revert: Allow 0-dimensional shape tensors (#972)

* Allow for arbitrary shape tensors, including 0-dimensional ("empty") shape tensors (floats)
   - Reverts PR #413
* Enforce optimizers return objective function as a scalar to harmonize return type
* Revert tests that enforce minimum shape
* Add tests for return structure shape of pyhf.infer.mle.fit
* Add docstring examples for astensor

Author: matthewfeickert

src/pyhf/infer/mle.py

@@ -40,7 +40,7 @@ def fit(data, pdf, init_pars=None, par_bounds=None, **kwargs):
         >>> bestfit_pars
         array([0.        , 1.0030512 , 0.96266961])
         >>> twice_nll
-        array([24.98393521])
+        array(24.98393521)
         >>> -2 * model.logpdf(bestfit_pars, data) == twice_nll
         array([ True])
 
@@ -86,7 +86,7 @@ def fixed_poi_fit(poi_val, data, pdf, init_pars=None, par_bounds=None, **kwargs)
         >>> bestfit_pars
         array([1.        , 0.97224597, 0.87553894])
         >>> twice_nll
-        array([28.92218013])
+        array(28.92218013)
         >>> -2 * model.logpdf(bestfit_pars, data) == twice_nll
         array([ True])
 

src/pyhf/infer/test_statistics.py

@@ -58,5 +58,5 @@ def qmu(mu, data, pdf, init_pars, par_bounds):
     qmu = fixed_poi_fit_lhood_val - unconstrained_fit_lhood_val
     qmu = tensorlib.where(
         muhatbhat[pdf.config.poi_index] > mu, tensorlib.astensor(0.0), qmu
-    )[0]
+    )
     return tensorlib.clip(qmu, 0, max_value=None)

src/pyhf/optimize/opt_numpy.py

@@ -27,6 +27,6 @@ def wrap_objective(objective, data, pdf, stitch_pars, do_grad=False, jit_pieces=
     def func(pars):
         pars = tensorlib.astensor(pars)
         constrained_pars = stitch_pars(pars)
-        return objective(constrained_pars, data, pdf)
+        return objective(constrained_pars, data, pdf)[0]
 
     return func

src/pyhf/optimize/opt_pytorch.py

@@ -29,14 +29,14 @@ def func(pars):
             constrained_pars = stitch_pars(pars)
             constr_nll = objective(constrained_pars, data, pdf)
             grad = torch.autograd.grad(constr_nll, pars)[0]
-            return constr_nll.detach().numpy(), grad
+            return constr_nll.detach().numpy()[0], grad
 
     else:
 
         def func(pars):
             pars = tensorlib.astensor(pars)
             constrained_pars = stitch_pars(pars)
             constr_nll = objective(constrained_pars, data, pdf)
-            return constr_nll
+            return constr_nll[0]
 
     return func

src/pyhf/optimize/opt_tflow.py

@@ -31,13 +31,13 @@ def func(pars):
             # when tf.gather is used and this needs to be converted back to a
             # tensor to be usable as a value
             grad = tape.gradient(constr_nll, pars)
-            return constr_nll.numpy(), tf.convert_to_tensor(grad)
+            return constr_nll.numpy()[0], tf.convert_to_tensor(grad)
 
     else:
 
         def func(pars):
             pars = tensorlib.astensor(pars)
             constrained_pars = stitch_pars(pars)
-            return objective(constrained_pars, data, pdf)
+            return objective(constrained_pars, data, pdf)[0]
 
     return func

src/pyhf/tensor/jax_backend.py

@@ -152,6 +152,17 @@ def astensor(self, tensor_in, dtype='float'):
         """
         Convert to a JAX ndarray.
 
+        Example:
+
+            >>> import pyhf
+            >>> pyhf.set_backend("jax")
+            >>> tensor = pyhf.tensorlib.astensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+            >>> tensor
+            DeviceArray([[1., 2., 3.],
+                         [4., 5., 6.]], dtype=float64)
+            >>> type(tensor)
+            <class 'jax.interpreters.xla.DeviceArray'>
+
         Args:
             tensor_in (Number or Tensor): Tensor object
 
@@ -163,13 +174,8 @@ def astensor(self, tensor_in, dtype='float'):
         except KeyError:
             log.error('Invalid dtype: dtype must be float, int, or bool.')
             raise
-        tensor = np.asarray(tensor_in, dtype=dtype)
-        # Ensure non-empty tensor shape for consistency
-        try:
-            tensor.shape[0]
-        except IndexError:
-            tensor = np.reshape(tensor, [1])
-        return np.asarray(tensor, dtype=dtype)
+
+        return np.asarray(tensor_in, dtype=dtype)
 
     def sum(self, tensor_in, axis=None):
         return np.sum(tensor_in, axis=axis)

src/pyhf/tensor/numpy_backend.py

@@ -145,6 +145,17 @@ def astensor(self, tensor_in, dtype='float'):
         """
         Convert to a NumPy array.
 
+        Example:
+
+            >>> import pyhf
+            >>> pyhf.set_backend("numpy")
+            >>> tensor = pyhf.tensorlib.astensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+            >>> tensor
+            array([[1., 2., 3.],
+                   [4., 5., 6.]])
+            >>> type(tensor)
+            <class 'numpy.ndarray'>
+
         Args:
             tensor_in (Number or Tensor): Tensor object
 
@@ -157,13 +168,7 @@ def astensor(self, tensor_in, dtype='float'):
             log.error('Invalid dtype: dtype must be float, int, or bool.')
             raise
 
-        tensor = np.asarray(tensor_in, dtype=dtype)
-        # Ensure non-empty tensor shape for consistency
-        try:
-            tensor.shape[0]
-        except IndexError:
-            tensor = tensor.reshape(1)
-        return tensor
+        return np.asarray(tensor_in, dtype=dtype)
 
     def sum(self, tensor_in, axis=None):
         return np.sum(tensor_in, axis=axis)

src/pyhf/tensor/pytorch_backend.py

@@ -112,6 +112,17 @@ def astensor(self, tensor_in, dtype='float'):
         """
         Convert to a PyTorch Tensor.
 
+        Example:
+
+            >>> import pyhf
+            >>> pyhf.set_backend("pytorch")
+            >>> tensor = pyhf.tensorlib.astensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+            >>> tensor
+            tensor([[1., 2., 3.],
+                    [4., 5., 6.]])
+            >>> type(tensor)
+            <class 'torch.Tensor'>
+
         Args:
             tensor_in (Number or Tensor): Tensor object
 
@@ -124,13 +135,7 @@ def astensor(self, tensor_in, dtype='float'):
             log.error('Invalid dtype: dtype must be float, int, or bool.')
             raise
 
-        tensor = torch.as_tensor(tensor_in, dtype=dtype)
-        # Ensure non-empty tensor shape for consistency
-        try:
-            tensor.shape[0]
-        except IndexError:
-            tensor = tensor.expand(1)
-        return tensor
+        return torch.as_tensor(tensor_in, dtype=dtype)
 
     def gather(self, tensor, indices):
         return tensor[indices.type(torch.LongTensor)]
@@ -222,6 +227,7 @@ def simple_broadcast(self, *args):
             list of Tensors: The sequence broadcast together.
         """
 
+        args = [arg.view(1) if not self.shape(arg) else arg for arg in args]
         max_dim = max(map(len, args))
         try:
             assert not [arg for arg in args if 1 < len(arg) < max_dim]

src/pyhf/tensor/tensorflow_backend.py

@@ -136,6 +136,18 @@ def astensor(self, tensor_in, dtype='float'):
         """
         Convert to a TensorFlow Tensor.
 
+        Example:
+
+            >>> import pyhf
+            >>> pyhf.set_backend("tensorflow")
+            >>> tensor = pyhf.tensorlib.astensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+            >>> tensor
+            <tf.Tensor: shape=(2, 3), dtype=float32, numpy=
+            array([[1., 2., 3.],
+                   [4., 5., 6.]], dtype=float32)>
+            >>> type(tensor)
+            <class 'tensorflow.python.framework.ops.EagerTensor'>
+
         Args:
             tensor_in (Number or Tensor): Tensor object
 
@@ -156,11 +168,6 @@ def astensor(self, tensor_in, dtype='float'):
             tensor.device
         except AttributeError:
             tensor = tf.convert_to_tensor(tensor_in)
-            # Ensure non-empty tensor shape for consistency
-            try:
-                tensor.shape[0]
-            except IndexError:
-                tensor = tf.reshape(tensor, [1])
         if tensor.dtype is not dtype:
             tensor = tf.cast(tensor, dtype)
         return tensor
@@ -276,22 +283,16 @@ def simple_broadcast(self, *args):
             list of Tensors: The sequence broadcast together.
 
         """
-        max_dim = max(map(lambda arg: arg.shape[0], args))
+
+        max_dim = max(map(tf.size, args))
         try:
-            assert not [arg for arg in args if 1 < arg.shape[0] < max_dim]
+            assert not [arg for arg in args if 1 < tf.size(arg) < max_dim]
         except AssertionError as error:
             log.error(
                 'ERROR: The arguments must be of compatible size: 1 or %i', max_dim
             )
             raise error
-
-        broadcast = [
-            arg
-            if arg.shape[0] > 1
-            else tf.tile(tf.slice(arg, [0], [1]), tf.stack([max_dim]))
-            for arg in args
-        ]
-        return broadcast
+        return [tf.broadcast_to(arg, (max_dim,)) for arg in args]
 
     def einsum(self, subscripts, *operands):
         """
@@ -452,7 +453,7 @@ def normal_cdf(self, x, mu=0.0, sigma=1):
             TensorFlow Tensor: The CDF
         """
         normal = tfp.distributions.Normal(
-            self.astensor(mu, dtype='float')[0], self.astensor(sigma, dtype='float')[0],
+            self.astensor(mu, dtype='float'), self.astensor(sigma, dtype='float'),
         )
         return normal.cdf(x)
 

tests/test_backend_consistency.py

@@ -117,7 +117,7 @@ def test_hypotest_q_mu(
         q_mu = pyhf.infer.test_statistics.qmu(
             1.0, data, pdf, pdf.config.suggested_init(), pdf.config.suggested_bounds(),
         )
-        test_statistic.append(pyhf.tensorlib.tolist(q_mu))
+        test_statistic.append(q_mu)
 
     # compare to NumPy/SciPy
     test_statistic = np.array(test_statistic)

tests/test_constraints.py

@@ -187,15 +187,15 @@ def test_batched_constraints(backend):
         )
     )
     assert np.isclose(
-        result[0],
+        result,
         sum(
             [
                 default_backend.poisson_logpdf(data, rate)
                 for data, rate in zip([12, 13, 14], [12, 13, 14])
             ]
         ),
     )
-    assert result.shape == (1,)
+    assert result.shape == ()
 
     suggested_pars = [1.1] * 3 + [0.0] * 5  # 2 pois 5 norm
     constraint = poisson_constraint_combined(config)
@@ -208,15 +208,15 @@ def test_batched_constraints(backend):
         )
     )
     assert np.isclose(
-        result[0],
+        result,
         sum(
             [
                 default_backend.poisson_logpdf(data, rate)
                 for data, rate in zip([12, 13, 14], [12 * 1.1, 13 * 1.1, 14 * 1.1])
             ]
         ),
     )
-    assert result.shape == (1,)
+    assert result.shape == ()
 
     constraint = poisson_constraint_combined(config, batch_size=10)
     result = constraint.logpdf(

tests/test_infer.py

@@ -20,6 +20,36 @@ def check_uniform_type(in_list):
     )
 
 
+def test_mle_fit_default(tmpdir, hypotest_args):
+    """
+    Check that the default return structure of pyhf.infer.mle.fit is as expected
+    """
+    tb = pyhf.tensorlib
+
+    _, data, model = hypotest_args
+    kwargs = {}
+    result = pyhf.infer.mle.fit(data, model, **kwargs)
+    # bestfit_pars
+    assert isinstance(result, type(tb.astensor(result)))
+    assert pyhf.tensorlib.shape(result) == (model.config.npars,)
+
+
+def test_mle_fit_return_fitted_val(tmpdir, hypotest_args):
+    """
+    Check that the return structure of pyhf.infer.mle.fit with the
+    return_fitted_val keyword arg is as expected
+    """
+    tb = pyhf.tensorlib
+
+    _, data, model = hypotest_args
+    kwargs = {"return_fitted_val": True}
+    result = pyhf.infer.mle.fit(data, model, **kwargs)
+    # bestfit_pars, twice_nll
+    assert pyhf.tensorlib.shape(result[0]) == (model.config.npars,)
+    assert isinstance(result[0], type(tb.astensor(result[0])))
+    assert pyhf.tensorlib.shape(result[1]) == ()
+
+
 def test_hypotest_default(tmpdir, hypotest_args):
     """
     Check that the default return structure of pyhf.infer.hypotest is as expected

tests/test_optim.py

@@ -318,7 +318,7 @@ def test_optim_with_value(backend, source, spec, mu):
         return_fitted_val=True,
     )
     assert pyhf.tensorlib.tolist(result)
-    assert pyhf.tensorlib.shape(fitted_val) == (1,)
+    assert pyhf.tensorlib.shape(fitted_val) == ()
 
 
 @pytest.mark.parametrize('mu', [1.0], ids=['mu=1'])

tests/test_tensor.py

@@ -41,16 +41,16 @@ def test_simple_tensor_ops(backend):
     assert tb.tolist(tb.abs(tb.astensor([-1, -2]))) == [1, 2]
     a = tb.astensor(1)
     b = tb.astensor(2)
-    assert tb.tolist(a < b)[0] is True
-    assert tb.tolist(b < a)[0] is False
-    assert tb.tolist(a < a)[0] is False
-    assert tb.tolist(a > b)[0] is False
-    assert tb.tolist(b > a)[0] is True
-    assert tb.tolist(a > a)[0] is False
+    assert tb.tolist(a < b) is True
+    assert tb.tolist(b < a) is False
+    assert tb.tolist(a < a) is False
+    assert tb.tolist(a > b) is False
+    assert tb.tolist(b > a) is True
+    assert tb.tolist(a > a) is False
     a = tb.astensor(4)
     b = tb.astensor(5)
-    assert tb.tolist(tb.conditional((a < b)[0], lambda: a + b, lambda: a - b)) == [9]
-    assert tb.tolist(tb.conditional((a > b)[0], lambda: a + b, lambda: a - b)) == [-1]
+    assert tb.tolist(tb.conditional((a < b), lambda: a + b, lambda: a - b)) == 9.0
+    assert tb.tolist(tb.conditional((a > b), lambda: a + b, lambda: a - b)) == -1.0
 
 
 def test_complex_tensor_ops(backend):
@@ -145,10 +145,9 @@ def test_shape(backend):
     tb = pyhf.tensorlib
     assert tb.shape(tb.ones((1, 2, 3, 4, 5))) == (1, 2, 3, 4, 5)
     assert tb.shape(tb.ones((0, 0))) == (0, 0)
+    assert tb.shape(tb.astensor(1.0)) == ()
     assert tb.shape(tb.astensor([])) == (0,)
     assert tb.shape(tb.astensor([1.0])) == (1,)
-    assert tb.shape(tb.astensor(1.0)) == tb.shape(tb.astensor([1.0]))
-    assert tb.shape(tb.astensor(0.0)) == tb.shape(tb.astensor([0.0]))
     assert tb.shape(tb.astensor((1.0, 1.0))) == tb.shape(tb.astensor([1.0, 1.0]))
     assert tb.shape(tb.astensor((0.0, 0.0))) == tb.shape(tb.astensor([0.0, 0.0]))
     with pytest.raises(