373 lines
14 KiB
Python
373 lines
14 KiB
Python
import copy
|
|
import os
|
|
import random
|
|
from functools import wraps
|
|
import kornia.augmentation as augs
|
|
|
|
import torch
|
|
import torch.nn.functional as F
|
|
import torchvision
|
|
from PIL import Image
|
|
from kornia import filters, apply_hflip
|
|
from torch import nn
|
|
from torchvision.transforms import ToTensor
|
|
|
|
from data.byol_attachment import RandomApply
|
|
from trainer.networks import register_model, create_model
|
|
from utils.util import checkpoint, opt_get
|
|
|
|
|
|
def default(val, def_val):
|
|
return def_val if val is None else val
|
|
|
|
|
|
def flatten(t):
|
|
return t.reshape(t.shape[0], -1)
|
|
|
|
|
|
def singleton(cache_key):
|
|
def inner_fn(fn):
|
|
@wraps(fn)
|
|
def wrapper(self, *args, **kwargs):
|
|
instance = getattr(self, cache_key)
|
|
if instance is not None:
|
|
return instance
|
|
|
|
instance = fn(self, *args, **kwargs)
|
|
setattr(self, cache_key, instance)
|
|
return instance
|
|
|
|
return wrapper
|
|
|
|
return inner_fn
|
|
|
|
|
|
def get_module_device(module):
|
|
return next(module.parameters()).device
|
|
|
|
|
|
def set_requires_grad(model, val):
|
|
for p in model.parameters():
|
|
p.requires_grad = val
|
|
|
|
|
|
# Specialized augmentor class that applies a set of image transformations on points as well, allowing one to track
|
|
# where a point in the src image is located in the dest image. Restricts transformation such that this is possible.
|
|
class PointwiseAugmentor(nn.Module):
|
|
def __init__(self, img_size=224):
|
|
super().__init__()
|
|
self.jitter = RandomApply(augs.ColorJitter(0.8, 0.8, 0.8, 0.2), p=0.8)
|
|
self.gray = augs.RandomGrayscale(p=0.2)
|
|
self.blur = RandomApply(filters.GaussianBlur2d((3, 3), (1.5, 1.5)), p=0.1)
|
|
self.rrc = augs.RandomResizedCrop((img_size, img_size), same_on_batch=True)
|
|
|
|
# Given a point in the source image, returns the same point in the source image, given the kornia RRC params.
|
|
def rrc_on_point(self, src_point, params):
|
|
dh, dw = params['dst'][:,2,1]-params['dst'][:,0,1], params['dst'][:,2,0] - params['dst'][:,0,0]
|
|
sh, sw = params['src'][:,2,1]-params['src'][:,0,1], params['src'][:,2,0] - params['src'][:,0,0]
|
|
scale_h, scale_w = sh.float() / dh.float(), sw.float() / dw.float()
|
|
t, l = src_point[0] - params['src'][0,0,1], src_point[1] - params['src'][0,0,0]
|
|
t = (t.float() / scale_h[0]).long()
|
|
l = (l.float() / scale_w[0]).long()
|
|
return torch.stack([t,l])
|
|
|
|
def flip_on_point(self, pt, input):
|
|
t, l = pt[0], pt[1]
|
|
center = input.shape[-1] // 2
|
|
return t, 2 * center - l
|
|
|
|
def forward(self, x, point):
|
|
d = self.jitter(x)
|
|
d = self.gray(d)
|
|
will_flip = random.random() > .5
|
|
if will_flip:
|
|
d = apply_hflip(d)
|
|
point = self.flip_on_point(point, x)
|
|
d = self.blur(d)
|
|
|
|
invalid = True
|
|
while invalid:
|
|
params = self.rrc.generate_parameters(d.shape)
|
|
potential = self.rrc_on_point(point, params)
|
|
# '10' is an arbitrary number: we want to provide some margin. Making predictions at the very edge of an image is not very useful.
|
|
if potential[0] <= 10 or potential[1] <= 10 or potential[0] > x.shape[-2]-10 or potential[1] > x.shape[-1]-10:
|
|
continue
|
|
d = self.rrc(d, params=params)
|
|
point = potential
|
|
invalid = False
|
|
|
|
return d, point
|
|
|
|
|
|
# loss fn
|
|
def loss_fn(x, y):
|
|
x = F.normalize(x, dim=-1, p=2)
|
|
y = F.normalize(y, dim=-1, p=2)
|
|
return 2 - 2 * (x * y).sum(dim=-1)
|
|
|
|
|
|
# exponential moving average
|
|
class EMA():
|
|
def __init__(self, beta):
|
|
super().__init__()
|
|
self.beta = beta
|
|
|
|
def update_average(self, old, new):
|
|
if old is None:
|
|
return new
|
|
return old * self.beta + (1 - self.beta) * new
|
|
|
|
|
|
def update_moving_average(ema_updater, ma_model, current_model):
|
|
for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
|
|
old_weight, up_weight = ma_params.data, current_params.data
|
|
ma_params.data = ema_updater.update_average(old_weight, up_weight)
|
|
|
|
|
|
# MLP class for projector and predictor
|
|
class MLP(nn.Module):
|
|
def __init__(self, dim, projection_size, hidden_size=4096):
|
|
super().__init__()
|
|
self.net = nn.Sequential(
|
|
nn.Linear(dim, hidden_size),
|
|
nn.BatchNorm1d(hidden_size),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(hidden_size, projection_size)
|
|
)
|
|
|
|
def forward(self, x):
|
|
x = flatten(x)
|
|
return self.net(x)
|
|
|
|
|
|
# A wrapper class for training against networks that do not collapse into a small-dimensioned latent.
|
|
class StructuralMLP(nn.Module):
|
|
def __init__(self, dim, projection_size, hidden_size=4096):
|
|
super().__init__()
|
|
b, c, h, w = dim
|
|
flattened_dim = c * h // 4 * w // 4
|
|
self.net = nn.Sequential(
|
|
nn.Conv2d(c, c, kernel_size=3, padding=1, stride=2),
|
|
nn.BatchNorm2d(c),
|
|
nn.ReLU(inplace=True),
|
|
nn.Conv2d(c, c, kernel_size=3, padding=1, stride=2),
|
|
nn.BatchNorm2d(c),
|
|
nn.ReLU(inplace=True),
|
|
nn.Flatten(),
|
|
nn.Linear(flattened_dim, hidden_size),
|
|
nn.BatchNorm1d(hidden_size),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(hidden_size, projection_size)
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.net(x)
|
|
|
|
|
|
# a wrapper class for the base neural network
|
|
# will manage the interception of the hidden layer output
|
|
# and pipe it into the projecter and predictor nets
|
|
class NetWrapper(nn.Module):
|
|
def __init__(self, net, projection_size, projection_hidden_size, layer=-2, use_structural_mlp=False):
|
|
super().__init__()
|
|
self.net = net
|
|
self.layer = layer
|
|
|
|
self.projector = None
|
|
self.projection_size = projection_size
|
|
self.projection_hidden_size = projection_hidden_size
|
|
self.structural_mlp = use_structural_mlp
|
|
|
|
self.hidden = None
|
|
self.hook_registered = False
|
|
|
|
def _find_layer(self):
|
|
if type(self.layer) == str:
|
|
modules = dict([*self.net.named_modules()])
|
|
return modules.get(self.layer, None)
|
|
elif type(self.layer) == int:
|
|
children = [*self.net.children()]
|
|
return children[self.layer]
|
|
return None
|
|
|
|
def _hook(self, _, __, output):
|
|
self.hidden = output
|
|
|
|
def _register_hook(self):
|
|
layer = self._find_layer()
|
|
assert layer is not None, f'hidden layer ({self.layer}) not found'
|
|
handle = layer.register_forward_hook(self._hook)
|
|
self.hook_registered = True
|
|
|
|
@singleton('projector')
|
|
def _get_projector(self, hidden):
|
|
if self.structural_mlp:
|
|
projector = StructuralMLP(hidden.shape, self.projection_size, self.projection_hidden_size)
|
|
else:
|
|
_, dim = hidden.flatten(1,-1).shape
|
|
projector = MLP(dim, self.projection_size, self.projection_hidden_size)
|
|
return projector.to(hidden)
|
|
|
|
def get_representation(self, **kwargs):
|
|
if self.layer == -1:
|
|
return self.net(**kwargs)
|
|
|
|
if not self.hook_registered:
|
|
self._register_hook()
|
|
|
|
unused = self.net(**kwargs)
|
|
hidden = self.hidden
|
|
self.hidden = None
|
|
assert hidden is not None, f'hidden layer {self.layer} never emitted an output'
|
|
return hidden
|
|
|
|
def forward(self, **kwargs):
|
|
representation = self.get_representation(**kwargs)
|
|
projector = self._get_projector(representation)
|
|
projection = checkpoint(projector, representation)
|
|
return projection
|
|
|
|
|
|
class BYOL(nn.Module):
|
|
def __init__(
|
|
self,
|
|
net,
|
|
image_size,
|
|
hidden_layer=-2,
|
|
projection_size=256,
|
|
projection_hidden_size=4096,
|
|
moving_average_decay=0.99,
|
|
use_momentum=True,
|
|
structural_mlp=False,
|
|
contrastive=False,
|
|
):
|
|
super().__init__()
|
|
|
|
self.online_encoder = NetWrapper(net, projection_size, projection_hidden_size, layer=hidden_layer,
|
|
use_structural_mlp=structural_mlp)
|
|
|
|
self.aug = PointwiseAugmentor(image_size)
|
|
self.use_momentum = use_momentum
|
|
self.contrastive = contrastive
|
|
self.target_encoder = None
|
|
self.target_ema_updater = EMA(moving_average_decay)
|
|
|
|
self.online_predictor = MLP(projection_size, projection_size, projection_hidden_size)
|
|
|
|
# get device of network and make wrapper same device
|
|
device = get_module_device(net)
|
|
self.to(device)
|
|
|
|
# send a mock image tensor to instantiate singleton parameters
|
|
self.forward(torch.randn(2, 3, image_size, image_size, device=device))
|
|
|
|
@singleton('target_encoder')
|
|
def _get_target_encoder(self):
|
|
target_encoder = copy.deepcopy(self.online_encoder)
|
|
set_requires_grad(target_encoder, False)
|
|
for p in target_encoder.parameters():
|
|
p.DO_NOT_TRAIN = True
|
|
return target_encoder
|
|
|
|
def reset_moving_average(self):
|
|
del self.target_encoder
|
|
self.target_encoder = None
|
|
|
|
def update_for_step(self, step, __):
|
|
assert self.use_momentum, 'you do not need to update the moving average, since you have turned off momentum for the target encoder'
|
|
assert self.target_encoder is not None, 'target encoder has not been created yet'
|
|
update_moving_average(self.target_ema_updater, self.target_encoder, self.online_encoder)
|
|
|
|
def get_debug_values(self, step, __):
|
|
# In the BYOL paper, this is made to increase over time. Not yet implemented, but still logging the value.
|
|
dbg = {'target_ema_beta': self.target_ema_updater.beta}
|
|
if self.contrastive and hasattr(self, 'logs_closs'):
|
|
dbg['contrastive_distance'] = self.logs_closs
|
|
dbg['byol_distance'] = self.logs_loss
|
|
return dbg
|
|
|
|
def visual_dbg(self, step, path):
|
|
torchvision.utils.save_image(self.im1.cpu().float(), os.path.join(path, "%i_image1.png" % (step,)))
|
|
torchvision.utils.save_image(self.im2.cpu().float(), os.path.join(path, "%i_image2.png" % (step,)))
|
|
|
|
def get_predictions_and_projections(self, image):
|
|
_, _, h, w = image.shape
|
|
point = torch.randint(h//8, 7*h//8, (2,)).long().to(image.device)
|
|
|
|
image_one, pt_one = self.aug(image, point)
|
|
image_two, pt_two = self.aug(image, point)
|
|
|
|
# Keep copies on hand for visual_dbg.
|
|
self.im1 = image_one.detach().clone()
|
|
self.im1[:,:,pt_one[0]-3:pt_one[0]+3,pt_one[1]-3:pt_one[1]+3] = 1
|
|
self.im2 = image_two.detach().clone()
|
|
self.im2[:,:,pt_two[0]-3:pt_two[0]+3,pt_two[1]-3:pt_two[1]+3] = 1
|
|
|
|
online_proj_one = self.online_encoder(img=image_one, pos=pt_one)
|
|
online_proj_two = self.online_encoder(img=image_two, pos=pt_two)
|
|
|
|
online_pred_one = self.online_predictor(online_proj_one)
|
|
online_pred_two = self.online_predictor(online_proj_two)
|
|
|
|
with torch.no_grad():
|
|
target_encoder = self._get_target_encoder() if self.use_momentum else self.online_encoder
|
|
target_proj_one = target_encoder(img=image_one, pos=pt_one).detach()
|
|
target_proj_two = target_encoder(img=image_two, pos=pt_two).detach()
|
|
return online_pred_one, online_pred_two, target_proj_one, target_proj_two
|
|
|
|
def forward_normal(self, image):
|
|
online_pred_one, online_pred_two, target_proj_one, target_proj_two = self.get_predictions_and_projections(image)
|
|
|
|
loss_one = loss_fn(online_pred_one, target_proj_two.detach())
|
|
loss_two = loss_fn(online_pred_two, target_proj_one.detach())
|
|
|
|
loss = loss_one + loss_two
|
|
return loss.mean()
|
|
|
|
def forward_contrastive(self, image):
|
|
online_pred_one_1, online_pred_two_1, target_proj_one_1, target_proj_two_1 = self.get_predictions_and_projections(image)
|
|
loss_one = loss_fn(online_pred_one_1, target_proj_two_1.detach())
|
|
loss_two = loss_fn(online_pred_two_1, target_proj_one_1.detach())
|
|
loss = loss_one + loss_two
|
|
|
|
online_pred_one_2, online_pred_two_2, target_proj_one_2, target_proj_two_2 = self.get_predictions_and_projections(image)
|
|
loss_one = loss_fn(online_pred_one_2, target_proj_two_2.detach())
|
|
loss_two = loss_fn(online_pred_two_2, target_proj_one_2.detach())
|
|
loss = (loss + loss_one + loss_two).mean()
|
|
|
|
contrastive_loss = torch.cat([loss_fn(online_pred_one_1, target_proj_two_2),
|
|
loss_fn(online_pred_two_1, target_proj_one_2),
|
|
loss_fn(online_pred_one_2, target_proj_two_1),
|
|
loss_fn(online_pred_two_2, target_proj_one_1)], dim=0)
|
|
k = contrastive_loss.shape[0] // 2 # Take half of the total contrastive loss predictions.
|
|
contrastive_loss = torch.topk(contrastive_loss, k, dim=0).values.mean()
|
|
|
|
self.logs_loss = loss.detach()
|
|
self.logs_closs = contrastive_loss.detach()
|
|
|
|
return loss - contrastive_los00s
|
|
|
|
def forward(self, image):
|
|
if self.contrastive:
|
|
return self.forward_contrastive(image)
|
|
else:
|
|
return self.forward_normal(image)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
pa = PointwiseAugmentor(256)
|
|
for j in range(100):
|
|
t = ToTensor()(Image.open('E:\\4k6k\\datasets\\ns_images\\imagesets\\000001_152761.jpg')).unsqueeze(0).repeat(8,1,1,1)
|
|
p = torch.randint(50,180,(2,))
|
|
augmented, dp = pa(t, p)
|
|
t, p = pa(t, p)
|
|
t[:,:,p[0]-3:p[0]+3,p[1]-3:p[1]+3] = 0
|
|
torchvision.utils.save_image(t, f"{j}_src.png")
|
|
augmented[:,:,dp[0]-3:dp[0]+3,dp[1]-3:dp[1]+3] = 0
|
|
torchvision.utils.save_image(augmented, f"{j}_dst.png")
|
|
|
|
|
|
@register_model
|
|
def register_pixel_local_byol(opt_net, opt):
|
|
subnet = create_model(opt, opt_net['subnet'])
|
|
return BYOL(subnet, opt_net['image_size'], opt_net['hidden_layer'], contrastive=opt_net['contrastive']) |