diff --git a/tf2jax/_src/numpy_compat.py b/tf2jax/_src/numpy_compat.py
index d52acf3..ecf0589 100644
--- a/tf2jax/_src/numpy_compat.py
+++ b/tf2jax/_src/numpy_compat.py
@@ -227,12 +227,22 @@ def broadcast_to(arr, shape):
 flip = lambda arr, axis: _get_np(arr).flip(arr, axis=axis)
 roll = lambda arr, shift, axis: _get_np(arr).roll(arr, shift=shift, axis=axis)
 split = lambda arr, sections, axis: _get_np(arr).split(arr, sections, axis=axis)
-squeeze = lambda arr, axis: _get_np(arr).squeeze(arr, axis=axis)
 stack = lambda arrs, axis: _get_np(*arrs).stack(arrs, axis=axis)
 tile = lambda arr, reps: _get_np(arr, reps).tile(arr, reps=reps)
 where = lambda cond, x, y: _get_np(cond, x, y).where(cond, x, y)
 
 
+def squeeze(arr, axis):
+  # tf.squeeze and np/jnp.squeeze have different behaviors when axis=().
+  # - tf.squeeze will squeeze all dimensions.
+  # - np/jnp.squeeze will not squeeze any dimensions.
+  # Here we change () to None to ensure that squeeze has the same behavior
+  # when converted from tf to np/jnp.
+  if axis == tuple():
+    axis = None
+  return _get_np(arr).squeeze(arr, axis=axis)
+
+
 def moveaxis(
     arr,
     source: Union[int, Sequence[int]],
diff --git a/tf2jax/_src/ops_test.py b/tf2jax/_src/ops_test.py
index 0d96158..7f40fd8 100644
--- a/tf2jax/_src/ops_test.py
+++ b/tf2jax/_src/ops_test.py
@@ -1458,8 +1458,9 @@ def roll_static():
     self._test_convert(roll_static, [])
 
   @chex.variants(with_jit=True, without_jit=True)
-  def test_squeeze(self):
-    inputs, dims = np.array([[[42], [47]]]), (0, 2)
+  @parameterized.parameters(((0, 2),), (tuple(),), (None,))
+  def test_squeeze(self, dims):
+    inputs = np.array([[[42], [47]]])
 
     def squeeze(x):
       return tf.raw_ops.Squeeze(input=x, axis=dims)