From 6277e268c1274594aff3171e205702b5d0a9e62e Mon Sep 17 00:00:00 2001 From: "Harle, Antoine (Contracteur)" Date: Wed, 5 Feb 2020 12:24:20 -0500 Subject: [PATCH] RandAugment --- augmentations_randaugment.py | 271 ++++++++++++++++++++++++++++++++++ higher/smart_aug/benchmark.py | 149 +++++++++++++------ higher/smart_aug/datasets.py | 60 ++++---- 3 files changed, 407 insertions(+), 73 deletions(-) create mode 100644 augmentations_randaugment.py diff --git a/augmentations_randaugment.py b/augmentations_randaugment.py new file mode 100644 index 0000000..b491942 --- /dev/null +++ b/augmentations_randaugment.py @@ -0,0 +1,271 @@ +# code in this file is adpated from rpmcruz/autoaugment +# https://github.com/rpmcruz/autoaugment/blob/master/transformations.py +import random + +import PIL, PIL.ImageOps, PIL.ImageEnhance, PIL.ImageDraw +import numpy as np +import torch +from PIL import Image + +def ShearX(img, v): # [-0.3, 0.3] + assert -0.3 <= v <= 0.3 + if random.random() > 0.5: + v = -v + return img.transform(img.size, PIL.Image.AFFINE, (1, v, 0, 0, 1, 0)) + + +def ShearY(img, v): # [-0.3, 0.3] + assert -0.3 <= v <= 0.3 + if random.random() > 0.5: + v = -v + return img.transform(img.size, PIL.Image.AFFINE, (1, 0, 0, v, 1, 0)) + + +def TranslateX(img, v): # [-150, 150] => percentage: [-0.45, 0.45] + assert -0.45 <= v <= 0.45 + if random.random() > 0.5: + v = -v + v = v * img.size[0] + return img.transform(img.size, PIL.Image.AFFINE, (1, 0, v, 0, 1, 0)) + + +def TranslateXabs(img, v): # [-150, 150] => percentage: [-0.45, 0.45] + assert 0 <= v + if random.random() > 0.5: + v = -v + return img.transform(img.size, PIL.Image.AFFINE, (1, 0, v, 0, 1, 0)) + + +def TranslateY(img, v): # [-150, 150] => percentage: [-0.45, 0.45] + assert -0.45 <= v <= 0.45 + if random.random() > 0.5: + v = -v + v = v * img.size[1] + return img.transform(img.size, PIL.Image.AFFINE, (1, 0, 0, 0, 1, v)) + + +def TranslateYabs(img, v): # [-150, 150] => percentage: [-0.45, 0.45] + assert 0 <= v + if random.random() > 0.5: + v = -v + return img.transform(img.size, PIL.Image.AFFINE, (1, 0, 0, 0, 1, v)) + + +def Rotate(img, v): # [-30, 30] + assert -30 <= v <= 30 + if random.random() > 0.5: + v = -v + return img.rotate(v) + + +def AutoContrast(img, _): + return PIL.ImageOps.autocontrast(img) + + +def Invert(img, _): + return PIL.ImageOps.invert(img) + + +def Equalize(img, _): + return PIL.ImageOps.equalize(img) + + +def Flip(img, _): # not from the paper + return PIL.ImageOps.mirror(img) + +def FlipLR(img, v): + return img.transpose(Image.FLIP_LEFT_RIGHT) + +def FlipUD(img, v): + return img.transpose(Image.FLIP_TOP_BOTTOM) + +def Solarize(img, v): # [0, 256] + assert 0 <= v <= 256 + return PIL.ImageOps.solarize(img, v) + + +def SolarizeAdd(img, addition=0, threshold=128): + img_np = np.array(img).astype(np.int) + img_np = img_np + addition + img_np = np.clip(img_np, 0, 255) + img_np = img_np.astype(np.uint8) + img = Image.fromarray(img_np) + return PIL.ImageOps.solarize(img, threshold) + + +def Posterize(img, v): # [4, 8] + v = int(v) + v = max(1, v) + return PIL.ImageOps.posterize(img, v) + + +def Contrast(img, v): # [0.1,1.9] + assert 0.1 <= v <= 1.9 + return PIL.ImageEnhance.Contrast(img).enhance(v) + + +def Color(img, v): # [0.1,1.9] + assert 0.1 <= v <= 1.9 + return PIL.ImageEnhance.Color(img).enhance(v) + + +def Brightness(img, v): # [0.1,1.9] + assert 0.1 <= v <= 1.9 + return PIL.ImageEnhance.Brightness(img).enhance(v) + + +def Sharpness(img, v): # [0.1,1.9] + assert 0.1 <= v <= 1.9 + return PIL.ImageEnhance.Sharpness(img).enhance(v) + + +def Cutout(img, v): # [0, 60] => percentage: [0, 0.2] + assert 0.0 <= v <= 0.2 + if v <= 0.: + return img + + v = v * img.size[0] + return CutoutAbs(img, v) + + +def CutoutAbs(img, v): # [0, 60] => percentage: [0, 0.2] + # assert 0 <= v <= 20 + if v < 0: + return img + w, h = img.size + x0 = np.random.uniform(w) + y0 = np.random.uniform(h) + + x0 = int(max(0, x0 - v / 2.)) + y0 = int(max(0, y0 - v / 2.)) + x1 = min(w, x0 + v) + y1 = min(h, y0 + v) + + xy = (x0, y0, x1, y1) + color = (125, 123, 114) + # color = (0, 0, 0) + img = img.copy() + PIL.ImageDraw.Draw(img).rectangle(xy, color) + return img + + +def SamplePairing(imgs): # [0, 0.4] + def f(img1, v): + i = np.random.choice(len(imgs)) + img2 = PIL.Image.fromarray(imgs[i]) + return PIL.Image.blend(img1, img2, v) + + return f + + +def Identity(img, v): + return img + + +def augment_list(): # 16 oeprations and their ranges + # https://github.com/google-research/uda/blob/master/image/randaugment/policies.py#L57 + l = [ + (Identity, 0., 1.0), + (FlipUD, 0., 1.0), + (FlipLR, 0., 1.0), + (Rotate, 0, 30), # 4 + (TranslateX, 0., 0.33), # 2 + (TranslateY, 0., 0.33), # 3 + (ShearX, 0., 0.3), # 0 + (ShearY, 0., 0.3), # 1 + #(AutoContrast, 0, 1), # 5 + #(Invert, 0, 1), # 6 + #(Equalize, 0, 1), # 7 + (Contrast, 0.1, 1.9), # 10 + (Color, 0.1, 1.9), # 11 + (Brightness, 0.1, 1.9), # 12 + (Sharpness, 0.1, 1.9), # 13 + (Posterize, 4, 8), # 9 + (Solarize, 1, 256), # 8 + + # (Cutout, 0, 0.2), # 14 + # (SamplePairing(imgs), 0, 0.4), # 15 + ] + + # https://github.com/tensorflow/tpu/blob/8462d083dd89489a79e3200bcc8d4063bf362186/models/official/efficientnet/autoaugment.py#L505 + #l = [ + # (AutoContrast, 0, 1), + # (Equalize, 0, 1), + # (Invert, 0, 1), + # (Rotate, 0, 30), + # (Posterize, 0, 4), + # (Solarize, 0, 256), + # (SolarizeAdd, 0, 110), + # (Color, 0.1, 1.9), + # (Contrast, 0.1, 1.9), + # (Brightness, 0.1, 1.9), + # (Sharpness, 0.1, 1.9), + # (ShearX, 0., 0.3), + # (ShearY, 0., 0.3), + # (CutoutAbs, 0, 40), + # (TranslateXabs, 0., 100), + # (TranslateYabs, 0., 100), + #] + + return l + + +class Lighting(object): + """Lighting noise(AlexNet - style PCA - based noise)""" + + def __init__(self, alphastd, eigval, eigvec): + self.alphastd = alphastd + self.eigval = torch.Tensor(eigval) + self.eigvec = torch.Tensor(eigvec) + + def __call__(self, img): + if self.alphastd == 0: + return img + + alpha = img.new().resize_(3).normal_(0, self.alphastd) + rgb = self.eigvec.type_as(img).clone() \ + .mul(alpha.view(1, 3).expand(3, 3)) \ + .mul(self.eigval.view(1, 3).expand(3, 3)) \ + .sum(1).squeeze() + + return img.add(rgb.view(3, 1, 1).expand_as(img)) + + +class CutoutDefault(object): + """ + Reference : https://github.com/quark0/darts/blob/master/cnn/utils.py + """ + def __init__(self, length): + self.length = length + + def __call__(self, img): + h, w = img.size(1), img.size(2) + mask = np.ones((h, w), np.float32) + y = np.random.randint(h) + x = np.random.randint(w) + + y1 = np.clip(y - self.length // 2, 0, h) + y2 = np.clip(y + self.length // 2, 0, h) + x1 = np.clip(x - self.length // 2, 0, w) + x2 = np.clip(x + self.length // 2, 0, w) + + mask[y1: y2, x1: x2] = 0. + mask = torch.from_numpy(mask) + mask = mask.expand_as(img) + img *= mask + return img + +PARAMETER_MAX = 1 +class RandAugment: + def __init__(self, n, m): + self.n = n + self.m = m # [0, PARAMETER_MAX] + self.augment_list = augment_list() + + def __call__(self, img): + ops = random.choices(self.augment_list, k=self.n) + for op, minval, maxval in ops: + val = (float(self.m) / PARAMETER_MAX) * float(maxval - minval) + minval + img = op(img, val) + + return img diff --git a/higher/smart_aug/benchmark.py b/higher/smart_aug/benchmark.py index bc19929..f2d453c 100644 --- a/higher/smart_aug/benchmark.py +++ b/higher/smart_aug/benchmark.py @@ -1,3 +1,6 @@ +""" Script to run series of experiments. + +""" from dataug import * #from utils import * from train_utils import * @@ -13,14 +16,16 @@ optim_param={ }, 'Inner':{ 'optim': 'SGD', - 'lr':1e-1, #1e-2 #1e-1 for ResNet + 'lr':1e-2, #1e-2 #1e-1 for ResNet 'momentum':0.9, #0.9 } } res_folder="../res/benchmark/CIFAR10/" -epochs= 150 +#res_folder="../res/HPsearch/" +epochs= 200 dataug_epoch_start=0 +nb_run= 3 # Use available TF (see transformations.py) tf_names = [ @@ -80,60 +85,107 @@ if __name__ == "__main__": ''' for model_type in model_list.keys(): for model_name in model_list[model_type]: - model = getattr(model_type, model_name)(pretrained=False) + for run in range(nb_run): - t0 = time.process_time() + torch.cuda.reset_max_memory_cached() #reset_peak_stats + t0 = time.perf_counter() - model = Higher_model(model) #run_dist_dataugV3 - if n_inner_iter!=0: - 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) - else: - aug_model = Augmented_model(RandAug(TF_dict=tf_dict, N_TF=n_tf), model).to(device) + model = getattr(model_type, model_name)(pretrained=False) - print("{} on {} for {} epochs - {} inner_it".format(str(aug_model), device_name, epochs, n_inner_iter)) - log= run_dist_dataugV3(model=aug_model, - epochs=epochs, - inner_it=n_inner_iter, - dataug_epoch_start=dataug_epoch_start, - opt_param=optim_param, - print_freq=epochs/4, - unsup_loss=1, - hp_opt=False, - save_sample_freq=None) + model = Higher_model(model, model_name) #run_dist_dataugV3 + if n_inner_iter!=0: + 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) + else: + aug_model = Augmented_model(RandAug(TF_dict=tf_dict, N_TF=n_tf), model).to(device) - exec_time=time.process_time() - t0 - #### - print('-'*9) - times = [x["time"] for x in log] - out = {"Accuracy": max([x["acc"] for x in log]), "Time": (np.mean(times),np.std(times), exec_time), 'Optimizer': optim_param, "Device": device_name, "Param_names": aug_model.TF_names(), "Log": log} - print(str(aug_model),": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1]) - filename = "{}-{} epochs (dataug:{})- {} in_it-{}".format(str(aug_model),epochs,dataug_epoch_start,n_inner_iter, run) - with open("../res/log/%s.json" % filename, "w+") as f: - try: - json.dump(out, f, indent=True) - print('Log :\"',f.name, '\" saved !') - except: - print("Failed to save logs :",f.name) + print("{} on {} for {} epochs - {} inner_it".format(str(aug_model), device_name, epochs, n_inner_iter)) + log= run_dist_dataugV3(model=aug_model, + epochs=epochs, + inner_it=n_inner_iter, + dataug_epoch_start=dataug_epoch_start, + opt_param=optim_param, + print_freq=epochs/4, + unsup_loss=1, + hp_opt=False, + save_sample_freq=None) - print('Execution Time : %.00f '%(exec_time)) - print('-'*9) - + exec_time=time.perf_counter() - t0 + max_cached = torch.cuda.max_memory_cached()/(1024.0 * 1024.0) #torch.cuda.max_memory_reserved() #MB + #### + print('-'*9) + times = [x["time"] for x in log] + out = {"Accuracy": max([x["acc"] for x in log]), + "Time": (np.mean(times),np.std(times), exec_time), + 'Optimizer': optim_param, + "Device": device_name, + "Memory": max_cached, + "Param_names": aug_model.TF_names(), + "Log": log} + print(str(aug_model),": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1]) + filename = "{}-{} epochs (dataug:{})- {} in_it-{}".format(str(aug_model),epochs,dataug_epoch_start,n_inner_iter, run) + with open(res_folder+"log/%s.json" % filename, "w+") as f: + try: + json.dump(out, f, indent=True) + print('Log :\"',f.name, '\" saved !') + except: + print("Failed to save logs :",f.name) + + print('Execution Time : %.00f '%(exec_time)) + print('-'*9) ''' + ### Benchmark - RandAugment ### + for model_type in model_list.keys(): + for model_name in model_list[model_type]: + for run in range(nb_run): + torch.cuda.reset_max_memory_cached() #reset_peak_stats + t0 = time.perf_counter() + + model = getattr(model_type, model_name)(pretrained=False).to(device) + + print("RandAugment(N{}-M{})-{} on {} for {} epochs".format(rand_aug['N'],rand_aug['M'],model_name, device_name, epochs)) + log= train_classic(model=model, opt_param=optim_param, epochs=epochs, print_freq=epochs/4) + + exec_time=time.perf_counter() - t0 + max_cached = torch.cuda.max_memory_cached()/(1024.0 * 1024.0) #torch.cuda.max_memory_reserved() #MB + #### + print('-'*9) + times = [x["time"] for x in log] + out = {"Accuracy": max([x["acc"] for x in log]), + "Time": (np.mean(times),np.std(times), exec_time), + 'Optimizer': optim_param, + "Device": device_name, + "Memory": max_cached, + "Rand_Aug": rand_aug, + "Log": log} + print("RandAugment-",model_name,": acc", out["Accuracy"], "in:", out["Time"][0], "+/-", out["Time"][1]) + filename = "RandAugment(N{}-M{})-{}-{} epochs -{}".format(rand_aug['N'],rand_aug['M'],model_name,epochs, run) + with open(res_folder+"log/%s.json" % filename, "w+") as f: + try: + json.dump(out, f, indent=True) + print('Log :\"',f.name, '\" saved !') + except: + print("Failed to save logs :",f.name) + + #plot_resV2(log, fig_name=res_folder+filename) + + print('Execution Time : %.00f '%(exec_time)) + print('-'*9) ### HP Search ### + ''' + from LeNet import * inner_its = [1] dist_mix = [0.0, 0.5, 0.8, 1.0] - N_seq_TF= [2, 3, 4] + N_seq_TF= [3, 2, 4] mag_setup = [(True,True), (False, False)] #(FxSh, Independant) #prob_setup = [True, False] - nb_run= 3 try: os.mkdir(res_folder) @@ -150,9 +202,10 @@ if __name__ == "__main__": p_setup=False for run in range(nb_run): - t0 = time.process_time() + t0 = time.perf_counter() - model = getattr(models.resnet, 'resnet18')(pretrained=False) + #model = getattr(models.resnet, 'resnet18')(pretrained=False) + 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) #aug_model = Augmented_model(RandAug(TF_dict=tf_dict, N_TF=2), model).to(device) @@ -168,7 +221,7 @@ if __name__ == "__main__": hp_opt=False, save_sample_freq=None) - exec_time=time.process_time() - t0 + exec_time=time.perf_counter() - t0 #### print('-'*9) times = [x["time"] for x in log] @@ -184,4 +237,4 @@ if __name__ == "__main__": print('Execution Time : %.00f '%(exec_time)) print('-'*9) - #''' + ''' diff --git a/higher/smart_aug/datasets.py b/higher/smart_aug/datasets.py index 1ec3241..de81d5d 100755 --- a/higher/smart_aug/datasets.py +++ b/higher/smart_aug/datasets.py @@ -1,6 +1,6 @@ """ Dataset definition. - MNIST / CIFAR10 + MNIST / CIFAR10 / CIFAR100 / SVHN / ImageNet """ import torch from torch.utils.data.dataset import ConcatDataset @@ -37,9 +37,16 @@ transform_train = torchvision.transforms.Compose([ #transforms.RandomVerticalFlip(), torchvision.transforms.ToTensor(), ]) -#from RandAugment import RandAugment + +## RandAugment ## +from RandAugment import RandAugment # Add RandAugment with N, M(hyperparameter) -#transform_train.transforms.insert(0, RandAugment(n=2, m=30)) +rand_aug={'N': 2, 'M': 1} +#rand_aug={'N': 2, 'M': 9./30} #RN-ImageNet +#rand_aug={'N': 3, 'M': 5./30} #WRN-CIFAR10 +#rand_aug={'N': 2, 'M': 14./30} #WRN-CIFAR100 +#rand_aug={'N': 3, 'M': 7./30} #WRN-SVHN +transform_train.transforms.insert(0, RandAugment(n=rand_aug['N'], m=rand_aug['M'])) ### Classic Dataset ### @@ -50,7 +57,7 @@ transform_train = torchvision.transforms.Compose([ #CIFAR data_train = torchvision.datasets.CIFAR10(dataroot, train=True, download=download_data, transform=transform_train) -#data_val = torchvision.datasets.CIFAR10(dataroot, train=True, download=download_data, transform=transform) +data_val = torchvision.datasets.CIFAR10(dataroot, train=True, download=download_data, transform=transform) data_test = torchvision.datasets.CIFAR10(dataroot, train=False, download=download_data, transform=transform) #data_train = torchvision.datasets.CIFAR100(dataroot, train=True, download=download_data, transform=transform_train) @@ -72,32 +79,18 @@ data_test = torchvision.datasets.CIFAR10(dataroot, train=False, download=downloa #Validation set size [0, 1] valid_size=0.1 -#train_subset_indices=range(int(len(data_train)*(1-valid_size))) -#val_subset_indices=range(int(len(data_train)*(1-valid_size)),len(data_train)) +train_subset_indices=range(int(len(data_train)*(1-valid_size))) +val_subset_indices=range(int(len(data_train)*(1-valid_size)),len(data_train)) #train_subset_indices=range(BATCH_SIZE*10) #val_subset_indices=range(BATCH_SIZE*10, BATCH_SIZE*20) -#from torch.utils.data import SubsetRandomSampler -#dl_train = torch.utils.data.DataLoader(data_train, batch_size=BATCH_SIZE, shuffle=False, sampler=SubsetRandomSampler(train_subset_indices), num_workers=num_workers, pin_memory=pin_memory) -#dl_val = torch.utils.data.DataLoader(data_train, batch_size=BATCH_SIZE, shuffle=False, sampler=SubsetRandomSampler(val_subset_indices), num_workers=num_workers, pin_memory=pin_memory) +from torch.utils.data import SubsetRandomSampler +dl_train = torch.utils.data.DataLoader(data_train, batch_size=BATCH_SIZE, shuffle=False, sampler=SubsetRandomSampler(train_subset_indices), num_workers=num_workers, pin_memory=pin_memory) +dl_val = torch.utils.data.DataLoader(data_val, batch_size=BATCH_SIZE, shuffle=False, sampler=SubsetRandomSampler(val_subset_indices), num_workers=num_workers, pin_memory=pin_memory) dl_test = torch.utils.data.DataLoader(data_test, batch_size=TEST_SIZE, shuffle=False, num_workers=num_workers, pin_memory=pin_memory) #Cross Validation ''' -from skorch.dataset import CVSplit -import numpy as np -cvs = CVSplit(cv=valid_size, stratified=True) #Stratified =True for unbalanced dataset #ShuffleSplit - -def next_CVSplit(): - - train_subset, val_subset = cvs(data_train, y=np.array(data_train.targets)) - dl_train = torch.utils.data.DataLoader(train_subset, batch_size=BATCH_SIZE, shuffle=True, num_workers=num_workers, pin_memory=pin_memory) - dl_val = torch.utils.data.DataLoader(val_subset, batch_size=BATCH_SIZE, shuffle=True, num_workers=num_workers, pin_memory=pin_memory) - - return dl_train, dl_val - -dl_train, dl_val = next_CVSplit() -''' import numpy as np from sklearn.model_selection import ShuffleSplit from sklearn.model_selection import StratifiedShuffleSplit @@ -134,7 +127,7 @@ class CVSplit(object): else: cv_cls = ShuffleSplit - self._cv= cv_cls(test_size=val_size, random_state=0) + self._cv= cv_cls(test_size=val_size, random_state=0) #Random state w/ fixed seed def next_split(self): """ Get next cross-validation split. @@ -157,4 +150,21 @@ class CVSplit(object): return dl_train, dl_val cvs = CVSplit(data_train, val_size=valid_size) -dl_train, dl_val = cvs.next_split() \ No newline at end of file +dl_train, dl_val = cvs.next_split() +''' + +''' +from skorch.dataset import CVSplit +import numpy as np +cvs = CVSplit(cv=valid_size, stratified=True) #Stratified =True for unbalanced dataset #ShuffleSplit + +def next_CVSplit(): + + train_subset, val_subset = cvs(data_train, y=np.array(data_train.targets)) + dl_train = torch.utils.data.DataLoader(train_subset, batch_size=BATCH_SIZE, shuffle=True, num_workers=num_workers, pin_memory=pin_memory) + dl_val = torch.utils.data.DataLoader(val_subset, batch_size=BATCH_SIZE, shuffle=True, num_workers=num_workers, pin_memory=pin_memory) + + return dl_train, dl_val + +dl_train, dl_val = next_CVSplit() +''' \ No newline at end of file