AntoineH/smart_augmentation
1
0
Fork 0
mirror of https://github.com/AntoineHX/smart_augmentation.git synced 2025-05-04 12:10:45 +02:00
Code Issues Projects Releases Packages Wiki Activity
smart_augmentation/higher/smart_aug/dataug.py
Antoine Harlé b89dac9084 Changes since Teledyne
2024-08-20 11:53:35 +02:00

1263 lines
No EOL
53 KiB
Python
Executable file
Raw Blame History

""" Data augmentation modules.
Features a custom implementaiton of RandAugment (RandAug), as well as a data augmentation modules allowing gradient propagation.
Typical usage:
aug_model = Augmented_model(Data_AugV5, model)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.distributions import *
#import kornia
#import random
import numpy as np
import copy
import transformations as TF
import torchvision
import higher
import higher_patch
from utils import clip_norm
from train_utils import compute_vaLoss
from datasets import MEAN, STD
norm = TF.Normalizer(MEAN, STD)
### Data augmenter ###
class Data_augV5(nn.Module): #Optimisation jointe (mag, proba)
"""Data augmentation module with learnable parameters.
Applies transformations (TF) to batch of data.
Each TF is defined by a (name, probability of application, magnitude of distorsion) tuple which can be learned. For the full definiton of the TF, see transformations.py.
The TF probabilities defines a distribution from which we sample the TF applied.
Be warry, that the order of sequential application of TF is not taken into account. See Data_augV7.
Attributes:
_data_augmentation (bool): Wether TF will be applied during forward pass.
_TF_dict (dict) : A dictionnary containing the data transformations (TF) to be applied.
_TF (list) : List of TF names.
_TF_ignore_mag (set): TF for which magnitude should be ignored (either it's fixed or unused).
_nb_tf (int) : Number of TF used.
_N_seqTF (int) : Number of TF to be applied sequentially to each inputs
_shared_mag (bool) : Wether to share a single magnitude parameters for all TF. Beware using shared mag with basic color TF as their lowest magnitude is at PARAMETER_MAX/2.
_fixed_mag (bool): Wether to lock the TF magnitudes.
_fixed_prob (bool): Wether to lock the TF probabilies.
_samples (list): Sampled TF index during last forward pass.
_temp (bool): Wether we use a mix of an uniform distribution and the real distribution (TF probabilites). If False, only a uniform distribution is used.
_fixed_temp (bool): Wether we lock the mix distribution factor.
_params (nn.ParameterDict): Learnable parameters.
_reg_tgt (Tensor): Target for the magnitude regularisation. Only used when _fixed_mag is set to false (ie. we learn the magnitudes).
_reg_mask (list): Mask selecting the TF considered for the regularisation.
"""
def __init__(self, TF_dict, N_TF=1, temp=0.5, fixed_prob=False, fixed_mag=True, shared_mag=True, TF_ignore_mag=TF.TF_ignore_mag):
"""Init Data_augv5.
Args:
TF_dict (dict): A dictionnary containing the data transformations (TF) to be applied. (default: use all available TF from transformations.py)
N_TF (int): Number of TF to be applied sequentially to each inputs. (default: 1)
temp (float): Proportion [0.0, 1.0] of the real distribution used for sampling/selection of the TF. Distribution = (1-temp)*Uniform_distribution + temp*Real_distribution. If None is given, try to learn this parameter. (default: 0.5)
fixed_prob (bool): Wether to lock the TF probabilies. (default: False)
fixed_mag (bool): Wether to lock the TF magnitudes. (default: True)
shared_mag (bool): Wether to share a single magnitude parameters for all TF. (default: True)
TF_ignore_mag (set): TF for which magnitude should be ignored (either it's fixed or unused).
"""
super(Data_augV5, self).__init__()
assert len(TF_dict)>0
assert N_TF>=0
self._data_augmentation = True
#TF
self._TF_dict = TF_dict
self._TF= list(self._TF_dict.keys())
self._TF_ignore_mag=TF_ignore_mag
self._nb_tf= len(self._TF)
self._N_seqTF = N_TF
#Mag
self._shared_mag = shared_mag
self._fixed_mag = fixed_mag
if not self._fixed_mag and len([tf for tf in self._TF if tf not in self._TF_ignore_mag])==0:
print("WARNING: Mag would be fixed as current TF doesn't allow gradient propagation:",self._TF)
self._fixed_mag=True
#Distribution
self._fixed_prob=fixed_prob
self._samples = []
# self._temp = False
# if temp != 0.0: #Mix dist
# self._temp = True
self._fixed_temp=True
if temp is None: #Learn Temp
print("WARNING: Learning Temperature parameter isn't working with this version (No grad)")
self._fixed_temp = False
temp=0.5
#Params
init_mag = float(TF.PARAMETER_MAX) if self._fixed_mag else float(TF.PARAMETER_MAX)/2
self._params = nn.ParameterDict({
#"prob": nn.Parameter(torch.ones(self._nb_tf)/self._nb_tf), #Distribution prob uniforme
"prob": nn.Parameter(torch.ones(self._nb_tf)),
"mag" : nn.Parameter(torch.tensor(init_mag) if self._shared_mag
else torch.tensor(init_mag).repeat(self._nb_tf)), #[0, PARAMETER_MAX]
"temp": nn.Parameter(torch.tensor(temp))#.clamp(min=0.0,max=0.999))
})
if not self._shared_mag:
for tf in self._TF_ignore_mag :
self._params['mag'].data[self._TF.index(tf)]=float(TF.PARAMETER_MAX) #TF fixe a max parameter
#for t in TF.TF_no_mag: self._params['mag'][self._TF.index(t)].data-=self._params['mag'][self._TF.index(t)].data #Mag inutile pour les TF ignore_mag
#Mag regularisation
if not self._fixed_mag:
if self._shared_mag :
self._reg_tgt = torch.tensor(TF.PARAMETER_MAX, dtype=torch.float) #Encourage amplitude max
else:
TF_mag=[t for t in self._TF if t not in self._TF_ignore_mag] #TF w/ differentiable mag
self._reg_mask=[self._TF.index(t) for t in TF_mag]
self._reg_tgt=torch.full(size=(len(self._reg_mask),), fill_value=TF.PARAMETER_MAX, dtype=self._params['mag'].dtype) #Encourage amplitude max
#Prevent Identity
#print(TF.TF_identity)
#self._reg_tgt=torch.full(size=(len(self._reg_mask),), fill_value=0.0)
#for val in TF.TF_identity.keys():
# idx=[self._reg_mask.index(self._TF.index(t)) for t in TF_mag if t in TF.TF_identity[val]]
# self._reg_tgt[idx]=val
#print(TF_mag, self._reg_tgt)
def forward(self, x):
""" Main method of the Data augmentation module.
Args:
x (Tensor): Batch of data.
Returns:
Tensor : Batch of tranformed data.
"""
self._samples = torch.Tensor([])
if self._data_augmentation:
device = x.device
batch_size, h, w = x.shape[0], x.shape[2], x.shape[3]
# x = copy.deepcopy(x) #Evite de modifier les echantillons par reference (Problematique pour des utilisations paralleles)
## Echantillonage ##
temp = self._params["temp"].detach() if self._fixed_temp else self._params["temp"]
prob = self._params["prob"].detach() if self._fixed_prob else self._params["prob"]
self._distrib = F.softmax(prob*temp, dim=0)
# prob = F.softmax(prob[1:], dim=0) #Bernouilli
cat_distrib= Categorical(probs=torch.ones((batch_size, self._nb_tf), device=device)*self._distrib)
self._samples=cat_distrib.sample([self._N_seqTF])
#Bernoulli (Requiert Identité en position 0)
#assert(self._TF[0]=="Identity")
# cat_distrib= Categorical(probs=torch.ones((batch_size, self._nb_tf-1), device=device)*self._distrib)
# bern_distrib = Bernoulli(torch.tensor([0.5], device=device))
# mask = bern_distrib.sample([self._N_seqTF, batch_size]).squeeze()
# self._samples=(cat_distrib.sample([self._N_seqTF])+1)*mask
for sample in self._samples:
## Transformations ##
x = self.apply_TF(x, sample)
# self._samples.to(device)
# for n in range(self._N_seqTF):
# # print('temp', (temp+0.3*n))
# self._distrib = F.softmax(prob*(temp+0.2*n), dim=0)
# # prob = F.softmax(prob[1:], dim=0) #Bernouilli
# cat_distrib= Categorical(probs=torch.ones((batch_size, self._nb_tf), device=device)*self._distrib)
# new_sample=cat_distrib.sample()
# self._samples=torch.cat((self._samples.to(device).to(new_sample.dtype), new_sample.unsqueeze(dim=0)), dim=0)
# x = self.apply_TF(x, new_sample)
# print('sample',self._samples.shape)
return x
def apply_TF(self, x, sampled_TF):
""" Applies the sampled transformations.
Args:
x (Tensor): Batch of data.
sampled_TF (Tensor): Indexes of the TF to be applied to each element of data.
Returns:
Tensor: Batch of tranformed data.
"""
device = x.device
batch_size, channels, h, w = x.shape
smps_x=[]
for tf_idx in range(self._nb_tf):
mask = sampled_TF==tf_idx #Create selection mask
smp_x = x[mask] #torch.masked_select() ? (Necessite d'expand le mask au meme dim)
if smp_x.shape[0]!=0: #if there's data to TF
magnitude=self._params["mag"] if self._shared_mag else self._params["mag"][tf_idx]
if self._fixed_mag: magnitude=magnitude.detach() #Fmodel tente systematiquement de tracker les gradient de tout les param
tf=self._TF[tf_idx]
#In place
#x[mask]=self._TF_dict[tf](x=smp_x, mag=magnitude)
#Out of place
smp_x = self._TF_dict[tf](x=smp_x, mag=magnitude)
idx= mask.nonzero()
idx= idx.expand(-1,channels).unsqueeze(dim=2).expand(-1,channels, h).unsqueeze(dim=3).expand(-1,channels, h, w) #Il y a forcement plus simple ...
x=x.scatter(dim=0, index=idx, src=smp_x)
return x
def adjust_param(self, soft=False): #Detach from gradient ?
""" Enforce limitations to the learned parameters.
Ensure that the parameters value stays in the right intevals. This should be called after each update of those parameters.
Args:
soft (bool): Wether to use a softmax function for TF probabilites. Tends to lock the probabilities if the learning rate is low, preventing them to be learned. (default: False)
"""
# if not self._fixed_prob:
# if soft :
# self._params['prob'].data=F.softmax(self._params['prob'].data, dim=0)
# else:
# self._params['prob'].data = self._params['prob'].data.clamp(min=1/(self._nb_tf*100),max=1.0)
# self._params['prob'].data = self._params['prob']/sum(self._params['prob']) #Contrainte sum(p)=1
if not self._fixed_mag:
self._params['mag'].data = self._params['mag'].data.clamp(min=TF.PARAMETER_MIN, max=TF.PARAMETER_MAX)
if not self._fixed_temp:
self._params['temp'].data = self._params['temp'].data.clamp(min=0.0, max=0.999)
def loss_weight(self, batch_norm=True):
""" Weights for the loss.
Compute the weights for the loss of each inputs depending on wich TF was applied to them.
Should be applied to the loss before reduction.
Do not take into account the order of application of the TF. See Data_augV7.
Args:
batch_norm (bool): Wether to normalize mean of the weights. (Default: True)
Returns:
Tensor : Loss weights.
"""
if len(self._samples)==0 : return torch.tensor(1, device=self._params["prob"].device) #Pas d'echantillon = pas de ponderation
prob = self._params["prob"].detach() if self._fixed_prob else self._params["prob"]
#prob = F.softmax(prob, dim=0)
#Plusieurs TF sequentielles (Attention ne prend pas en compte ordre !)
w_loss = torch.zeros((self._samples[0].shape[0],self._nb_tf), device=self._samples[0].device)
for sample in self._samples.to(torch.long):
tmp_w = torch.zeros(w_loss.size(),device=w_loss.device)
tmp_w.scatter_(dim=1, index=sample.view(-1,1), value=1/self._N_seqTF)
w_loss += tmp_w
#w_loss=w_loss/w_loss.sum(dim=1, keepdim=True) #Bernoulli
#Normalizing by mean, would lend an exact normalization but can lead to unstable behavior of probabilities.
w_loss = w_loss * prob
w_loss = torch.sum(w_loss,dim=1)
if batch_norm:
w_min = w_loss.min()
w_loss = w_loss-w_min if w_min<0 else w_loss
w_loss = w_loss/w_loss.mean() #mean(w_loss)=1
#Normalizing by distribution is a statistical approximation of the exact normalization. It lead to more smooth probabilities evolution but will only return 1 if temp=1.
# w_loss = w_loss * prob/self._distrib #Ponderation par les proba (divisee par la distrib pour pas diminuer la loss)
# w_loss = torch.sum(w_loss,dim=1)
return w_loss
def reg_loss(self, reg_factor=0.005):
""" Regularisation term used to learn the magnitudes.
Use an L2 loss to encourage high magnitudes TF.
Args:
reg_factor (float): Factor by wich the regularisation loss is multiplied. (default: 0.005)
Returns:
Tensor containing the regularisation loss value.
"""
if self._fixed_mag or self._fixed_prob: #Not enough DOF
return torch.tensor(0)
else:
#return reg_factor * F.l1_loss(self._params['mag'][self._reg_mask], target=self._reg_tgt, reduction='mean')
mags = self._params['mag'] if self._params['mag'].shape==torch.Size([]) else self._params['mag'][self._reg_mask]
max_mag_reg = reg_factor * F.mse_loss(mags, target=self._reg_tgt.to(mags.device), reduction='mean') #Close to target ?
#max_mag_reg = - reg_factor * F.mse_loss(mags, target=self._reg_tgt.to(mags.device), reduction='mean') #Far from target ?
return max_mag_reg
def train(self, mode=True):
""" Set the module training mode.
Args:
mode (bool): Wether to learn the parameter of the module. None would not change mode. (default: None)
"""
#if mode is None :
# mode=self._data_augmentation
self.augment(mode=mode) #Inutile si mode=None
super(Data_augV5, self).train(mode)
return self
def eval(self):
""" Set the module to evaluation mode.
"""
return self.train(mode=False)
def augment(self, mode=True):
""" Set the augmentation mode.
Args:
mode (bool): Wether to perform data augmentation on the forward pass. (default: True)
"""
self._data_augmentation=mode
def is_augmenting(self):
""" Return wether data augmentation is applied.
Returns:
bool : True if data augmentation is applied.
"""
return self._data_augmentation
def __getitem__(self, key):
"""Access to the learnable parameters
Args:
key (string): Name of the learnable parameter to access.
Returns:
nn.Parameter.
"""
if key == 'prob': #Override prob access
return F.softmax(self._params["prob"]*self._params["temp"], dim=0)
return self._params[key]
def __str__(self):
"""Name of the module
Returns:
String containing the name of the module as well as the higher levels parameters.
"""
dist_param=''
if self._fixed_prob: dist_param+='Fx'
mag_param='Mag'
if self._fixed_mag: mag_param+= 'Fx'
if self._shared_mag: mag_param+= 'Sh'
# if not self._temp:
# return "Data_augV5(Uniform%s-%dTFx%d-%s)" % (dist_param, self._nb_tf, self._N_seqTF, mag_param)
if self._fixed_temp:
return "Data_augV5(T%.1f%s-%dTFx%d-%s)" % (self._params['temp'].item(),dist_param, self._nb_tf, self._N_seqTF, mag_param)
else:
return "Data_augV5(T%s-%dTFx%d-%s)" % (dist_param, self._nb_tf, self._N_seqTF, mag_param)
class Data_augV8(nn.Module): #Apprentissage proba sequentielles
"""Data augmentation module with learnable parameters.
Applies transformations (TF) to batch of data.
Each TF is defined by a (name, probability of application, magnitude of distorsion) tuple which can be learned. For the full definiton of the TF, see transformations.py.
The TF probabilities defines a distribution from which we sample the TF applied.
Attributes:
_data_augmentation (bool): Wether TF will be applied during forward pass.
_TF_dict (dict) : A dictionnary containing the data transformations (TF) to be applied.
_TF (list) : List of TF names.
_TF_ignore_mag (set): TF for which magnitude should be ignored (either it's fixed or unused).
_nb_tf (int) : Number of TF used.
_N_seqTF (int) : Number of TF to be applied sequentially to each inputs
_shared_mag (bool) : Wether to share a single magnitude parameters for all TF. Beware using shared mag with basic color TF as their lowest magnitude is at PARAMETER_MAX/2.
_fixed_mag (bool): Wether to lock the TF magnitudes.
_fixed_prob (bool): Wether to lock the TF probabilies.
_samples (list): Sampled TF index during last forward pass.
_temp (bool): Wether we use a mix of an uniform distribution and the real distribution (TF probabilites). If False, only a uniform distribution is used.
_fixed_temp (bool): Wether we lock the mix distribution factor.
_params (nn.ParameterDict): Learnable parameters.
_reg_tgt (Tensor): Target for the magnitude regularisation. Only used when _fixed_mag is set to false (ie. we learn the magnitudes).
_reg_mask (list): Mask selecting the TF considered for the regularisation.
"""
def __init__(self, TF_dict, N_TF=1, temp=0.5, fixed_prob=False, fixed_mag=True, shared_mag=True, TF_ignore_mag=TF.TF_ignore_mag):
"""Init Data_augv8.
Args:
TF_dict (dict): A dictionnary containing the data transformations (TF) to be applied. (default: use all available TF from transformations.py)
N_TF (int): Number of TF to be applied sequentially to each inputs. (default: 1)
temp (float): Proportion [0.0, 1.0] of the real distribution used for sampling/selection of the TF. Distribution = (1-temp)*Uniform_distribution + temp*Real_distribution. If None is given, try to learn this parameter. (default: 0.5)
fixed_prob (bool): Wether to lock the TF probabilies. (default: False)
fixed_mag (bool): Wether to lock the TF magnitudes. (default: True)
shared_mag (bool): Wether to share a single magnitude parameters for all TF. (default: True)
TF_ignore_mag (set): TF for which magnitude should be ignored (either it's fixed or unused).
"""
super(Data_augV8, self).__init__()
assert len(TF_dict)>0
assert N_TF>=0
self._data_augmentation = True
#TF
self._TF_dict = TF_dict
self._TF= list(self._TF_dict.keys())
self._TF_ignore_mag=TF_ignore_mag
self._nb_tf= len(self._TF)
self._N_seqTF = N_TF
#Mag
self._shared_mag = shared_mag
self._fixed_mag = fixed_mag
if not self._fixed_mag and len([tf for tf in self._TF if tf not in self._TF_ignore_mag])==0:
print("WARNING: Mag would be fixed as current TF doesn't allow gradient propagation:",self._TF)
self._fixed_mag=True
#Distribution
self._fixed_prob=fixed_prob
self._samples = []
# self._temp = False
# if temp != 0.0: #Mix dist
# self._temp = True
self._fixed_temp=True
if temp is None: #Learn temp
print("WARNING: Learning Temperature parameter isn't working with this version (No grad)")
self._fixed_temp = False
temp=0.5
#Params
init_mag = float(TF.PARAMETER_MAX) if self._fixed_mag else float(TF.PARAMETER_MAX)/2
self._params = nn.ParameterDict({
#"prob": nn.Parameter(torch.ones(self._nb_tf)/self._nb_tf), #Distribution prob uniforme
# "prob": nn.Parameter(torch.ones([self._nb_tf for _ in range(self._N_seqTF)])),
"prob": nn.Parameter(torch.ones(self._nb_tf**self._N_seqTF)),
"mag" : nn.Parameter(torch.tensor(init_mag) if self._shared_mag
else torch.tensor(init_mag).repeat(self._nb_tf)), #[0, PARAMETER_MAX]
"temp": nn.Parameter(torch.tensor(temp))#.clamp(min=0.0,max=0.999))
})
self._prob_mem=torch.zeros(self._nb_tf**self._N_seqTF)
if not self._shared_mag:
for tf in self._TF_ignore_mag :
self._params['mag'].data[self._TF.index(tf)]=float(TF.PARAMETER_MAX) #TF fixe a max parameter
#for t in TF.TF_no_mag: self._params['mag'][self._TF.index(t)].data-=self._params['mag'][self._TF.index(t)].data #Mag inutile pour les TF ignore_mag
#Mag regularisation
if not self._fixed_mag:
if self._shared_mag :
self._reg_tgt = torch.tensor(TF.PARAMETER_MAX, dtype=torch.float) #Encourage amplitude max
else:
TF_mag=[t for t in self._TF if t not in self._TF_ignore_mag] #TF w/ differentiable mag
self._reg_mask=[self._TF.index(t) for t in TF_mag]
self._reg_tgt=torch.full(size=(len(self._reg_mask),), fill_value=TF.PARAMETER_MAX, dtype=self._params['mag'].dtype) #Encourage amplitude max
#Prevent Identity
#print(TF.TF_identity)
#self._reg_tgt=torch.full(size=(len(self._reg_mask),), fill_value=0.0)
#for val in TF.TF_identity.keys():
# idx=[self._reg_mask.index(self._TF.index(t)) for t in TF_mag if t in TF.TF_identity[val]]
# self._reg_tgt[idx]=val
#print(TF_mag, self._reg_tgt)
def forward(self, x):
""" Main method of the Data augmentation module.
Args:
x (Tensor): Batch of data.
Returns:
Tensor : Batch of tranformed data.
"""
self._samples = torch.Tensor([])
if self._data_augmentation:
device = x.device
batch_size, h, w = x.shape[0], x.shape[2], x.shape[3]
# x = copy.deepcopy(x) #Evite de modifier les echantillons par reference (Problematique pour des utilisations paralleles)
## Echantillonage ##
# if not self._temp:
# self._distrib = torch.ones(1,self._nb_tf**self._N_seqTF,device=device).softmax(dim=1)
# else:
# prob = self._params["prob"].detach() if self._fixed_prob else self._params["prob"] #Uniform dist
# # print(prob.shape)
# # prob = prob.view(1, -1)
# # prob = F.softmax(prob, dim=0)
# temp = self._params["temp"].detach() if self._fixed_temp else self._params["temp"] #Temperature
# self._distrib = F.softmax(temp*prob, dim=0)
# # self._distrib = (temp*prob+(1-temp)*uniforme_dist)#.softmax(dim=1) #Mix distrib reel / uniforme avec mix_factor
# # print(prob.shape)
prob = self._params["prob"].detach() if self._fixed_prob else self._params["prob"]
temp = self._params["temp"].detach() if self._fixed_temp else self._params["temp"] #Temperature
self._distrib = F.softmax(temp*prob, dim=0)
cat_distrib= Categorical(probs=torch.ones((self._nb_tf**self._N_seqTF), device=device)*self._distrib)
samples=cat_distrib.sample([batch_size]) # (batch_size)
# print(samples.shape)
samples=torch.zeros((batch_size, self._nb_tf**self._N_seqTF), dtype=torch.bool, device=device).scatter_(dim=1, index=samples.unsqueeze(dim=1), value=1)
self._samples=samples
# print(samples.shape)
# print(samples)
samples=samples.view((batch_size,)+tuple([self._nb_tf for _ in range(self._N_seqTF)]))
# print(samples.shape)
# print(samples)
samples= torch.nonzero(samples)[:,1:].T #Find indexes (TF sequence) => (N_seqTF, batch_size)
# print(samples.shape)
#Bernoulli (Requiert Identité en position 0)
#assert(self._TF[0]=="Identity")
# cat_distrib= Categorical(probs=torch.ones((batch_size, self._nb_tf-1), device=device)*self._distrib)
# bern_distrib = Bernoulli(torch.tensor([0.5], device=device))
# mask = bern_distrib.sample([self._N_seqTF, batch_size]).squeeze()
# self._samples=(cat_distrib.sample([self._N_seqTF])+1)*mask
for sample in samples:
## Transformations ##
x = self.apply_TF(x, sample)
return x
def apply_TF(self, x, sampled_TF):
""" Applies the sampled transformations.
Args:
x (Tensor): Batch of data.
sampled_TF (Tensor): Indexes of the TF to be applied to each element of data.
Returns:
Tensor: Batch of tranformed data.
"""
device = x.device
batch_size, channels, h, w = x.shape
smps_x=[]
for tf_idx in range(self._nb_tf):
mask = sampled_TF==tf_idx #Create selection mask
smp_x = x[mask] #torch.masked_select() ? (Necessite d'expand le mask au meme dim)
if smp_x.shape[0]!=0: #if there's data to TF
magnitude=self._params["mag"] if self._shared_mag else self._params["mag"][tf_idx]
if self._fixed_mag: magnitude=magnitude.detach() #Fmodel tente systematiquement de tracker les gradient de tout les param
tf=self._TF[tf_idx]
#In place
#x[mask]=self._TF_dict[tf](x=smp_x, mag=magnitude)
#Out of place
smp_x = self._TF_dict[tf](x=smp_x, mag=magnitude)
idx= mask.nonzero()
idx= idx.expand(-1,channels).unsqueeze(dim=2).expand(-1,channels, h).unsqueeze(dim=3).expand(-1,channels, h, w) #Il y a forcement plus simple ...
x=x.scatter(dim=0, index=idx, src=smp_x)
return x
def adjust_param(self, soft=False): #Detach from gradient ?
""" Enforce limitations to the learned parameters.
Ensure that the parameters value stays in the right intevals. This should be called after each update of those parameters.
Args:
soft (bool): Wether to use a softmax function for TF probabilites. Tends to lock the probabilities if the learning rate is low, preventing them to be learned. (default: False)
"""
# if not self._fixed_prob:
# if soft :
# self._params['prob'].data=F.softmax(self._params['prob'].data, dim=0)
# else:
# self._params['prob'].data = self._params['prob'].data.clamp(min=1/(self._nb_tf*100),max=1.0)
# self._params['prob'].data = self._params['prob']/sum(self._params['prob']) #Contrainte sum(p)=1
if not self._fixed_mag:
self._params['mag'].data = self._params['mag'].data.clamp(min=TF.PARAMETER_MIN, max=TF.PARAMETER_MAX)
if not self._fixed_temp:
self._params['temp'].data = self._params['temp'].data.clamp(min=0.0, max=0.999)
def loss_weight(self, batch_norm=True):
""" Weights for the loss.
Compute the weights for the loss of each inputs depending on wich TF was applied to them.
Should be applied to the loss before reduction.
Args:
batch_norm (bool): Wether to normalize mean of the weights. (Default: True)
Returns:
Tensor : Loss weights.
"""
device=self._params["prob"].device
if len(self._samples)==0 : return torch.tensor(1, device=device) #Pas d'echantillon = pas de ponderation
prob = self._params["prob"].detach() if self._fixed_prob else self._params["prob"]
# print("prob",prob.shape)
# print(self._samples.shape)
#w_loss=w_loss/w_loss.sum(dim=1, keepdim=True) #Bernoulli
#Normalizing by mean, would lend an exact normalization but can lead to unstable behavior of probabilities.
w_loss = self._samples * prob
w_loss = torch.sum(w_loss,dim=1)
# print("W_loss",w_loss.shape)
# print(w_loss)
if batch_norm:
w_min = w_loss.min()
w_loss = w_loss-w_min if w_min<0 else w_loss
w_loss = w_loss/w_loss.mean() #mean(w_loss)=1
#Normalizing by distribution is a statistical approximation of the exact normalization. It lead to more smooth probabilities evolution but will only return 1 if temp=1.
# w_loss = w_loss * prob/self._distrib #Ponderation par les proba (divisee par la distrib pour pas diminuer la loss)
# w_loss = torch.sum(w_loss,dim=1)
return w_loss
def reg_loss(self, reg_factor=0.005):
""" Regularisation term used to learn the magnitudes.
Use an L2 loss to encourage high magnitudes TF.
Args:
reg_factor (float): Factor by wich the regularisation loss is multiplied. (default: 0.005)
Returns:
Tensor containing the regularisation loss value.
"""
if self._fixed_mag or self._fixed_prob: #Not enough DOF
return torch.tensor(0)
else:
#return reg_factor * F.l1_loss(self._params['mag'][self._reg_mask], target=self._reg_tgt, reduction='mean')
mags = self._params['mag'] if self._params['mag'].shape==torch.Size([]) else self._params['mag'][self._reg_mask]
max_mag_reg = reg_factor * F.mse_loss(mags, target=self._reg_tgt.to(mags.device), reduction='mean') #Close to target ?
#max_mag_reg = - reg_factor * F.mse_loss(mags, target=self._reg_tgt.to(mags.device), reduction='mean') #Far from target ?
return max_mag_reg
def TF_prob(self):
""" Gives an estimation of the individual TF probabilities.
Returms:
Tensor containing the single TF probabilities of applications.
"""
# if not torch.all(self._params['prob']==self._prob_mem.to(self._params['prob'].device)): #Prevent recompute if originial prob didn't changed
# self._prob_mem=self._params['prob'].data.detach_()
# p = self._params['prob'].view([self._nb_tf for _ in range(self._N_seqTF)])
# # print('prob',p)
# self._single_TF_prob=p.mean(dim=[i+1 for i in range(self._N_seqTF-1)]) #Reduce to 1D tensor
# # print(self._single_TF_prob)
# self._single_TF_prob=F.softmax(self._single_TF_prob, dim=0)
# print('Soft',self._single_TF_prob)
p=F.softmax(self._params['prob']*self._params["temp"], dim=0) #Sampling dist
p=p.view([self._nb_tf for _ in range(self._N_seqTF)])
p=p.mean(dim=[i+1 for i in range(self._N_seqTF-1)]) #Reduce to 1D tensor
#Means over each dim
# dim_idx=tuple(range(self._N_seqTF))
# means=[]
# for d in dim_idx:
# dim_mean=list(dim_idx)
# dim_mean.remove(d)
# means.append(p.mean(dim=dim_mean).unsqueeze(dim=1))
# means=torch.cat(means,dim=1)
# print(means)
# p=means.mean(dim=1)
# print(p)
return p
def train(self, mode=True):
""" Set the module training mode.
Args:
mode (bool): Wether to learn the parameter of the module. None would not change mode. (default: None)
"""
#if mode is None :
# mode=self._data_augmentation
self.augment(mode=mode) #Inutile si mode=None
super(Data_augV8, self).train(mode)
return self
def eval(self):
""" Set the module to evaluation mode.
"""
return self.train(mode=False)
def augment(self, mode=True):
""" Set the augmentation mode.
Args:
mode (bool): Wether to perform data augmentation on the forward pass. (default: True)
"""
self._data_augmentation=mode
def is_augmenting(self):
""" Return wether data augmentation is applied.
Returns:
bool : True if data augmentation is applied.
"""
return self._data_augmentation
def __getitem__(self, key):
"""Access to the learnable parameters
Args:
key (string): Name of the learnable parameter to access.
Returns:
nn.Parameter.
"""
if key == 'prob': #Override prob access
return self.TF_prob()
return self._params[key]
def __str__(self):
"""Name of the module
Returns:
String containing the name of the module as well as the higher levels parameters.
"""
dist_param=''
if self._fixed_prob: dist_param+='Fx'
mag_param='Mag'
if self._fixed_mag: mag_param+= 'Fx'
if self._shared_mag: mag_param+= 'Sh'
# if not self._temp:
# return "Data_augV8(Uniform%s-%dTFx%d-%s)" % (dist_param, self._nb_tf, self._N_seqTF, mag_param)
if self._fixed_temp:
return "Data_augV8(T%.1f%s-%dTFx%d-%s)" % (self._params['temp'].item(),dist_param, self._nb_tf, self._N_seqTF, mag_param)
else:
return "Data_augV8(T%s-%dTFx%d-%s)" % (dist_param, self._nb_tf, self._N_seqTF, mag_param)
class RandAug(nn.Module): #RandAugment = UniformFx-MagFxSh + rapide
"""RandAugment implementation.
Applies transformations (TF) to batch of data.
Each TF is defined by a (name, probability of application, magnitude of distorsion) tuple. For the full definiton of the TF, see transformations.py.
The TF probabilities are ignored and, instead selected randomly.
Attributes:
_data_augmentation (bool): Wether TF will be applied during forward pass.
_TF_dict (dict) : A dictionnary containing the data transformations (TF) to be applied.
_TF (list) : List of TF names.
_nb_tf (int) : Number of TF used.
_N_seqTF (int) : Number of TF to be applied sequentially to each inputs
_shared_mag (bool) : Wether to share a single magnitude parameters for all TF. Should be True.
_fixed_mag (bool): Wether to lock the TF magnitudes. Should be True.
_params (nn.ParameterDict): Data augmentation parameters.
"""
def __init__(self, TF_dict, N_TF=1, mag=TF.PARAMETER_MAX):
"""Init RandAug.
Args:
TF_dict (dict): A dictionnary containing the data transformations (TF) to be applied. (default: use all available TF from transformations.py)
N_TF (int): Number of TF to be applied sequentially to each inputs. (default: 1)
mag (float): Magnitude of the TF. Should be between [PARAMETER_MIN, PARAMETER_MAX] defined in transformations.py. (default: PARAMETER_MAX)
"""
super(RandAug, self).__init__()
self._data_augmentation = True
self._TF_dict = TF_dict
self._TF= list(self._TF_dict.keys())
self._nb_tf= len(self._TF)
self._N_seqTF = N_TF
self.mag=nn.Parameter(torch.tensor(float(mag)))
self._params = nn.ParameterDict({
"prob": nn.Parameter(torch.ones(self._nb_tf)/self._nb_tf), #Ignored
"mag" : nn.Parameter(torch.tensor(float(mag))),
})
self._shared_mag = True
self._fixed_mag = True
self._fixed_prob=True
self._fixed_temp=True
self._params['mag'].data = self._params['mag'].data.clamp(min=TF.PARAMETER_MIN, max=TF.PARAMETER_MAX)
def forward(self, x):
""" Main method of the Data augmentation module.
Args:
x (Tensor): Batch of data.
Returns:
Tensor : Batch of tranformed data.
"""
if self._data_augmentation:
device = x.device
batch_size, h, w = x.shape[0], x.shape[2], x.shape[3]
# x = copy.deepcopy(x) #Evite de modifier les echantillons par reference (Problematique pour des utilisations paralleles)
## Echantillonage ## == sampled_ops = np.random.choice(transforms, N)
uniforme_dist = torch.ones(1,self._nb_tf,device=device).softmax(dim=1)
cat_distrib= Categorical(probs=torch.ones((batch_size, self._nb_tf), device=device)*uniforme_dist)
self._samples=cat_distrib.sample([self._N_seqTF])
#Bernoulli (Requiert Identité en position 0)
# uniforme_dist = torch.ones(1,self._nb_tf-1,device=device).softmax(dim=1)
# cat_distrib= Categorical(probs=torch.ones((batch_size, self._nb_tf-1), device=device)*uniforme_dist)
# bern_distrib = Bernoulli(torch.tensor([0.5], device=device))
# mask = bern_distrib.sample([self._N_seqTF, batch_size]).squeeze()
# self._samples=(cat_distrib.sample([self._N_seqTF])+1)*mask
for sample in self._samples:
## Transformations ##
x = self.apply_TF(x, sample)
return x
def apply_TF(self, x, sampled_TF):
""" Applies the sampled transformations.
Args:
x (Tensor): Batch of data.
sampled_TF (Tensor): Indexes of the TF to be applied to each element of data.
Returns:
Tensor: Batch of tranformed data.
"""
smps_x=[]
for tf_idx in range(self._nb_tf):
mask = sampled_TF==tf_idx #Create selection mask
smp_x = x[mask] #torch.masked_select() ? (NEcessite d'expand le mask au meme dim)
if smp_x.shape[0]!=0: #if there's data to TF
magnitude=self._params["mag"].detach()
tf=self._TF[tf_idx]
#print(magnitude)
#In place
x[mask]=self._TF_dict[tf](x=smp_x, mag=magnitude)
return x
def adjust_param(self, soft=False):
"""Not used
"""
pass #Pas de parametre a opti
def loss_weight(self, batch_norm=False):
"""Not used
"""
return torch.tensor([1], device=self._params["prob"].device) #Pas d'echantillon = pas de ponderation
def reg_loss(self, reg_factor=0.005):
"""Not used
"""
return torch.tensor(0) #Pas de regularisation
def train(self, mode=None):
""" Set the module training mode.
Args:
mode (bool): Wether to learn the parameter of the module. None would not change mode. (default: None)
"""
if mode is None :
mode=self._data_augmentation
self.augment(mode=mode) #Inutile si mode=None
super(RandAug, self).train(mode)
def eval(self):
""" Set the module to evaluation mode.
"""
self.train(mode=False)
def augment(self, mode=True):
""" Set the augmentation mode.
Args:
mode (bool): Wether to perform data augmentation on the forward pass. (default: True)
"""
self._data_augmentation=mode
def is_augmenting(self):
""" Return wether data augmentation is applied.
Returns:
bool : True if data augmentation is applied.
"""
return self._data_augmentation
def __getitem__(self, key):
"""Access to the learnable parameters
Args:
key (string): Name of the learnable parameter to access.
Returns:
nn.Parameter.
"""
return self._params[key]
def __str__(self):
"""Name of the module
Returns:
String containing the name of the module as well as the higher levels parameters.
"""
return "RandAug(%dTFx%d-Mag%d)" % (self._nb_tf, self._N_seqTF, self.mag)
### Models ###
class Higher_model(nn.Module):
"""Model wrapper for higher gradient tracking.
Keep in memory the orginial model and it's functionnal, higher, version.
Might not be needed anymore if Higher implement detach for fmodel.
see : https://github.com/facebookresearch/higher
TODO: Get rid of the original model if not needed by user.
Attributes:
_name (string): Name of the model.
_mods (nn.ModuleDict): Models (Orginial and Higher version).
"""
def __init__(self, model, model_name=None):
"""Init Higher_model.
Args:
model (nn.Module): Network for which higher gradients can be tracked.
model_name (string): Model name. (Default: Class name of model)
"""
super(Higher_model, self).__init__()
self._name = model_name if model_name else model.__class__.__name__ #model.__str__()
self._mods = nn.ModuleDict({
'original': model,
'functional': higher.patch.monkeypatch(model, device=None, copy_initial_weights=True)
})
def get_diffopt(self, opt, grad_callback=None, track_higher_grads=True):
"""Get a differentiable version of an Optimizer.
Higher/Differentiable optimizer required to be used for higher gradient tracking.
Usage : diffopt.step(loss) == (opt.zero_grad, loss.backward, opt.step)
Be warry that if track_higher_grads is set to True, a new state of the model would be saved each time diffopt.step() is called.
Thus increasing memory consumption. The detach_() method should be called to reset the gradient tape and prevent memory saturation.
Args:
opt (torch.optim): Optimizer to make differentiable.
grad_callback (fct(grads)=grads): Function applied to the list of gradients parameters (ex: clipping). (default: None)
track_higher_grads (bool): Wether higher gradient are tracked. If True, the graph/states will be retained to allow backpropagation. (default: True)
Returns:
(Higher.DifferentiableOptimizer): Differentiable version of the optimizer.
"""
return higher.optim.get_diff_optim(opt,
self._mods['original'].parameters(),
fmodel=self._mods['functional'],
grad_callback=grad_callback,
track_higher_grads=track_higher_grads)
def forward(self, x):
""" Main method of the model.
Args:
x (Tensor): Batch of data.
Returns:
Tensor : Output of the network. Should be logits.
"""
return self._mods['functional'](x)
def detach_(self):
"""Detach from the graph.
Needed to limit the number of state kept in memory.
"""
tmp = self._mods['functional'].fast_params
self._mods['functional']._fast_params=[]
self._mods['functional'].update_params(tmp)
for p in self._mods['functional'].fast_params:
p.detach_().requires_grad_()
def state_dict(self):
"""Returns a dictionary containing a whole state of the module.
"""
return self._mods['functional'].state_dict()
def __getitem__(self, key):
"""Access to modules
Args:
key (string): Name of the module to access.
Returns:
nn.Module.
"""
return self._mods[key]
def __str__(self):
"""Name of the module
Returns:
String containing the name of the module.
"""
return self._name
class Augmented_model(nn.Module):
"""Wrapper for a Data Augmentation module and a model.
Attributes:
_mods (nn.ModuleDict): A dictionary containing the modules.
_data_augmentation (bool): Wether data augmentation should be used.
"""
def __init__(self, data_augmenter, model):
"""Init Augmented Model.
By default, data augmentation will be performed.
Args:
data_augmenter (nn.Module): Data augmentation module.
model (nn.Module): Network.
"""
super(Augmented_model, self).__init__()
self._mods = nn.ModuleDict({
'data_aug': data_augmenter,
'model': model
})
self.augment(mode=True)
def forward(self, x, copy=False):
""" Main method of the Augmented model.
Perform the forward pass of both modules.
Args:
x (Tensor): Batch of data.
copy (Bool): Wether to alter a copy or the original input. It's recommended to use a copy for parallel use of the input. (Default: False)
Returns:
Tensor : Output of the networks. Should be logits.
"""
if copy:
x = copy.deepcopy(x) #Evite de modifier les echantillons par reference (Problematique pour des utilisations paralleles)
return self._mods['model'](norm(self._mods['data_aug'](x)))
# return self._mods['model'](self._mods['data_aug'](x))
def augment(self, mode=True):
""" Set the augmentation mode.
Args:
mode (bool): Wether to perform data augmentation on the forward pass. (default: True)
"""
self._data_augmentation=mode
self._mods['data_aug'].augment(mode)
#ABN
# if mode :
# self._mods['model']['functional'].set_mode('augmented')
# else :
# self._mods['model']['functional'].set_mode('clean')
#### Encapsulation Meta Opt ####
def start_bilevel_opt(self, inner_it, hp_list, opt_param, dl_val):
""" Set up Augmented Model for bi-level optimisation.
Create and keep in Augmented Model the necessary objects for meta-optimisation.
This allow for an almost transparent use by just hiding the bi-level optimisation (see ''run_dist_dataugV3'') by ::
model.step(loss)
Does not support LR scheduler.
See ''run_simple_smartaug'' for a complete example.
Args:
inner_it (int): Number of inner iteration before a meta-step. 0 inner iteration means there's no meta-step.
hp_list (list): List of hyper-parameters to be learned.
opt_param (dict): Dictionnary containing optimizers parameters.
dl_val (DataLoader): Data loader of validation data.
"""
self._it_count=0
self._in_it=inner_it
self._opt_param=opt_param
#Inner Opt
inner_opt = torch.optim.SGD(self._mods['model']['original'].parameters(),
lr=opt_param['Inner']['lr'],
momentum=opt_param['Inner']['momentum'],
weight_decay=opt_param['Inner']['decay'],
nesterov=opt_param['Inner']['nesterov']) #lr=1e-2 / momentum=0.9
#Validation data
self._dl_val=dl_val
self._dl_val_it=iter(dl_val)
self._val_loss=0.
if inner_it==0 or len(hp_list)==0: #No meta-opt
print("No meta optimization")
#Inner Opt
self._diffopt = self._mods['model'].get_diffopt(
inner_opt,
grad_callback=(lambda grads: clip_norm(grads, max_norm=10)),
track_higher_grads=False)
self._meta_opt=None
else: #Bi-level opt
print("Bi-Level optimization")
#Inner Opt
self._diffopt = self._mods['model'].get_diffopt(
inner_opt,
grad_callback=(lambda grads: clip_norm(grads, max_norm=10)),
track_higher_grads=True)
#Meta Opt
self._meta_opt = torch.optim.Adam(hp_list, lr=opt_param['Meta']['lr'])
self._meta_opt.zero_grad()
def step(self, loss):
""" Perform a model update.
''start_bilevel_opt'' method needs to be called once before using this method.
Perform a step of inner optimization and, if needed, a step of meta optimization.
Replace ::
opt.zero_grad()
loss.backward()
opt.step()
val_loss=...
val_loss.backward()
meta_opt.step()
adjust_param()
detach()
meta_opt.zero_grad()
By ::
model.step(loss)
Args:
loss (Tensor): the training loss tensor.
"""
self._it_count+=1
self._diffopt.step(loss) #(opt.zero_grad, loss.backward, opt.step)
if(self._meta_opt and self._it_count>0 and self._it_count%self._in_it==0): #Perform Meta step
#print("meta")
self._val_loss = compute_vaLoss(model=self._mods['model'], dl_it=self._dl_val_it, dl=self._dl_val) + self._mods['data_aug'].reg_loss()
#print_graph(val_loss) #to visualize computational graph
self._val_loss.backward()
torch.nn.utils.clip_grad_norm_(self._mods['data_aug'].parameters(), max_norm=10, norm_type=2) #Prevent exploding grad with RNN
self._meta_opt.step()
#Adjust Hyper-parameters
self._mods['data_aug'].adjust_param(soft=False) #Contrainte sum(proba)=1
#For optimizer parameters, if needed
#for param_group in self._diffopt.param_groups:
# for param in list(self._opt_param['Inner'].keys())[1:]:
# param_group[param].data = param_group[param].data.clamp(min=1e-4)
#Reset gradients
self._diffopt.detach_()
self._mods['model'].detach_()
self._meta_opt.zero_grad()
self._it_count=0
def val_loss(self):
""" Get the validation loss.
Compute, if needed, the validation loss and returns it.
''start_bilevel_opt'' method needs to be called once before using this method.
Returns:
(Tensor) Validation loss on a single batch of data.
"""
if(self._meta_opt): #Bilevel opti
return self._val_loss
else:
return compute_vaLoss(model=self._mods['model'], dl_it=self._dl_val_it, dl=self._dl_val)
##########################
def train(self, mode=True):
""" Set the module training mode.
Args:
mode (bool): Wether to learn the parameter of the module. (default: None)
"""
#if mode is None :
# mode=self._data_augmentation
super(Augmented_model, self).train(mode)
self._mods['data_aug'].augment(mode=self._data_augmentation) #Restart if needed data augmentation
return self
def eval(self):
""" Set the module to evaluation mode.
"""
#return self.train(mode=False)
super(Augmented_model, self).train(mode=False)
self._mods['data_aug'].augment(mode=False)
return self
def items(self):
"""Return an iterable of the ModuleDict key/value pairs.
"""
return self._mods.items()
def update(self, modules):
"""Update the module dictionnary.
The new dictionnary should keep the same structure.
"""
assert(self._mods.keys()==modules.keys())
self._mods.update(modules)
def is_augmenting(self):
""" Return wether data augmentation is applied.
Returns:
bool : True if data augmentation is applied.
"""
return self._data_augmentation
def TF_names(self):
""" Get the transformations names used by the data augmentation module.
Returns:
list : names of the transformations of the data augmentation module.
"""
try:
return self._mods['data_aug']._TF
except:
return None
def __getitem__(self, key):
"""Access to the modules.
Args:
key (string): Name of the module to access.
Returns:
nn.Module.
"""
return self._mods[key]
def __str__(self):
"""Name of the module
Returns:
String containing the name of the module as well as the higher levels parameters.
"""
return "Aug_mod("+str(self._mods['data_aug'])+"-"+str(self._mods['model'])+")"
Reference in a new issue View git blame Copy permalink