diff --git a/README.md b/README.md
index 181968b..66d8259 100644
--- a/README.md
+++ b/README.md
@@ -1,6 +1,6 @@
# Visual Question Reasoning on General Dependency Tree
-This is the code for the paper
+This is the code for the paper on CLEVR
**Visual Question Reasoning on General Dependency Tree**
@@ -28,11 +28,36 @@ If you find this code useful in your research then please cite
}
```
+# Requirement
+ * pytorch 0.3.0.post4
+ * python 3.5
+ * tensorboardX
+ * skimage
+ * scipy
+ * numpy
+ * torchvision
+ * h5py
+ * tqdm
+
+# Data Preprocessing
+Before you can train any models, you need to download the datasets; you also need to preprocess questions, and extract features for the images.
+
+## Step 1: Download the data
+You can follow the [instructions](https://github.com/jcjohnson/clevr-iep/blob/master/TRAINING.md) to download CLEVR dataset. Or you can skip this step by using our preprossed questions data(Step 2) and extracted image features(Step 3).
+
+## Step 2: Preprocess Questions
+Codes for preprocessing would be available soon. For now you can download our preprocessed data with the following commands:
+```
+sh ./data/download_preprocessed_questions.sh
+```
+## Step 3: Extract Image Features
+Codes for image feature extraction and our extracted image features would be available soon. For now you can follow the [instructions](https://github.com/jcjohnson/clevr-iep/blob/master/TRAINING.md) to extract image features.
+We assume the extracted features `features_train.h5`, `features_val.h5`, `features_test.h5` are placed in `./data/clevr/clevr_res101/`.
# TODO
```
-1. data preprocess
-2. image feature extract
+1. data preprocess code
+2. upload image feature
3. train and evaluation
```
\ No newline at end of file
diff --git a/data/download_preprocessed_questions.sh b/data/download_preprocessed_questions.sh
new file mode 100644
index 0000000..e69de29
diff --git a/requrements.txt b/requrements.txt
index aa9a784..2902ffb 100644
--- a/requrements.txt
+++ b/requrements.txt
@@ -3,7 +3,6 @@ tensorboardX
skimage
scipy
numpy
-scipy
torchvision
h5py
tqdm
\ No newline at end of file