Add kolors eval

siliconflow · Jul 20, 2024 · 3b942b5 · 3b942b5
1 parent 0d56572
commit 3b942b5
Show file tree

Hide file tree

Showing 18 changed files with 21 additions and 1 deletion.
diff --git a/metrics/fid.py b/metrics/fid.py
@@ -2,6 +2,6 @@
 from T2IBenchmark.datasets import get_coco_fid_stats
 
 fid, _ = calculate_fid(
-    'output_images',
+    '/home/lixiang/data/fid_kolors_nexfort',
     get_coco_fid_stats()
 )
diff --git a/models/kolors/README.md b/models/kolors/README.md
@@ -0,0 +1,18 @@
+# Using onediff to accelerate the quality evaluation of kolors.
+
+
+## Qualitative evaluation
+
+We used several typical prompts to visualize generated images, proving that the compilation acceleration of onediff is almost lossless.
+
+| Prompt | Reference Image | OnFlow Backend | NexFort Backend |
+|--------|-----------------|----------------|-----------------|
+| 穿着黑色T恤的可爱小狗，T恤上面中文绿色大字写着“可图”。 | <img src="./asset/kolors_01.png" width="150px"> | <img src="./asset/kolors_oneflow_compile_01.png" width="150px"> | <img src="./asset/kolors_nexfort_compile_01.png" width="150px"> |
+| 一张兔子的特写照片，春天的森林中，有雾、晕影、开花、戏剧性的氛围，以三分法为中心，200mm 1.4f的微距镜头拍摄。 | <img src="./asset/kolors_02.png" width="150px"> | <img src="./asset/kolors_oneflow_compile_02.png" width="150px"> | <img src="./asset/kolors_nexfort_compile_02.png" width="150px"> |
+| 一条繁忙的街道，车辆在两个方向上行驶，街道上有几辆双层巴士和周围的人们。 | <img src="./asset/kolors_03.png" width="150px"> | <img src="./asset/kolors_oneflow_compile_03.png" width="150px"> | <img src="./asset/kolors_nexfort_compile_03.png" width="150px"> |
+| 动漫风格，一名女孩在室内，坐在客厅的沙发上，拥有粉红色的头发、白色的袜子、蓝色的眼睛，从背后、从上方看，面向观众，正在玩视频游戏，手持控制器，穿着黑色丝绸，嘴唇微张。 | <img src="./asset/kolors_04.png" width="150px"> | <img src="./asset/kolors_oneflow_compile_04.png" width="150px"> | <img src="./asset/kolors_nexfort_compile_04.png" width="150px"> |
+| 一张年轻中国女性在公园的肖像照片。她的长黑发轻轻飘动。背景中柔和的樱花粉增添了画面的美感。 | <img src="./asset/kolors_05.png" width="150px"> | <img src="./asset/kolors_oneflow_compile_05.png" width="150px"> | <img src="./asset/kolors_nexfort_compile_05.png" width="150px"> |
+
+
+
+## Quantitative evaluation
diff --git a/models/kolors/asset/kolors_01.png b/models/kolors/asset/kolors_01.png
diff --git a/models/kolors/asset/kolors_02.png b/models/kolors/asset/kolors_02.png
diff --git a/models/kolors/asset/kolors_03.png b/models/kolors/asset/kolors_03.png
diff --git a/models/kolors/asset/kolors_04.png b/models/kolors/asset/kolors_04.png
diff --git a/models/kolors/asset/kolors_05.png b/models/kolors/asset/kolors_05.png
diff --git a/models/kolors/asset/kolors_nexfort_compile_01.png b/models/kolors/asset/kolors_nexfort_compile_01.png
diff --git a/models/kolors/asset/kolors_nexfort_compile_02.png b/models/kolors/asset/kolors_nexfort_compile_02.png
diff --git a/models/kolors/asset/kolors_nexfort_compile_03.png b/models/kolors/asset/kolors_nexfort_compile_03.png
diff --git a/models/kolors/asset/kolors_nexfort_compile_04.png b/models/kolors/asset/kolors_nexfort_compile_04.png
diff --git a/models/kolors/asset/kolors_nexfort_compile_05.png b/models/kolors/asset/kolors_nexfort_compile_05.png
diff --git a/models/kolors/asset/kolors_oneflow_compile_01.png b/models/kolors/asset/kolors_oneflow_compile_01.png
diff --git a/models/kolors/asset/kolors_oneflow_compile_02.png b/models/kolors/asset/kolors_oneflow_compile_02.png
diff --git a/models/kolors/asset/kolors_oneflow_compile_03.png b/models/kolors/asset/kolors_oneflow_compile_03.png
diff --git a/models/kolors/asset/kolors_oneflow_compile_04.png b/models/kolors/asset/kolors_oneflow_compile_04.png
diff --git a/models/kolors/asset/kolors_oneflow_compile_05.png b/models/kolors/asset/kolors_oneflow_compile_05.png
diff --git a/models/kolors/test.py b/models/kolors/test.py
@@ -162,11 +162,13 @@ def main():
     }
 
     kolors.warmup(gen_args, args.warmup_iterations)
+    # torch.cuda.empty_cache()
 
     for idx, prompt in enumerate(prompts):
         gen_args["prompt"] = prompt
         output_path = os.path.join(args.output_dir, f"image_{idx+1}.png")
         image, inference_time = kolors.generate(gen_args, output_path)
+        # torch.cuda.empty_cache()
         print(f"Generated image saved to {output_path} in {inference_time:.2f} seconds.")
         cuda_mem_after_used = torch.cuda.max_memory_allocated() / (1024 ** 3)
         print(f"Max used CUDA memory: {cuda_mem_after_used:.3f} GiB")