TF: use the correct config with `(...)EncoderDecoder` models (#18097)

src/transformers/modeling_tf_utils.py

@@ -403,8 +403,13 @@ def unpack_inputs(func):
         # move any arg into kwargs, if they exist
         fn_args_and_kwargs.update(dict(zip(func.__code__.co_varnames[1:], args)))
 
-        # process the inputs and call the wrapped function
-        unpacked_inputs = input_processing(func, self.config, **fn_args_and_kwargs)
+        # Encoder Decoder models delegate the application of the configuration options to their inner models.
+        if "encoder_decoder" in str(self).lower():
+            config = None
+        else:
+            config = self.config
+
+        unpacked_inputs = input_processing(func, config, **fn_args_and_kwargs)
         return func(self, **unpacked_inputs)
 
     # Keras enforces the first layer argument to be passed, and checks it through `inspect.getfullargspec()`. This
@@ -559,6 +564,7 @@ def input_processing(func, config, **kwargs):
     if "kwargs" in output:
         del output["kwargs"]
 
+    if config is not None:
         boolean_dict = {
             k: v
             for k, v in output.items()

src/transformers/models/encoder_decoder/modeling_tf_encoder_decoder.py

@@ -630,13 +630,13 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss):
             warnings.warn(DEPRECATION_WARNING, FutureWarning)
             loss = self.hf_compute_loss(labels, logits)
 
+        if not return_dict:
             past_key_values = None
-        if decoder_inputs["use_cache"]:
+            if use_cache:
                 past_key_values = decoder_outputs[1]
             # The starting index of the remaining elements in `decoder_outputs`
             start_index = sum([1 if x is not None else 0 for x in (loss, logits, past_key_values)])
 
-        if not decoder_inputs["return_dict"]:
             if not isinstance(encoder_outputs, tuple):
                 encoder_outputs = encoder_outputs.to_tuple()
             output = (loss, logits, past_key_values) + decoder_outputs[start_index:] + encoder_outputs
@@ -646,7 +646,7 @@ class TFEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLoss):
         return TFSeq2SeqLMOutput(
             loss=loss,
             logits=decoder_outputs.logits,
-            past_key_values=past_key_values,
+            past_key_values=decoder_outputs.past_key_values,
             decoder_hidden_states=decoder_outputs.hidden_states,
             decoder_attentions=decoder_outputs.attentions,
             cross_attentions=decoder_outputs.cross_attentions,

src/transformers/models/vision_encoder_decoder/modeling_tf_vision_encoder_decoder.py

@@ -663,13 +663,13 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
             warnings.warn(DEPRECATION_WARNING, FutureWarning)
             loss = self.hf_compute_loss(labels, logits)
 
+        if not return_dict:
             past_key_values = None
-        if decoder_inputs["use_cache"]:
+            if use_cache:
                 past_key_values = decoder_outputs[1]
             # The starting index of the remaining elements in `decoder_outputs`
             start_index = sum([1 if x is not None else 0 for x in (loss, logits, past_key_values)])
 
-        if not decoder_inputs["return_dict"]:
             if not isinstance(encoder_outputs, tuple):
                 encoder_outputs = encoder_outputs.to_tuple()
             output = (loss, logits, past_key_values) + decoder_outputs[start_index:] + encoder_outputs
@@ -679,7 +679,7 @@ class TFVisionEncoderDecoderModel(TFPreTrainedModel, TFCausalLanguageModelingLos
         return TFSeq2SeqLMOutput(
             loss=loss,
             logits=decoder_outputs.logits,
-            past_key_values=past_key_values,
+            past_key_values=decoder_outputs.past_key_values,
             decoder_hidden_states=decoder_outputs.hidden_states,
             decoder_attentions=decoder_outputs.attentions,
             cross_attentions=decoder_outputs.cross_attentions,

tests/models/encoder_decoder/test_modeling_encoder_decoder.py

@@ -351,6 +351,40 @@ class EncoderDecoderMixin:
             outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
         )
 
+    def _check_output_with_attentions(
+        self, outputs_encoder_decoder, config, input_ids, decoder_config, decoder_input_ids
+    ):
+        encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
+        self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
+
+        self.assertEqual(
+            encoder_attentions[0].shape[-3:], (config.num_attention_heads, input_ids.shape[-1], input_ids.shape[-1])
+        )
+
+        decoder_attentions = outputs_encoder_decoder["decoder_attentions"]
+        num_decoder_layers = (
+            decoder_config.num_decoder_layers
+            if hasattr(decoder_config, "num_decoder_layers")
+            else decoder_config.num_hidden_layers
+        )
+        self.assertEqual(len(decoder_attentions), num_decoder_layers)
+
+        self.assertEqual(
+            decoder_attentions[0].shape[-3:],
+            (decoder_config.num_attention_heads, decoder_input_ids.shape[-1], decoder_input_ids.shape[-1]),
+        )
+
+        cross_attentions = outputs_encoder_decoder["cross_attentions"]
+        self.assertEqual(len(cross_attentions), num_decoder_layers)
+
+        cross_attention_input_seq_len = decoder_input_ids.shape[-1] * (
+            1 + (decoder_config.ngram if hasattr(decoder_config, "ngram") else 0)
+        )
+        self.assertEqual(
+            cross_attentions[0].shape[-3:],
+            (decoder_config.num_attention_heads, cross_attention_input_seq_len, input_ids.shape[-1]),
+        )
+
     def check_encoder_decoder_model_output_attentions(
         self,
         config,
@@ -376,36 +410,58 @@ class EncoderDecoderMixin:
             decoder_attention_mask=decoder_attention_mask,
             output_attentions=True,
         )
+        self._check_output_with_attentions(
+            outputs_encoder_decoder, config, input_ids, decoder_config, decoder_input_ids
+        )
 
-        encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
-        self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
+    def check_encoder_decoder_model_output_attentions_from_config(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        encoder_hidden_states,
+        decoder_config,
+        decoder_input_ids,
+        decoder_attention_mask,
+        labels,
+        **kwargs
+    ):
+        # Similar to `check_encoder_decoder_model_output_attentions`, but with `output_attentions` triggered from the
+        # config file. Contrarily to most models, changing the model's config won't work -- the defaults are loaded
+        # from the inner models' configurations.
 
-        self.assertEqual(
-            encoder_attentions[0].shape[-3:], (config.num_attention_heads, input_ids.shape[-1], input_ids.shape[-1])
+        decoder_input_ids = decoder_input_ids[:, :-1]
+        decoder_attention_mask = decoder_attention_mask[:, :-1]
+        encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
+        enc_dec_model = EncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
+        enc_dec_model.config.output_attentions = True  # model config -> won't work
+        enc_dec_model.to(torch_device)
+        outputs_encoder_decoder = enc_dec_model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
         )
-
-        decoder_attentions = outputs_encoder_decoder["decoder_attentions"]
-        num_decoder_layers = (
-            decoder_config.num_decoder_layers
-            if hasattr(decoder_config, "num_decoder_layers")
-            else decoder_config.num_hidden_layers
+        self.assertTrue(
+            all(
+                key not in outputs_encoder_decoder
+                for key in ["encoder_attentions", "decoder_attentions", "cross_attentions"]
             )
-        self.assertEqual(len(decoder_attentions), num_decoder_layers)
-
-        self.assertEqual(
-            decoder_attentions[0].shape[-3:],
-            (decoder_config.num_attention_heads, decoder_input_ids.shape[-1], decoder_input_ids.shape[-1]),
         )
 
-        cross_attentions = outputs_encoder_decoder["cross_attentions"]
-        self.assertEqual(len(cross_attentions), num_decoder_layers)
-
-        cross_attention_input_seq_len = decoder_input_ids.shape[-1] * (
-            1 + (decoder_config.ngram if hasattr(decoder_config, "ngram") else 0)
+        config.output_attentions = True  # inner model config -> will work
+        decoder_config.output_attentions = True
+        encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
+        enc_dec_model = EncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
+        enc_dec_model.to(torch_device)
+        outputs_encoder_decoder = enc_dec_model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
         )
-        self.assertEqual(
-            cross_attentions[0].shape[-3:],
-            (decoder_config.num_attention_heads, cross_attention_input_seq_len, input_ids.shape[-1]),
+        self._check_output_with_attentions(
+            outputs_encoder_decoder, config, input_ids, decoder_config, decoder_input_ids
         )
 
     def check_encoder_decoder_model_generate(self, input_ids, config, decoder_config, **kwargs):
@@ -543,6 +599,10 @@ class EncoderDecoderMixin:
         input_ids_dict = self.prepare_config_and_inputs()
         self.check_encoder_decoder_model_output_attentions(**input_ids_dict)
 
+    def test_encoder_decoder_model_output_attentions_from_config(self):
+        input_ids_dict = self.prepare_config_and_inputs()
+        self.check_encoder_decoder_model_output_attentions_from_config(**input_ids_dict)
+
     def test_encoder_decoder_model_generate(self):
         input_ids_dict = self.prepare_config_and_inputs()
         self.check_encoder_decoder_model_generate(**input_ids_dict)

tests/models/encoder_decoder/test_modeling_tf_encoder_decoder.py

@@ -255,6 +255,40 @@ class TFEncoderDecoderMixin:
             outputs_encoder_decoder["encoder_last_hidden_state"].shape, (input_ids.shape + (config.hidden_size,))
         )
 
+    def _check_output_with_attentions(
+        self, outputs_encoder_decoder, config, input_ids, decoder_config, decoder_input_ids
+    ):
+        encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
+        self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
+
+        self.assertEqual(
+            encoder_attentions[0].shape[-3:], (config.num_attention_heads, input_ids.shape[-1], input_ids.shape[-1])
+        )
+
+        decoder_attentions = outputs_encoder_decoder["decoder_attentions"]
+        num_decoder_layers = (
+            decoder_config.num_decoder_layers
+            if hasattr(decoder_config, "num_decoder_layers")
+            else decoder_config.num_hidden_layers
+        )
+        self.assertEqual(len(decoder_attentions), num_decoder_layers)
+
+        self.assertEqual(
+            decoder_attentions[0].shape[-3:],
+            (decoder_config.num_attention_heads, decoder_input_ids.shape[-1], decoder_input_ids.shape[-1]),
+        )
+
+        cross_attentions = outputs_encoder_decoder["cross_attentions"]
+        self.assertEqual(len(cross_attentions), num_decoder_layers)
+
+        cross_attention_input_seq_len = decoder_input_ids.shape[-1] * (
+            1 + (decoder_config.ngram if hasattr(decoder_config, "ngram") else 0)
+        )
+        self.assertEqual(
+            cross_attentions[0].shape[-3:],
+            (decoder_config.num_attention_heads, cross_attention_input_seq_len, input_ids.shape[-1]),
+        )
+
     def check_encoder_decoder_model_output_attentions(
         self,
         config,
@@ -279,36 +313,57 @@ class TFEncoderDecoderMixin:
             output_attentions=True,
             kwargs=kwargs,
         )
+        self._check_output_with_attentions(
+            outputs_encoder_decoder, config, input_ids, decoder_config, decoder_input_ids
+        )
 
-        encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
-        self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
+    def check_encoder_decoder_model_output_attentions_from_config(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        encoder_hidden_states,
+        decoder_config,
+        decoder_input_ids,
+        decoder_attention_mask,
+        **kwargs
+    ):
+        # Similar to `check_encoder_decoder_model_output_attentions`, but with `output_attentions` triggered from the
+        # config file. Contrarily to most models, changing the model's config won't work -- the defaults are loaded
+        # from the inner models' configurations.
 
-        self.assertEqual(
-            encoder_attentions[0].shape[-3:], (config.num_attention_heads, input_ids.shape[-1], input_ids.shape[-1])
+        decoder_input_ids = decoder_input_ids[:, :-1]
+        decoder_attention_mask = decoder_attention_mask[:, :-1]
+        encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
+        enc_dec_model = TFEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
+        enc_dec_model.config.output_attentions = True  # model config -> won't work
+        outputs_encoder_decoder = enc_dec_model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            kwargs=kwargs,
         )
-
-        decoder_attentions = outputs_encoder_decoder["decoder_attentions"]
-        num_decoder_layers = (
-            decoder_config.num_decoder_layers
-            if hasattr(decoder_config, "num_decoder_layers")
-            else decoder_config.num_hidden_layers
+        self.assertTrue(
+            all(
+                key not in outputs_encoder_decoder
+                for key in ["encoder_attentions", "decoder_attentions", "cross_attentions"]
             )
-        self.assertEqual(len(decoder_attentions), num_decoder_layers)
-
-        self.assertEqual(
-            decoder_attentions[0].shape[-3:],
-            (decoder_config.num_attention_heads, decoder_input_ids.shape[-1], decoder_input_ids.shape[-1]),
         )
 
-        cross_attentions = outputs_encoder_decoder["cross_attentions"]
-        self.assertEqual(len(cross_attentions), num_decoder_layers)
-
-        cross_attention_input_seq_len = decoder_input_ids.shape[-1] * (
-            1 + (decoder_config.ngram if hasattr(decoder_config, "ngram") else 0)
+        config.output_attentions = True  # inner model config -> will work
+        decoder_config.output_attentions = True
+        encoder_model, decoder_model = self.get_encoder_decoder_model(config, decoder_config)
+        enc_dec_model = TFEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
+        outputs_encoder_decoder = enc_dec_model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            kwargs=kwargs,
         )
-        self.assertEqual(
-            cross_attentions[0].shape[-3:],
-            (decoder_config.num_attention_heads, cross_attention_input_seq_len, input_ids.shape[-1]),
+        self._check_output_with_attentions(
+            outputs_encoder_decoder, config, input_ids, decoder_config, decoder_input_ids
         )
 
     def check_encoder_decoder_model_generate(self, input_ids, config, decoder_config, **kwargs):
@@ -570,6 +625,10 @@ class TFEncoderDecoderMixin:
         input_ids_dict = self.prepare_config_and_inputs()
         self.check_encoder_decoder_model_output_attentions(**input_ids_dict)
 
+    def test_encoder_decoder_model_output_attentions_from_config(self):
+        input_ids_dict = self.prepare_config_and_inputs()
+        self.check_encoder_decoder_model_output_attentions_from_config(**input_ids_dict)
+
     def test_encoder_decoder_model_generate(self):
         input_ids_dict = self.prepare_config_and_inputs()
         self.check_encoder_decoder_model_generate(**input_ids_dict)