some fixes

Author: GreenWizard2015

NN/Renderer.py

@@ -28,15 +28,15 @@ def batched(self, ittr, B, N, batchSize=None, training=False):
     for i in tf.range(NBatches):
       index = i * stepBy
       data = ittr(index, stepBy)
-      V = self._restorator.reverse(**data, training=training)
+      V = self._restorator.reverse(**data, training=training, index=index)
       C = tf.shape(V)[-1]
       res = res.write(i, tf.reshape(V, (B, stepBy, C)))
       continue
     #################
     index = NBatches * stepBy
 
     data = ittr(index, N - index)
-    V = self._restorator.reverse(**data, training=training)
+    V = self._restorator.reverse(**data, training=training, index=index)
     C = tf.shape(V)[-1]
 
     w = N - index

NN/RestorationModel/CRepeatedRestorator.py

@@ -26,14 +26,14 @@ def call(self, latents, pos, T, V, residual, training=None):
       continue
     return V
 
-  def reverse(self, latents, pos, reverseArgs, training, value, residual):
+  def reverse(self, latents, pos, reverseArgs, training, value, residual, index):
     for i in range(self._N):
       if tf.is_tensor(value): value = tf.stop_gradient(value)
       value = self._restorator.reverse(
         latents=self._withID(latents, i, training),
         pos=pos, reverseArgs=reverseArgs,
         residual=residual,
-        training=training, value=value
+        training=training, value=value, index=index
       )
       continue
     return value

NN/RestorationModel/CRestorationModel.py

@@ -49,7 +49,7 @@ def call(self, latents, pos, T, V, residual, training=None):
     res = self._decoder(condition=latents, coords=EPos, timestep=t, V=V, training=training)
     return self._withResidual(res, residual)
 
-  def reverse(self, latents, pos, reverseArgs, training, value, residual):
+  def reverse(self, latents, pos, reverseArgs, training, value, residual, index):
     EPos = self._encodePos(pos, training=training, args=reverseArgs.get('decoder', {}))
 
     def denoiser(x, t, mask=None):
@@ -65,7 +65,8 @@ def denoiser(x, t, mask=None):
     return self._restorator.reverse(
       value=value, denoiser=denoiser, 
       modelT=lambda t: self._encodeTime(t, training=training),
-      **reverseArgs
+      **reverseArgs,
+      index=index
     )
 
   def train_step(self, x0, latents, positions, params, xT=None):

NN/RestorationModel/CSequentialRestorator.py

@@ -15,13 +15,13 @@ def call(self, latents, pos, T, V, residual, training=None):
       continue
     return V
 
-  def reverse(self, latents, pos, reverseArgs, training, value, residual):
+  def reverse(self, latents, pos, reverseArgs, training, value, residual, index):
     for restorator in self._restorators:
       if tf.is_tensor(value): value = tf.stop_gradient(value)
       value = restorator.reverse(
         latents=latents, pos=pos, reverseArgs=reverseArgs,
         residual=residual,
-        training=training, value=value
+        training=training, value=value, index=index
       )
       continue
     return value

NN/restorators/CARProcess.py

@@ -39,12 +39,12 @@ def denoiser(x, t=None, **kwargs):
       return model(x=x, t=T, mask=kwargs.get('mask', None))[:, :self.predictions]
     return denoiser
 
-  def reverse(self, value, denoiser, modelT=None, **kwargs):
+  def reverse(self, value, denoiser, modelT=None, index=0, **kwargs):
     if isinstance(value, tuple):
       value = self._sourceDistribution.initialValueFor(value + (self.predictions, ))
 
     denoiser = self._makeDenoiser(denoiser, modelT)
-    res = self._sampler.sample(value=value, model=denoiser, **kwargs)
+    res = self._sampler.sample(value=value, model=denoiser, index=index, **kwargs)
     tf.assert_equal(tf.shape(res), tf.shape(value))
     return res
 

NN/restorators/samplers/CBasicInterpolantSampler.py

@@ -12,8 +12,9 @@ def __init__(self, interpolant, algorithm):
   def interpolant(self): return self._interpolant
 
   @tf.function
-  def sample(self, value, model, **kwargs):
-    kwargs = dict(**kwargs, interpolant=self._interpolant) # add interpolant to kwargs
+  def sample(self, value, model, index=0, **kwargs):
+    # add interpolant to kwargs and index
+    kwargs = dict(**kwargs, interpolant=self._interpolant, index=index)
     # wrap algorithm with hook, if provided
     algorithm = kwargs.get('algorithmInterceptor', lambda x: x)( self._algorithm )
     assert isinstance(algorithm, ISamplingAlgorithm), f'Algorithm must be an instance of ISamplingAlgorithm, but got {type(algorithm)}'

NN/restorators/samplers/CSamplerWatcher.py

@@ -39,26 +39,27 @@ def tracked(self, name):
       res = tf.concat([self._initialValue[None], res], axis=0)
     return res
 
-  def _updateTracked(self, name, value, mask=None):
+  def _updateTracked(self, name, value, mask=None, index=None):
     tracked = self._tracked.get(name, None)
     if tracked is None: return
-    value = self._withIndices(value)
+    value, idx = self._withIndices(value, index, mask=mask)
 
     iteration = self._iteration
-    if (mask is None) or ('value' == name): # 'value' is always unmasked
-      tracked[iteration].assign(value)
-      return
-    
-    mask = self._withIndices(mask)
-    prev = tracked[iteration - 1]
-    # expand mask to match the value shape by copying values from the previous iteration
-    indices = tf.where(mask)
-    value = tf.tensor_scatter_nd_update(prev, indices, value)
-    tf.assert_equal(tf.shape(prev), tf.shape(value), 'Must be the same shape')
-    tracked[iteration].assign(value)
+    if (mask is not None) and not ('value' == name): # 'value' is always unmasked
+      mask, _ = self._withIndices(mask, index, mask=mask)
+      prev = self._tracked[name][iteration]
+      # expand mask to match the value shape by copying values from the previous iteration
+      indices = tf.where(mask)
+      sz = tf.shape(value)[0]
+      value = tf.tensor_scatter_nd_update(prev[idx:idx+sz], indices, value)
+      pass
+
+    sz = tf.shape(value)[0]
+    tracked[iteration, idx:idx+sz].assign(value)
     return
 
   def _onNextStep(self, iteration, kwargs):
+    index = kwargs['index']
     self._iteration.assign(iteration)
     # track also solution
     solution = kwargs['solution']
@@ -68,17 +69,36 @@ def _onNextStep(self, iteration, kwargs):
     mask = step.mask if hasattr(step, 'mask') else None
     # iterate over all fields
     for name in solution._fields:
-      self._updateTracked(name, getattr(solution, name), mask=mask)
+      self._updateTracked(name, getattr(solution, name), mask=mask, index=index)
       continue
     return
 
   def _onStart(self, value, kwargs):
+    index = kwargs['index']
     self._iteration.assign(0)
     if 'value' in self._tracked: # save initial value
-      self._initialValue.assign( self._withIndices(value) )
+      value, idx = self._withIndices(value, index)
+      # update slice [index:index+sz] with the value
+      sz = tf.shape(value)[0]
+      self._initialValue[idx:idx+sz].assign(value)
     return
 
-  def _withIndices(self, value):
-    if self._indices is None: return value
-    return tf.gather(value, self._indices, axis=0)
+  def _withIndices(self, value, index, mask=None):
+    if self._indices is None: return value, index
+    # find subset of indices
+    sz = tf.shape(value)[0]
+
+    validMask = tf.logical_and(index <= self._indices, self._indices < index + sz)
+    if mask is not None:
+      maskedIndices = tf.range(sz)
+      maskedIndices = tf.boolean_mask(maskedIndices, mask) + index
+      # exclude masked indices
+      maskedIndices = tf.reduce_any(maskedIndices[:, None] == self._indices[None], axis=0)
+      validMask = tf.logical_and(validMask, maskedIndices)
+      pass
+    
+    startIndex = tf.reduce_min(tf.where(validMask))
+    startIndex = tf.cast(startIndex, tf.int32)
+    indices = tf.boolean_mask(self._indices, validMask) - index
+    return tf.gather(value, indices, axis=0), startIndex
 # End of CSamplerWatcher

huggingface/HF/NN/CInterpolantVisualization.py

@@ -158,7 +158,7 @@ def _collectSteps(self, points, initialValues, kwargs):
       initialValues=initialValues,
       reverseArgs=dict(
         **reverseArgs,
-        algorithmInterceptor=watcher.interceptor(),
+        algorithmInterceptor=watcher,
       ),
     )
     N = watcher.iteration + 1

tests/test_CSamplerWatcher.py

@@ -41,7 +41,7 @@ def _fake_AR(threshold, timesteps=10):
   shape = (32, 3)
   fakeNoise = tf.random.normal(shape)
   def fakeModel(x, t, mask, **kwargs):
-    s = tf.boolean_mask(fakeNoise, mask)
+    s = tf.boolean_mask(fakeNoise, mask) if mask is not None else fakeNoise
     return s + tf.cast(t, tf.float32) * x
 
   x = tf.random.normal(shape)
@@ -153,4 +153,45 @@ def test_trackSolutionWithMask_value():
   )
   _ = fake.interpolant.sample(value=fake.x, model=fake.model, algorithmInterceptor=watcher.interceptor())
   _checkTracked(watcher.tracked('value'), N=11)
+  return
+
+# test multiple calls with index
+def test_multipleCallsWithIndex():
+  fake = _fake_sampler()
+  watcher = CSamplerWatcher(
+    steps=10,
+    tracked=dict(value=(32*3, 3))
+  )
+  arg = dict(value=fake.x, model=fake.model, algorithmInterceptor=watcher.interceptor())
+  _ = fake.interpolant.sample(**arg, index=0)
+  _ = fake.interpolant.sample(**arg, index=32)
+  _ = fake.interpolant.sample(**arg, index=64)
+
+  collectedSteps = watcher.tracked('value')
+  tf.debugging.assert_equal(tf.shape(collectedSteps)[1], 96, 'Must collect 96 values')
+  tf.debugging.assert_equal(watcher.iteration, 10, 'Must collect 10 steps')
+  # values must be same across (0..32), (32..64), (64..96)
+  tf.debugging.assert_equal(collectedSteps[:, :32], collectedSteps[:, 32:64])
+  tf.debugging.assert_equal(collectedSteps[:, 32:64], collectedSteps[:, 64:])
+  return
+
+# TODO: find out why this test fails, but previous one passes
+# test multiple calls with index and mask
+def test_multipleCallsWithIndexAndMask():
+  fake = _fake_AR(threshold=0.1)
+  watcher = CSamplerWatcher(
+    steps=10,
+    tracked=dict(value=(3,)),
+    indices=[0, 32, 64, 65]
+  )
+  arg = dict(value=fake.x, model=fake.model, algorithmInterceptor=watcher.interceptor())
+  _ = fake.interpolant.sample(**arg, index=0)
+  _ = fake.interpolant.sample(**arg, index=32)
+  _ = fake.interpolant.sample(**arg, index=64)
+
+  collectedSteps = watcher.tracked('value')
+  tf.assert_equal(tf.shape(collectedSteps), (11, 4, 3), 'Must be (11, 4, 3)')
+  tf.debugging.assert_equal(watcher.iteration, 10, 'Must collect 10 steps')
+  tf.debugging.assert_equal(collectedSteps[:, 0:1], collectedSteps[:, 1:2])
+  tf.debugging.assert_equal(collectedSteps[:, 1:2], collectedSteps[:, 2:3])
   return