Implemented creep stage and finishing details

Author: davidcortesortuno

fidimag/common/chain_method_integrators.py

@@ -150,26 +150,40 @@ def _step(self, t, y, h, f):
 class FSIntegrator(BaseIntegrator):
     """A step integrator considering the action of the band
     """
-    def __init__(self, band, forces, action, rhs_fun, action_fun, n_images, n_dofs_image,
-                 max_steps=1000,
-                 maxCreep=5, eta_scale=1.0, stopping_dE=1e-6, dEta=2,
-                 etaMin=0.001,
+    def __init__(self, band, forces, rhs_fun, action_fun, n_images, n_dofs_image,
+                 maxSteps=1000,
+                 maxCreep=5, actionTol=1e-10, forcesTol=1e-6,
+                 etaScale=1.0, dEta=2, minEta=0.001,
                  # perturbSeed=42, perturbFactor=0.1,
-                 nTrail=10, resetMax=20, mXgradE_tol=0.1
+                 nTrail=10, resetMax=20
                  ):
         super(FSIntegrator, self).__init__(band, rhs_fun)
 
         self.action_fun = action_fun
 
+        # Integration parameters
+        # TODO: move to run function?
+        self.dEta = dEta
+        self.minEta = minEta
+        self.resetMax = resetMax
+        self.maxCreep = maxCreep
+        self.etaScale = etaScale
+        self.forcesTol = forcesTol
+        self.actionTol = actionTol
+        self.maxSteps = maxSteps
+
         self.i_step = 0
         self.n_images = n_images
         self.n_dofs_image = n_dofs_image
-        self.forces_prev = np.zeros_like(band).reshape(n_images, -1)
+        # self.forces_prev = np.zeros_like(band).reshape(n_images, -1)
         # self.G :
         self.forces = forces
-        self.max_steps = max_steps
+        self.forces_old = np.zeros_like(forces)
+
+        # Rename y to band
+        self.band = self.y
+        self.band_old = np.zeros_like(self.band)  # y -> band
 
-        self.y_last = np.zeros_like(self.y)  # y -> band
         self.step = 0
         self.nTrail = nTrail
 
@@ -184,18 +198,21 @@ def run_for(self, n_steps):
         totalRestart = True
         resetCount = 0
         creepCount = 0
-        self.trailAction = np.zeros(nTrail)
-        trailPool = cycle(range(nTrail))  # cycle through 0,1,...,(nTrail-1),0,1,...
+        self.trailAction = np.zeros(self.nTrail)
+        trailPool = cycle(range(self.nTrail))  # cycle through 0,1,...,(nTrail-1),0,1,...
         eta = 1.0
 
+        # In __init__:
+        # self.band_last[:] = self.band
+
         INNER_DOFS = slice(self.n_dofs_image, -self.n_dofs_image)
 
         while not exitFlag:
 
             if totalRestart:
                 if self.step > 0:
                     print('Restarting')
-                self.y[:] = self.y_last
+                self.band[:] = self.band_old
 
                 # Compute from self.band. Do not update the step at this stage:
                 # This step updates the forces in the G array of the nebm module,
@@ -206,10 +223,10 @@ def run_for(self, n_steps):
                 self.action = self.action_fun()
 
                 # self.step += 1
-                self.forces_last[:] = self.forces  # Scale field??
+                self.forces_old[:] = self.forces  # Scale field??
                 # self.gradE_last[~_material] = 0.0
                 # self.gradE[:] = self.gradE_last
-                self.action_last = self.action
+                self.action_old = self.action
                 self.trailAction[nStart] = self.action
                 nStart = next(trailPool)
                 eta = 1.0
@@ -218,10 +235,10 @@ def run_for(self, n_steps):
             creepCount = 0
 
             # Creep stage: minimise with a fixed eta
-            while creepCount < maxCreep:
+            while creepCount < self.maxCreep:
                 # Update spin. Avoid pinned or zero-Ms sites
-                self.y[:] = self.y_last + eta * eta_scale * self.forces
-                normalise_spins(self.y)
+                self.band[:] = self.band_old + eta * self.etaScale * self.forces
+                normalise_spins(self.band)
 
                 self.trailAction
 
@@ -231,88 +248,72 @@ def run_for(self, n_steps):
                 self.trailAction[nStart] = self.action
                 nStart = next(trailPool)
 
-                bandNormSquared = 0.0
                 # Getting averages of forces from the INNER images in the band (no extrema)
                 # (forces are given by vector G in the chain method code)
+                # TODO: we might use all band images, not only inner ones
                 G_norms = np.linalg.norm(self.forces[INNER_DOFS].reshape(-1, 3), axis=1)
                 # Compute the root mean square per image
                 rms_G_norms_per_image = np.sqrt(np.mean(G_norms.reshape(self.n_images - 2, -1), axis=1))
                 mean_rms_G_norms_per_image = np.mean(rms_G_norms_per_image)
 
+                # Average step difference between trailing action and new action
+                deltaAction = (np.abs(self.trailAction[nStart] - self.action)) / self.nTrail
 
                 # 10 seems like a magic number; we set here a minimum number of evaulations
-                if (nStart > nTrail * 10) and :
-
-
-
-        # while abs(self.i_step - steps) > EPSILON:
-        st = 1
-        while st < n_steps:
-            self.i_step = self._step(self.t, self.y, self.stepsize, self.rhs)
-
-            self.rhs_evals_nb += 0
-
-            if self.i_step > self.max_steps:
-                break
-
-            st += 1
-        return 0
+                if (nStart > self.nTrail * 10) and (deltaAction < self.actionTol):
+                    print('Change in action is negligible')
+                    exitFlag = True
+                    break  # creep loop
+
+                if (n_steps >= self.maxSteps):
+                    print('Number of steps reached maximum')
+                    exitFlag = True
+                    break  # creep loop
+
+                # If action increases compared to last step, decrease eta and start creeping again
+                if (self.action_old < self.action):
+                    print('Action increased. Start new creep stage from last step')
+                    creepCount = 0
+                    eta = eta / (self.dEta * self.dEta)
+
+                    # If eta is too small, reset and start again
+                    if (eta < 1e-3):
+                        print('')
+                        resetCount += 1
+                        bestAction = self.action_old
+                        RefinePath(self.band)  # Resets the path to equidistant structures (smoothing  kinks?)
+                        # PathChanged[:] = True
+
+                        if resetCount > self.resetMax:
+                            print('Failed to converge!')
+                    # Otherwise, just 
+                    else:
+                        self.band[:] = self.band_old
+                        self.forces[:] = self.forces_old
+                # If action decreases, move to next creep step
+                else:
+                    creepCount += 1
+                    self.action_old = self.action
+                    self.band_old[:] = self.band
+                    self.forces_old[:] = self.forces
+
+                    if (mean_rms_G_norms_per_image < self.forcesTol):
+                        print('Change of mean of the force RMSquares negligible')
+                        exitFlag = True
+                        break  # creep loop
+            # END creep while loop
+
+            # After creep loop:
+            self.eta = self.eta * self.dEta
+            resetCount = 0
 
     def set_options(self):
         pass
 
     def _step(self, t, y, h, f):
         """
         """
-
-        f(t, y)
-        force_images = self.forces
-        force_images.shape = (self.n_images, -1)
-        # In this case f represents the force: a = dy/dt = f/m
-        # * self.m_inv[:, np.newaxis]
-        y.shape = (self.n_images, -1)
-        # print(force_images[2])
-
-        # Loop through every image in the band
-        for i in range(1, self.n_images - 1):
-
-            force = force_images[i]
-            velocity = self.velocity[i]
-            velocity_new = self.velocity_new[i]
-
-            # Update coordinates from Newton eq, which uses the "acceleration"
-            # At step 0 velocity is zero
-            y[i][:] = y[i] + h * (velocity + (h / (2 * self.mass)) * force)
-
-            # Update the velocity from a mean with the prev step
-            # (Velocity Verlet)
-            velocity[:] = velocity_new[:] + (h / (2 * self.mass)) * (self.forces_prev[i] + force)
-
-            # Project the force of the image into the velocity vector: < v | F >
-            velocity_proj = np.einsum('i,i', force, velocity)
-
-            # Set velocity to zero when the proj = 0
-            if velocity_proj <= 0:
-                velocity_new[:] = 0.0
-            else:
-                # Norm of the force squared <F | F>
-                force_norm_2 = np.einsum('i,i', force, force)
-                factor = velocity_proj / force_norm_2
-                # Set updated velocity: v = v * (<v|F> / |F|^2)
-                velocity_new[:] = factor * force
-
-            # New velocity from Newton equation (old Verlet)
-            # velocity[:] = velocity_new + (h / self.mass) * force
-
-        # Store the force for the Velocity Verlet algorithm
-        self.forces_prev[:] = force_images
-
-        y.shape = (-1,)
-        force_images.shape = (-1,)
-        normalise_spins(y)
-        tp = t + h
-        return tp
-
+        pass
 
 def normalise_spins(y):
     # Normalise an array of spins y with 3 * N elements