diff --git a/README.md b/README.md
new file mode 100644
index 0000000..ffd11ef
--- /dev/null
+++ b/README.md
@@ -0,0 +1,16 @@
+# L3
+
+To use:
+- download the data from http://nlp.stanford.edu/data/muj/shapeworld_4k.tar.gz, and untar inside `data` folder, to create `data/shapeworld` folder
+- use python 2, e.g.
+```
+virtualenv -p python2 .venv
+source .venv/bin/activate
+```
+- install requirements.txt, e.g.
+```
+pip install -r requirements.txt
+```
+- choose a command-line from `exp` folder, eg `cls_hint` is for L3
+    - `ex` is Meta, `gold` is ground-truth descriptions, `hint` is L^3, `joint` is Meta+Joint, `sim` is another baseline that didn't make it into the paper, and `vis` is just visualization.
+    - For the other two experiments, evaluation is done in a different script from training; those have the `_eval` suffix.
diff --git a/models.py b/models.py
index 9d8d015..1329907 100644
--- a/models.py
+++ b/models.py
@@ -5,8 +5,10 @@
 import gflags
 import numpy as np
 import sys
-import tensorflow as tf
 import os
+import tensorflow.compat.v1 as tf
+
+tf.disable_v2_behavior()
 
 FLAGS = gflags.FLAGS
 
@@ -41,7 +43,7 @@ def _set_flags():
 def _encode(name, t_input, t_len, t_vecs, t_init=None):
     multi = len(t_input.get_shape()) == 3
     assert multi or len(t_input.get_shape()) == 2
-    cell = tf.contrib.rnn.GRUCell(N_HIDDEN)
+    cell = tf.compat.v1.nn.rnn_cell.GRUCell(N_HIDDEN)
     if multi:
         t_shape = tf.shape(t_input)
         t_n_batch, t_n_multi, t_n_toks = t_shape[0], t_shape[1], t_shape[2]
@@ -108,7 +110,7 @@ def __init__(self, name, t_init, t_target, t_last, t_last_hidden, t_vecs):
         multi = len(t_init.get_shape()) == 3
 
         assert multi or len(t_init.get_shape()) == 2
-        cell = tf.contrib.rnn.GRUCell(N_HIDDEN)
+        cell = tf.compat.v1.nn.rnn_cell.GRUCell(N_HIDDEN)
         if multi:
             t_shape = tf.shape(t_target)
             t_n_batch, t_n_multi, t_n_toks = t_shape[0], t_shape[1], t_shape[2]
diff --git a/net.py b/net.py
index 1a1105f..fac0316 100644
--- a/net.py
+++ b/net.py
@@ -1,4 +1,6 @@
-import tensorflow as tf
+import tensorflow.compat.v1 as tf
+
+tf.disable_v2_behavior()
 
 INIT_SCALE = 1.43
 
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..b5760da
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,3 @@
+tensorflow~=2.1.0
+pillow
+python-gflags
diff --git a/tasks/shapes.py b/tasks/shapes.py
index 2b5a427..f18fe00 100644
--- a/tasks/shapes.py
+++ b/tasks/shapes.py
@@ -42,9 +42,9 @@ def __init__(self):
 
         data = {}
         for fold in ("train", "val", "test", "val_same", "test_same"):
-            examples = np.load(os.path.join(sw_path, fold, "examples.npy"))
-            inputs = np.load(os.path.join(sw_path, fold, "inputs.npy"))
-            labels = np.load(os.path.join(sw_path, fold, "labels.npy"))
+            examples = np.load(os.path.join(sw_path, fold, "examples.npz"))['arr_0']
+            inputs = np.load(os.path.join(sw_path, fold, "inputs.npz"))['arr_0']
+            labels = np.load(os.path.join(sw_path, fold, "labels.npz"))['arr_0']
 
             with open(os.path.join(sw_path, fold, "hints.json")) as hint_f:
                 hints = json.load(hint_f)
@@ -82,8 +82,8 @@ def __init__(self):
             #            i_feat = self.feature_index[tuple(feature)]
             #            if i_feat is not None:
             #                inp_features[i_datum, i_feat] = 1
-            ex_features = np.load(os.path.join(sw_path, fold, "examples.feats.npy"))
-            inp_features = np.load(os.path.join(sw_path, fold, "inputs.feats.npy"))
+            ex_features = np.load(os.path.join(sw_path, fold, "examples.feats.npz"))['arr_0']
+            inp_features = np.load(os.path.join(sw_path, fold, "inputs.feats.npz"))['arr_0']
 
             fold_data = []