tfds csp

official/vision/beta/projects/yolo/tasks/image_classification.py

@@ -54,4 +54,65 @@ class ImageClassificationTask(image_classification.ImageClassificationTask):
     dataset = reader.read(input_context=input_context)
     return dataset
 
+  def train_step(self, inputs, model, optimizer, metrics=None):
+    """Does forward and backward.
+
+    Args:
+      inputs: a dictionary of input tensors.
+      model: the model, forward pass definition.
+      optimizer: the optimizer for this training step.
+      metrics: a nested structure of metrics objects.
+
+    Returns:
+      A dictionary of logs.
+    """
+    features, labels = inputs
+    if self.task_config.losses.one_hot:
+      labels = tf.one_hot(labels, self.task_config.model.num_classes)
+
+    num_replicas = tf.distribute.get_strategy().num_replicas_in_sync
+    with tf.GradientTape() as tape:
+      outputs = model(features, training=True)
+      # Casting output layer as float32 is necessary when mixed_precision is
+      # mixed_float16 or mixed_bfloat16 to ensure output is casted as float32.
+      outputs = tf.nest.map_structure(
+          lambda x: tf.cast(x, tf.float32), outputs)
+
+      # Computes per-replica loss.
+      loss = self.build_losses(
+          model_outputs=outputs, labels=labels, aux_losses=model.losses)
+      # Scales loss as the default gradients allreduce performs sum inside the
+      # optimizer.
+      tf.print(loss)
+      scaled_loss = loss / num_replicas
+
+      # For mixed_precision policy, when LossScaleOptimizer is used, loss is
+      # scaled for numerical stability.
+      if isinstance(
+          optimizer, tf.keras.mixed_precision.experimental.LossScaleOptimizer):
+        scaled_loss = optimizer.get_scaled_loss(scaled_loss)
+
+    tvars = model.trainable_variables
+    grads = tape.gradient(scaled_loss, tvars)
+    # Scales back gradient before apply_gradients when LossScaleOptimizer is
+    # used.
+    if isinstance(
+        optimizer, tf.keras.mixed_precision.experimental.LossScaleOptimizer):
+      grads = optimizer.get_unscaled_gradients(grads)
+
+    # Apply gradient clipping.
+    if self.task_config.gradient_clip_norm > 0:
+      grads, _ = tf.clip_by_global_norm(
+          grads, self.task_config.gradient_clip_norm)
+    optimizer.apply_gradients(list(zip(grads, tvars)))
+
+    logs = {self.loss: loss}
+    if metrics:
+      self.process_metrics(metrics, labels, outputs)
+      logs.update({m.name: m.result() for m in metrics})
+    elif model.compiled_metrics:
+      self.process_compiled_metrics(model.compiled_metrics, labels, outputs)
+      logs.update({m.name: m.result() for m in model.metrics})
+    return logs
+