mirror of
https://github.com/AntoineHX/smart_augmentation.git
synced 2025-05-03 11:40:46 +02:00
Fix Translate + TF loader
This commit is contained in:
Harle, Antoine (Contracteur)
parent
79de0191a8
commit
b170af076f
9 changed files with 674 additions and 40 deletions
193
higher/config/bad_tf_config.json
Normal file
193
higher/config/bad_tf_config.json
Normal file
|
|
@ -0,0 +1,193 @@
|
|||
[
|
||||
{
|
||||
"name": "Identity",
|
||||
"function": "identity"
|
||||
},
|
||||
{
|
||||
"name": "FlipUD",
|
||||
"function": "flip",
|
||||
"param": {
|
||||
"axis": "Y"
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "FlipLR",
|
||||
"function": "flip",
|
||||
"param": {
|
||||
"axis": "X"
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Rotate",
|
||||
"function": "rotate",
|
||||
"param": {
|
||||
"min": null,
|
||||
"max": 30,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "TranslateX",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": null,
|
||||
"max": 0.33,
|
||||
"absolute": false,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "TranslateY",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": null,
|
||||
"max": 0.33,
|
||||
"absolute": false,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "ShearX",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": null,
|
||||
"max": 0.3,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "ShearY",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": null,
|
||||
"max": 0.3,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Contrast",
|
||||
"function": "contrast",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Color",
|
||||
"function": "color",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Brightness",
|
||||
"function": "brightness",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Sharpness",
|
||||
"function": "sharpness",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Posterize",
|
||||
"function": "posterize",
|
||||
"param": {
|
||||
"min": 4.0,
|
||||
"max": 8.0,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Solarize",
|
||||
"function": "solarize",
|
||||
"param": {
|
||||
"min": 0.00390625,
|
||||
"max": 1.0,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "BTranslateX",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": 25,
|
||||
"max": 30,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "BTranslateX-",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": -25,
|
||||
"max": -30,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "BTranslateY",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": 25,
|
||||
"max": 30,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "BTranslateY-",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": -25,
|
||||
"max": -30,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "BShearX",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": 0.9,
|
||||
"max": 1.2,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "BShearY",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": 0.9,
|
||||
"max": 1.2,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
}
|
||||
]
|
||||
127
higher/config/base_tf_config.json
Normal file
127
higher/config/base_tf_config.json
Normal file
|
|
@ -0,0 +1,127 @@
|
|||
[
|
||||
{
|
||||
"name": "Identity",
|
||||
"function": "identity"
|
||||
},
|
||||
{
|
||||
"name": "FlipUD",
|
||||
"function": "flip",
|
||||
"param": {
|
||||
"axis": "Y"
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "FlipLR",
|
||||
"function": "flip",
|
||||
"param": {
|
||||
"axis": "X"
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Rotate",
|
||||
"function": "rotate",
|
||||
"param": {
|
||||
"min": null,
|
||||
"max": 30,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "TranslateX",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": null,
|
||||
"max": 0.33,
|
||||
"absolute": false,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "TranslateY",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": null,
|
||||
"max": 0.33,
|
||||
"absolute": false,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "ShearX",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": null,
|
||||
"max": 0.3,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "ShearY",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": null,
|
||||
"max": 0.3,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Contrast",
|
||||
"function": "contrast",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Color",
|
||||
"function": "color",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Brightness",
|
||||
"function": "brightness",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Sharpness",
|
||||
"function": "sharpness",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Posterize",
|
||||
"function": "posterize",
|
||||
"param": {
|
||||
"min": 4.0,
|
||||
"max": 8.0,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Solarize",
|
||||
"function": "solarize",
|
||||
"param": {
|
||||
"min": 0.00390625,
|
||||
"max": 1.0,
|
||||
"invScale": false
|
||||
}
|
||||
}
|
||||
]
|
||||
163
higher/config/invScale_tf_config.json
Normal file
163
higher/config/invScale_tf_config.json
Normal file
|
|
@ -0,0 +1,163 @@
|
|||
[
|
||||
{
|
||||
"name": "Identity",
|
||||
"function": "identity"
|
||||
},
|
||||
{
|
||||
"name": "FlipUD",
|
||||
"function": "flip",
|
||||
"param": {
|
||||
"axis": "Y"
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "FlipLR",
|
||||
"function": "flip",
|
||||
"param": {
|
||||
"axis": "X"
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Rotate",
|
||||
"function": "rotate",
|
||||
"param": {
|
||||
"min": null,
|
||||
"max": 30,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "TranslateX",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": null,
|
||||
"max": 0.33,
|
||||
"absolute": false,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "TranslateY",
|
||||
"function": "translate",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": null,
|
||||
"max": 0.33,
|
||||
"absolute": false,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "ShearX",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "X",
|
||||
"min": null,
|
||||
"max": 0.3,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "ShearY",
|
||||
"function": "shear",
|
||||
"param": {
|
||||
"axis": "Y",
|
||||
"min": null,
|
||||
"max": 0.3,
|
||||
"absolute": true,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "+Contrast",
|
||||
"function": "contrast",
|
||||
"param": {
|
||||
"min": 1.0,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "+Color",
|
||||
"function": "color",
|
||||
"param": {
|
||||
"min": 1.0,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "+Brightness",
|
||||
"function": "brightness",
|
||||
"param": {
|
||||
"min": 1.0,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "+Sharpness",
|
||||
"function": "sharpness",
|
||||
"param": {
|
||||
"min": 1.0,
|
||||
"max": 1.9,
|
||||
"invScale": false
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "-Contrast",
|
||||
"function": "contrast",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.0,
|
||||
"invScale": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "-Color",
|
||||
"function": "color",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.0,
|
||||
"invScale": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "-Brightness",
|
||||
"function": "brightness",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.0,
|
||||
"invScale": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "-Sharpness",
|
||||
"function": "sharpness",
|
||||
"param": {
|
||||
"min": 0.1,
|
||||
"max": 1.0,
|
||||
"invScale": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Posterize",
|
||||
"function": "posterize",
|
||||
"param": {
|
||||
"min": 4.0,
|
||||
"max": 8.0,
|
||||
"invScale": true
|
||||
}
|
||||
},
|
||||
{
|
||||
"name": "Solarize",
|
||||
"function": "solarize",
|
||||
"param": {
|
||||
"min": 0.00390625,
|
||||
"max": 1.0,
|
||||
"invScale": true
|
||||
}
|
||||
}
|
||||
]
|
||||
|
|
@ -16,8 +16,11 @@ optim_param={
|
|||
},
|
||||
'Inner':{
|
||||
'optim': 'SGD',
|
||||
'lr':1e-2, #1e-2 #1e-1 for ResNet
|
||||
'lr':1e-1, #1e-2/1e-1 (ResNet)
|
||||
'momentum':0.9, #0.9
|
||||
'decay':0.0005, #0.0005
|
||||
'nesterov':True,
|
||||
'scheduler':'cosine', #None, 'cosine', 'multiStep', 'exponential'
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -28,6 +31,7 @@ dataug_epoch_start=0
|
|||
nb_run= 3
|
||||
|
||||
# Use available TF (see transformations.py)
|
||||
'''
|
||||
tf_names = [
|
||||
## Geometric TF ##
|
||||
'Identity',
|
||||
|
|
@ -63,7 +67,10 @@ tf_names = [
|
|||
#'RandBlend'
|
||||
]
|
||||
tf_dict = {k: TF.TF_dict[k] for k in tf_names}
|
||||
|
||||
'''
|
||||
tf_config='../config/base_tf_config.json'
|
||||
TF_loader=TF_loader()
|
||||
tf_dict, tf_ignore_mag =TF_loader.load_TF_dict(tf_config)
|
||||
|
||||
device = torch.device('cuda')
|
||||
|
||||
|
|
@ -82,11 +89,11 @@ np.random.seed(0)
|
|||
if __name__ == "__main__":
|
||||
|
||||
### Benchmark ###
|
||||
'''
|
||||
#'''
|
||||
n_inner_iter = 1
|
||||
dist_mix = [0.5]#[0.5, 1.0]
|
||||
N_seq_TF= [3, 4]
|
||||
mag_setup = [(True, True), (False, False)] #(FxSh, Independant)
|
||||
mag_setup = [(False, False)] #[(True, True), (False, False)] #(FxSh, Independant)
|
||||
|
||||
for model_type in model_list.keys():
|
||||
for model_name in model_list[model_type]:
|
||||
|
|
@ -100,7 +107,7 @@ if __name__ == "__main__":
|
|||
torch.cuda.reset_max_memory_cached() #reset_peak_stats
|
||||
t0 = time.perf_counter()
|
||||
|
||||
model = getattr(model_type, model_name)(pretrained=False)
|
||||
model = getattr(model_type, model_name)(pretrained=False, num_classes=len(dl_train.dataset.classes))
|
||||
|
||||
model = Higher_model(model, model_name) #run_dist_dataugV3
|
||||
if n_inner_iter!=0:
|
||||
|
|
@ -137,6 +144,7 @@ if __name__ == "__main__":
|
|||
'Optimizer': optim_param,
|
||||
"Device": device_name,
|
||||
"Memory": [max_allocated, max_cached],
|
||||
"TF_config": tf_config,
|
||||
"Param_names": aug_model.TF_names(),
|
||||
"Log": log}
|
||||
print(str(aug_model),": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1])
|
||||
|
|
@ -150,9 +158,9 @@ if __name__ == "__main__":
|
|||
|
||||
print('Execution Time : %.00f '%(exec_time))
|
||||
print('-'*9)
|
||||
'''
|
||||
### Benchmark - RandAugment/Vanilla ###
|
||||
#'''
|
||||
### Benchmark - RandAugment/Vanilla ###
|
||||
'''
|
||||
for model_type in model_list.keys():
|
||||
for model_name in model_list[model_type]:
|
||||
for run in range(nb_run):
|
||||
|
|
@ -160,7 +168,7 @@ if __name__ == "__main__":
|
|||
torch.cuda.reset_max_memory_cached() #reset_peak_stats
|
||||
t0 = time.perf_counter()
|
||||
|
||||
model = getattr(model_type, model_name)(pretrained=False).to(device)
|
||||
model = getattr(model_type, model_name)(pretrained=False, num_classes=len(dl_train.dataset.classes)).to(device)
|
||||
|
||||
print("{} on {} for {} epochs".format(model_name, device_name, epochs))
|
||||
#print("RandAugment(N{}-M{:.2f})-{} on {} for {} epochs".format(rand_aug['N'],rand_aug['M'],model_name, device_name, epochs))
|
||||
|
|
@ -180,7 +188,7 @@ if __name__ == "__main__":
|
|||
#"Rand_Aug": rand_aug,
|
||||
"Log": log}
|
||||
print(model_name,": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1])
|
||||
filename = "{} epochs -{}".format(model_name,epochs, run)
|
||||
filename = "{}-{} epochs -{}".format(model_name,epochs, run)
|
||||
#print("RandAugment-",model_name,": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1])
|
||||
#filename = "RandAugment(N{}-M{:.2f})-{}-{} epochs -{}".format(rand_aug['N'],rand_aug['M'],model_name,epochs, run)
|
||||
with open(res_folder+"log/%s.json" % filename, "w+") as f:
|
||||
|
|
@ -189,12 +197,13 @@ if __name__ == "__main__":
|
|||
print('Log :\"',f.name, '\" saved !')
|
||||
except:
|
||||
print("Failed to save logs :",f.name)
|
||||
print(sys.exc_info()[1])
|
||||
|
||||
#plot_resV2(log, fig_name=res_folder+filename)
|
||||
|
||||
print('Execution Time : %.00f '%(exec_time))
|
||||
print('-'*9)
|
||||
#'''
|
||||
'''
|
||||
### HP Search ###
|
||||
'''
|
||||
from LeNet import *
|
||||
|
|
@ -221,7 +230,7 @@ if __name__ == "__main__":
|
|||
|
||||
t0 = time.perf_counter()
|
||||
|
||||
model = getattr(models.resnet, 'resnet18')(pretrained=False)
|
||||
model = getattr(models.resnet, 'resnet18')(pretrained=False, num_classes=len(dl_train.dataset.classes))
|
||||
#model = LeNet(3,10)
|
||||
model = Higher_model(model) #run_dist_dataugV3
|
||||
aug_model = Augmented_model(Data_augV5(TF_dict=tf_dict, N_TF=n_tf, mix_dist=dist, fixed_prob=p_setup, fixed_mag=m_setup[0], shared_mag=m_setup[1]), model).to(device)
|
||||
|
|
|
|||
|
|
@ -39,6 +39,7 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba)
|
|||
_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.
|
||||
|
|
@ -51,7 +52,7 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba)
|
|||
_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=TF.TF_dict, N_TF=1, mix_dist=0.0, fixed_prob=False, fixed_mag=True, shared_mag=True):
|
||||
def __init__(self, TF_dict, N_TF=1, mix_dist=0.0, fixed_prob=False, fixed_mag=True, shared_mag=True, TF_ignore_mag=TF.TF_ignore_mag):
|
||||
"""Init Data_augv5.
|
||||
|
||||
Args:
|
||||
|
|
@ -61,6 +62,7 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba)
|
|||
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
|
||||
|
|
@ -71,13 +73,14 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba)
|
|||
#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 TF.TF_ignore_mag])==0:
|
||||
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
|
||||
|
||||
|
|
@ -112,7 +115,7 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba)
|
|||
if self._shared_mag :
|
||||
self._reg_tgt = torch.tensor(TF.PARAMETER_MAX, dtype=torch.float) #Encourage amplitude max
|
||||
else:
|
||||
self._reg_mask=[self._TF.index(t) for t in self._TF if t not in TF.TF_ignore_mag]
|
||||
self._reg_mask=[self._TF.index(t) for t in self._TF if t not in self._TF_ignore_mag]
|
||||
self._reg_tgt=torch.full(size=(len(self._reg_mask),), fill_value=TF.PARAMETER_MAX) #Encourage amplitude max
|
||||
|
||||
def forward(self, x):
|
||||
|
|
@ -324,6 +327,7 @@ class Data_augV7(nn.Module): #Proba sequentielles
|
|||
_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.
|
||||
|
|
@ -336,7 +340,7 @@ class Data_augV7(nn.Module): #Proba sequentielles
|
|||
_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=TF.TF_dict, N_TF=2, mix_dist=0.0, fixed_prob=False, fixed_mag=True, shared_mag=True):
|
||||
def __init__(self, TF_dict, N_TF=2, mix_dist=0.0, fixed_prob=False, fixed_mag=True, shared_mag=True, TF_ignore_mag=TF.TF_ignore_mag):
|
||||
"""Init Data_augv7.
|
||||
|
||||
Args:
|
||||
|
|
@ -346,6 +350,7 @@ class Data_augV7(nn.Module): #Proba sequentielles
|
|||
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_augV7, self).__init__()
|
||||
assert len(TF_dict)>0
|
||||
|
|
@ -359,13 +364,14 @@ class Data_augV7(nn.Module): #Proba sequentielles
|
|||
#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 TF.TF_ignore_mag])==0:
|
||||
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
|
||||
|
||||
|
|
@ -423,7 +429,7 @@ class Data_augV7(nn.Module): #Proba sequentielles
|
|||
if self._shared_mag :
|
||||
self._reg_tgt = torch.FloatTensor(TF.PARAMETER_MAX) #Encourage amplitude max
|
||||
else:
|
||||
self._reg_mask=[idx for idx,t in enumerate(self._TF) if t not in TF.TF_ignore_mag]
|
||||
self._reg_mask=[idx for idx,t in enumerate(self._TF) if t not in self._TF_ignore_mag]
|
||||
self._reg_tgt=torch.full(size=(len(self._reg_mask),), fill_value=TF.PARAMETER_MAX) #Encourage amplitude max
|
||||
|
||||
def forward(self, x):
|
||||
|
|
@ -657,7 +663,7 @@ class RandAug(nn.Module): #RandAugment = UniformFx-MagFxSh + rapide
|
|||
_fixed_mag (bool): Wether to lock the TF magnitudes. Should be True.
|
||||
_params (nn.ParameterDict): Data augmentation parameters.
|
||||
"""
|
||||
def __init__(self, TF_dict=TF.TF_dict, N_TF=1, mag=TF.PARAMETER_MAX):
|
||||
def __init__(self, TF_dict, N_TF=1, mag=TF.PARAMETER_MAX):
|
||||
"""Init RandAug.
|
||||
|
||||
Args:
|
||||
|
|
|
|||
|
|
@ -8,6 +8,7 @@ from dataug import *
|
|||
from train_utils import *
|
||||
|
||||
# Use available TF (see transformations.py)
|
||||
'''
|
||||
tf_names = [
|
||||
## Geometric TF ##
|
||||
'Identity',
|
||||
|
|
@ -57,7 +58,8 @@ tf_names = [
|
|||
#'Random',
|
||||
#'RandBlend'
|
||||
]
|
||||
|
||||
'''
|
||||
TF_loader=TF_loader()
|
||||
|
||||
device = torch.device('cuda') #Select device to use
|
||||
|
||||
|
|
@ -77,12 +79,12 @@ if __name__ == "__main__":
|
|||
|
||||
#Task to perform
|
||||
tasks={
|
||||
'classic',
|
||||
#'aug_model'
|
||||
#'classic',
|
||||
'aug_model'
|
||||
}
|
||||
#Parameters
|
||||
n_inner_iter = 1
|
||||
epochs = 200
|
||||
epochs = 20
|
||||
dataug_epoch_start=0
|
||||
optim_param={
|
||||
'Meta':{
|
||||
|
|
@ -91,11 +93,11 @@ if __name__ == "__main__":
|
|||
},
|
||||
'Inner':{
|
||||
'optim': 'SGD',
|
||||
'lr':1e-1, #1e-2/1e-1
|
||||
'lr':1e-1, #1e-2/1e-1 (ResNet)
|
||||
'momentum':0.9, #0.9
|
||||
'decay':0.0005, #0.0005
|
||||
'nesterov':True,
|
||||
'scheduler':'exponential', #None, 'cosine', 'multiStep', 'exponential'
|
||||
'scheduler':'cosine', #None, 'cosine', 'multiStep', 'exponential'
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -137,7 +139,7 @@ if __name__ == "__main__":
|
|||
print(model_name,": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1])
|
||||
filename = "{}-{} epochs".format(model_name,epochs)
|
||||
#print("RandAugment-",model_name,": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1])
|
||||
#filename = "RandAugment(N{}-M{:.2f})-{}-{} epochs".format(rand_aug['N'],rand_aug['M'],model_name,epochs)
|
||||
#filename = "RandAugment(N{}-M{:.2f})-{}-{} epochs".format(rand_aug['N'],rand_aug['M'],model_name,epochs)+'-cosine'
|
||||
with open("../res/log/%s.json" % filename, "w+") as f:
|
||||
try:
|
||||
json.dump(out, f, indent=True)
|
||||
|
|
@ -157,13 +159,23 @@ if __name__ == "__main__":
|
|||
|
||||
#### Augmented Model ####
|
||||
if 'aug_model' in tasks:
|
||||
tf_config='../config/base_tf_config.json'
|
||||
tf_dict, tf_ignore_mag =TF_loader.load_TF_dict(tf_config)
|
||||
#tf_dict = {k: TF_dict[k] for k in tf_names}
|
||||
|
||||
torch.cuda.reset_max_memory_allocated() #reset_peak_stats
|
||||
torch.cuda.reset_max_memory_cached() #reset_peak_stats
|
||||
t0 = time.perf_counter()
|
||||
|
||||
tf_dict = {k: TF.TF_dict[k] for k in tf_names}
|
||||
model = Higher_model(model, model_name) #run_dist_dataugV3
|
||||
aug_model = Augmented_model(Data_augV5(TF_dict=tf_dict, N_TF=1, mix_dist=0.5, fixed_prob=False, fixed_mag=False, shared_mag=False), model).to(device)
|
||||
aug_model = Augmented_model(
|
||||
Data_augV5(TF_dict=tf_dict,
|
||||
N_TF=3,
|
||||
mix_dist=0.5,
|
||||
fixed_prob=False,
|
||||
fixed_mag=False,
|
||||
shared_mag=False,
|
||||
TF_ignore_mag=tf_ignore_mag), model).to(device)
|
||||
#aug_model = Augmented_model(RandAug(TF_dict=tf_dict, N_TF=2), model).to(device)
|
||||
|
||||
print("{} on {} for {} epochs - {} inner_it".format(str(aug_model), device_name, epochs, n_inner_iter))
|
||||
|
|
@ -175,7 +187,7 @@ if __name__ == "__main__":
|
|||
print_freq=1,
|
||||
unsup_loss=1,
|
||||
hp_opt=False,
|
||||
save_sample_freq=None)
|
||||
save_sample_freq=0)
|
||||
|
||||
exec_time=time.perf_counter() - t0
|
||||
max_allocated = torch.cuda.max_memory_allocated()/(1024.0 * 1024.0)
|
||||
|
|
@ -188,6 +200,7 @@ if __name__ == "__main__":
|
|||
'Optimizer': optim_param,
|
||||
"Device": device_name,
|
||||
"Memory": [max_allocated, max_cached],
|
||||
"TF_config": tf_config,
|
||||
"Param_names": aug_model.TF_names(),
|
||||
"Log": log}
|
||||
print(str(aug_model),": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1])
|
||||
|
|
|
|||
|
|
@ -272,15 +272,15 @@ def run_dist_dataugV3(model, opt_param, epochs=1, inner_it=1, dataug_epoch_start
|
|||
#Scheduler
|
||||
inner_scheduler=None
|
||||
if opt_param['Inner']['scheduler']=='cosine':
|
||||
inner_scheduler=torch.optim.lr_scheduler.CosineAnnealingLR(optim, T_max=epochs, eta_min=0.)
|
||||
inner_scheduler=torch.optim.lr_scheduler.CosineAnnealingLR(inner_opt, T_max=epochs, eta_min=0.)
|
||||
elif opt_param['Inner']['scheduler']=='multiStep':
|
||||
#Multistep milestones inspired by AutoAugment
|
||||
inner_scheduler=torch.optim.lr_scheduler.MultiStepLR(optim,
|
||||
inner_scheduler=torch.optim.lr_scheduler.MultiStepLR(inner_opt,
|
||||
milestones=[int(epochs/3), int(epochs*2/3), int(epochs*2.7/3)],
|
||||
gamma=0.1)
|
||||
elif opt_param['Inner']['scheduler']=='exponential':
|
||||
#inner_scheduler=torch.optim.lr_scheduler.ExponentialLR(optim, gamma=0.1) #Wrong gamma
|
||||
inner_scheduler=torch.optim.lr_scheduler.LambdaLR(optim, lambda epoch: (1 - epoch / epochs) ** 0.9)
|
||||
inner_scheduler=torch.optim.lr_scheduler.LambdaLR(inner_opt, lambda epoch: (1 - epoch / epochs) ** 0.9)
|
||||
elif opt_param['Inner']['scheduler'] is not None:
|
||||
raise ValueError("Lr scheduler unknown : %s"%opt_param['Inner']['scheduler'])
|
||||
|
||||
|
|
|
|||
|
|
@ -19,9 +19,9 @@ import kornia
|
|||
import random
|
||||
|
||||
#TF that don't have use for magnitude parameter.
|
||||
TF_no_mag={'Identity', 'FlipUD', 'FlipLR', 'Random', 'RandBlend'}
|
||||
TF_no_mag={'Identity', 'FlipUD', 'FlipLR', 'Random', 'RandBlend', 'identity', 'flipUD', 'flipLR'}
|
||||
#TF which implemetation doesn't allow gradient propagaition.
|
||||
TF_no_grad={'Solarize', 'Posterize', '=Solarize', '=Posterize'}
|
||||
TF_no_grad={'Solarize', 'Posterize', '=Solarize', '=Posterize', 'posterize','solarize'}
|
||||
#TF for which magnitude should be ignored (Magnitude fixed).
|
||||
TF_ignore_mag= TF_no_mag | TF_no_grad
|
||||
|
||||
|
|
@ -30,6 +30,7 @@ PARAMETER_MAX = 1
|
|||
# What is the min 'level' a transform could be predicted
|
||||
PARAMETER_MIN = 0.1
|
||||
|
||||
'''
|
||||
# Dictionnary mapping tranformations identifiers to their function.
|
||||
# Each value of the dict should be a lambda function taking a (batch of data, magnitude of transformations) tuple as input and returns a batch of data.
|
||||
TF_dict={ #Dataugv5+
|
||||
|
|
@ -38,8 +39,8 @@ TF_dict={ #Dataugv5+
|
|||
'FlipUD' : (lambda x, mag: flipUD(x)),
|
||||
'FlipLR' : (lambda x, mag: flipLR(x)),
|
||||
'Rotate': (lambda x, mag: rotate(x, angle=rand_floats(size=x.shape[0], mag=mag, maxval=30))),
|
||||
'TranslateX': (lambda x, mag: translate(x, translation=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=x.shape[1]*0.33), zero_pos=0))),
|
||||
'TranslateY': (lambda x, mag: translate(x, translation=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=x.shape[2]*0.33), zero_pos=1))),
|
||||
'TranslateX': (lambda x, mag: translate(x, translation=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=x.shape[2]*0.33), zero_pos=0))),
|
||||
'TranslateY': (lambda x, mag: translate(x, translation=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=x.shape[3]*0.33), zero_pos=1))),
|
||||
'TranslateXabs': (lambda x, mag: translate(x, translation=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=20), zero_pos=0))),
|
||||
'TranslateYabs': (lambda x, mag: translate(x, translation=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=20), zero_pos=1))),
|
||||
'ShearX': (lambda x, mag: shear(x, shear=zero_stack(rand_floats(size=(x.shape[0],), mag=mag, maxval=0.3), zero_pos=0))),
|
||||
|
|
@ -49,7 +50,7 @@ TF_dict={ #Dataugv5+
|
|||
'Contrast': (lambda x, mag: contrast(x, contrast_factor=rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.9))),
|
||||
'Color':(lambda x, mag: color(x, color_factor=rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.9))),
|
||||
'Brightness':(lambda x, mag: brightness(x, brightness_factor=rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.9))),
|
||||
'Sharpness':(lambda x, mag: sharpeness(x, sharpness_factor=rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.9))),
|
||||
'Sharpness':(lambda x, mag: sharpness(x, sharpness_factor=rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.9))),
|
||||
'Posterize': (lambda x, mag: posterize(x, bits=rand_floats(size=x.shape[0], mag=mag, minval=4., maxval=8.))),#Perte du gradient
|
||||
'Solarize': (lambda x, mag: solarize(x, thresholds=rand_floats(size=x.shape[0], mag=mag, minval=1/256., maxval=256/256.))), #Perte du gradient #=>Image entre [0,1]
|
||||
|
||||
|
|
@ -57,11 +58,11 @@ TF_dict={ #Dataugv5+
|
|||
'+Contrast': (lambda x, mag: contrast(x, contrast_factor=rand_floats(size=x.shape[0], mag=mag, minval=1.0, maxval=1.9))),
|
||||
'+Color':(lambda x, mag: color(x, color_factor=rand_floats(size=x.shape[0], mag=mag, minval=1.0, maxval=1.9))),
|
||||
'+Brightness':(lambda x, mag: brightness(x, brightness_factor=rand_floats(size=x.shape[0], mag=mag, minval=1.0, maxval=1.9))),
|
||||
'+Sharpness':(lambda x, mag: sharpeness(x, sharpness_factor=rand_floats(size=x.shape[0], mag=mag, minval=1.0, maxval=1.9))),
|
||||
'+Sharpness':(lambda x, mag: sharpness(x, sharpness_factor=rand_floats(size=x.shape[0], mag=mag, minval=1.0, maxval=1.9))),
|
||||
'-Contrast': (lambda x, mag: contrast(x, contrast_factor=invScale_rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.0))),
|
||||
'-Color':(lambda x, mag: color(x, color_factor=invScale_rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.0))),
|
||||
'-Brightness':(lambda x, mag: brightness(x, brightness_factor=invScale_rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.0))),
|
||||
'-Sharpness':(lambda x, mag: sharpeness(x, sharpness_factor=invScale_rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.0))),
|
||||
'-Sharpness':(lambda x, mag: sharpness(x, sharpness_factor=invScale_rand_floats(size=x.shape[0], mag=mag, minval=0.1, maxval=1.0))),
|
||||
'=Posterize': (lambda x, mag: posterize(x, bits=invScale_rand_floats(size=x.shape[0], mag=mag, minval=4., maxval=8.))),#Perte du gradient
|
||||
'=Solarize': (lambda x, mag: solarize(x, thresholds=invScale_rand_floats(size=x.shape[0], mag=mag, minval=1/256., maxval=256/256.))), #Perte du gradient #=>Image entre [0,1]
|
||||
|
||||
|
|
@ -86,7 +87,7 @@ TF_dict={ #Dataugv5+
|
|||
#'Auto_Contrast': (lambda mag: None), #Pas opti pour des batch (Super lent)
|
||||
#'Equalize': (lambda mag: None),
|
||||
}
|
||||
|
||||
'''
|
||||
## Image type cast ##
|
||||
def int_image(float_image):
|
||||
"""Convert a float Tensor/Image to an int Tensor/Image.
|
||||
|
|
@ -304,7 +305,7 @@ def brightness(x, brightness_factor):
|
|||
|
||||
return blend(torch.zeros(x.size(), device=device), x, brightness_factor).clamp(min=0.0,max=1.0) #Expect image in the range of [0, 1]
|
||||
|
||||
def sharpeness(x, sharpness_factor):
|
||||
def sharpness(x, sharpness_factor):
|
||||
"""Adjust sharpness of images.
|
||||
|
||||
Args:
|
||||
|
|
|
|||
|
|
@ -14,6 +14,128 @@ import torch.nn.functional as F
|
|||
|
||||
import time
|
||||
|
||||
import transformations as TF
|
||||
class TF_loader(object):
|
||||
""" Transformations builder.
|
||||
|
||||
See 'config' folder for pre-defined config files.
|
||||
|
||||
Attributes:
|
||||
_filename (str): Path to config file (JSON) used.
|
||||
_TF_dict (dict): Transformations dictionnary built from config file.
|
||||
_TF_ignore_mag (set): Ensemble of transformations names for which magnitude should be ignored.
|
||||
_TF_names (list): List of transformations names/keys.
|
||||
"""
|
||||
def __init__(self):
|
||||
""" Initialize TF_loader.
|
||||
|
||||
"""
|
||||
self._filename=''
|
||||
self._TF_dict={}
|
||||
self._TF_ignore_mag=set()
|
||||
self._TF_names=[]
|
||||
|
||||
def load_TF_dict(self, filename):
|
||||
""" Build a TF dictionnary.
|
||||
|
||||
Args:
|
||||
filename (str): Path to config file (JSON) defining the transformations.
|
||||
Returns:
|
||||
(dict, set) : TF dicttionnary built and ensemble of TF names for which mag should be ignored.
|
||||
"""
|
||||
self._filename=filename
|
||||
self._TF_names=[]
|
||||
self._TF_dict={}
|
||||
self._TF_ignore_mag=set()
|
||||
|
||||
with open(filename) as json_file:
|
||||
TF_params = json.load(json_file)
|
||||
|
||||
for tf in TF_params:
|
||||
self._TF_names.append(tf['name'])
|
||||
if tf['function'] in TF.TF_ignore_mag:
|
||||
self._TF_ignore_mag.add(tf['name'])
|
||||
|
||||
if tf['function'] == 'identity':
|
||||
self._TF_dict[tf['name']]=(lambda x, mag: x)
|
||||
|
||||
elif tf['function'] == 'flip':
|
||||
#Inverser axes ?
|
||||
if tf['param']['axis'] == 'X':
|
||||
self._TF_dict[tf['name']]=(lambda x, mag: TF.flipLR(x))
|
||||
elif tf['param']['axis'] == 'Y':
|
||||
self._TF_dict[tf['name']]=(lambda x, mag: TF.flipUD(x))
|
||||
else:
|
||||
raise Exception("Unknown TF axis : %s in %s"%(tf['function'], self._filename))
|
||||
|
||||
elif tf['function'] in {'translate', 'shear'}:
|
||||
rand_fct= 'invScale_rand_floats' if tf['param']['invScale'] else 'rand_floats'
|
||||
self._TF_dict[tf['name']]=self.build_lambda(tf['function'], rand_fct, tf['param']['min'], tf['param']['max'], tf['param']['absolute'], tf['param']['axis'])
|
||||
|
||||
else:
|
||||
rand_fct= 'invScale_rand_floats' if tf['param']['invScale'] else 'rand_floats'
|
||||
self._TF_dict[tf['name']]=self.build_lambda(tf['function'], rand_fct, tf['param']['min'], tf['param']['max'])
|
||||
|
||||
return self._TF_dict, self._TF_ignore_mag
|
||||
|
||||
def build_lambda(self, fct_name, rand_fct_name, minval, maxval, absolute=True, axis=None):
|
||||
""" Build a lambda function performing transformations.
|
||||
|
||||
Args:
|
||||
fct_name (str): Name of the transformations to use (see transformations.py).
|
||||
rand_fct_name (str): Name of the random mapping function to use (see transformations.py).
|
||||
minval (float): minimum magnitude value of the TF.
|
||||
maxval (float): maximum magnitude value of the TF.
|
||||
absolute (bool): Wether the maxval should be relative (absolute=False) to the image size. (default: True)
|
||||
axis (str): Axis ('X' / 'Y') of the TF, if relevant. Should be used for (flip)/translate/shear functions. (default: None)
|
||||
|
||||
Returns:
|
||||
(function) transformations function : Tensor=f(Tensor, magnitude)
|
||||
"""
|
||||
if absolute:
|
||||
max_val_fct=(lambda x: maxval)
|
||||
else: #Relative to img size
|
||||
max_val_fct=(lambda x: x*maxval)
|
||||
|
||||
if axis is None:
|
||||
return (lambda x, mag:
|
||||
getattr(TF, fct_name)(
|
||||
x,
|
||||
getattr(TF, rand_fct_name)(
|
||||
size=x.shape[0],
|
||||
mag=mag,
|
||||
minval=minval,
|
||||
maxval=maxval)))
|
||||
elif axis =='X':
|
||||
return (lambda x, mag:
|
||||
getattr(TF, fct_name)(
|
||||
x,
|
||||
TF.zero_stack(
|
||||
getattr(TF, rand_fct_name)(
|
||||
size=(x.shape[0],),
|
||||
mag=mag,
|
||||
minval=minval,
|
||||
maxval=max_val_fct(x.shape[2])),
|
||||
zero_pos=0)))
|
||||
elif axis == 'Y':
|
||||
return (lambda x, mag:
|
||||
getattr(TF, fct_name)(
|
||||
x,
|
||||
TF.zero_stack(
|
||||
getattr(TF, rand_fct_name)(
|
||||
size=(x.shape[0],),
|
||||
mag=mag,
|
||||
minval=minval,
|
||||
maxval=max_val_fct(x.shape[3])),
|
||||
zero_pos=1)))
|
||||
else:
|
||||
raise Exception("Unknown TF axis : %s in %s"%(fct_name, self._filename))
|
||||
|
||||
def get_TF_names(self):
|
||||
return self._TF_names
|
||||
def get_TF_dict(self):
|
||||
return self._TF_dict
|
||||
|
||||
class ConfusionMatrix(object):
|
||||
""" Confusion matrix.
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue