Enable squared error loss in classifier_loss_and_stats.

PiperOrigin-RevId: 730457825

examples/classifier_mnist/experiment.py

@@ -66,7 +66,7 @@ def classifier_loss(
   logits = convolutional_classifier().apply(params, batch["images"])
 
   loss, stats = losses.classifier_loss_and_stats(
-      logits=logits,
+      predictions=logits,
       labels_as_int=batch["labels"],
       params=params,
       l2_reg=l2_reg if is_training else 0.0,

examples/losses.py

@@ -123,7 +123,7 @@ def softmax_cross_entropy(
   log_z = special.logsumexp(logits, axis=-1)
 
   if logits.shape == labels.shape:
-    # Labels are encoded as 1-hot vectors
+    # Labels are encoded as (possibly smoothed) 1-hot vectors
     loss = -jnp.sum(logits * labels, axis=-1) + log_z
 
   elif logits.ndim == labels.ndim + 1:
@@ -156,20 +156,28 @@ def squared_error(
     targets: Array,
     weight: float = 1.0,
     register_loss: bool = True,
+    mask: Array | None = None,
     extra_registration_kwargs: dict[str, Any] | None = None,
     registration_module: types.ModuleType = kfac_jax,
 ) -> Array:
   """Squared error loss."""
   extra_registration_kwargs = extra_registration_kwargs or {}
 
-  if prediction.shape != targets.shape:
-    raise ValueError("prediction and targets should have the same shape.")
-
   if register_loss:
     registration_module.register_squared_error_loss(
         prediction, targets, weight, **extra_registration_kwargs)
 
-  return weight * jnp.sum(jnp.square(prediction - targets), axis=-1)
+  if prediction.shape != targets.shape:
+    raise ValueError("prediction and targets should have the same shape.")
+
+  squared_residuals = jnp.square(prediction - targets)
+
+  loss = jnp.sum(squared_residuals, axis=-1)
+
+  if mask:
+    loss = loss * mask
+
+  return weight * loss
 
 
 def top_k_accuracy(
@@ -226,7 +234,7 @@ def add_label_smoothing(
 
 
 def classifier_loss_and_stats(
-    logits: Array,
+    predictions: Array,
     labels_as_int: Array,
     params: Params,
     l2_reg: Numeric,
@@ -240,26 +248,67 @@ def classifier_loss_and_stats(
     normalization_mode: str = "batch_size_only",
     extra_registration_kwargs: dict[str, Any] | None = None,
     registration_module: types.ModuleType = kfac_jax,
+    loss_type: str = "cross_entropy",
 ) -> tuple[Array, dict[str, Array]]:
-  """Softmax cross-entropy with regularizer and accuracy statistics."""
-
-  batch_size = logits.shape[0]
-
-  if labels_as_int.shape[0] != batch_size:
-    raise ValueError(f"Size of first dimension of logits ({batch_size}) "
-                     f"(i.e. batch size) doesn't match that of labels "
-                     f"({labels_as_int.shape[0]})")
-
-  if mask is not None and mask.shape[0] != batch_size:
-    raise ValueError(f"Size of first dimension of logits ({batch_size}) "
-                     f"(i.e. batch size) doesn't match that of mask "
-                     f"({mask.shape[0]})")
+  """Classification loss with regularizer and accuracy statistics.
+
+  Args:
+    predictions: The output of the model. Logits for loss_type="cross_entropy",
+      or quantity to be compared to (possibly smoothed) one-hot targets for
+      loss_type="squared_error". Predictions will have shape (batch_size, ...,
+      num_classes).
+    labels_as_int: The labels to be used in the loss regarded as integers. Must
+      be of shape predictions.shape[:-1].
+    params: The parameters of the model.
+    l2_reg: The L2 regularization coefficient.
+    haiku_exclude_batch_norm: Whether to exclude batch norm parameters from the
+      L2 regularization (assumes models are Haiku models).
+    haiku_exclude_biases: Whether to exclude biases from the L2 regularization
+      (assumes models are Haiku models).
+    label_smoothing: The label smoothing coefficient.
+    top_k_stats: The top-k accuracies to compute.
+    average_loss: Whether to average the loss over the batch.
+    register_loss: Whether to register the loss.
+    mask: If not None, a binary mask of shape predictions.shape[:-1]. It's
+      nonzero values determine which predictions are used in the loss
+      computation.
+    normalization_mode: The normalization mode to use for the returned loss, one
+      of "batch_size_only", "all_dims", or "all_dims_nonmasked".
+      "batch_size_only" means the loss is normalized by the batch size.
+      "all_dims" means the loss is normalized by the product of the dimensions
+      of the predictions array, excluding the last dimension.
+      "all_dims_nonmasked" means the loss is normalized by the number of
+      nonzero entries of the mask if it is not None.
+    extra_registration_kwargs: Extra kwargs to pass to the registration
+      functions.
+    registration_module: The module containing the loss registration functions
+      that will be used. These are 'register_softmax_cross_entropy_loss' and
+      'register_squared_error_loss".
+    loss_type: The type of loss to use ("cross_entropy" or "squared_error").
+
+  Returns:
+    The regularized loss and a dictionary of statistics.
+  """
+
+  batch_size = predictions.shape[0]
+
+  if labels_as_int.shape != predictions.shape[:-1]:
+    raise ValueError(
+        f"Shape mismatch: labels_as_int shape ({labels_as_int.shape}) "
+        f"not compatible with predictions shape {predictions.shape}"
+    )
+
+  if mask is not None and mask.shape != predictions.shape[:-1]:
+    raise ValueError(
+        f"Shape mismatch: mask shape ({mask.shape}) "
+        f"not compatible with predictions shape {predictions.shape}"
+    )
 
   if normalization_mode == "batch_size_only":
     weight = 1.0
 
   elif normalization_mode == "all_dims":
-    weight = 1.0 / kfac_jax.utils.product(logits.shape[1:-1])
+    weight = 1.0 / kfac_jax.utils.product(predictions.shape[1:-1])
 
   elif normalization_mode == "all_dims_nonmasked":
     assert mask is not None
@@ -269,16 +318,40 @@ def classifier_loss_and_stats(
     raise ValueError(f"Unrecognized value for normalization_mode: "
                      f"{normalization_mode}")
 
-  labels = add_label_smoothing(labels_as_int, label_smoothing, logits.shape[-1])
+  labels = add_label_smoothing(
+      labels_as_int, label_smoothing, predictions.shape[-1]
+  )
 
-  softmax_loss = softmax_cross_entropy(
-      logits, labels, weight=weight, register_loss=register_loss, mask=mask,
-      extra_registration_kwargs=extra_registration_kwargs,
-      registration_module=registration_module)
+  if loss_type == "cross_entropy":
+    raw_loss = softmax_cross_entropy(
+        predictions,
+        labels,
+        weight=weight,
+        register_loss=register_loss,
+        mask=mask,
+        extra_registration_kwargs=extra_registration_kwargs,
+        registration_module=registration_module,
+    )
+  elif loss_type == "squared_error":
+
+    if predictions.ndim == labels.ndim + 1:
+      labels = jax.nn.one_hot(labels, predictions.shape[-1])
+
+    raw_loss = squared_error(
+        predictions,
+        labels,
+        weight=weight,
+        register_loss=register_loss,
+        mask=mask,
+        extra_registration_kwargs=extra_registration_kwargs,
+        registration_module=registration_module,
+    )
+  else:
+    raise ValueError(f"Unknown loss type: {loss_type}")
 
-  averaged_raw_loss = jnp.sum(softmax_loss, axis=0) / batch_size
+  averaged_raw_loss = jnp.sum(raw_loss, axis=0) / batch_size
 
-  loss = averaged_raw_loss if average_loss else softmax_loss
+  loss = averaged_raw_loss if average_loss else raw_loss
 
   l2_reg_val = l2_regularizer(
       params, haiku_exclude_batch_norm, haiku_exclude_biases)
@@ -290,6 +363,6 @@ def classifier_loss_and_stats(
       l2_reg_val=l2_reg_val,
   )
   for k in top_k_stats:
-    stats[f"top_{k}_accuracy"] = top_k_accuracy(logits, labels_as_int, k)
+    stats[f"top_{k}_accuracy"] = top_k_accuracy(predictions, labels_as_int, k)
 
   return regularized_loss, stats

examples/lrelunet101_imagenet/experiment.py

@@ -360,7 +360,7 @@ def lrelunet_loss(
       params, rng, batch["images"], is_training)
 
   return losses.classifier_loss_and_stats(
-      logits=logits,
+      predictions=logits,
       labels_as_int=batch["labels"],
       params=params,
       l2_reg=l2_reg if is_training else 0.0,

examples/resnet50_imagenet/experiment.py

@@ -79,7 +79,7 @@ def resnet50_loss(
   ).apply(params, state, batch["images"], is_training=is_training)
 
   loss, stats = losses.classifier_loss_and_stats(
-      logits=logits,
+      predictions=logits,
       labels_as_int=batch["labels"],
       params=params,
       l2_reg=l2_reg if is_training else 0.0,