Simplify LR test, train and eval pipeline

README.md

@@ -23,10 +23,10 @@ dataset
         ├── ILSVRC2012_val_00000002.JPEG
         └── ...
 ```
-If you have already downloaded the ILSVRC 2012 dataset and don't want to change its location, edit the paths in the file scripts/bash scripts/pc_info.sh to the location of your dataset.
+If you have already downloaded the ILSVRC 2012 dataset and don't want to change its location, edit the paths in the file `scripts/bash scripts/pc_info.sh` to the location of your dataset.
 
 ## Initial Setup
-Check that you have a working Python 3 and pip installation before proceeding.
+Check that you have a working cuda, Python 3 and pip installation before proceeding.
 - Clone this repository:
 ~~~
 git clone https://github.com/bmezaris/TAME
@@ -44,21 +44,23 @@ python3 -m venv ./venv
 ~~~
 pip install -r requirements.txt
 ~~~
-Note: you may have to install the libraries torch and torchvision separately. Follow the pip instructions [here](https://pytorch.org/get-started/locally/).
+> _**Note**_: you may have to install the libraries torch and torchvision separately. Follow the pip instructions [here](https://pytorch.org/get-started/locally/). Also, make sure you have a working matplotlib gui backend.
 
 ## Usage
 You can generate explanation maps with the pretrained attention module:
 ~~~
 cd "TAME/scripts/bash scripts" 
 . get_mask.sh <image_name> <label>
 ~~~
-The image will be searched in the images directory. `image_name` should contain the file extension as well. If no `image_name` and `label` are provided, runs default test with image `162_166.JPEG` and label `162`.
+The image will be searched in the `TAME/images` directory. `image_name` should contain the file extension as well. If no `image_name` and `label` are provided, runs default test with image `TAME/images/162_166.JPEG` and label `162`.
+> _**Note**_:
+If you want to load a model from a different directory, add the path to the checkpoint directory on the flag `--restore-dir`
 
 ## Training
-- To train TAME on VGG-16 or ResNet-50 from scratch, run:
+- To train TAME on VGG-16 or ResNet-50 from scratch, and generate the evaluation metrics AD, IC for the 100%, 50% and 15% masks, run:
 ~~~
 cd "scripts/bash scripts"
-. job.sh 
+. train_eval.sh 
 ~~~
 
 **OR**, for the ResNet-50 backbone with the selected loss function:

scripts/bash scripts/experiment_script.sh

@@ -0,0 +1,36 @@
+#!/bin/sh
+
+
+
+(
+cd ../
+
+
+CUDA_VISIBLE_DEVICES=0 python train_script.py \
+	--img-dir=$IMGDIR \
+	--restore-dir=$RESTORE \
+	--train-list=${TRAIN:=VGG16_train.txt} \
+  --num-workers=4 \
+	--model=${MODEL:=vgg16} \
+  --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} \
+
+
+CUDA_VISIBLE_DEVICES=0 python eval_script.py \
+	--val-dir=$VALDIR \
+  --restore-dir=$RESTORE \
+	--test-list=${TEST:="Evaluation_2000.txt"} \
+	--num-workers=4 \
+  --model=${MODEL:="vgg16"} \
+  --version${VERSION:="TAME"} \
+	--layers="${LAYERS:="features.16 features.23 features.30"}" \
+	--start-epoch=1 \
+	--end-epoch=${EPOCHS:=8}
+
+
+exit 0
+)

scripts/bash scripts/get_mask.sh

@@ -6,6 +6,7 @@ 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

@@ -1,23 +1,17 @@
 #!/bin/sh
 
 (
-. pc_info.sh
-
-ARCH="_f30"
-LAYERS="features.16 features.23 features.30"
-#BN="true"
-#WD="5e-4"
-MODE="all"
-#B2="0.99"
-BSIZE=8
-
-#MODEL="resnet50"
-#LAYERS="layer2 layer3 layer4"
-VER="V5"
-
 cd ../
 
-. lr_finder.sh
+CUDA_VISIBLE_DEVICES=0 python lr_finder.py \
+	--img-dir=$IMGDIR \
+	--num-workers=4 \
+  --train-list=${TRAIN:=VGG16_train.txt} \
+	--model=${MODEL:=vgg16} \
+	--layers="${LAYERS:="features.16 features.23 features.30"}" \
+	--wd=${WD:=5e-4} \
+	--version="${VERSION:="TAME"}" \
+	--batch-size=${BSIZE:=8} \
 
 exit 0
-)
+)