diff --git a/higher/compare_res.py b/higher/compare_res.py index c7ccd6e..25d5e6b 100755 --- a/higher/compare_res.py +++ b/higher/compare_res.py @@ -2,13 +2,12 @@ from utils import * if __name__ == "__main__": - ''' + #''' files=[ - #"res/good_TF_tests/log/Aug_mod(Data_augV5(Mix0.5-14TFx2-MagFxSh)-LeNet)-100 epochs (dataug:0)- 0 in_it.json", - #"res/good_TF_tests/log/Aug_mod(Data_augV5(Uniform-14TFx2-MagFxSh)-LeNet)-100 epochs (dataug:0)- 0 in_it.json", - "res/brutus-tests/log/Aug_mod(Data_augV5(Uniform-14TFx3-MagFxSh)-LeNet)-150epochs(dataug:0)-10in_it-0.json", - "res/brutus-tests/log/Aug_mod(Data_augV5(Uniform-14TFx3-MagFxSh)-LeNet)-150epochs(dataug:0)-10in_it-1.json", - "res/brutus-tests/log/Aug_mod(Data_augV5(Uniform-14TFx3-MagFxSh)-LeNet)-150epochs(dataug:0)-10in_it-2.json", + "res/log/Aug_mod(Data_augV5(Mix0.8-23TFx4-Mag)-LeNet)-100 epochs (dataug:0)- 1 in_it.json", + #"res/brutus-tests/log/Aug_mod(Data_augV5(Uniform-14TFx3-MagFxSh)-LeNet)-150epochs(dataug:0)-10in_it-0.json", + #"res/brutus-tests/log/Aug_mod(Data_augV5(Uniform-14TFx3-MagFxSh)-LeNet)-150epochs(dataug:0)-10in_it-1.json", + #"res/brutus-tests/log/Aug_mod(Data_augV5(Uniform-14TFx3-MagFxSh)-LeNet)-150epochs(dataug:0)-10in_it-2.json", #"res/log/Aug_mod(RandAugUDA(18TFx2-Mag1)-LeNet)-100 epochs (dataug:0)- 0 in_it.json", ] @@ -18,7 +17,7 @@ if __name__ == "__main__": data = json.load(json_file) plot_resV2(data['Log'], fig_name=file.replace('.json','').replace('log/',''), param_names=data['Param_names']) #plot_TF_influence(data['Log'], param_names=data['Param_names']) - ''' + #''' ## Loss , Acc, Proba = f(epoch) ## #plot_compare(filenames=files, fig_name="res/compare") @@ -78,7 +77,7 @@ if __name__ == "__main__": ''' #Res print - #''' + ''' nb_run=3 accs = [] times = [] @@ -93,4 +92,4 @@ if __name__ == "__main__": print(idx, data['Accuracy']) print(files[0], np.mean(accs), np.std(accs), np.mean(times)) - #''' \ No newline at end of file + ''' \ No newline at end of file diff --git a/higher/dataug.py b/higher/dataug.py index f492f2a..3f731ac 100755 --- a/higher/dataug.py +++ b/higher/dataug.py @@ -531,7 +531,7 @@ class Data_augV4(nn.Module): #Transformations avec mask return "Data_augV4(Mix %.1f-%d TF x %d)" % (self._mix_factor, self._nb_tf, self._N_seqTF) class Data_augV5(nn.Module): #Optimisation jointe (mag, proba) - 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=TF.TF_dict, N_TF=1, mix_dist=0.0, fixed_prob=False, fixed_mag=True, shared_mag=True, ): super(Data_augV5, self).__init__() assert len(TF_dict)>0 @@ -548,8 +548,8 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba) #self._fixed_mag=5 #[0, PARAMETER_MAX] self._params = nn.ParameterDict({ "prob": nn.Parameter(torch.ones(self._nb_tf)/self._nb_tf), #Distribution prob uniforme - "mag" : nn.Parameter(torch.tensor(float(TF.PARAMETER_MAX)) if self._shared_mag - else torch.tensor(float(TF.PARAMETER_MAX)).expand(self._nb_tf)), #[0, PARAMETER_MAX] + "mag" : nn.Parameter(torch.tensor(float(TF.PARAMETER_MAX)/2) if self._shared_mag + else torch.tensor(float(TF.PARAMETER_MAX)/2).expand(self._nb_tf)), #[0, PARAMETER_MAX] }) #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 @@ -633,7 +633,7 @@ class Data_augV5(nn.Module): #Optimisation jointe (mag, proba) 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) #Bloque une fois au extreme + self._params['mag'].data = self._params['mag'].data.clamp(min=TF.PARAMETER_MIN, max=TF.PARAMETER_MAX) #self._params['mag'].data = F.relu(self._params['mag'].data) - F.relu(self._params['mag'].data - TF.PARAMETER_MAX) def loss_weight(self): diff --git a/higher/test_brutus.py b/higher/test_brutus.py index 046df89..e488e0f 100755 --- a/higher/test_brutus.py +++ b/higher/test_brutus.py @@ -93,15 +93,15 @@ if __name__ == "__main__": json.dump(out, f, indent=True) print('Log :\"',f.name, '\" saved !') ''' - res_folder="res/brutus-tests/" + res_folder="res/brutus-tests2/" epochs= 150 inner_its = [1] dist_mix = [0.0, 0.5, 0.8, 1.0] dataug_epoch_starts= [0] tf_dict = {k: TF.TF_dict[k] for k in tf_names} TF_nb = [len(tf_dict)] #range(10,len(TF.TF_dict)+1) #[len(TF.TF_dict)] - N_seq_TF= [2, 3] - mag_setup = [(True,True), (False, False)] + N_seq_TF= [2, 3, 4] + mag_setup = [(True,True), (False, False)] #(Fixed, Shared) #prob_setup = [True, False] nb_run= 3 @@ -118,12 +118,14 @@ if __name__ == "__main__": #for i in TF_nb: for m_setup in mag_setup: #for p_setup in prob_setup: - p_setup=True + p_setup=False for run in range(nb_run): - if n_inner_iter == 0 and (m_setup!=(True,True) and p_setup!=True): continue #Autres setup inutiles sans meta-opti + if (n_inner_iter == 0 and (m_setup!=(True,True) and p_setup!=True)) or (p_setup and dist!=0.0): continue #Autres setup inutiles sans meta-opti #keys = list(TF.TF_dict.keys())[0:i] #ntf_dict = {k: TF.TF_dict[k] for k in keys} + t0 = time.process_time() + 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) @@ -143,9 +145,9 @@ if __name__ == "__main__": 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) - with open("res/log/%s.json" % filename, "w+") as f: + 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: json.dump(out, f, indent=True) print('Log :\"',f.name, '\" saved !') print('-'*9) - ''' + #''' diff --git a/higher/test_dataug.py b/higher/test_dataug.py index 137eb02..8278431 100755 --- a/higher/test_dataug.py +++ b/higher/test_dataug.py @@ -19,8 +19,8 @@ tf_names = [ 'Color', 'Brightness', 'Sharpness', - 'Posterize', - 'Solarize', #=>Image entre [0,1] #Pas opti pour des batch + #'Posterize', + #'Solarize', #=>Image entre [0,1] #Pas opti pour des batch #Color TF (Common mag scale) #'+Contrast', @@ -66,7 +66,7 @@ if __name__ == "__main__": #'aug_dataset', 'aug_model' } - n_inner_iter = 10 + n_inner_iter = 1 epochs = 100 dataug_epoch_start=0 optim_param={ @@ -168,7 +168,7 @@ if __name__ == "__main__": t0 = time.process_time() tf_dict = {k: TF.TF_dict[k] for k in tf_names} - aug_model = Augmented_model(Data_augV5(TF_dict=tf_dict, N_TF=2, mix_dist=0.0, 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=0.8, fixed_prob=False, fixed_mag=False, shared_mag=False), 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)) @@ -187,7 +187,7 @@ if __name__ == "__main__": 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) + filename = "{}-{} epochs (dataug:{})- {} in_it".format(str(aug_model),epochs,dataug_epoch_start,n_inner_iter)+"demi_mag" with open("res/log/%s.json" % filename, "w+") as f: json.dump(out, f, indent=True) print('Log :\"',f.name, '\" saved !') diff --git a/higher/utils.py b/higher/utils.py index 068d2e1..6e5675d 100755 --- a/higher/utils.py +++ b/higher/utils.py @@ -90,6 +90,7 @@ def plot_resV2(log, fig_name='res', param_names=None): ax[0, 2].set_title('Mag =f(epoch)') ax[0, 2].stackplot(epochs, mag, labels=param_names) + #ax[0, 2].plot(epochs, np.array(mag).T, label=param_names) ax[0, 2].legend(param_names, loc='center left', bbox_to_anchor=(1, 0.5)) ax[1, 2].set_title('Mag =f(TF)') diff --git a/salvador/checkpoint.pt b/salvador/checkpoint.pt index 02f739f..6252c9e 100644 Binary files a/salvador/checkpoint.pt and b/salvador/checkpoint.pt differ diff --git a/salvador/train.py b/salvador/train.py index a5cb4a5..f481d3f 100644 --- a/salvador/train.py +++ b/salvador/train.py @@ -31,12 +31,12 @@ tf_names = [ 'ShearY', ## Color TF (Expect image in the range of [0, 1]) ## - 'Contrast', - 'Color', - 'Brightness', - 'Sharpness', - 'Posterize', - 'Solarize', #=>Image entre [0,1] #Pas opti pour des batch + #'Contrast', + #'Color', + #'Brightness', + #'Sharpness', + #'Posterize', + #'Solarize', #=>Image entre [0,1] #Pas opti pour des batch ] class Lambda(nn.Module): @@ -95,6 +95,7 @@ def train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, mas unsupp_coeff = 1 loss = sup_loss + (aug_loss + KL_loss) * unsupp_coeff + #print(sup_loss.item(), (aug_loss + KL_loss).item()) optimizer.zero_grad() loss.backward() @@ -210,7 +211,7 @@ def get_train_valid_loader(args, augment, random_seed, valid_size=0.1, shuffle=T split = int(np.floor(valid_size * num_train)) if shuffle: - #np.random.seed(random_seed) + np.random.seed(random_seed) np.random.shuffle(indices) train_idx, valid_idx = indices[split:], indices[:split] @@ -277,6 +278,8 @@ def main(args): model = Augmented_model(RandAug(TF_dict=tf_dict, N_TF=2), model).to(device) if args.augment=='RandKL': Kldiv=True + + model['data_aug']['mag'].data = model['data_aug']['mag'].data * args.magnitude print("Augmodel") # model.fc = nn.Linear(model.fc.in_features, 2) @@ -294,7 +297,7 @@ def main(args): optimizer, lambda x: (1 - x / (len(data_loader) * args.epochs)) ** 0.9) - es = utils.EarlyStopping() + es = utils.EarlyStopping() if not (args.augment=='Rand' or args.augment=='RandKL') else utils.EarlyStopping(augmented_model=True) if args.test_only: model.load_state_dict(torch.load('checkpoint.pt', map_location=lambda storage, loc: storage)) @@ -324,8 +327,8 @@ def main(args): # print('Train') # print(train_confmat) - print('Valid') - print(valid_confmat) + #print('Valid') + #print(valid_confmat) # if es.early_stop: # break @@ -339,9 +342,9 @@ def parse_args(): import argparse parser = argparse.ArgumentParser(description='PyTorch Classification Training') - parser.add_argument('--data-path', default='/Salvador', help='dataset') + parser.add_argument('--data-path', default='/github/smart_augmentation/salvador/data', help='dataset') parser.add_argument('--model', default='resnet18', help='model') #'resnet18' - parser.add_argument('--device', default='cuda:1', help='device') + parser.add_argument('--device', default='cuda:0', help='device') parser.add_argument('-b', '--batch-size', default=8, type=int) parser.add_argument('--epochs', default=3, type=int, metavar='N', help='number of total epochs to run') @@ -364,6 +367,10 @@ def parse_args(): parser.add_argument('-a', '--augment', default='None', type=str, metavar='N', help='Data augment', dest='augment') + parser.add_argument('-m', '--magnitude', default=1.0, type=float, + metavar='N', help='Augmentation magnitude', + dest='magnitude') + args = parser.parse_args() diff --git a/salvador/train_dataug.py b/salvador/train_dataug.py index c8a7dde..a867167 100644 --- a/salvador/train_dataug.py +++ b/salvador/train_dataug.py @@ -549,10 +549,10 @@ def parse_args(): import argparse parser = argparse.ArgumentParser(description='PyTorch Classification Training') - parser.add_argument('--data-path', default='/Salvador', help='dataset') - parser.add_argument('--model', default='resnet50', help='model') - parser.add_argument('--device', default='cuda:1', help='device') - parser.add_argument('-b', '--batch-size', default=4, type=int) + parser.add_argument('--data-path', default='/github/smart_augmentation/salvador/data', help='dataset') + parser.add_argument('--model', default='resnet18', help='model') #'resnet18' + parser.add_argument('--device', default='cuda:0', help='device') + parser.add_argument('-b', '--batch-size', default=8, type=int) parser.add_argument('--epochs', default=3, type=int, metavar='N', help='number of total epochs to run') parser.add_argument('-j', '--workers', default=0, type=int, metavar='N', diff --git a/salvador/utils.py b/salvador/utils.py index 9b4bd0a..c8b1050 100644 --- a/salvador/utils.py +++ b/salvador/utils.py @@ -157,7 +157,7 @@ def accuracy(output, target, topk=(1,)): class EarlyStopping: """Early stops the training if validation loss doesn't improve after a given patience.""" - def __init__(self, patience=7, verbose=False, delta=0): + def __init__(self, patience=7, verbose=False, delta=0, augmented_model=False): """ Args: patience (int): How long to wait after last time validation loss improved. @@ -175,6 +175,8 @@ class EarlyStopping: self.val_loss_min = np.Inf self.delta = delta + self.augmented_model = augmented_model + def __call__(self, val_loss, model): score = -val_loss @@ -196,5 +198,5 @@ class EarlyStopping: '''Saves model when validation loss decrease.''' if self.verbose: print(f'Validation loss decreased ({self.val_loss_min:.6f} --> {val_loss:.6f}). Saving model ...') - torch.save(model.state_dict(), 'checkpoint.pt') + torch.save(model.state_dict(), 'checkpoint.pt') if not self.augmented_model else torch.save(model['model'].state_dict(), 'checkpoint.pt') self.val_loss_min = val_loss \ No newline at end of file