Add VP test in Stable diffusion pipeline (#19300)

### Description
1. Add visual parity test based on openai clip model
2. Add trigger rules

### Motivation and Context
1. check generated image is expected
2. reduce unnecessary triggers

onnxruntime/python/tools/transformers/models/stable_diffusion/demo_txt2img.py

@@ -61,6 +61,7 @@ if __name__ == "__main__":
             controlnet_scales=controlnet_scale,
             show_latency=not warmup,
             output_type="pil",
+            deterministic=args.deterministic,
         )
 
     if not args.disable_cuda_graph:

onnxruntime/python/tools/transformers/models/stable_diffusion/demo_utils.py

@@ -239,6 +239,7 @@ def parse_arguments(is_xl: bool, parser):
     )
     parser.add_argument("--nvtx-profile", action="store_true", help="Enable NVTX markers for performance profiling.")
     parser.add_argument("--seed", type=int, default=None, help="Seed for random generator to get consistent results.")
+    parser.add_argument("--deterministic", action="store_true", help="use deterministic algorithms.")
     parser.add_argument("-dc", "--disable-cuda-graph", action="store_true", help="Disable cuda graph.")
 
     group = parser.add_argument_group("Options for ORT_CUDA engine only")

onnxruntime/python/tools/transformers/models/stable_diffusion/pipeline_stable_diffusion.py

@@ -754,6 +754,7 @@ class StableDiffusionPipeline:
         controlnet_scales: Optional[torch.Tensor] = None,
         show_latency: bool = False,
         output_type: str = "pil",
+        deterministic: bool = False,
     ):
         """
         Run the diffusion pipeline.
@@ -783,6 +784,9 @@ class StableDiffusionPipeline:
             output_type (str):
                 It can be "latent", "pt" or "pil".
         """
+        if deterministic:
+            torch.use_deterministic_algorithms(True)
+
         if self.is_backend_tensorrt():
             import tensorrt as trt
             from trt_utilities import TRT_LOGGER

onnxruntime/python/tools/transformers/models/stable_diffusion/test/check_image.py

@@ -0,0 +1,68 @@
+import argparse
+import os
+
+import cv2
+import open_clip
+import torch
+from PIL import Image
+from sentence_transformers import util
+
+
+def arg_parser():
+    parser = argparse.ArgumentParser(description="Options for Compare 2 image")
+    parser.add_argument("--image1", type=str, help="Path to image 1")
+    parser.add_argument("--image2", type=str, help="Path to image 2")
+    args = parser.parse_args()
+    return args
+
+
+def image_encoder(img: Image.Image):  # -> torch.Tensor:
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model, _, preprocess = open_clip.create_model_and_transforms("ViT-B-16-plus-240", pretrained="laion400m_e32")
+    model.to(device)
+
+    img1 = Image.fromarray(img).convert("RGB")
+    img1 = preprocess(img1).unsqueeze(0).to(device)
+    img1 = model.encode_image(img1)
+    return img1
+
+
+def load_image(image_path: str):  # -> Image.Image:
+    # cv2.imread() can silently fail when the path is too long
+    # https://stackoverflow.com/questions/68716321/how-to-use-absolute-path-in-cv2-imread
+    if os.path.isabs(image_path):
+        directory = os.path.dirname(image_path)
+        current_directory = os.getcwd()
+        os.chdir(directory)
+        img = cv2.imread(os.path.basename(image_path), cv2.IMREAD_UNCHANGED)
+        os.chdir(current_directory)
+    else:
+        img = cv2.imread(image_path, cv2.IMREAD_UNCHANGED)
+    return img
+
+
+def generate_score(image1: str, image2: str):  # -> float:
+    test_img = load_image(image1)
+    data_img = load_image(image2)
+    img1 = image_encoder(test_img)
+    img2 = image_encoder(data_img)
+    cos_scores = util.pytorch_cos_sim(img1, img2)
+    score = round(float(cos_scores[0][0]) * 100, 2)
+    return score
+
+
+def main():
+    args = arg_parser()
+    image1 = args.image1
+    image2 = args.image2
+    score = round(generate_score(image1, image2), 2)
+    print("similarity Score: ", {score})
+    if score < 99:
+        print(f"{image1} and {image2} are different")
+        raise SystemExit(1)
+    else:
+        print(f"{image1} and {image2} are same")
+
+
+if __name__ == "__main__":
+    main()

onnxruntime/python/tools/transformers/models/stable_diffusion/test/requirements.txt

@@ -0,0 +1,4 @@
+git+https://github.com/openai/CLIP.git
+open_clip_torch
+sentence_transformers
+pillow

tools/ci_build/github/azure-pipelines/bigmodels-ci-pipeline.yml

@@ -1,3 +1,32 @@
+##### start trigger Don't edit it manually, Please do edit set-trigger-rules.py ####
+trigger:
+  branches:
+    include:
+    - main
+    - rel-*
+  paths:
+    exclude:
+    - docs/**
+    - README.md
+    - CONTRIBUTING.md
+    - BUILD.md
+    - 'js/web'
+    - 'onnxruntime/core/providers/js'
+pr:
+  branches:
+    include:
+    - main
+    - rel-*
+  paths:
+    exclude:
+    - docs/**
+    - README.md
+    - CONTRIBUTING.md
+    - BUILD.md
+    - 'js/web'
+    - 'onnxruntime/core/providers/js'
+#### end trigger ####parameters:
+
 # reference: https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/transformers/models/stable_diffusion/README.md
 parameters:
 - name: specificArtifact
@@ -143,7 +172,6 @@ stages:
   - job: Stable_Diffusion
     variables:
       skipComponentGovernanceDetection: true
-      CCACHE_DIR: $(Pipeline.Workspace)/ccache
     workspace:
       clean: all
     pool: onnxruntime-Linux-GPU-A10-12G
@@ -162,7 +190,7 @@ stages:
 
     - script: |
         docker run --rm --gpus all -v $PWD:/workspace -v $(Build.BinariesDirectory)/Release:/Release nvcr.io/nvidia/pytorch:22.11-py3 \
-          bash -c "
+          bash -c '
             set -ex; \
             python3 --version; \
             python3 -m pip install --upgrade pip; \
@@ -171,12 +199,31 @@ stages:
             python3 -m pip install -r requirements-cuda11.txt; \
             python3 -m pip install --upgrade polygraphy onnx-graphsurgeon --extra-index-url https://pypi.ngc.nvidia.com; \
             echo Generate an image guided by a text prompt; \
-            python3 demo_txt2img.py 'astronaut riding a horse on mars'; \
-            popd; \
-          "
+            python3 demo_txt2img.py --seed 1 --deterministic "astronaut riding a horse on mars" ; \
+            find $(pwd) -name "*.png" ; \
+            popd ; \
+            '
       displayName: 'Run stable diffusion demo'
       workingDirectory: $(Build.SourcesDirectory)
 
+    - script: |
+        docker run --rm --gpus all -v $PWD:/workspace nvcr.io/nvidia/pytorch:22.11-py3 \
+          bash -c '
+            set -ex; \
+            python3 --version; \
+            python3 -m pip install --upgrade pip; \
+            pushd /workspace/onnxruntime/python/tools/transformers/models/stable_diffusion/; \
+            image2=$(find $(pwd) -name "astronaut_riding_a_h*.png") ; \
+            pushd test; \
+            python3 -m pip install -r requirements.txt; \
+            echo check demo_txt2image.py generate image; \
+            python3 -u check_image.py --image1 astronaut_riding_txt2image-DDIM-50.png --image2 $image2; \
+            popd ; \
+            popd ; \
+            '
+      displayName: 'Check the generated image'
+      workingDirectory: $(Build.SourcesDirectory)
+
 - stage: Llama2_ONNX_FP16
   dependsOn:
   - Build_Onnxruntime_Cuda

tools/ci_build/set-trigger-rules.py

@@ -14,6 +14,7 @@ skip_doc_changes = ["web-ci-pipeline.yml"]
 skip_js_changes = [
     "android-arm64-v8a-QNN-crosscompile-ci-pipeline.yml",
     "android-x86_64-crosscompile-ci-pipeline.yml",
+    "bigmodels-ci-pipeline.yml",
     "linux-ci-pipeline.yml",
     "linux-cpu-aten-pipeline.yml",
     "linux-cpu-eager-pipeline.yml",
@@ -31,7 +32,6 @@ skip_js_changes = [
     "orttraining-linux-ci-pipeline.yml",
     "orttraining-linux-gpu-ci-pipeline.yml",
     "orttraining-linux-gpu-ortmodule-distributed-test-ci-pipeline.yml",
-    "orttraining-linux-gpu-training-apis.yml",
     "orttraining-mac-ci-pipeline.yml",
     "win-ci-pipeline.yml",
     "win-gpu-ci-pipeline.yml",