fixxed some bugs and typos

drawrect.py

@@ -2,6 +2,7 @@
 from dv.init import *
 from dv.MyImageFolderWithPaths import CarsDataset
 from dv.DFL import DFL_ResNet50
+from dv.util import *
 from train import *
 from validate import *
 import sys
@@ -21,7 +22,6 @@
 import os
 import re
 import numpy as np
-from utils.util import *
 
 parser = argparse.ArgumentParser(description='Discriminative Filter Learning within a CNN')
 parser.add_argument('--dataroot', default = './dataset', metavar='DIR',

mDFL-CNN/utils/save.py → dv/save.py

@@ -2,7 +2,8 @@
 import os
 import shutil
 import datetime
-from utils.util import *
+from dv.util import *
+
 class Log(object):
     def save_train_info(self, epoch, batch, maxbatch, losses, top1, top5):
         """
@@ -13,10 +14,10 @@ def save_train_info(self, epoch, batch, maxbatch, losses, top1, top5):
         loss2 = losses[2]
         loss3 = losses[3]
         root_dir = os.path.abspath('./')
-        log_dir = os.path.join(root_dir, 'log') 
+        log_dir = os.path.join(root_dir, 'log')
         if not os.path.exists(log_dir):
             os.mkdir(log_dir)
-        
+
         log_file = os.path.join(log_dir, 'log_train.txt')
         if not os.path.exists(log_file):
             os.mknod(log_file)
@@ -30,11 +31,11 @@ def save_train_info(self, epoch, batch, maxbatch, losses, top1, top5):
                     'Prec@1 ({top1.avg:.3f})\t'
                     'Prec@5 ({top5.avg:.3f})\n'.format(epoch, batch, maxbatch,loss = loss,loss1 = loss1,loss2 = loss2, loss3=loss3, top1=top1, top5=top5))
 
-    
+
     def save_test_info(self, epoch, top1, top5):
         root_dir = os.path.abspath('./')
-        log_dir = os.path.join(root_dir, 'log') 
-        # check log_dir 
+        log_dir = os.path.join(root_dir, 'log')
+        # check log_dir
         if not os.path.exists(log_dir):
             os.mkdir(log_dir)
         log_file = os.path.join(log_dir, 'log_test.txt')
@@ -48,33 +49,28 @@ def save_test_info(self, epoch, top1, top5):
 # this is for weight
 def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
     """[summary]
-    
+
     [description]
-    
+
     Arguments:
         state {[type]} -- [description] a dict describe some params
         is_best {bool} -- [description] a bool value
-    
+
     Keyword Arguments:
         filename {str} -- [description] (default: {'checkpoint.pth.tar'})
     """
     root_dir = get_root_dir()
     weight_dir = os.path.join(root_dir, 'weight')
     if not os.path.exists(weight_dir):
         os.mkdir(weight_dir)
-    
+
     epoch = state['epoch']
     prec1 = state['prec1']
 
-    file_path = os.path.join(weight_dir, 'epoch_{:04d}_top1_{:02d}_{}'.format(int(epoch), int(prec1), filename))  
+    file_path = os.path.join(weight_dir, 'epoch_{:04d}_top1_{:02d}_{}'.format(int(epoch), int(prec1), filename))
     torch.save(state, file_path)
-    
+
     best_path = os.path.join(weight_dir, 'model_best.pth.tar')
-    
+
     if is_best:
         shutil.copyfile(file_path, best_path)
-
-
-
-
-

main.py

@@ -2,9 +2,10 @@
 from dv.init import *
 from dv.MyImageFolderWithPaths import CarsDataset, CUB_2011
 from dv.transform import *
-from utils.util import *
+from dv.util import *
 from train import *
 from validate import *
+from drawrect import *
 import sys
 import argparse
 import os
@@ -18,7 +19,6 @@
 import torch.utils.data
 import torchvision.transforms as transforms
 import torchvision.datasets as datasets
-from drawrect import *
 
 parser = argparse.ArgumentParser(description='Discriminative Filter Learning within a CNN')
 parser.add_argument('--dataroot', metavar='DIR',
@@ -174,15 +174,15 @@ def main():
         train_dataset, batch_size=args.gpu * args.train_batchsize_per_gpu, shuffle=True,
         num_workers=args.workers, pin_memory=True, drop_last = False)
     test_loader = torch.utils.data.DataLoader(
-        test_dataset, batch_size=args.testbatch, shuffle=True,
+        test_dataset, batch_size=args.test_batch_size, shuffle=True,
         num_workers=args.workers, pin_memory=True, drop_last = False)
     test_loader_simple = torch.utils.data.DataLoader(
-        test_dataset_simple, batch_size=args.testbatch, shuffle=True,
+        test_dataset_simple, batch_size=args.test_batch_size, shuffle=True,
         num_workers=args.workers, pin_memory=True, drop_last = False)
     print('Deep Vision <==> Part3 : loading dataset  <==> Done')
 
 
-    print('DFL-CNN <==> Part4 : model training  <==> Begin')
+    print('Deep Vision <==> Part4 : model training  <==> Begin')
 
     if args.gpu is not None:
         torch.cuda.empty_cache()

run.sh

@@ -0,0 +1,15 @@
+CUDA_VISIBLE_DEVICES=0   python main.py \
+--gpu 1 \
+--dataroot '/export/home/dv/dv029/DFL-CNN/dataset' \
+--result './vis_result' \
+--lr 0.005 \
+--epochs 40 \
+--workers 4 \
+--train_batchsize_per_gpu 8 \
+--init_type 'xavier' \
+--print-freq 1 \
+--vis_epoch 1 \
+--eval_epoch 1 \
+--num_filters 4 \
+--nclass 196 \
+--dataset 'cars'

train.py

@@ -1,12 +1,12 @@
+from dv.util import *
+from dv.save import *
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from PIL import Image
 import os
 import sys
 import time
-from utils.util import *
-from utils.save import *
 from torchvision import datasets, transforms, utils
 import torchvision.models as models
 import numpy as np
@@ -18,17 +18,17 @@ def train(args, train_loader, model, criterion, optimizer, epoch):
     top1 = AverageMeter()
     top5 = AverageMeter()
     log = Log()
-    
+
     losses1 = AverageMeter()
     losses2 = AverageMeter()
     losses3 = AverageMeter()
     # switch to train mode
     model.train()
 
     for i, (data, target) in enumerate(train_loader):
-        
+
         target = target.type(torch.LongTensor)
-        
+
         if args.gpu is not None:
             device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
             data = data.to(device)
@@ -40,9 +40,9 @@ def train(args, train_loader, model, criterion, optimizer, epoch):
         loss1 = criterion(out1, target)
         loss2 = criterion(out2, target)
         loss3 = criterion(out3, target)
-        
+
         loss = loss1 + loss2 + 0.1 * loss3
-        
+
         # measure accuracy and record loss
         prec1, prec5 = accuracy(out, target, topk=(1, 5))  # this is metric on trainset
         batchsize = data.size(0)
@@ -56,12 +56,12 @@ def train(args, train_loader, model, criterion, optimizer, epoch):
         losses3.update(loss3.item(), batchsize)
         top1.update(prec1[0], batchsize)
         top5.update(prec5[0], batchsize)
-        
+
         # compute gradient and do SGD step
         optimizer.zero_grad()
         loss.backward()
         optimizer.step()
-        
+
         if i % args.print_freq == 0:
             print('DFL-CNN <==> Train Epoch: [{0}][{1}/{2}]\n'
                 'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
@@ -71,9 +71,6 @@ def train(args, train_loader, model, criterion, optimizer, epoch):
                 'Top1 {top1.val:.3f} ({top1.avg:.3f})\t'
                 'Top5 {top5.val:.3f} ({top5.avg:.3f})'.format(
                 epoch, i, len(train_loader), loss=losses, loss1=losses1, loss2=losses2, loss3=losses3, top1=top1, top5=top5))
-            
+
             totalloss = [losses, losses1, losses2, losses3]
             log.save_train_info(epoch, i, len(train_loader), totalloss, top1, top5)
-
-
-

validate.py

@@ -1,19 +1,19 @@
 import torch
 import time
 import sys
-from utils.util import *
-from utils.save import *
+from dv.util import *
+from dv.save import *
 from tqdm import tqdm
 
 def validate_dv(args, val_loader, model, criterion, epoch):
     print('Deep Vision module: validating on the test set...')
-    
+
     batch_time = AverageMeter()
     losses = AverageMeter()
     top1 = AverageMeter()
     top5 = AverageMeter()
     log = Log()
-    model.eval()   
+    model.eval()
     end = time.time()
 
     total_output= []
@@ -40,7 +40,7 @@ def validate_dv(args, val_loader, model, criterion, epoch):
         else:
            total_output = torch.cat((total_output, output.data.float()) , 0)
            total_label = torch.cat((total_label , target.data.float()) , 0)
-        
+
     _,predict = torch.max(total_output,1)
 
     acc = torch.sum(torch.squeeze(predict).float() == total_label).item() / float(total_label.size()[0])
@@ -54,7 +54,7 @@ def validate(args, val_loader, model, criterion, epoch):
     top1 = AverageMeter()
     top5 = AverageMeter()
     log = Log()
-    model.eval()   
+    model.eval()
     end = time.time()
 
     # we may have ten d in data
@@ -88,13 +88,13 @@ def validate_simple(args, val_loader, model, criterion, epoch):
     top1 = AverageMeter()
     top5 = AverageMeter()
     log = Log()
-    model.eval()   
+    model.eval()
     end = time.time()
 
     # we may have ten d in data
     for i, (data, target) in enumerate(val_loader):
         target = target.type(torch.LongTensor)
-        
+
         if args.gpu is not None:
             device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
             data = data.to(device)
@@ -116,4 +116,4 @@ def validate_simple(args, val_loader, model, criterion, epoch):
 
     print('DFL-CNN <==> Test Total <==> Top1 {:.3f}% Top5 {:.3f}%'.format(top1.avg, top5.avg))
     log.save_test_info(epoch, top1, top5)
-    return top1.avg
+    return top1.avg