not None:
param.grad.detach_()
param.grad.zero_()
def model_grads_to_master_grads(param_groups_and_shapes, master_params):
"""
Copy the gradients from the model parameters into the master parameters
from make_master_params().
"""
for master_param, (param_group, shape) in zip(
master_params, param_groups_and_shapes
):
master_param.grad = _flatten_dense_tensors(
[param_grad_or_zeros(param) for (_, param) in param_group]
).view(shape)
def param_grad_or_zeros(param):
if param.grad is not None:
return param.grad.data.detach()
else:
return th.zeros_like(param)
def master_params_to_model_params(param_groups_and_shapes, master_params):
"""
Copy the master parameter data back into the model parameters.
"""
# Without copying to a list, if a generator is passed, this will
# silently not copy any parameters.
for master_param, (param_group, _) in zip(master_params, param_groups_and_shapes):
for (_, param), unflat_master_param in zip(
param_group, unflatten_master_params(param_group, master_param.view(-1))
):
param.detach().copy_(unflat_master_param)
def unflatten_master_params(param_group, master_param):
return _unflatten_dense_tensors(master_param, [param for (_, param) in param_group])
def master_params_to_state_dict(
model, param_groups_and_shapes, master_params, use_fp16
):
if use_fp16:
state_dict = model.state_dict()
for master_param, (param_group, _) in zip(
master_params, param_groups_and_shapes
):
for (name, _), unflat_master_param in zip(
param_group, unflatten_master_params(param_group, master_param.view(-1))
):
assert name in state_dict
state_dict[name] = unflat_master_param
else:
state_dict = model.state_dict()
for i, (name, _value) in enumerate(model.named_parameters()):
assert name in state_dict
state_dict[name] = master_params[i]
return state_dict
def state_dict_to_master_params(model, state_dict, use_fp16):
if use_fp16:
named_model_params = [
(name, state_dict[name].to(dev)) for name, _ in model.named_parameters()
]
param_groups_and_shapes = get_param_groups_and_shapes(named_model_params)
master_params = make_master_params(param_groups_and_shapes)
else:
master_params = [state_dict[name] for name, _ in model.named_parameters()]
return master_params
def check_overflow(value):
return (value == float("inf")) or (value == -float("inf")) or (value != value)
class MixedPrecisionTrainer:
def __init__(
self,
*,
model,
use_fp16=False,
fp16_scale_growth=1e-3,
initial_lg_loss_scale=INITIAL_LOG_LOSS_SCALE,
):
self.model = model
self.use_fp16 = use_fp16
self.fp16_scale_growth = fp16_scale_growth
self.model_params = list(self.model.parameters())
self.master_params = self.model_params
self.param_groups_and_shapes = None
self.lg_loss_scale = initial_lg_loss_scale
#for name, param in self.model.named_parameters():
# print(name, param.requires_grad)
if self.use_fp16:
self.param_groups_and_shapes = get_param_groups_and_shapes(
self.model.named_parameters()
)
self.master_params = make_master_params(self.param_groups_and_shapes)
self.model.convert_to_fp16()
def zero_grad(self):
zero_grad(self.model_params)
def backward(self, loss: th.Tensor):
if self.use_fp16:
loss_scale = 2**self.lg_loss_scale
#print("loss value: ", (loss*loss_scale).item())
(loss * loss_scale).backward()
else:
loss.backward()
def optimize(self, opt: th.optim.Optimizer):
if self.use_fp16:
return self._optimize_fp16(opt)
else:
return self._optimize_normal(opt)
def _optimize_fp16(self, opt: th.optim.Optimizer):
logkv_mean("lg_loss_scale", self.lg_loss_scale)
model_grads_to_master_grads(self.param_groups_and_shapes, self.master_params)
grad_norm, param_norm = self._compute_norms(grad_scale=2**self.lg_loss_scale)
if check_overflow(grad_norm):
self.lg_loss_scale -= 1
log(f"Found NaN, decreased lg_loss_scale to {self.lg_loss_scale}")
zero_master_grads(self.master_params)
return False
logkv_mean("grad_norm", grad_norm)
logkv_mean("param_norm", param_norm)
for p in self.master_params:
p.grad.mul_(1.0 / (2**self.lg_loss_scale))
opt.step()
zero_master_grads(self.master_params)
master_params_to_model_params(self.param_groups_and_shapes, self.master_params)
self.lg_loss_scale += self.fp16_scale_growth
return True
def _optimize_normal(self, opt: th.optim.Optimizer):
# model_grads_to_master_grads(self.param_groups_and_shapes, self.master_params)
grad_norm, param_norm = self._compute_norms()
logkv_mean("grad_norm", grad_norm)
logkv_mean("param_norm", param_norm)
opt.step()
return True
def _compute_norms(self, grad_scale=1.0):
grad_norm = 0.0
param_norm = 0.0
for p in self.master_params:
with th.no_grad():
param_norm += th.norm(p, p=2, dtype=th.float32).item() ** 2
if p.grad is not None:
grad_norm += th.norm(p.grad, p=2, dtype=th.float32).item() ** 2
return np.sqrt(grad_norm) / grad_scale, np.sqrt(param_norm)
def master_params_to_state_dict(self, master_params):
return master_params_to_state_dict(
self.model, self.param_groups_and_shapes, master_params, self.use_fp16
)
def state_dict_to_master_params(self, state_dict):
return state_dict_to_master_params(self.model, state_dict, self.use_fp16)
class DummyGenerator:
def randn(self, *args, **kwargs):
return th.randn(*args, **kwargs)
def randint(self, *args, **kwargs):
return th.randint(*args, **kwargs)
def randn_like(self, *args, **kwargs):
return th.randn_like(*args, **kwargs)
class DeterministicGenerator:
"""
RNG to deterministically sample num_samples samples that does not depend on batch_size or mpi_machines
Uses a single rng and samples num_samples sized randomness and subsamples the current indices
"""
def __init__(self, num_samples, seed=0):
self.num_samples = num_samples
self.done_samples = 0
self.seed = seed
self.rng_cpu = th.Generator()
if th.cuda.is_available():
self.rng_cuda = th.Generator(dev)
self.set_seed(seed)
def get_global_size_and_indices(self, size):
global_size = (self.num_samples, *size[1:])
indices = th.arange(self.done_samples, self.done_samples + size[0])
indices = th.clamp(indices, 0, self.num_samples - 1)
return global_size, indices
def get_generator(self, device):
return self.rng_cpu if th.device(device).type == "cpu" else self.rng_cuda
def randn(self, *size, dtype=th.float, device="cpu"):
global_size, indices = self.get_global_size_and_indices(size)
generator = self.get_generator(device)
return th.randn(*global_size, generator=generator, dtype=dtype, device=device)[indices]
def randint(self, low, high, size, dtype=th.long, device="cpu"):
global_size, indices = self.get_global_size_and_indices(size)
generator = self.get_generator(device)
return th.randint(low, high, generator=generator, size=global_size, dtype=dtype, device=device)[indices]
def randn_like(self, tensor):
size, dtype, device = tensor.size(), tensor.dtype, tensor.device
return self.randn(*size, dtype=dtype, device=device)
def set_done_samples(self, done_samples):
self.done_samples = done_samples
self.set_seed(self.seed)
def get_seed(self):
return self.seed
def set_seed(self, seed):
self.rng_cpu.manual_seed(seed)
if th.cuda.is_available():
self.rng_cuda.manual_seed(seed)
def get_generator(generator, num_samples=0, seed=0):
if generator == "dummy":
return DummyGenerator()
elif generator == "determ":
return DeterministicGenerator(num_samples, seed)
else:
print("check if the necessary module has been removed. check in cm.random_util.py")
raise NotImplementedError
def parse_resume_step_from_filename(filename):
"""
Parse filenames of the form path/to/modelNNNNNN.pt, where NNNNNN is the
checkpoint's number of steps.
"""
split = filename.split("model")
if len(split) < 2:
return 0
split1 = split[-1].split(".")[0]
try:
return int(split1)
except ValueError:
return 0
def find_ema_checkpoint(main_checkpoint, step, rate):
if main_checkpoint is None:
return None
filename = f"ema_{rate}_{(step):06d}.pt"
path = bf.join(bf.dirname(main_checkpoint), filename)
if bf.exists(path):
return path
return None
def get_target_param_groups_and_shapes(named_model_params, source_named_model_params):
named_model_params = list(named_model_params)
source_named_model_params = {n: p for (n,p) in list(source_named_model_params)}
scalar = []
matrix = []
for n, p in named_model_params:
if p.ndim <= 1 and source_named_model_params[n].requires_grad:
scalar.append((n, p))
if p.ndim > 1 and source_named_model_params[n].requires_grad:
matrix.append((n, p))
scalar_vector_named_params = (
scalar,
(-1),
)
matrix_named_params = (
matrix,
(1, -1),
)
return [scalar_vector_named_params, matrix_named_params]
def karras_sample(
diffusion,
model,
shape,
steps,
clip_denoised=True,
progress=False,
callback=None,
model_kwargs=None,
device=None,
sigma_min=0.002,
sigma_max=80, # higher for highres?
rho=7.0,
sampler="heun",
s_churn=0.0,
s_tmin=0.0,
s_tmax=float("inf"),
s_noise=1.0,
generator=None,
ts=None,
x_T=None,
ctm=False,
teacher=False,
clip_output=True,
train=False,
ind_1=0,
ind_2=0,
gamma=0.5,
):
if generator is None:
generator = get_generator("dummy")
if sampler in ["progdist", 'euler', 'exact', 'cm_multistep', 'gamma_multistep']:
sigmas = get_sigmas_karras(steps + 1, sigma_min, sigma_max, rho, device=device)
else:
sigmas = get_sigmas_karras(steps, sigma_min, sigma_max, rho, device=device)
if x_T == None:
x_T = generator.randn(*shape, device=device) * sigma_max
sample_fn = {
"heun": sample_heun,
# "dpm": sample_dpm,
# "ancestral": sample_euler_ancestral,
# "onestep": sample_onestep,
"exact": sample_exact,
# "gamma": sample_gamma,
# "gamma_multistep": sample_gamma_multistep,
# "progdist": sample_progdist,
# "euler": sample_euler,
# "multistep": stochastic_iterative_sampler,
# "cm_multistep": sample_multistep,
}[sampler]
# print(sampler)
if sampler in ["heun", "dpm"]:
sampler_args = dict(
s_churn=s_churn, s_tmin=s_tmin, s_tmax=s_tmax, s_noise=s_noise
)
elif sampler in ["multistep", "exact", "cm_multistep"]:
sampler_args = dict(
ts=ts, t_min=sigma_min, t_max=sigma_max, rho=rho, steps=steps
)
elif sampler in ["gamma"]:
sampler_args = dict(ind_1=ind_1, ind_2=ind_2)
elif sampler in ["gamma_multistep"]:
sampler_args = dict(
ts=ts, t_min=sigma_min, t_max=sigma_max, rho=rho, steps=steps, gamma=gamma,
)
else:
sampler_args = {}
if sampler in ['heun']:
sampler_args['teacher'] = False if train else teacher
sampler_args['ctm'] = ctm
#print("clip_denoised, clip_output: ", clip_denoised, clip_output)
def denoiser(x_t, t, s=th.ones(x_T.shape[0], device=device)):
denoised, G_theta = diffusion.get_denoised_and_G(model, x_t, t, s, ctm, teacher, **model_kwargs)
if sampler in ['exact', 'cm_multistep', 'onestep', 'gamma', 'gamma_multistep']:
denoised = G_theta
if clip_denoised:
#print("clip denoised!!!")
denoised = denoised.clamp(-1, 1)
return denoised
x_0 = sample_fn(
denoiser,
x_T,
sigmas,
generator,
progress=progress,
callback=callback,
**sampler_args,
)
if clip_output:
#print("clip output")
return x_0.clamp(-1, 1)
return x_0
def to_d(x, sigma, denoised):
"""Converts a denoiser output to a Karras ODE derivative."""
return (x - denoised) / append_dims(sigma, x.ndim)
@th.no_grad()
def sample_heun(
denoiser,
x,
sigmas,
generator,
progress=False,
callback=None,
s_churn=0.0,
s_tmin=0.0,
s_tmax=float("inf"),
s_noise=1.0,
teacher=False,
ctm=False,
):
"""Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
s_in = x.new_ones([x.shape[0]])
indices = range(len(sigmas) - 1)
if progress:
from tqdm.auto import tqdm
indices = tqdm(indices)
for i in indices:
# print("sigmas: ", sigmas[i], ctm, teacher)
gamma = (
min(s_churn / (len(sigmas) - 1), 2**0.5 - 1)
if s_tmin <= sigmas[i] <= s_tmax
else 0.0
)
eps = generator.randn_like(x) * s_noise
sigma_hat = sigmas[i] * (gamma + 1)
if gamma > 0:
x = x + eps * (sigma_hat**2 - sigmas[i] ** 2) ** 0.5
if ctm:
denoised = denoiser(x, sigma_hat * s_in, s=sigma_hat * s_in)
else:
#if teacher:
denoised = denoiser(x, sigma_hat * s_in, s=None)
#else:
# denoised = denoiser(x, sigma_hat * s_in, s=sigma_hat * s_in)
#print("denoised: ", denoised[0][0][0][:3])
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback(
{
"x": x,
"i": i,
"sigma": sigmas[i],
"sigma_hat": sigma_hat,
"denoised": denoised,
}
)
dt = sigmas[i + 1] - sigma_hat
if sigmas[i + 1] == 0:
# Euler method
x = x + d * dt
#print("last")
else:
#print("no last")
# Heun's method
x_2 = x + d * dt
if ctm:
denoised_2 = denoiser(x_2, sigmas[i + 1] * s_in, s=sigmas[i + 1] * s_in)
else:
#if teacher:
denoised_2 = denoiser(x_2, sigmas[i + 1] * s_in, s=None)
#else:
# denoised_2 = denoiser(x_2, sigmas[i + 1] * s_in, s=sigmas[i + 1] * s_in)
d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
d_prime = (d + d_2) / 2
x = x + d_prime * dt
return x
@th.no_grad()
def sample_exact(
denoiser,
x,
sigmas,
generator,
progress=False,
callback=None,
ts=[],
t_min=0.002,
t_max=80.0,
rho=7.0,
steps=40,
):
"""Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
s_in = x.new_ones([x.shape[0]])
if ts != [] and ts != None:
sigmas = []
t_max_rho = t_max ** (1 / rho)
t_min_rho = t_min ** (1 / rho)
s_in = x.new_ones([x.shape[0]])
for i in range(len(ts)):
sigmas.append((t_max_rho + ts[i] / (steps - 1) * (t_min_rho - t_max_rho)) ** rho)
sigmas = th.tensor(sigmas)
sigmas = append_zero(sigmas).to(x.device)
indices = range(len(sigmas) - 1)
if progress:
from tqdm.auto import tqdm
indices = tqdm(indices)
for i in indices[:-1]:
sigma = sigmas[i]
# print(sigma, sigmas[i+1])
if sigmas[i+1] != 0:
denoised = denoiser(x, sigma * s_in, s=sigmas[i + 1] * s_in)
x = denoised
else:
denoised = denoiser(x, sigma * s_in, s=sigma * s_in)
d = to_d(x, sigma, denoised)
dt = sigmas[i + 1] - sigma
x = x + d * dt
#else:
# denoised = denoiser(x, sigma * s_in)
if callback is not None:
callback(
{
"x": x,
"i": i,
"sigma": sigmas[i],
"denoised": denoised,
}
)
#x = denoised
return x
def get_sigmas_karras(n, sigma_min, sigma_max, rho=7.0, device="cpu"):
"""Constructs the noise schedule of Karras et al. (2022)."""
ramp = th.linspace(0, 1, n)
min_inv_rho = sigma_min ** (1 / rho)
max_inv_rho = sigma_max ** (1 / rho)
sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
return append_zero(sigmas).to(device)
def append_zero(x):
return th.cat([x, x.new_zeros([1])])
class TrainLoop:
def __init__(
self,
#*,
model,
discriminator,
diffusion,
data,
batch_size,
args=None,
):
self.args = args
self.model = model
if self.args.sanity_check:
for name, param in self.model.named_parameters():
log("check and understand how consistency-type models override model parameters")
log("model parameter before overriding: ", param.data.cpu().detach().reshape(-1)[:3])
break
self.discriminator = discriminator
self.diffusion = diffusion
self.data = data
self.batch_size = batch_size
self.microbatch = args.microbatch if args.microbatch > 0 else batch_size
self.lr = args.lr
self.ema_rate = (
[args.ema_rate]
if isinstance(args.ema_rate, float)
else [float(x) for x in args.ema_rate.split(",")]
)
self.step = 0
self.resume_step = 0
self.global_batch = self.batch_size #* dist.get_world_size()
self.fids = []
self.generator = get_generator('determ', self.args.eval_num_samples, self.args.eval_seed)
self.x_T = self.generator.randn(*(self.args.sampling_batch, self.args.in_channels, self.args.image_size, self.args.image_size),
device='cpu') * self.args.sigma_max #.to(dist_util.dev())
if self.args.class_cond:
self.classes = self.generator.randint(0, self.args.num_classes, (self.args.sampling_batch,), device='cpu')
if self.args.data_name.lower() == 'cifar10':
self.classes.sort()
self.sync_cuda = th.cuda.is_available()
self._load_and_sync_parameters()
if self.args.sanity_check:
for name, param in self.model.named_parameters():
log("model parameter after overriding: ", param.data.cpu().detach().reshape(-1)[:3])
break
if self.discriminator != None:
if self.args.d_opt_load:
if self.args.sanity_check:
for name, param in self.discriminator.named_parameters():
log("discriminator parameter before overriding: ", param.data.cpu().detach().reshape(-1)[:3])
break
self._load_and_sync_discriminator_parameters()
if self.args.sanity_check:
for name, param in self.discriminator.named_parameters():
log("discriminator parameter after overriding: ", param.data.cpu().detach().reshape(-1)[:3])
break
print("loaded discriminator")
self.mp_trainer = MixedPrecisionTrainer(
model=self.model,
use_fp16=args.use_fp16,
fp16_scale_growth=args.fp16_scale_growth,
)
print("loaded mixed precision trainer")
if self.args.sanity_check:
log("mp trainer master parameter (should same to the model parameter if no linear_probing): ", self.mp_trainer.master_params[1].reshape(-1)[:3])
self.opt = RAdam(
self.mp_trainer.master_params, lr=self.lr, weight_decay=self.args.weight_decay
)
#if self.args.sanity_check:
# print("opt state dict before overriding: ", self.opt.state_dict())
if self.discriminator != None:
self.d_mp_trainer = MixedPrecisionTrainer(
model=self.discriminator,
use_fp16=args.use_d_fp16,
fp16_scale_growth=args.fp16_scale_growth,
)
self.d_opt = RAdam(
self.d_mp_trainer.master_params, lr=args.d_lr, weight_decay=self.args.weight_decay, betas=(0.5, 0.9)
)
print('going to resume step')
if self.resume_step:
self._load_optimizer_state()
if self.discriminator != None:
try:
#if self.args.sanity_check:
# print("discriminator opt state dict before overriding: ", self.d_opt.state_dict())
self._load_d_optimizer_state()
#if self.args.sanity_check:
# print("discriminator opt state dict after overriding: ", self.d_opt.state_dict())
except:
print("!!!!!!!!!!!!!!!!!!!!!!!!!!!! warning !!!!!!!!!!!!!!!!!!!!!!!!!!!! discriminator optimizer not loaded successfully")
# Model was resumed, either due to a restart or a checkpoint
# being specified at the command line.
self.ema_params = [
self._load_ema_parameters(rate) for rate in self.ema_rate
]
else:
self.ema_params = [
copy.deepcopy(self.mp_trainer.master_params)
for _ in range(len(self.ema_rate))
]
if th.cuda.is_available():
self.use_ddp = True
self.ddp_model = self.model #DDP(
# self.model,
# device_ids=[dist_util.dev()],
# output_device=dist_util.dev(),
# broadcast_buffers=False,
# bucket_cap_mb=128,
# find_unused_parameters=False,
# )
self.ddp_discriminator = None
if self.args.gan_training:
self.ddp_discriminator = self.discriminator #DDP(
# self.discriminator,
# device_ids=[dist_util.dev()],
# output_device=dist_util.dev(),
# broadcast_buffers=False,
# bucket_cap_mb=128,
# find_unused_parameters=False,
# )
else:
# if dist.get_world_size() > 1:
# logger.warn(
# "Distributed training requires CUDA. "
# "Gradients will not be synchronized properly!"
# )
self.use_ddp = False
self.ddp_model = self.model
self.step = self.resume_step
def _load_and_sync_parameters(self):
resume_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
if resume_checkpoint:
self.resume_step = parse_resume_step_from_filename(resume_checkpoint)
log(f"loading pretrained model from checkpoint: {resume_checkpoint}...")
state_dict = th.load(resume_checkpoint, map_location=dev)#"cpu")
self.model.load_state_dict(state_dict, strict=False)
log(f"end loading pretrained model from checkpoint: {resume_checkpoint}...")
# dist_util.sync_params(self.model.parameters())
# dist_util.sync_params(self.model.buffers())
log(f"end synchronizing pretrained model from GPU0 to all GPUs")
def _load_and_sync_discriminator_parameters(self):
resume_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
if resume_checkpoint:
self.resume_step = parse_resume_step_from_filename(resume_checkpoint)
resume_checkpoint = bf.join(bf.dirname(resume_checkpoint), f"d_model{self.resume_step:06}.pt")
# if dist.get_rank() == 0:
if os.path.exists(resume_checkpoint):
log(f"loading discriminator model from checkpoint: {resume_checkpoint}...")
#try:
#if self.args.map_location == 'cuda':
# state_dict = dist_util.load_state_dict(
# resume_checkpoint, map_location=dist_util.dev()
# )
#else:
state_dict = th.load(resume_checkpoint, map_location="cpu")
self.discriminator.load_state_dict(state_dict)
log(f"end loading discriminator model from checkpoint: {resume_checkpoint}...")
# dist_util.sync_params(self.discriminator.parameters())
# dist_util.sync_params(self.discriminator.buffers())
# log(f"end synchronizing discriminator from GPU0 to all GPUs")
def _load_ema_parameters(self, rate):
ema_params = copy.deepcopy(self.mp_trainer.master_params)
if self.args.sanity_check:
log(f"{rate} ema param before overriding: ", ema_params[1].reshape(-1)[:3])
main_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
ema_checkpoint = find_ema_checkpoint(main_checkpoint, self.resume_step, rate)
if ema_checkpoint:
# if dist.get_rank() == 0:
log(f"loading EMA from checkpoint: {ema_checkpoint}...")
if self.args.map_location == 'cuda':
state_dict = th.load(ema_checkpoint, map_location= dev)#"cpu")
else:
state_dict = load_state_dict(
ema_checkpoint, map_location='cpu'
)
ema_params = self.mp_trainer.state_dict_to_master_params(state_dict)
log(f"end loading EMA from checkpoint: {ema_checkpoint}...")
# dist_util.sync_params(ema_params)
log(f"end synchronizing EMA from GPU0 to all GPUs")
if self.args.sanity_check:
log(f"{rate} ema param after overriding: ", ema_params[1].reshape(-1)[:3])
return ema_params
def _load_optimizer_state(self):
main_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
opt_checkpoint = bf.join(
bf.dirname(main_checkpoint), f"opt{self.resume_step:06}.pt"
)
if bf.exists(opt_checkpoint):
log(f"loading optimizer state from checkpoint: {opt_checkpoint}")
#if self.args.map_location == 'cuda':
# state_dict = dist_util.load_state_dict(
# opt_checkpoint, map_location=dist_util.dev()
# )
#else:
state_dict = th.load(opt_checkpoint, map_location="cpu")
self.opt.load_state_dict(state_dict)
log(f"end loading optimizer state from checkpoint: {opt_checkpoint}")
if self.args.sanity_check:
print("opt state dict after overriding: ", self.opt.state_dict()['state'])
def _load_d_optimizer_state(self):
main_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
opt_checkpoint = bf.join(
bf.dirname(main_checkpoint), f"d_opt{self.resume_step:06}.pt"
)
if bf.exists(opt_checkpoint):
log(f"loading d_optimizer state from checkpoint: {opt_checkpoint}")
if os.path.exists(opt_checkpoint):
if self.args.map_location == 'cuda':
state_dict = load_state_dict(
opt_checkpoint, map_location=dev
)
else:
state_dict = th.load(opt_checkpoint, map_location="cpu")
self.d_opt.load_state_dict(state_dict)
log(f"end loading d_optimizer state from checkpoint: {opt_checkpoint}")
def _update_ema(self):
for rate, params in zip(self.ema_rate, self.ema_params):
update_ema(params, self.mp_trainer.master_params, rate=rate)
def _anneal_lr(self):
if not self.args.lr_anneal_steps:
return
frac_done = (self.step + self.resume_step) / self.args.lr_anneal_steps
lr = self.lr * (1 - frac_done)
for param_group in self.opt.param_groups:
param_group["lr"] = lr
def log_step(self):
logkv("step", self.step + self.resume_step)
logkv("samples", (self.step + self.resume_step + 1) * self.global_batch)
def sampling(self, model, sampler, ctm=None, teacher=False, step=-1, num_samples=-1, batch_size=-1, rate=0.999,
png=False, resize=True, generator=None, class_generator=None, sample_dir=''):
if not teacher:
model.eval()
if step == -1:
step = self.args.sampling_steps
if batch_size == -1:
batch_size = self.args.sampling_batch
number = 0
while num_samples > number:
with th.no_grad():
model_kwargs = {}
if self.args.class_cond:
if self.args.train_classes >= 0:
classes = th.ones(size=(batch_size,), device=dev, dtype=int) * self.args.train_classes
model_kwargs["y"] = classes
elif self.args.train_classes == -2:
classes = [0, 1, 9, 11, 29, 31, 33, 55, 76, 89, 90, 130, 207, 250, 279, 281, 291, 323, 386, 387,
388, 417, 562, 614, 759, 789, 800, 812, 848, 933, 973, 980]
assert batch_size % len(classes) == 0
model_kwargs["y"] = th.tensor([x for x in classes for _ in range(batch_size // len(classes))], device=dev)
else:
if class_generator != None:
model_kwargs["y"] = class_generator.randint(0, self.args.num_classes, (batch_size,), device=dev)
else:
if num_samples == -1:
model_kwargs["y"] = self.classes.to(dev)
else:
model_kwargs["y"] = th.randint(0, self.args.num_classes, size=(batch_size, ), device=dev)
if generator != None:
x_T = generator.randn(*(batch_size, self.args.in_channels, self.args.image_size, self.args.image_size),
device=dev) * self.args.sigma_max
if self.args.large_log:
print("x_T: ", x_T[0][0][0][:3])
else:
x_T = None
sample = karras_sample(
diffusion=self.diffusion,
model=model,
shape=(batch_size, self.args.in_channels, self.args.image_size, self.args.image_size),
steps=step,
model_kwargs=model_kwargs,
device=dev,
clip_denoised=True if teacher else self.args.clip_denoised,
sampler=sampler,
generator=None,
teacher=teacher,
ctm=ctm if ctm != None else True if self.args.training_mode.lower() == 'ctm' else False,
x_T=x_T if generator != None else self.x_T.to(dev) if num_samples == -1 else None,
clip_output=self.args.clip_output,
sigma_min=self.args.sigma_min,
sigma_max=self.args.sigma_max,
train=False,
)
if resize:
sample = F.interpolate(sample, size=224, mode="bilinear")
sample = ((sample + 1) * 127.5).clamp(0, 255).to(th.uint8)
sample = sample.permute(0, 2, 3, 1)
sample = sample.contiguous()
# gathered_samples = [th.zeros_like(sample) for _ in range(dist.get_world_size())]
# dist.all_gather(gathered_samples, sample)
# all_images = [sample.cpu().numpy() for sample in gathered_samples]
# arr = np.concatenate(all_images, axis=0)
arr = sample.cpu().numpy()
# if dist.get_rank() == 0:
os.makedirs(bf.join(get_blob_logdir(), f"{sample_dir}"), exist_ok=True)
if self.args.large_log:
print(f"saving to {bf.join(get_blob_logdir(), sample_dir)}")
nrow = int(np.sqrt(arr.shape[0]))
image_grid = make_grid(th.tensor(arr).permute(0, 3, 1, 2) / 255., nrow, padding=2)
if num_samples == -1:
print("1")
with bf.BlobFile(bf.join(get_blob_logdir(), f"{'teacher_' if teacher else ''}sample_{sampler}_sampling_step_{step}_step_{self.step}.png"), "wb") as fout:
save_image(image_grid, fout)
else:
print("2")
if generator != None:
print("3")
os.makedirs(bf.join(get_blob_logdir(), sample_dir),
exist_ok=True)
np.savez(bf.join(get_blob_logdir(), f"{sample_dir}/sample_{number // arr.shape[0]}.npz"),
arr)
if png and number <= 3000:
print("4")
with bf.BlobFile(bf.join(get_blob_logdir(),
f"{sample_dir}/sample_{number // arr.shape[0]}.png"), "wb") as fout:
save_image(image_grid, fout)
else:
print("5")
r = np.random.randint(100000000)
# if self.args.large_log:
# log(f'{dist.get_rank()} number {number}')
log(bf.join(get_blob_logdir(), f"{sample_dir}"))
os.makedirs(bf.join(get_blob_logdir(), f"{sample_dir}"),
exist_ok=True)
np.savez(bf.join(get_blob_logdir(), f"{sample_dir}/sample_{r}.npz"),
arr)
if png and number <= 1000:
print("6")
with bf.BlobFile(bf.join(get_blob_logdir(),
f"{sample_dir}/sample_{r}.png"), "wb") as fout:
save_image(image_grid, fout)
number += arr.shape[0]
print(f"{number} number samples complete")
if not teacher:
model.train()
def calculate_similarity_metrics(self, image_path, num_samples=50000, step=1, batch_size=100, rate=0.999, sampler='exact', log=True):
files = glob.glob(os.path.join(image_path, 'sample*.npz'))
files.sort()
count = 0
psnr = 0
ssim = 0
for i, file in enumerate(files):
images = np.load(file)['arr_0']
for k in range((images.shape[0] - 1) // batch_size + 1):
#ref_img = self.ref_images[count + k * batch_size: count + (k + 1) * batch_size]
if count + batch_size > num_samples:
remaining_num_samples = num_samples - count
else:
remaining_num_samples = batch_size
img = images[k * batch_size: k * batch_size + remaining_num_samples]
ref_img = self.ref_images[count: count + remaining_num_samples]
psnr += cv2.PSNR(img, ref_img) * remaining_num_samples
ssim += SSIM_(img,ref_img,multichannel=True,channel_axis=3,data_range=255) * remaining_num_samples
count = count + remaining_num_samples
print(count)
if count >= num_samples:
break
if count >= num_samples:
break
assert count == num_samples
print(count)
psnr /= num_samples
ssim /= num_samples
assert num_samples % 1000 == 0
if log:
log(f"{self.step}-th step {sampler} sampler (NFE {step}) EMA {rate} PSNR-{num_samples // 1000}k: {psnr}, SSIM-{num_samples // 1000}k: {ssim}")
else:
return psnr, ssim
def calculate_inception_stats(self, data_name, image_path, num_samples=50000, batch_size=100, device=th.device('cuda')):
if data_name.lower() == 'cifar10':
print(f'Loading images from "{image_path}"...')
mu = th.zeros([self.feature_dim], dtype=th.float64, device=device)
sigma = th.zeros([self.feature_dim, self.feature_dim], dtype=th.float64, device=device)
files = glob.glob(os.path.join(image_path, 'sample*.npz'))
count = 0
for file in files:
images = np.load(file)['arr_0'] # [0]#["samples"]
for k in range((images.shape[0] - 1) // batch_size + 1):
mic_img = images[k * batch_size: (k + 1) * batch_size]
mic_img = th.tensor(mic_img).permute(0, 3, 1, 2).to(device)
features = self.detector_net(mic_img, **self.detector_kwargs).to(th.float64)
if count + mic_img.shape[0] > num_samples:
remaining_num_samples = num_samples - count
else:
remaining_num_samples = mic_img.shape[0]
mu += features[:remaining_num_samples].sum(0)
sigma += features[:remaining_num_samples].T @ features[:remaining_num_samples]
count = count + remaining_num_samples
print(count)
if count >= num_samples:
break
if count >= num_samples:
break
assert count == num_samples
print(count)
mu /= num_samples
sigma -= mu.ger(mu) * num_samples
sigma /= num_samples - 1
mu = mu.cpu().numpy()
sigma = sigma.cpu().numpy()
return mu, sigma
else:
filenames = glob.glob(os.path.join(image_path, '*.npz'))
imgs = []
for file in filenames:
try:
img = np.load(file) # ['arr_0']
try:
img = img['data']
except:
img = img['arr_0']
imgs.append(img)
except:
pass
imgs = np.concatenate(imgs, axis=0)
os.makedirs(os.path.join(image_path, 'single_npz'), exist_ok=True)
np.savez(os.path.join(os.path.join(image_path, 'single_npz'), f'data'),
imgs) # , labels)
log("computing sample batch activations...")
sample_acts = self.evaluator.read_activations(
os.path.join(os.path.join(image_path, 'single_npz'), f'data.npz'))
log("computing/reading sample batch statistics...")
sample_stats, sample_stats_spatial = tuple(self.evaluator.compute_statistics(x) for x in sample_acts)
with open(os.path.join(os.path.join(image_path, 'single_npz'), f'stats'), 'wb') as f:
pickle.dump({'stats': sample_stats, 'stats_spatial': sample_stats_spatial}, f)
with open(os.path.join(os.path.join(image_path, 'single_npz'), f'acts'), 'wb') as f:
pickle.dump({'acts': sample_acts[0], 'acts_spatial': sample_acts[1]}, f)
return sample_acts, sample_stats, sample_stats_spatial
def compute_fid(self, mu, sigma, ref_mu=None, ref_sigma=None):
if np.array(ref_mu == None).sum():
ref_mu = self.mu_ref
assert ref_sigma == None
ref_sigma = self.sigma_ref
m = np.square(mu - ref_mu).sum()
s, _ = scipy.linalg.sqrtm(np.dot(sigma, ref_sigma), disp=False)
fid = m + np.trace(sigma + ref_sigma - s * 2)
fid = float(np.real(fid))
return fid
def calculate_inception_stats_npz(self, image_path, num_samples=50000, step=1, batch_size=100, device=th.device('cuda'),
rate=0.999):
print(f'Loading images from "{image_path}"...')
mu = th.zeros([self.feature_dim], dtype=th.float64, device=device)
sigma = th.zeros([self.feature_dim, self.feature_dim], dtype=th.float64, device=device)
files = glob.glob(os.path.join(image_path, 'sample*.npz'))
count = 0
for file in files:
images = np.load(file)['arr_0'] # [0]#["samples"]
for k in range((images.shape[0] - 1) // batch_size + 1):
mic_img = images[k * batch_size: (k + 1) * batch_size]
mic_img = th.tensor(mic_img).permute(0, 3, 1, 2).to(device)
features = self.detector_net(mic_img, **self.detector_kwargs).to(th.float64)
if count + mic_img.shape[0] > num_samples:
remaining_num_samples = num_samples - count
else:
remaining_num_samples = mic_img.shape[0]
mu += features[:remaining_num_samples].sum(0)
sigma += features[:remaining_num_samples].T @ features[:remaining_num_samples]
count = count + remaining_num_samples
log(count)
if count >= num_samples:
break
assert count == num_samples
print(count)
mu /= num_samples
sigma -= mu.ger(mu) * num_samples
sigma /= num_samples - 1
mu = mu.cpu().numpy()
sigma = sigma.cpu().numpy()
m = np.square(mu - self.mu_ref).sum()
s, _ = scipy.linalg.sqrtm(np.dot(sigma, self.sigma_ref), disp=False)
fid = m + np.trace(sigma + self.sigma_ref - s * 2)
fid = float(np.real(fid))
assert num_samples % 1000 == 0
log(f"{self.step}-th step exact sampler (NFE {step}) EMA {rate} FID-{num_samples // 1000}k: {fid}")
def update_ema(target_params, source_params, rate=0.99):
"""
Update target parameters to be closer to those of source parameters using
an exponential moving average.
:param target_params: the target parameter sequence.
:param source_params: the source parameter sequence.
:param rate: the EMA rate (closer to 1 means slower).
"""
for targ, src in zip(target_params, source_params):
targ.detach().mul_(rate).add_(src, alpha=1 - rate)
import nvidia_smi
import functools
import gc
import datetime
def log_loss_dict(losses):
for key, values in losses.items():
logkv_mean(f"{key} mean", values.mean().item())
# Log the quantiles (four quartiles, in particular).
logkv_mean(f"{key} std", values.std().item())
#for sub_t, sub_loss in zip(ts.cpu().numpy(), values.detach().cpu().numpy()):
# quartile = int(4 * sub_t / diffusion.num_timesteps)
# logkv_mean(f"{key}_q{quartile}", sub_loss)
def get_blob_logdir():
# You can change this to be a separate path to save checkpoints to
# a blobstore or some external drive.
return get_dir()
class CTMTrainLoop(TrainLoop):
def __init__(
self,
*,
target_model,
teacher_model,
ema_scale_fn,
**kwargs,
):
super().__init__(**kwargs)
self.training_mode = self.args.training_mode
self.ema_scale_fn = ema_scale_fn
self.target_model = target_model
self.teacher_model = teacher_model
self.total_training_steps = self.args.total_training_steps
if target_model:
if self.args.sanity_check:
for name, param in self.target_model.named_parameters():
log("target model parameter before overriding: ", param.data.cpu().detach().reshape(-1)[:3])
break
self._load_and_sync_ema_parameters_to_target_parameters()
if self.args.sanity_check:
# print("doing sanity check ---------------->>>>>>>>>")
for name, param in self.target_model.named_parameters():
log("target model parameter after overriding: ", param.data.cpu().detach().reshape(-1)[:3])
break
self.target_model.requires_grad_(False)
self.target_model.train()
if self.args.use_fp16:
# print("usinig args.usefp16 ----------------->>>>>>>>>>>>>")
self.target_model_param_groups_and_shapes = get_target_param_groups_and_shapes(
self.target_model.named_parameters(), self.model.named_parameters()
)
self.target_model_master_params = make_master_params(
self.target_model_param_groups_and_shapes
)
else:
# print("In the else part of args.usefp16 ----------------->>>>")
self.target_model_param_groups_and_shapes = list(self.target_model.named_parameters())
self.target_model_master_params = list(target_model.parameters())
for rate, params in zip(self.ema_rate, self.ema_params):
if rate == 0.999:
log(f"loading target model from 0.999 ema...")
update_ema(
self.target_model_master_params,
params,
rate=0.0,
)
if self.args.use_fp16:
master_params_to_model_params(
self.target_model_param_groups_and_shapes,
self.target_model_master_params,
)
if self.args.sanity_check:
for name, param in self.target_model.named_parameters():
log("target model parameter after all: ", param.data.cpu().detach().reshape(-1)[:3])
break
if teacher_model:
#self._load_and_sync_teacher_parameters()
self.teacher_model.requires_grad_(False)
self.teacher_model.eval()
self.diffusion.teacher_model = teacher_model
# print("printing the teacher model:::::::::: ", type(teacher_model))
# print("printing the target_model model:::::::::: ", type(target_model))
self.global_step = self.step
self.initial_step = copy.deepcopy(self.step)
if self.args.gpu_usage:
nvidia_smi.nvmlInit()
self.deviceCount = nvidia_smi.nvmlDeviceGetCount()
self.print_gpu_usage('Before everything')
# print('self.args.check_dm_performance', self.args.check_dm_performance)
# exit()check_dm_performance
# self.args.check_dm_performance = False # -------------------------->>>>>>>>> i have hardcoded this to false. but it was true for uncond
if self.args.check_dm_performance and False: # ------------>>>>>> it is throwing error at the sampling part
if not os.path.exists(self.args.dm_sample_path_seed_42):
self.sampling(model=self.teacher_model, sampler='heun', teacher=True, step=18 if self.args.data_name.lower() == 'cifar10' else 40,
num_samples=self.args.eval_num_samples, batch_size=self.args.eval_batch,
rate=0.0, ctm=False, png=False, resize=False,
generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
sample_dir=self.args.dm_sample_path_seed_42)
print(self.args.data_name.lower())
if self.args.data_name.lower() == 'cifar10':
print('Loading Inception-v3 model...')
detector_url = 'https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/metrics/inception-2015-12-05.pkl'
self.detector_kwargs = dict(return_features=True)
self.feature_dim = 2048
with open_url(detector_url, verbose=(0 == 0)) as f:
self.detector_net = pickle.load(f).to(dev)
with open_url(self.args.ref_path) as f:
ref = dict(np.load(f))
self.mu_ref = ref['mu']
self.sigma_ref = ref['sigma']
def print_gpu_usage(self, prefix=''):
for i in range(self.deviceCount):
handle = nvidia_smi.nvmlDeviceGetHandleByIndex(i)
util = nvidia_smi.nvmlDeviceGetUtilizationRates(handle)
mem = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)
log(
f"{prefix} |Device {i}| Mem Free: {mem.free / 1024 ** 2:5.2f}MB / {mem.total / 1024 ** 2:5.2f}MB | gpu-util: {util.gpu / 100.0:3.1%} | gpu-mem: {util.memory / 100.0:3.1%} |")
def _load_and_sync_ema_parameters_to_target_parameters(self):
for rate, params in zip(self.ema_rate, self.ema_params):
if rate == self.args.start_ema: # 0.999
log(f"loading target model from {self.args.start_ema} ema...")
state_dict = self.mp_trainer.master_params_to_state_dict(params)
self.target_model.load_state_dict(state_dict)
log(f"end loading target model from {self.args.start_ema} ema...")
def _load_and_sync_target_parameters(self):
resume_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
if resume_checkpoint:
path, name = os.path.split(resume_checkpoint)
target_name = name.replace("model", "target_model")
resume_target_checkpoint = os.path.join(path, target_name)
if bf.exists(resume_target_checkpoint) == 0:
log(
f"loading target model from checkpoint: {resume_target_checkpoint}..."
)
state_dict = load_state_dict(
resume_target_checkpoint, map_location=dev
)
self.target_model.load_state_dict(state_dict, strict=False)
def _load_and_sync_teacher_parameters(self):
resume_checkpoint = find_resume_checkpoint() or self.args.resume_checkpoint
if resume_checkpoint:
path, name = os.path.split(resume_checkpoint)
teacher_name = name.replace("model", "teacher_model")
resume_teacher_checkpoint = os.path.join(path, teacher_name)
if bf.exists(resume_teacher_checkpoint) == 0:
log(
f"loading teacher model from checkpoint: {resume_teacher_checkpoint}..."
)
state_dict = load_state_dict(
resume_teacher_checkpoint, map_location=dev
)
self.teacher_model.load_state_dict(state_dict)#, strict=False)
def run_loop(self):
if self.args.gpu_usage:
self.print_gpu_usage('Before training')
saved = False
while (
self.step < self.args.lr_anneal_steps
or self.global_step < self.total_training_steps
):
batch, cond = next(self.data)
if self.args.large_log:
print("batch size: ", batch.shape)
# print("rank: ", dist.get_rank())
if self.args.intermediate_samples:
#print("!!: ", self.step, dist.get_rank(), self.args.sample_interval, self.args.training_mode)
#if dist.get_rank() == 0:
if self.step == self.initial_step + 10 or (self.step % self.args.sample_interval == self.args.sample_interval - 1):
if self.args.training_mode.lower() == 'ctm':
if self.args.consistency_weight > 0.:
self.sampling(model=self.ddp_model, sampler='exact', num_samples=self.args.sampling_batch, png=True)
self.sampling(model=self.ddp_model, sampler='exact', step=2, num_samples=self.args.sampling_batch, png=True)
self.sampling(model=self.ddp_model, sampler='exact', step=1, num_samples=self.args.sampling_batch, png=True)
else:
self.sampling(model=self.ddp_model, sampler='heun', ctm=True, teacher=True, num_samples=self.args.sampling_batch, png=True)
elif self.args.training_mode.lower() == 'cd':
self.sampling(model=self.ddp_model, sampler='onestep', step=1, num_samples=self.args.sampling_batch, png=True)
elif self.args.training_mode.lower() == 'edm':
self.sampling(model=self.ddp_model, sampler='heun', step=40, num_samples=self.args.sampling_batch, png=True)
if self.step == self.initial_step + 10 and self.teacher_model != None:
self.sampling(model=self.teacher_model, sampler='heun', ctm=False, teacher=True, num_samples=self.args.sampling_batch, png=True)
self.run_step(batch, cond)
if self.args.gpu_usage:
self.print_gpu_usage('After one step training')
if self.args.large_log:
print("mp trainer master parameter after one step update: ", self.mp_trainer.master_params[1].reshape(-1)[:3])
for name, param in self.model.named_parameters():
print("model parameter after one step update: ", param.data.cpu().detach().reshape(-1)[:3])
break
for name, param in self.target_model.named_parameters():
print("target model parameter after one step update: ", param.data.cpu().detach().reshape(-1)[:3])
break
if self.args.check_ctm_denoising_ability:
self.eval(step=18, sampler='heun', teacher=True, ctm=True, rate=0.0)
if (
self.global_step
and self.args.eval_interval != -1
and self.global_step % self.args.eval_interval == self.args.eval_interval - 1
#and self.step - self.initial_step > 10
or self.step == self.args.lr_anneal_steps - 1
or self.global_step == self.total_training_steps - 1
):
if self.args.gpu_usage:
self.print_gpu_usage('Before emptying cache in evaluation 1')
gc.collect()
th.cuda.empty_cache()
if self.args.gpu_usage:
self.print_gpu_usage('After emptying cache in evaluation 1')
model_state_dict = self.model.state_dict()
if self.args.linear_probing:
self.eval(step=18, sampler='heun', teacher=True, ctm=True, rate=0.0,
generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
delete=True)
self.evaluation(0.0)
log('Evaluation with model parameter end')
for rate, params in zip(self.ema_rate, self.ema_params):
if not self.args.compute_ema_fids:
if rate != 0.999:
continue
state_dict = self.mp_trainer.master_params_to_state_dict(params)
self.model.load_state_dict(state_dict, strict=False)
self.evaluation(rate)
log(f'Evaluation with {rate}-EMA model parameter end')
self.model.load_state_dict(model_state_dict, strict=True)
del model_state_dict, state_dict
if self.args.gpu_usage:
self.print_gpu_usage('Before emptying cache in evaluation 2')
gc.collect()
th.cuda.empty_cache()
if self.args.gpu_usage:
self.print_gpu_usage('After emptying cache in evaluation 2')
# dist.barrier()
if (
self.global_step
and self.args.eval_interval != -1
and self.global_step % self.args.save_check_period == self.args.save_check_period - 1
#and self.step - self.initial_step > 10000
or self.step == self.args.lr_anneal_steps - 1
or self.global_step == self.total_training_steps - 1
):
gc.collect()
th.cuda.empty_cache()
model_state_dict = self.model.state_dict()
for rate, params in zip(self.ema_rate, self.ema_params):
if rate == 0.999:
state_dict = self.mp_trainer.master_params_to_state_dict(params)
self.model.load_state_dict(state_dict, strict=False)
fid = self.save_check(rate)
# if dist.get_rank() == 0:
assert fid != None
self.fids.append(fid)
save_ckpt = (self.fids[-1] == np.min(self.fids))
log("FID by iteration (NFE 1, EMA 0.999): ", self.fids)
self.model.load_state_dict(model_state_dict, strict=True)
del model_state_dict, state_dict
# if dist.get_rank() == 0:
if save_ckpt:
self.save(save_full=False)
gc.collect()
th.cuda.empty_cache()
if self.args.large_log:
print("mp trainer master parameter after sampling: ",
self.mp_trainer.master_params[1].reshape(-1)[:3])
for name, param in self.model.named_parameters():
print("model parameter after sampling: ", param.data.cpu().detach().reshape(-1)[:3])
break
for name, param in self.target_model.named_parameters():
print("target model parameter after sampling: ", param.data.cpu().detach().reshape(-1)[:3])
break
saved = False
if (
self.global_step
and self.args.save_interval != -1
and self.global_step % self.args.save_interval == 0
):
self.save()
if self.discriminator != None:
self.d_save()
saved = True
gc.collect()
th.cuda.empty_cache()
# Run for a finite amount of time in integration tests.
if os.environ.get("DIFFUSION_TRAINING_TEST", "") and self.step > 0:
return
if self.global_step % self.args.log_interval == 0:
dumpkvs()
log(datetime.datetime.now().strftime("SONY-%Y-%m-%d-%H-%M-%S"))
if self.args.large_log:
print("mp trainer master parameter after saving: ",
self.mp_trainer.master_params[1].reshape(-1)[:3])
for name, param in self.model.named_parameters():
print("model parameter after saving: ", param.data.cpu().detach().reshape(-1)[:3])
break
for name, param in self.target_model.named_parameters():
print("target model parameter after saving: ", param.data.cpu().detach().reshape(-1)[:3])
break
print(f"0.999 ema param after overriding (should be same to the target parameter): ",
self.ema_params[0][1].reshape(-1)[:3])
# Save the last checkpoint if it wasn't already saved.
if not saved:
self.save()
if self.discriminator != None:
self.d_save()
def save_check(self, rate):
if self.args.training_mode.lower() == 'ctm':
assert rate == 0.999
#fid = self.eval(step=1, rate=rate, ctm=True, delete=True, out=True)
fid = self.eval(step=1, rate=rate, ctm=True, generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
metric='similarity', delete=True, out=True)
return fid
def evaluation(self, rate):
if self.args.training_mode.lower() == 'ctm':
if self.args.eval_fid:
self.eval(step=1, rate=rate, ctm=True, delete=True)
if self.args.eval_similarity:
self.eval(step=1, rate=rate, ctm=True, generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
metric='similarity', delete=True)
if self.args.eval_fid:
self.eval(step=2, rate=rate, ctm=True, delete=True)
if self.args.eval_similarity:
self.eval(step=2, rate=rate, ctm=True,
generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
metric='similarity', delete=True)
if self.args.compute_ema_fids:
if self.args.eval_fid:
self.eval(step=4, rate=rate, ctm=True, delete=True)
if self.args.eval_similarity:
self.eval(step=4, rate=rate, ctm=True, generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
metric='similarity', delete=True)
step = 18 if self.args.data_name.lower() == 'cifar10' else 40
if self.args.eval_large_nfe:
if self.args.eval_fid:
self.eval(step=step, rate=rate, ctm=True, delete=True)
if self.args.eval_similarity:
self.eval(step=step, rate=rate, ctm=True, generator=get_generator('determ', self.args.eval_num_samples, self.args.eval_seed),
class_generator=get_generator('determ', self.args.eval_num_samples, 0),
metric='similarity', delete=True)
elif self.args.training_mode.lower() == 'cm':
if self.args.eval_fid:
self.eval(step=1, sampler='onestep', rate=rate, ctm=False, delete=True)
def run_step(self, batch, cond):
if self.args.large_log:
print("mp trainer master parameter before update: ", self.mp_trainer.master_params[1].reshape(-1)[:3])
for name, param in self.model.named_parameters():
print("model parameter before update: ", param.data.cpu().detach().reshape(-1)[:3])
break
for name, param in self.target_model.named_parameters():
print("target model parameter before update: ", param.data.cpu().detach().reshape(-1)[:3])
break
self.forward_backward(batch, cond)
if self.discriminator == None:
took_step = self.mp_trainer.optimize(self.opt)
if took_step:
self._update_ema()
if self.target_model:
self._update_target_ema()
self.step += 1
self.global_step += 1
else:
if self.step % self.args.g_learning_period == 0:
took_step = self.mp_trainer.optimize(self.opt)
else:
took_step = self.d_mp_trainer.optimize(self.d_opt)
# print(self.step, took_step)
if took_step:
if self.step % self.args.g_learning_period == 0:
self._update_ema()
if self.target_model:
self._update_target_ema()
self.step += 1
self.global_step += 1
self._anneal_lr()
self.log_step()
def _update_target_ema(self):
target_ema, scales = self.ema_scale_fn(self.global_step)
with th.no_grad():
update_ema(
self.target_model_master_params,
self.mp_trainer.master_params,
rate=target_ema,
)
if self.args.use_fp16:
master_params_to_model_params(
self.target_model_param_groups_and_shapes,
self.target_model_master_params,
)
def forward_backward(self, batch, cond):
self.mp_trainer.zero_grad()
if self.discriminator != None:
self.d_mp_trainer.zero_grad()
num_heun_step = [self.diffusion.get_num_heun_step(num_heun_step=self.args.num_heun_step)]
if self.args.gan_specific_time:
gan_num_heun_step = [self.diffusion.get_num_heun_step(num_heun_step=self.args.gan_num_heun_step,
heun_step_strategy=self.args.gan_heun_step_strategy)]
diffusion_training_ = [np.random.rand() < self.args.diffusion_training_frequency]
gan_training_ = [np.random.rand() < self.args.gan_training_frequency]
# dist.broadcast_object_list(num_heun_step, 0)
# if self.args.gan_specific_time:
# dist.broadcast_object_list(gan_num_heun_step, 0)
# dist.broadcast_object_list(diffusion_training_, 0)
# dist.broadcast_object_list(gan_training_, 0)
num_heun_step = num_heun_step[0]
if self.args.gan_specific_time:
gan_num_heun_step = gan_num_heun_step[0]
else:
gan_num_heun_step = -1
diffusion_training_ = diffusion_training_[0]
gan_training_ = gan_training_[0]
for i in range(0, batch.shape[0], self.microbatch):
micro = batch[i : i + self.microbatch].to(dev)
micro_cond = {
k: v[i : i + self.microbatch].to(dist_util.dev())
for k, v in cond.items()
}
last_batch = (i + self.microbatch) >= batch.shape[0]
compute_losses = functools.partial(
self.diffusion.ctm_losses,
step=self.step,
model=self.ddp_model,
x_start=micro,
model_kwargs=micro_cond,
target_model=self.target_model,
discriminator=self.ddp_discriminator,
init_step=self.initial_step,
ctm=True if self.training_mode.lower() == 'ctm' else False,
num_heun_step=num_heun_step,
gan_num_heun_step=gan_num_heun_step,
diffusion_training_=diffusion_training_,
gan_training_=gan_training_,
)
if last_batch or not self.use_ddp:
losses = compute_losses()
else:
if self.step % self.args.g_learning_period == 0:
# with self.ddp_model.no_sync():
losses = compute_losses()
else:
# with self.ddp_discriminator.no_sync():
losses = compute_losses()
if 'consistency_loss' in list(losses.keys()):
# print("Consistency learning")
loss = self.args.consistency_weight * losses["consistency_loss"].mean()
if 'd_loss' in list(losses.keys()):
if self.args.large_log:
print("GAN learning, ", self.args.discriminator_weight, losses['d_loss'].mean())
loss = loss + self.args.discriminator_weight * losses['d_loss'].mean()
if 'denoising_loss' in list(losses.keys()):
loss = loss + self.args.denoising_weight * losses['denoising_loss'].mean()
log_loss_dict({k: v.view(-1) for k, v in losses.items()})
if self.args.sanity_check:
# print("rank: ", dist.get_rank())
for name, param in self.model.named_parameters():
print("model parameter gradient for current microbatch: ",
th.autograd.grad(outputs=(2**self.mp_trainer.lg_loss_scale) * loss, inputs=param, retain_graph=True)[0].reshape(-1)[:3])
break
self.mp_trainer.backward(loss)
elif 'd_loss' in list(losses.keys()):
assert self.step % self.args.g_learning_period != 0
loss = (losses["d_loss"]).mean()
self.d_mp_trainer.backward(loss)
if self.args.large_log:
for param in self.discriminator.parameters():
try:
print("discriminator param data, grad: ", param.grad.reshape(-1)[:3])
except:
print("discriminator param grad: ", param.grad)
break
elif 'denoising_loss' in list(losses.keys()):
loss = losses['denoising_loss'].mean()
log_loss_dict({k: v.view(-1) for k, v in losses.items()})
self.mp_trainer.backward(loss)
if self.args.sanity_check:
# print("rank: ", dist.get_rank())
for name, param in self.model.named_parameters():
print("model parameter gradient across all microbatch: ", param.grad.cpu().detach().reshape(-1)[:3])
break
@th.no_grad()
def eval(self, step=1, sampler='exact', teacher=False, ctm=False, rate=0.999, generator=None, class_generator=None, metric='fid', delete=False, out=False):
model = self.model
sample_dir = f"{self.step}_{sampler}_{step}_{rate}"
if generator != None:
sample_dir = sample_dir + "_seed_42"
self.sampling(model=model, sampler=sampler, teacher=teacher, step=step,
num_samples=self.args.eval_num_samples, batch_size=self.args.eval_batch,
rate=rate, ctm=ctm, png=True, resize=False, generator=generator,
class_generator=class_generator, sample_dir=sample_dir)
gc.collect()
th.cuda.empty_cache()
# if dist.get_rank() == 0:
if self.args.data_name.lower() == 'cifar10':
if metric == 'fid':
mu, sigma = self.calculate_inception_stats(self.args.data_name,
os.path.join(get_blob_logdir(), sample_dir),
num_samples=self.args.eval_num_samples)
log(f"{self.step}-th step {sampler} sampler (NFE {step}) EMA {rate}"
f" FID-{self.args.eval_num_samples // 1000}k: {self.compute_fid(mu, sigma)}")
if metric == 'similarity':
mu, sigma = self.calculate_inception_stats(self.args.data_name,
os.path.join(get_blob_logdir(), sample_dir),
num_samples=self.args.eval_num_samples)
log(f"{self.step}-th step {sampler} sampler (NFE {step}) seed 42 EMA {rate}"
f" FID-{self.args.eval_num_samples // 1000}k: {self.compute_fid(mu, sigma)}")
if self.args.check_dm_performance and False:
log(f"{self.step}-th step {sampler} sampler (NFE {step}) EMA {rate}"
f" FID-{self.args.eval_num_samples // 1000}k compared with DM: {self.compute_fid(mu, sigma, self.dm_mu, self.dm_sigma)}")
self.calculate_similarity_metrics(os.path.join(get_blob_logdir(), sample_dir),
num_samples=self.args.eval_num_samples, step=step, rate=rate)
if delete:
shutil.rmtree(os.path.join(get_blob_logdir(), sample_dir))
if out:
return self.compute_fid(mu, sigma)
else:
sample_acts, sample_stats, sample_stats_spatial = self.calculate_inception_stats(self.args.data_name,
bf.join(get_blob_logdir(), sample_dir),
num_samples=self.args.eval_num_samples)
log(f"Inception Score-{self.args.eval_num_samples // 1000}k:", self.evaluator.compute_inception_score(sample_acts[0]))
log(f"FID-{self.args.eval_num_samples // 1000}k:", sample_stats.frechet_distance(self.ref_stats))
log(f"sFID-{self.args.eval_num_samples // 1000}k:", sample_stats_spatial.frechet_distance(self.ref_stats_spatial))
prec, recall = self.evaluator.compute_prec_recall(self.ref_acts[0], sample_acts[0])
log("Precision:", prec)
log("Recall:", recall)
#self.evaluator.sess.close()
#tf.reset_default_graph()
def save(self, save_full=True):
def save_checkpoint(rate, params):
state_dict = self.mp_trainer.master_params_to_state_dict(params)
# if dist.get_rank() == 0:
log(f"saving model {rate}...")
if not rate:
filename = f"model{self.global_step:06d}.pt"
else:
filename = f"ema_{rate}_{self.global_step:06d}.pt"
with bf.BlobFile(bf.join(get_blob_logdir(), filename), "wb") as f:
th.save(state_dict, f)
for rate, params in zip(self.ema_rate, self.ema_params):
if not save_full:
if rate == 0.999:
save_checkpoint(rate, params)
else:
save_checkpoint(rate, params)
if save_full:
log("saving optimizer state...")
# if dist.get_rank() == 0:
with bf.BlobFile(
bf.join(get_blob_logdir(), f"opt{self.global_step:06d}.pt"),
"wb",
) as f:
th.save(self.opt.state_dict(), f)
# if dist.get_rank() == 0:
if self.target_model:
log("saving target model state")
filename = f"target_model{self.global_step:06d}.pt"
with bf.BlobFile(bf.join(get_blob_logdir(), filename), "wb") as f:
th.save(self.target_model.state_dict(), f)
if self.teacher_model and self.training_mode == "progdist":
log("saving teacher model state")
filename = f"teacher_model{self.global_step:06d}.pt"
with bf.BlobFile(bf.join(get_blob_logdir(), filename), "wb") as f:
th.save(self.teacher_model.state_dict(), f)
# Save model parameters last to prevent race conditions where a restart
# loads model at step N, but opt/ema state isn't saved for step N.
save_checkpoint(0, self.mp_trainer.master_params)
# dist.barrier()
def d_save(self):
log("saving d_optimizer state...")
# if dist.get_rank() == 0:
with bf.BlobFile(
bf.join(get_blob_logdir(), f"d_opt{self.global_step:06d}.pt"),
"wb",
) as f:
th.save(self.d_opt.state_dict(), f)
with bf.BlobFile(bf.join(get_blob_logdir(), f"d_model{self.global_step:06d}.pt"), "wb") as f:
th.save(self.d_mp_trainer.master_params_to_state_dict(self.d_mp_trainer.master_params), f)
# Save model parameters last to prevent race conditions where a restart
# loads model at step N, but opt/ema state isn't saved for step N.
# dist.barrier()
def log_step(self):
step = self.global_step
logkv("step", step)
logkv("samples", (step + 1) * self.global_batch)