Finish README.md file

Co-authored-by: bmezaris <[email protected]>

.gitignore

@@ -1,3 +1,7 @@
-/scripts/pc2/
-/scripts/pc3/
-/.idea/
+.idea/
+images/heatmaps/
+dataset/ILSVRC2012_img_train
+dataset/ILSVRC2012_img_val
+snapshots/data/
+*.pyc
+venv/

requirements.txt

@@ -3,6 +3,6 @@ numpy~=1.23.2
 opencv_python~=4.6.0.66
 Pillow~=9.2.0
 scikit_learn~=1.1.2
-torch~=1.12.1+cu113
-torchvision~=0.13.1+cu113
+torch~=1.12.1
+torchvision~=0.13.1
 tqdm~=4.64.0

scripts/bash scripts/experiment_script.sh

@@ -8,28 +8,28 @@ cd ../
 
 CUDA_VISIBLE_DEVICES=0 python train_script.py \
 	--img-dir=$IMGDIR \
-	--restore-dir=$RESTORE \
+	--restore-from=$RESTORE \
 	--train-list=${TRAIN:=VGG16_train.txt} \
   --num-workers=4 \
 	--model=${MODEL:=vgg16} \
-  --version"${VERSION:="TAME"}" \
+  --version="${VERSION:="TAME"}" \
 	--layers="${LAYERS:="features.16 features.23 features.30"}" \
   --wd=${WD:=5e-4} \
 	--max-lr=${MLR:=1e-2} \
 	--epoch=${EPOCHS:=8} \
-	--batch-size=${BSIZE:=8} \
+	--batch-size=${BSIZE:=32}
 
 
 CUDA_VISIBLE_DEVICES=0 python eval_script.py \
 	--val-dir=$VALDIR \
-  --restore-dir=$RESTORE \
+  --restore-from=$RESTORE \
 	--test-list=${TEST:="Evaluation_2000.txt"} \
 	--num-workers=4 \
   --model=${MODEL:="vgg16"} \
-  --version${VERSION:="TAME"} \
+  --version=${VERSION:="TAME"} \
 	--layers="${LAYERS:="features.16 features.23 features.30"}" \
 	--start-epoch=1 \
-	--end-epoch=${EPOCHS:=8}
+	--end-epoch=${EPOCHS:=32}
 
 
 exit 0

scripts/bash scripts/get_mask_ResNet-50.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+
+(
+cd ../
+name=$1
+label=$2
+
+CUDA_VISIBLE_DEVICES=0 python masked_print.py \
+  --model=resnet50 \
+  --layers="layer2 layer3 layer4" \
+  --name=${name:="162_166.JPEG"} \
+  --label=${label:="162"}
+)

scripts/bash scripts/get_mask.sh → scripts/bash scripts/get_mask_VGG-16.sh

@@ -6,7 +6,6 @@ name=$1
 label=$2
 
 CUDA_VISIBLE_DEVICES=0 python masked_print.py \
-  --restore-dir=''\
   --name=${name:="162_166.JPEG"} \
   --label=${label:="162"}
 )

scripts/bash scripts/lr_finder.sh

@@ -11,7 +11,7 @@ CUDA_VISIBLE_DEVICES=0 python lr_finder.py \
 	--layers="${LAYERS:="features.16 features.23 features.30"}" \
 	--wd=${WD:=5e-4} \
 	--version="${VERSION:="TAME"}" \
-	--batch-size=${BSIZE:=8} \
+	--batch-size=${BSIZE:=32} \
 
 exit 0
 )

scripts/bash scripts/lr_finder_Generic.sh

@@ -0,0 +1,22 @@
+#!/bin/sh
+
+(
+. pc_info.sh
+
+export MODEL=resnet50
+export LAYERS="layer2 layer3 layer4"
+export TRAIN=ResNet50_train.txt
+
+export BSIZE=32
+
+# export VERSION="TAME"
+# export WD=
+# export LAYERS=
+# export RESTORE=
+# export TEST="Validation_2000.txt"
+
+
+. lr_finder.sh
+
+exit 0
+)

scripts/bash scripts/lr_finder_ResNet-50.sh

@@ -7,9 +7,8 @@ export MODEL=resnet50
 export LAYERS="layer2 layer3 layer4"
 export TRAIN=ResNet50_train.txt
 
-export EPOCHS=8
+export BSIZE=32
 
-# export BSIZE=8
 # export VERSION="TAME"
 # export WD=
 # export LAYERS=

scripts/bash scripts/lr_finder_VGG-16.sh

@@ -3,9 +3,8 @@
 (
 . pc_info.sh
 
-export EPOCHS=8
+export BSIZE=32
 
-# export BSIZE=8
 # export VERSION="TAME"
 # export WD=
 # export LAYERS=

scripts/bash scripts/pc_info.sh

@@ -1,4 +1,4 @@
 #!/bin/sh
 
-export IMGDIR="../../dataset/ILSVRC2012_img_train"
-export VALDIR="../../dataset/ILSVRC2012_img_val"
+export IMGDIR="dataset/ILSVRC2012_img_train"
+export VALDIR="dataset/ILSVRC2012_img_val"

scripts/bash scripts/train_eval_Generic.sh

@@ -9,8 +9,8 @@ export TRAIN=ResNet50_train.txt
 
 export MLR=1e-4
 export EPOCHS=8
+export BSIZE=32
 
-# export BSIZE=8
 # export VERSION="TAME"
 # export WD=
 # export LAYERS=

scripts/bash scripts/train_eval_ResNet-50.sh

@@ -8,9 +8,9 @@ export LAYERS="layer2 layer3 layer4"
 export TRAIN=ResNet50_train.txt
 
 export MLR=1e-4
-export EPOCHS=8
+export EPOCHS=4
+export BSIZE=32
 
-# export BSIZE=8
 # export VERSION="TAME"
 # export WD=
 # export LAYERS=

scripts/bash scripts/train_eval_VGG-16.sh

@@ -3,10 +3,10 @@
 (
 . pc_info.sh
 
-export MLR="1e-3"
+export MLR=1e-3
 export EPOCHS=8
+export BSIZE=32
 
-# export BSIZE=8
 # export VERSION="TAME"
 # export WD=
 # export LAYERS=

scripts/eval_script.py

@@ -1,20 +1,20 @@
 # checked, should be working correctly
 
 import argparse
+import json
 import os
 
 import numpy as np
 import torch
 import torch.nn.functional as F
-import torchvision.models as models
 from tqdm import tqdm
 
-from composite_models import Generic
-import metrics
-from avg_meter import AverageMeter
-from load_data import data_loader
-from model_prep import model_prep
-from restore import restore
+from utilities import metrics
+from utilities.avg_meter import AverageMeter
+from utilities.composite_models import Generic
+from utilities.load_data import data_loader
+from utilities.model_prep import model_prep
+from utilities.restore import restore
 
 # Paths
 os.chdir('../')
@@ -34,10 +34,10 @@ def get_arguments():
     parser.add_argument("--batch-size", type=int, default=1)
     parser.add_argument("--input-size", type=int, default=256)
     parser.add_argument("--crop-size", type=int, default=224)
-    parser.add_argument("--dataset", type=str, default='imagenet')
     parser.add_argument("--num-workers", type=int, default=0)
     parser.add_argument("--model", type=str, default='vgg16')
-    parser.add_argument("--version", type=str, default='TAME')
+    parser.add_argument("--version", type=str, default='TAME',
+                        choices=['TAME', 'Noskipconnection', 'NoskipNobatchnorm', 'Sigmoidinfeaturebranch'])
     parser.add_argument("--layers", type=str, default='features.16 features.23 features.30')
     parser.add_argument("--global-counter", type=int, default=0)
     parser.add_argument("--start-epoch", type=int, default=1)
@@ -55,6 +55,8 @@ def get_model(args):
 
 def main():
     args = get_arguments()
+    print('Running parameters:\n')
+    print(json.dumps(vars(args), indent=4))
 
     args.test_list = os.path.join(ROOT_DIR, 'datalist', 'ILSVRC', args.test_list)
 

scripts/lr_finder.py

@@ -8,8 +8,8 @@
 import torchvision.models as models
 from tqdm.auto import tqdm
 
-from composite_models import Generic
-from load_data import data_loader
+from utilities.composite_models import Generic
+from utilities.load_data import data_loader
 
 # Paths
 from utilities.model_prep import model_prep
@@ -37,7 +37,8 @@ def get_arguments():
     parser.add_argument("--model", type=str, default='vgg16')
     parser.add_argument("--layers", type=str, default='features.16 features.23 features.30')
     parser.add_argument("--wd", type=float, default=5e-4)
-    parser.add_argument("--version", type=str, default='TAME')
+    parser.add_argument("--version", type=str, default='TAME',
+                        choices=['TAME', 'Noskipconnection', 'NoskipNobatchnorm', 'Sigmoidinfeaturebranch'])
     return parser.parse_args()
 
 
@@ -77,56 +78,56 @@ def find_lr(args, init_value=1e-8, final_value=10., beta=0.999):
     batch_num = 0
     losses = []
     lrs = []
-    try:
-        tq_loader = tqdm(trn_loader,
-                         unit='batches',
-                         bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt} [ETA: {remaining}, {rate_fmt}{postfix}]')
-        for idx, data in enumerate(tq_loader):
-            batch_num += 1
-            # As before, get the loss for this mini-batch of inputs/outputs
-            inputs, labels = data
-            inputs, labels = inputs.cuda(), labels.cuda()
-            optimizer.zero_grad()
-
-            logits = net(inputs, labels)
-            masks = net.get_a(labels.long())
-            loss, a, b, c = net.get_loss(logits, labels, masks)
-
-            # Compute the smoothed loss
-            avg_loss = beta * avg_loss + (1-beta) * loss.item()
-            smoothed_loss = avg_loss / (1 - beta**batch_num)
-            # Stop if the loss is exploding
-            if batch_num > 1 and smoothed_loss > 4 * best_loss:
-                return lrs, losses
-            # Record the best loss
-            if smoothed_loss < best_loss or batch_num == 1:
-                best_loss = smoothed_loss
-            losses.append(smoothed_loss)
-            postfix = {'Current Loss': smoothed_loss, 'Min Loss': best_loss,
-                       'Current LR': lr}
-
-            tq_loader.set_postfix(postfix)
-
-            lrs.append(lr)
-            # Do the SGD step
-            loss.backward()
-            optimizer.step()
-            # Update the lr for the next step
-            lr *= mult
-
-            optimizer.param_groups[0]['lr'] = lr
-            optimizer.param_groups[1]['lr'] = 2 * lr
-
-    finally:
-        return lrs, losses
+
+    tq_loader = tqdm(trn_loader,
+                     unit='batches',
+                     bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt} [ETA: {remaining}, {rate_fmt}{postfix}]')
+    for idx, data in enumerate(tq_loader):
+        batch_num += 1
+        # As before, get the loss for this mini-batch of inputs/outputs
+        inputs, labels = data
+        inputs, labels = inputs.cuda(), labels.cuda()
+        optimizer.zero_grad()
+
+        logits = net(inputs, labels)
+        masks = net.get_a(labels.long())
+        loss, a, b, c = net.get_loss(logits, labels, masks)
+
+        # Compute the smoothed loss
+        avg_loss = beta * avg_loss + (1-beta) * loss.item()
+        smoothed_loss = avg_loss / (1 - beta**batch_num)
+        # Stop if the loss is exploding
+        if batch_num > 1 and smoothed_loss > 4 * best_loss:
+            return lrs, losses
+        # Record the best loss
+        if smoothed_loss < best_loss or batch_num == 1:
+            best_loss = smoothed_loss
+        losses.append(smoothed_loss)
+        postfix = {'Current Loss': smoothed_loss, 'Min Loss': best_loss,
+                   'Current LR': lr}
+
+        tq_loader.set_postfix(postfix)
+
+        lrs.append(lr)
+        # Do the SGD step
+        loss.backward()
+        optimizer.step()
+        # Update the lr for the next step
+        lr *= mult
+
+        optimizer.param_groups[0]['lr'] = lr
+        optimizer.param_groups[1]['lr'] = 2 * lr
+
+    return lrs, losses
 
 
 def main():
     args = get_arguments()
     print('Running parameters:\n')
     print(json.dumps(vars(args), indent=4))
+    args.train_list = os.path.join(ROOT_DIR, 'datalist', 'ILSVRC', args.train_list)
+    os.makedirs(data_dir, exist_ok=True)
     lrs, losses = find_lr(args)
-
     json_dir = os.path.join(args.data_dir, f'{args.model}_{args.version}_{args.batch_size}.json')
 
     with open(json_dir, mode='x') as loss_samples_file:

scripts/masked_print.py

@@ -24,12 +24,14 @@
 snapshot_dir = os.path.join(ROOT_DIR, 'snapshots')
 img_dir = os.path.join(ROOT_DIR, 'images')
 
+
 def get_arguments():
     parser = argparse.ArgumentParser(description='Generate mask for a given image-label pair with a given model')
     parser.add_argument("--snapshot_dir", type=str, default=snapshot_dir)
-    parser.add_argument("--restore-dir", type=str, default='')
+    parser.add_argument("--restore-from", type=str, default='')
     parser.add_argument("--model", type=str, default='vgg16')
-    parser.add_argument("--version", type=str, default='TAME')
+    parser.add_argument("--version", type=str, default='TAME',
+                        choices=['TAME', 'Noskipconnection', 'NoskipNobatchnorm', 'Sigmoidinfeaturebranch'])
     parser.add_argument("--layers", type=str, default='features.16 features.23 features.30')
     parser.add_argument("--name", type=str, default='162_166.JPEG')
     parser.add_argument("--label", type=int, default=162)
@@ -75,7 +77,7 @@ def main():
     im = tsfm_val(img).unsqueeze(0).cuda()
     img = tsfm(img)
     img_name = os.path.splitext(img_name)
-    img.save(os.path.join(heatmap_dir, f"{img_name[0]}_{label}{img_name[1]}"))
+    img.save(os.path.join(heatmap_dir, f"{img_name[0]}{img_name[1]}"))
     # to take care of png imgs
     im = im[:, 0:3, :, :]
 
@@ -87,7 +89,7 @@ def main():
         cam_map = model.get_c(label)
         cam_map = metrics.normalizeMinMax(cam_map)
         cam_map = F.interpolate(cam_map.unsqueeze(0), size=(224, 224), mode='bilinear', align_corners=False)
-        cam_map = metrics.drop_Npercent(cam_map, 0)
+        # cam_map = metrics.drop_Npercent(cam_map, 0)
         opencvImage = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
         mask = np.array(cam_map.squeeze().cpu().numpy() * 255, dtype=np.uint8)
         mask = cv2.applyColorMap(mask, cv2.COLORMAP_JET)
@@ -99,5 +101,3 @@ def main():
 
 if __name__ == '__main__':
     main()
-
-