Improve Swin for VisionEncoderDecoder (#16070)

* Add Swin2Bart test

* Fix swin tests
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

src/transformers/models/swin/configuration_swin.py

@@ -94,6 +94,7 @@ class SwinConfig(PretrainedConfig):
 
     attribute_map = {
         "num_attention_heads": "num_heads",
+        "num_hidden_layers": "num_layers",
     }
 
     def __init__(
@@ -141,4 +142,4 @@ class SwinConfig(PretrainedConfig):
         self.encoder_stride = encoder_stride
         # we set the hidden_size attribute in order to make Swin work with VisionEncoderDecoderModel
         # this indicates the channel dimension after the last stage of the model
-        self.hidden_size = embed_dim * 8
+        self.hidden_size = int(embed_dim * 2 ** (len(depths) - 1))

tests/swin/test_modeling_swin.py

@@ -56,8 +56,8 @@ class SwinModelTester:
         patch_size=2,
         num_channels=3,
         embed_dim=16,
-        depths=[1],
-        num_heads=[2],
+        depths=[1, 2, 1],
+        num_heads=[2, 2, 4],
         window_size=2,
         mlp_ratio=2.0,
         qkv_bias=True,
@@ -73,7 +73,7 @@ class SwinModelTester:
         scope=None,
         use_labels=True,
         type_sequence_label_size=10,
-        encoder_stride=2,
+        encoder_stride=8,
     ):
         self.parent = parent
         self.batch_size = batch_size
@@ -139,8 +139,7 @@ class SwinModelTester:
         model.eval()
         result = model(pixel_values)
 
-        # since the model we're testing only consists of a single layer, expected_seq_len = number of patches
-        expected_seq_len = (config.image_size // config.patch_size) ** 2
+        expected_seq_len = ((config.image_size // config.patch_size) ** 2) // (4 ** (len(config.depths) - 1))
         expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
 
         self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))

tests/vision_encoder_decoder/test_modeling_vision_encoder_decoder.py

@@ -22,8 +22,10 @@ from datasets import load_dataset
 from transformers.file_utils import cached_property, is_torch_available, is_vision_available
 from transformers.testing_utils import require_torch, require_vision, slow, torch_device
 
+from ..bart.test_modeling_bart import BartModelTester
 from ..bert.test_modeling_bert import BertModelTester
 from ..deit.test_modeling_deit import DeiTModelTester
+from ..swin.test_modeling_swin import SwinModelTester
 from ..test_modeling_common import floats_tensor, ids_tensor, random_attention_mask
 from ..trocr.test_modeling_trocr import TrOCRStandaloneDecoderModelTester
 from ..vit.test_modeling_vit import ViTModelTester
@@ -35,8 +37,10 @@ if is_torch_available():
 
     from transformers import (
         AutoTokenizer,
+        BartForCausalLM,
         BertLMHeadModel,
         DeiTModel,
+        SwinModel,
         TrOCRForCausalLM,
         VisionEncoderDecoderConfig,
         VisionEncoderDecoderModel,
@@ -514,6 +518,90 @@ class ViT2BertModelTest(EncoderDecoderMixin, unittest.TestCase):
         }
 
 
+@require_torch
+class Swin2BartModelTest(EncoderDecoderMixin, unittest.TestCase):
+    def get_encoder_decoder_model(self, config, decoder_config):
+        encoder_model = SwinModel(config).eval()
+        decoder_model = BartForCausalLM(decoder_config).eval()
+        return encoder_model, decoder_model
+
+    def prepare_config_and_inputs(self):
+        model_tester_encoder = SwinModelTester(self, batch_size=13, embed_dim=32)
+        model_tester_decoder = BartModelTester(self, batch_size=13, hidden_size=32, max_position_embeddings=512)
+        encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
+        decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs()
+        config, pixel_values, _ = encoder_config_and_inputs
+        decoder_config, decoder_inputs_dict = decoder_config_and_inputs
+
+        # make sure that cross attention layers are added
+        decoder_config.add_cross_attention = True
+        #  disable cache for now
+        decoder_config.use_cache = False
+        return {
+            "config": config,
+            "pixel_values": pixel_values,
+            "decoder_config": decoder_config,
+            **decoder_inputs_dict,
+        }
+
+    def check_encoder_decoder_model_output_attentions(
+        self,
+        config,
+        decoder_config,
+        decoder_input_ids,
+        decoder_attention_mask,
+        labels=None,
+        pixel_values=None,
+        **kwargs
+    ):
+        # make the decoder inputs a different shape from the encoder inputs to harden the test
+        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 = VisionEncoderDecoderModel(encoder=encoder_model, decoder=decoder_model)
+        enc_dec_model.to(torch_device)
+        outputs_encoder_decoder = enc_dec_model(
+            pixel_values=pixel_values,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            output_attentions=True,
+        )
+
+        encoder_attentions = outputs_encoder_decoder["encoder_attentions"]
+        self.assertEqual(len(encoder_attentions), config.num_hidden_layers)
+
+        # in Swin, the seq_len equals:
+        seq_len = encoder_model.config.window_size**2
+        self.assertEqual(encoder_attentions[0].shape[-3:], (config.num_attention_heads[0], seq_len, seq_len))
+
+        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)
+
+        encoder_seq_len = ((config.image_size // config.patch_size) ** 2) // (4 ** (len(config.depths) - 1))
+        cross_attention_input_seq_len = decoder_input_ids.shape[-1]
+        self.assertEqual(
+            cross_attentions[0].shape[-3:],
+            (decoder_config.num_attention_heads, cross_attention_input_seq_len, encoder_seq_len),
+        )
+
+    # there are no published pretrained BART-causal checkpoints for now
+    def test_real_model_save_load_from_pretrained(self):
+        pass
+
+
 @require_torch
 class ViT2TrOCR(EncoderDecoderMixin, unittest.TestCase):
     def get_encoder_decoder_model(self, config, decoder_config):