-
Notifications
You must be signed in to change notification settings - Fork 7
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
3 changed files
with
337 additions
and
1 deletion.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,29 @@ | ||
| { | ||
| "wandb_project":"YOUR_PROJECT", | ||
| "wandb_entity":"YOUR_ENTITY", | ||
| "task":"classification", | ||
| "num_classes":2, | ||
| "device":"cuda:1", | ||
| "wandb":false, | ||
| "ssl_encoder":null, | ||
| "contrastive":false, | ||
| "split_path":"./train_test_split.json", | ||
| "supervised_data_path":"/mnt/nvme1/nbountos/datasets/HephaestusSSL", | ||
| "data_path":"/mnt/nvme1/nbountos/datasets/Hephaestus_Raw", | ||
| "batch_size":64, | ||
| "num_workers":12, | ||
| "lr":0.0001, | ||
| "weight_decay":1e-4, | ||
| "epochs":10, | ||
| "architecture":"ResNet18", | ||
| "linear_evaluation":false, | ||
| "oversampling":true, | ||
| "threshold":null, | ||
| "multilabel":false, | ||
| "augment":false, | ||
| "num_channels":3, | ||
| "class_weights":[1.0,1.0], | ||
| "synthetic_probability":0, | ||
| "synth_path":"/mnt/nvme1/nbountos/datasets/synth/", | ||
| "seed":999 | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,190 @@ | ||
| import torch | ||
| import torch.nn as nn | ||
| import numpy as np | ||
| from utilities import utils as utils | ||
| import json | ||
| from tqdm import tqdm | ||
| from pathlib import Path | ||
| import random | ||
| import timm | ||
| from torchmetrics.classification import BinaryStatScores | ||
| import wandb | ||
| import pprint | ||
|
|
||
| np.random.seed(999) | ||
| torch.manual_seed(999) | ||
| random.seed(999) | ||
|
|
||
| def define_checkpoint(configs): | ||
| checkpoint_path = ( | ||
| Path("finetuning_checkpoints") / | ||
| configs["task"].lower() / | ||
| configs["architecture"].lower() | ||
| ) | ||
| checkpoint_path.mkdir(parents=True, exist_ok=True) | ||
| return checkpoint_path | ||
|
|
||
| def eval_cls(configs,loader,criterion,mode='Val',model=None, epoch=-1): | ||
| print('\n\nInitializing Evaluation: ' + mode+'\n\n\n') | ||
| if model is None: | ||
| checkpoint_path = 'best_model_task_'+configs['task']+'.pt' | ||
| model = torch.load(configs['checkpoint_path'] / checkpoint_path)['model'] | ||
| model.eval() | ||
| accuracy, fscore, precision, recall, _, average_precision = utils.initialize_metrics(configs) | ||
| metric = BinaryStatScores().to(configs['device']) | ||
| loss = 0.0 | ||
| prediction_list = [] | ||
| ground_truth = [] | ||
|
|
||
| with torch.no_grad(): | ||
| for batch in tqdm(loader): | ||
| insar, _ , label = batch | ||
| insar = insar.to(configs['device']) | ||
|
|
||
| label = label.to(configs['device']) | ||
| out = model(insar) | ||
| if configs['num_classes']==1: | ||
| label = label.unsqueeze(1) | ||
| if configs['multilabel'] or configs['num_classes']==1: | ||
| predictions = torch.sigmoid(out) | ||
| tmp_pred = predictions.clone() | ||
|
|
||
| predictions[tmp_pred>=configs['threshold']] = 1 | ||
| predictions[tmp_pred<configs['threshold']] = 0 | ||
| else: | ||
| predictions = out.argmax(1) | ||
|
|
||
| prediction_list.extend(predictions.cpu().detach().numpy()) | ||
| ground_truth.extend(label.cpu().detach().numpy()) | ||
|
|
||
| accuracy(predictions,label) | ||
| fscore(predictions,label) | ||
| precision(predictions,label) | ||
| recall(predictions,label) | ||
| average_precision(out[:,1].float(), label) | ||
|
|
||
| metric(predictions, label) | ||
| loss += criterion(out,label) | ||
| print('TP, FP, TN, FN, Support') | ||
| print(metric.compute()) | ||
|
|
||
| if configs['wandb']: | ||
| wandb.log({mode+ ' F-Score':fscore.compute(),mode+' Acc':accuracy.compute(),mode + ' Precision':precision.compute(),mode + ' Recall': recall.compute(), mode + ' Avg. Precision':average_precision.compute(), mode + 'Loss: ':loss/len(loader.dataset)}) | ||
| print({mode+ ' F-Score':fscore.compute(),mode+' Acc':accuracy.compute(),mode + ' Precision':precision.compute(),mode + ' Recall': recall.compute(), mode + ' Avg. Precision':average_precision.compute()}, mode + 'Loss: ',loss/len(loader.dataset),'Epoch: ',epoch) | ||
|
|
||
| return fscore.compute() | ||
|
|
||
|
|
||
| def train_cls(configs): | ||
| train_loader, val_loader, test_loader = utils.prepare_supervised_learning_loaders(configs) | ||
|
|
||
| accuracy, fscore, precision, recall, _, avg = utils.initialize_metrics(configs) | ||
| class_weights = torch.tensor(configs['class_weights']).to(configs['device']) | ||
| if configs['num_classes']>=2 or configs['num_classes']==1: | ||
| criterion = nn.CrossEntropyLoss(weight=class_weights) | ||
| else: | ||
| criterion = nn.BCEWithLogitsLoss() | ||
|
|
||
| if configs['ssl_encoder'] is None: | ||
| print('='*20) | ||
| print('Creating model pretrained on Imagenet') | ||
| print('='*20) | ||
| model = timm.create_model(configs['architecture'].lower(),num_classes=configs['num_classes'],pretrained=True) | ||
| if configs['linear_evaluation']: | ||
| for param in model.parameters(): | ||
| param.requires_grad = False | ||
| model.eval() | ||
| if 'vit' in configs['architecture'] or 'swin' in configs['architecture']: | ||
| model.head = nn.Linear(model.head.in_features,configs['num_classes']) | ||
| else: | ||
| model.fc = nn.Linear(model.fc.in_features,configs['num_classes']) | ||
| else: | ||
| print('='*20) | ||
| print('Creating SSL model pretrained on Hephaestus') | ||
| print('='*20) | ||
|
|
||
| model = torch.load('pretrained_encoder_resnet50.pt',map_location=configs['device']) | ||
| out_dim = 2048 | ||
|
|
||
| if configs['linear_evaluation']: | ||
| for param in model.parameters(): | ||
| param.requires_grad = False | ||
| model.eval() | ||
| else: | ||
| for param in model.parameters(): | ||
| param.requires_grad = False | ||
|
|
||
| model.fc = nn.Linear(out_dim,configs['num_classes']) | ||
| model.to(configs['device']) | ||
| #print(model) | ||
| optimizer = torch.optim.AdamW(model.parameters(),lr=configs['lr'],weight_decay=configs['weight_decay']) | ||
|
|
||
| best_val = 0.0 | ||
| best_epoch = 0.0 | ||
| best_stats = {} | ||
| for epoch in range(configs['epochs']): | ||
| for idx, batch in tqdm(enumerate(train_loader)): | ||
| if idx>200: | ||
| break | ||
| optimizer.zero_grad() | ||
| if not configs['linear_evaluation']: | ||
| model.train() | ||
| insar, _, label = batch | ||
| insar = insar.to(configs['device']) | ||
| label = label.to(configs['device']) | ||
| if configs['num_classes']==1: | ||
| label = label.unsqueeze(1) | ||
| out = model(insar) | ||
| if configs['multilabel'] or configs['num_classes']==1: | ||
| predictions = torch.sigmoid(out) | ||
| tmp_pred = predictions.clone() | ||
| predictions[tmp_pred>=configs['threshold']] = 1 | ||
| predictions[tmp_pred<configs['threshold']] = 0 | ||
| else: | ||
| predictions = out.argmax(1) | ||
| accuracy(predictions,label) | ||
|
|
||
| loss = criterion(out,label) | ||
| loss.backward() | ||
| optimizer.step() | ||
|
|
||
| if idx%100 == 0: | ||
| log_dict = {'Loss: ':loss.mean().item(),' Train accuracy: ':accuracy.compute(),'Epoch':epoch} | ||
| print(log_dict) | ||
| if configs['wandb']: | ||
| wandb.log(log_dict) | ||
|
|
||
| val_loss = eval_cls(configs,val_loader,criterion,model=model, epoch=epoch) | ||
| if val_loss >= best_val: | ||
| best_val = val_loss | ||
| best_epoch = epoch | ||
| best_stats = {'val_loss':best_val,'epoch':best_epoch} | ||
| print('New Best model: ',best_stats) | ||
|
|
||
| model_path = 'best_model_task_'+configs['task']+'.pt' | ||
| print('Saving model to: ',configs['checkpoint_path'] / model_path ) | ||
| torch.save({'model':model,'stats':best_stats},configs['checkpoint_path'] / model_path ) | ||
|
|
||
| print('='*20) | ||
| print('Start Testing') | ||
| print('='*20) | ||
| eval_cls(configs,test_loader,criterion,mode='Test') | ||
|
|
||
| def train_segmentation(configs): | ||
| pass | ||
|
|
||
| if __name__== '__main__': | ||
|
|
||
| config_path = 'configs/supervised_configs.json' | ||
| configs =json.load(open(config_path,'r')) | ||
| augmentation_cfg = json.load(open("configs/augmentations/augmentation.json", "r")) | ||
| configs.update(augmentation_cfg) | ||
| configs['checkpoint_path'] = define_checkpoint(configs) | ||
| if configs['wandb']: | ||
| wandb.init( | ||
| project=configs['wandb_project'], | ||
| entity=configs['wandb_entity'], | ||
| config=configs | ||
| ) | ||
| pprint.pprint(configs) | ||
| train_cls(configs) |