Merge pull request CompVis#127 from CompVis/scene-images-oi

Added scene image generation for Open Images

Author: rromb

assets/scene_images_samples.svg

@@ -48,13 +48,12 @@ data:
   target: main.DataModuleFromConfig
   params:
     batch_size: 6
-    num_workers: 12
     train:
       target: taming.data.annotated_objects_coco.AnnotatedObjectsCoco
       params:
-        data_path: data/coco_annotations_100
+        data_path: data/coco_annotations_100  # substitute with path to full dataset
         split: train
-        keys: [image, objects_bbox, file_name]
+        keys: [image, objects_bbox, file_name, annotations]
         no_tokens: 8192
         target_image_size: 256
         min_object_area: 0.00001
@@ -67,9 +66,9 @@ data:
     validation:
       target: taming.data.annotated_objects_coco.AnnotatedObjectsCoco
       params:
-        data_path: data/coco_annotations_100
+        data_path: data/coco_annotations_100  # substitute with path to full dataset
         split: validation
-        keys: [image, objects_bbox, file_name]
+        keys: [image, objects_bbox, file_name, annotations]
         no_tokens: 8192
         target_image_size: 256
         min_object_area: 0.00001

configs/coco_scene_images_transformer.yaml

@@ -0,0 +1,86 @@
+model:
+  base_learning_rate: 4.5e-06
+  target: taming.models.cond_transformer.Net2NetTransformer
+  params:
+    cond_stage_key: objects_bbox
+    transformer_config:
+      target: taming.modules.transformer.mingpt.GPT
+      params:
+        vocab_size: 8192
+        block_size: 348  # = 256 + 92 = dim(vqgan_latent_space,16x16) + dim(conditional_builder.embedding_dim)
+        n_layer: 36
+        n_head: 16
+        n_embd: 1536
+        embd_pdrop: 0.1
+        resid_pdrop: 0.1
+        attn_pdrop: 0.1
+    first_stage_config:
+      target: taming.models.vqgan.VQModel
+      params:
+        ckpt_path: /path/to/coco_oi_epoch12.ckpt  # https://heibox.uni-heidelberg.de/f/461d9a9f4fcf48ab84f4/
+        embed_dim: 256
+        n_embed: 8192
+        ddconfig:
+          double_z: false
+          z_channels: 256
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 1
+          - 2
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions:
+          - 16
+          dropout: 0.0
+        lossconfig:
+          target: taming.modules.losses.DummyLoss
+    cond_stage_config:
+      target: taming.models.dummy_cond_stage.DummyCondStage
+      params:
+        conditional_key: objects_bbox
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 6
+    train:
+      target: taming.data.annotated_objects_open_images.AnnotatedObjectsOpenImages
+      params:
+        data_path: data/open_images_annotations_100  # substitute with path to full dataset
+        split: train
+        keys: [image, objects_bbox, file_name, annotations]
+        no_tokens: 8192
+        target_image_size: 256
+        category_allow_list_target: taming.data.open_images_helper.top_300_classes_plus_coco_compatibility
+        category_mapping_target: taming.data.open_images_helper.open_images_unify_categories_for_coco
+        min_object_area: 0.0001
+        min_objects_per_image: 2
+        max_objects_per_image: 30
+        crop_method: random-2d
+        random_flip: true
+        use_group_parameter: true
+        use_additional_parameters: true
+        encode_crop: true
+    validation:
+      target: taming.data.annotated_objects_open_images.AnnotatedObjectsOpenImages
+      params:
+        data_path: data/open_images_annotations_100  # substitute with path to full dataset
+        split: validation
+        keys: [image, objects_bbox, file_name, annotations]
+        no_tokens: 8192
+        target_image_size: 256
+        category_allow_list_target: taming.data.open_images_helper.top_300_classes_plus_coco_compatibility
+        category_mapping_target: taming.data.open_images_helper.open_images_unify_categories_for_coco
+        min_object_area: 0.0001
+        min_objects_per_image: 2
+        max_objects_per_image: 30
+        crop_method: center
+        random_flip: false
+        use_group_parameter: true
+        use_additional_parameters: true
+        encode_crop: true