Merge pull request #521 from vroulet:no_grad_eval_backtrack_linesearch

PiperOrigin-RevId: 569193154

jaxopt/_src/backtracking_linesearch.py

@@ -18,29 +18,34 @@
 from typing import Callable
 from typing import NamedTuple
 from typing import Optional
+from typing import Union
 
 from dataclasses import dataclass
 
 import jax
 import jax.numpy as jnp
 
 from jaxopt._src import base
+from jaxopt._src.cond import cond
 from jaxopt.tree_util import tree_add_scalar_mul
+from jaxopt.tree_util import tree_conj
 from jaxopt.tree_util import tree_scalar_mul
 from jaxopt.tree_util import tree_vdot_real
-from jaxopt.tree_util import tree_conj
+
 
 class BacktrackingLineSearchState(NamedTuple):
   """Named tuple containing state information."""
   iter_num: int
+  params: Any
   value: float
+  grad: Any  # either initial or final for armijo or glodstein
+  value_init: float 
+  grad_init: Any
   error: float
   done: bool
-  params: Any
-  grad: Any
+  failed: bool
   num_fun_eval: int
   num_grad_eval: int
-  failed: bool
   aux: Optional[Any] = None
 
 
@@ -56,7 +61,8 @@ class BacktrackingLineSearch(base.IterativeLineSearch):
       ``*args`` and ``**kwargs`` are additional arguments.
     value_and_grad: if ``False``, ``fun`` should return the function value only.
       If ``True``, ``fun`` should return both the function value and the
-      gradient.
+      gradient. If it is a Callable, fun should return the value while value_and_grad
+      returns value and gradient of the objective.
     has_aux: if ``False``, ``fun`` should return the function value only.
       If ``True``, ``fun`` should return a pair ``(value, aux)`` where ``aux``
       is a pytree of auxiliary values.
@@ -77,7 +83,7 @@ class BacktrackingLineSearch(base.IterativeLineSearch):
     unroll: whether to unroll the optimization loop (default: "auto").
   """
   fun: Callable
-  value_and_grad: bool = False
+  value_and_grad: Union[bool, Callable] = False
   has_aux: bool = False
 
   maxiter: int = 30
@@ -117,30 +123,30 @@ def init_state(
       state
     """
     del descent_direction  # Not used.
+    del init_stepsize  # Not used.
 
-    num_fun_eval = 0
-    num_grad_eval = 0
+    num_fun_eval = jnp.asarray(0, base.NUM_EVAL_DTYPE)
+    num_grad_eval = jnp.asarray(0, base.NUM_EVAL_DTYPE)
     if value is None or grad is None:
-      if self.has_aux:
-        (value, _), grad = self._value_and_grad_fun(
-            params, *fun_args, **fun_kwargs
-        )
-      else:
-        value, grad = self._value_and_grad_fun(params, *fun_args, **fun_kwargs)
+      (value, _), grad = self._value_and_grad_fun_with_aux(
+          params, *fun_args, **fun_kwargs
+      )
       num_fun_eval += 1
       num_grad_eval += 1
 
     return BacktrackingLineSearchState(iter_num=jnp.asarray(0),
+                                       params=params,
                                        value=value,
+                                       grad=grad,
+                                       value_init=value,
+                                       grad_init=grad,
                                        aux=None,  # we do not need to have aux
                                        # in the initial state
                                        error=jnp.asarray(jnp.inf),
-                                       params=params,
-                                       num_fun_eval=num_fun_eval,
-                                       num_grad_eval=num_grad_eval,
                                        done=jnp.asarray(False),
-                                       grad=grad,
-                                       failed=jnp.asarray(False))
+                                       failed=jnp.asarray(False),
+                                       num_fun_eval=num_fun_eval,
+                                       num_grad_eval=num_grad_eval)
 
   def update(
       self,
@@ -159,8 +165,8 @@ def update(
       stepsize: current estimate of the step size.
       state: named tuple containing the line search state.
       params: current parameters.
-      value: current function value (recomputed if None).
-      grad: current gradient (recomputed if None).
+      value: current function value (computed at initialization, unused here).
+      grad: current gradient (computed at initialization, unused here).
       descent_direction: descent direction (negative gradient if None).
       fun_args: additional positional arguments to be passed to ``fun``.
       fun_kwargs: additional keyword arguments to be passed to ``fun``.
@@ -172,43 +178,45 @@ def update(
     num_fun_eval = state.num_fun_eval
     num_grad_eval = state.num_grad_eval
 
-    if value is None or grad is None:
-      if self.has_aux:
-        (value, _), grad = self._value_and_grad_fun(params, *fun_args, **fun_kwargs)
-      else:
-        value, grad = self._value_and_grad_fun(params, *fun_args, **fun_kwargs)
-      num_fun_eval += 1
-      num_grad_eval += 1
+    # Grab value and grad from initialization and avoid recomputing them
+    del value
+    del grad
+    value = state.value_init
+    grad = state.grad_init
 
     if descent_direction is None:
       descent_direction = tree_scalar_mul(-1, tree_conj(grad))
 
-    gd_vdot = tree_vdot_real(tree_conj(grad), descent_direction)
+    slope = tree_vdot_real(tree_conj(grad), descent_direction)
+
     # For backtracking linesearches, we want to compute the next point
     # from the basepoint. i.e. x_i  = x_0 + s_i * p
     new_params = tree_add_scalar_mul(params, stepsize, descent_direction)
-    if self.has_aux:
-      (new_value, new_aux), new_grad = self._value_and_grad_fun(
+    # Every condition requires the new function value, but not every one
+    # requires the new gradient value (we'll assume that this code is called
+    # under `jit`).
+    num_fun_eval += 1
+    if self.condition in ["armijo", "goldstein"]:
+      new_value, new_aux = self._fun_with_aux(
           new_params, *fun_args, **fun_kwargs
       )
+      # For those conditions, no need to compute a new grad
+      # We recompute a new_grad only once we have found the right stepsize,
+      # see below
+      new_grad, new_slope = grad, slope
     else:
-      new_value, new_grad = self._value_and_grad_fun(
+      (new_value, new_aux), new_grad = self._value_and_grad_fun_with_aux(
           new_params, *fun_args, **fun_kwargs
       )
-      new_aux = None
-    new_gd_vdot = tree_vdot_real(tree_conj(new_grad), descent_direction)
-
-    # Every condition requires the new function value, but not every one
-    # requires the new gradient value (we'll assume that this code is called
-    # under `jit`).
-    num_fun_eval += 1
+      new_slope = tree_vdot_real(tree_conj(new_grad), descent_direction)
+      num_grad_eval += 1
 
     # Armijo condition (upper bound on admissible step size).
-    # cond1 = new_value <= value + self.c1 * stepsize * gd_vdot
+    # cond1 = new_value <= value + self.c1 * stepsize * slope
     # See equation (3.6a), Numerical Optimization, Second edition.
-    diff_cond1 = new_value - (value + self.c1 * stepsize * gd_vdot)
-    error_cond1 = jnp.maximum(diff_cond1, 0.0)
-
+    diff_cond1 = new_value - (value + self.c1 * stepsize * slope)
+    error_cond1 = jnp.where(jnp.isnan(diff_cond1), jnp.inf, diff_cond1)
+    error_cond1 = jnp.maximum(error_cond1, 0.0)
     error = error_cond1
 
     if self.condition == "armijo":
@@ -217,54 +225,71 @@ def update(
       pass
 
     elif self.condition == "strong-wolfe":
-      # cond2 = abs(new_gd_vdot) <= c2 * abs(gd_vdot)
+      # cond2 = abs(new_slope) <= c2 * abs(slope)
       # See equation (3.7b), Numerical Optimization, Second edition.
-      diff_cond2 = jnp.abs(new_gd_vdot) - self.c2 * jnp.abs(gd_vdot)
-      error_cond2 = jnp.maximum(diff_cond2, 0.0)
+      diff_cond2 = jnp.abs(new_slope) - self.c2 * jnp.abs(slope)
+      error_cond2 = jnp.where(jnp.isnan(diff_cond2), jnp.inf, diff_cond2)
+      error_cond2 = jnp.maximum(error_cond2, 0.0)
       error = jnp.maximum(error_cond1, error_cond2)
-      num_grad_eval += 1
 
     elif self.condition == "wolfe":
-      # cond2 = new_gd_vdot >= c2 * gd_vdot
+      # cond2 = new_slope >= c2 * slope
       # See equation (3.6b), Numerical Optimization, Second edition.
-      diff_cond2 = self.c2 * gd_vdot - new_gd_vdot
-      error_cond2 = jnp.maximum(diff_cond2, 0.0)
+      diff_cond2 = self.c2 * slope - new_slope
+      error_cond2 = jnp.where(jnp.isnan(diff_cond2), jnp.inf, diff_cond2)
+      error_cond2 = jnp.maximum(error_cond2, 0.0)
       error = jnp.maximum(error_cond1, error_cond2)
-      num_grad_eval += 1
 
     elif self.condition == "goldstein":
-      # cond2 = new_value >= value + (1 - self.c1) * stepsize * gd_vdot
-      diff_cond2 = value + (1 - self.c1) * stepsize * gd_vdot - new_value
-      error_cond2 = jnp.maximum(diff_cond2, 0.0)
+      # cond2 = new_value >= value + (1 - self.c1) * stepsize * slope
+      diff_cond2 = value + (1 - self.c1) * stepsize * slope - new_value
+      error_cond2 = jnp.where(jnp.isnan(diff_cond2), jnp.inf, diff_cond2)
+      error_cond2 = jnp.maximum(error_cond2, 0.0)
       error = jnp.maximum(error_cond1, error_cond2)
 
     else:
       raise ValueError("condition should be one of "
                        "'armijo', 'goldstein', 'strong-wolfe' or 'wolfe'.")
 
-    new_stepsize = jnp.where(error <= self.tol,
-                             stepsize,
-                             stepsize * self.decrease_factor)
     done = state.done | (error <= self.tol)
     failed = state.failed | ((state.iter_num + 1 == self.maxiter) & ~done)
 
-    new_state = BacktrackingLineSearchState(iter_num=state.iter_num + 1,
-                                            value=new_value,
-                                            aux=new_aux,
-                                            grad=new_grad,
-                                            params=new_params,
-                                            num_fun_eval=num_fun_eval,
-                                            num_grad_eval=num_grad_eval,
-                                            done=done,
-                                            error=error,
-                                            failed=failed)
+    new_stepsize = jnp.where(done | failed,
+                             stepsize,
+                             stepsize * self.decrease_factor)
+
+    if self.condition in ["armijo", "goldstein"]:
+      # If we are done for the armijo or the goldstein conditions,
+      # we compute the final gradient (we had not computed it before since 
+      # these conditions did not require it)
+      new_grad = cond(done | failed,
+                      self._compute_final_grad,
+                      lambda *_: grad, 
+                      new_params, fun_args, fun_kwargs,
+                      jit=self.jit)
+      maybe_additional_eval = jnp.asarray(done | failed, dtype=base.NUM_EVAL_DTYPE)
+      num_grad_eval = num_grad_eval + maybe_additional_eval
+      # We a priori always access the function value when computing the gradient
+      num_fun_eval = num_fun_eval + maybe_additional_eval
+
+    new_state = state._replace(iter_num=state.iter_num + 1,
+                               params=new_params,
+                               value=new_value,
+                               grad=new_grad,
+                               aux=new_aux,
+                               done=done,
+                               error=error,
+                               failed=failed,
+                               num_fun_eval=num_fun_eval,
+                               num_grad_eval=num_grad_eval)
 
     return base.LineSearchStep(stepsize=new_stepsize, state=new_state)
+
+  def _compute_final_grad(self, params, fun_args, fun_kwargs):
+    return self._grad_with_aux(params, *fun_args, **fun_kwargs)[0]
 
   def __post_init__(self):
-    if self.value_and_grad:
-      self._value_and_grad_fun = self.fun
-    else:
-      self._value_and_grad_fun = jax.value_and_grad(
-          self.fun, has_aux=self.has_aux
-      )
+    self._fun_with_aux, self._grad_with_aux, self._value_and_grad_fun_with_aux = \
+        base._make_funs_with_aux(fun=self.fun,
+                                 value_and_grad=self.value_and_grad,
+                                 has_aux=self.has_aux)

jaxopt/_src/base.py

@@ -266,7 +266,7 @@ def _cond_fun(self, inputs):
     _, state = inputs[0]
     if self.verbose:
       name = self.__class__.__name__
-      jax.debug.print("Sovler: %s, Error: {error}" % name, error=state.error)
+      jax.debug.print("Solver: %s, Error: {error}" % name, error=state.error)
     return state.error > self.tol
 
   def _body_fun(self, inputs):

jaxopt/_src/bfgs.py

@@ -273,7 +273,7 @@ def _value_and_grad_fun(self, params, *args, **kwargs):
   def __post_init__(self):
     super().__post_init__()
 
-    _, _, self._value_and_grad_with_aux = \
+    _fun_with_aux, _, self._value_and_grad_with_aux = \
       base._make_funs_with_aux(fun=self.fun,
                                value_and_grad=self.value_and_grad,
                                has_aux=self.has_aux)
@@ -282,8 +282,8 @@ def __post_init__(self):
     unroll = self._get_unroll_option()
     self.linesearch_solver = _setup_linesearch(
         linesearch=self.linesearch,
-        fun=self._value_and_grad_with_aux,
-        value_and_grad=True,
+        fun=_fun_with_aux,
+        value_and_grad=self._value_and_grad_with_aux,
         has_aux=True,
         maxlsiter=self.maxls,
         max_stepsize=self.max_stepsize,