No public description

PiperOrigin-RevId: 559849502

official/projects/pix2seq/modeling/pix2seq_model.py

@@ -335,6 +335,7 @@ class Pix2Seq(tf.keras.Model):
       inputs: tf.Tensor,
       targets: Optional[tf.Tensor] = None,
       training: bool = None,
+      use_teacher_forcing_for_eval: bool = False
   ) -> List[Any]:
     features = self._backbone(inputs)[self._backbone_endpoint_name]
     mask = tf.ones_like(features)
@@ -350,22 +351,18 @@ class Pix2Seq(tf.keras.Model):
     pos_emb = tf.cast(pos_emb, features.dtype)
 
     tokens = None
+    inputs = {
+        "inputs": features,
+        "tokens": targets,
+        "pos_emb": pos_emb,
+    }
     if training:
-      logits = self._transformer(
-          {
-              "inputs": features,
-              "tokens": targets,
-              "pos_emb": pos_emb,
-          },
-          training,
-      )
+      logits = self._transformer(inputs, training=True)
+    elif use_teacher_forcing_for_eval:
+      logits = self._transformer(inputs, training=False)
     else:
       tokens, logits = self._transformer.infer(
-          {
-              "inputs": features,
-              "tokens": targets,
-              "pos_emb": pos_emb,
-          },
+          inputs,
           top_k=self._top_k,
           top_p=self._top_p,
       )

official/projects/pix2seq/modeling/pix2seq_model_test.py

@@ -30,7 +30,11 @@ class Pix2SeqTest(tf.test.TestCase):
     backbone = resnet.ResNet(50, bn_trainable=False)
     backbone_endpoint_name = '5'
     model = pix2seq_model.Pix2Seq(
-        backbone, backbone_endpoint_name, max_seq_len, vocab_size, hidden_size,
+        backbone,
+        backbone_endpoint_name,
+        max_seq_len,
+        vocab_size,
+        hidden_size,
         num_heads=num_heads,
     )
     _, outs = model(
@@ -41,6 +45,32 @@ class Pix2SeqTest(tf.test.TestCase):
 
     self.assertLen(outs, 2)  # intermediate decoded outputs.
 
+  def test_forward_infer_teacher_forcing(self):
+    hidden_size = 256
+    num_heads = 8
+    max_seq_len = 50
+    vocab_size = 164
+    image_size = 224
+    batch_size = 2
+    backbone = resnet.ResNet(50, bn_trainable=False)
+    backbone_endpoint_name = '5'
+    model = pix2seq_model.Pix2Seq(
+        backbone,
+        backbone_endpoint_name,
+        max_seq_len,
+        vocab_size,
+        hidden_size,
+        num_heads=num_heads,
+    )
+    _, outs = model(
+        tf.ones((batch_size, image_size, image_size, 3)),
+        tf.ones((batch_size, max_seq_len), tf.int64),
+        training=False,
+        use_teacher_forcing_for_eval=True,
+    )
+
+    self.assertLen(outs, 2)  # intermediate decoded outputs.
+
   def test_forward_infer(self):
     hidden_size = 256
     num_heads = 8
@@ -51,7 +81,11 @@ class Pix2SeqTest(tf.test.TestCase):
     backbone = resnet.ResNet(50, bn_trainable=False)
     backbone_endpoint_name = '5'
     model = pix2seq_model.Pix2Seq(
-        backbone, backbone_endpoint_name, max_seq_len, vocab_size, hidden_size,
+        backbone,
+        backbone_endpoint_name,
+        max_seq_len,
+        vocab_size,
+        hidden_size,
         num_heads=num_heads,
     )
     tokens, _ = model(