Add SD2 inpainting integration tests (#1412)

SD2 inpainting integration tests

Author: anton-l

tests/pipelines/stable_diffusion_2/test_stable_diffusion_inpaint.py

@@ -0,0 +1,345 @@
+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import random
+import unittest
+
+import numpy as np
+import torch
+
+from diffusers import AutoencoderKL, PNDMScheduler, StableDiffusionInpaintPipeline, UNet2DConditionModel
+from diffusers.utils import floats_tensor, load_image, load_numpy, torch_device
+from diffusers.utils.testing_utils import require_torch_gpu
+from PIL import Image
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class StableDiffusionInpaintPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    @property
+    def dummy_image(self):
+        batch_size = 1
+        num_channels = 3
+        sizes = (32, 32)
+
+        image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device)
+        return image
+
+    @property
+    def dummy_cond_unet_inpaint(self):
+        torch.manual_seed(0)
+        model = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=9,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=32,
+            # SD2-specific config below
+            attention_head_dim=(2, 4, 8, 8),
+            use_linear_projection=True,
+        )
+        return model
+
+    @property
+    def dummy_vae(self):
+        torch.manual_seed(0)
+        model = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+        )
+        return model
+
+    @property
+    def dummy_text_encoder(self):
+        torch.manual_seed(0)
+        config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            # SD2-specific config below
+            hidden_act="gelu",
+            projection_dim=512,
+        )
+        return CLIPTextModel(config)
+
+    @property
+    def dummy_extractor(self):
+        def extract(*args, **kwargs):
+            class Out:
+                def __init__(self):
+                    self.pixel_values = torch.ones([0])
+
+                def to(self, device):
+                    self.pixel_values.to(device)
+                    return self
+
+            return Out()
+
+        return extract
+
+    def test_stable_diffusion_inpaint(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        unet = self.dummy_cond_unet_inpaint
+        scheduler = PNDMScheduler(skip_prk_steps=True)
+        vae = self.dummy_vae
+        text_encoder = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
+        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+        mask_image = Image.fromarray(np.uint8(image + 4)).convert("RGB").resize((64, 64))
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionInpaintPipeline(
+            unet=unet,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            safety_checker=None,
+            feature_extractor=None,
+        )
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=device).manual_seed(0)
+        output = sd_pipe(
+            [prompt],
+            generator=generator,
+            guidance_scale=6.0,
+            num_inference_steps=2,
+            output_type="np",
+            image=init_image,
+            mask_image=mask_image,
+        )
+
+        image = output.images
+
+        generator = torch.Generator(device=device).manual_seed(0)
+        image_from_tuple = sd_pipe(
+            [prompt],
+            generator=generator,
+            guidance_scale=6.0,
+            num_inference_steps=2,
+            output_type="np",
+            image=init_image,
+            mask_image=mask_image,
+            return_dict=False,
+        )[0]
+
+        image_slice = image[0, -3:, -3:, -1]
+        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 64, 64, 3)
+        expected_slice = np.array([0.4727, 0.5735, 0.3941, 0.5446, 0.5926, 0.4394, 0.5062, 0.4654, 0.4476])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
+
+    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
+    def test_stable_diffusion_inpaint_fp16(self):
+        """Test that stable diffusion inpaint works with fp16"""
+        unet = self.dummy_cond_unet_inpaint
+        scheduler = PNDMScheduler(skip_prk_steps=True)
+        vae = self.dummy_vae
+        text_encoder = self.dummy_text_encoder
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
+        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
+        mask_image = Image.fromarray(np.uint8(image + 4)).convert("RGB").resize((64, 64))
+
+        # put models in fp16
+        unet = unet.half()
+        vae = vae.half()
+        text_encoder = text_encoder.half()
+
+        # make sure here that pndm scheduler skips prk
+        sd_pipe = StableDiffusionInpaintPipeline(
+            unet=unet,
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            safety_checker=None,
+            feature_extractor=None,
+        )
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        prompt = "A painting of a squirrel eating a burger"
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        image = sd_pipe(
+            [prompt],
+            generator=generator,
+            num_inference_steps=2,
+            output_type="np",
+            image=init_image,
+            mask_image=mask_image,
+        ).images
+
+        assert image.shape == (1, 64, 64, 3)
+
+
+# @slow
+@require_torch_gpu
+class StableDiffusionInpaintPipelineIntegrationTests(unittest.TestCase):
+    def tearDown(self):
+        # clean up the VRAM after each test
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_stable_diffusion_inpaint_pipeline(self):
+        init_image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+            "/sd2-inpaint/init_image.png"
+        )
+        mask_image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint/mask.png"
+        )
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint"
+            "/yellow_cat_sitting_on_a_park_bench.npy"
+        )
+
+        model_id = "stabilityai/stable-diffusion-2-inpainting"
+        pipe = StableDiffusionInpaintPipeline.from_pretrained(model_id, safety_checker=None)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = pipe(
+            prompt=prompt,
+            image=init_image,
+            mask_image=mask_image,
+            generator=generator,
+            output_type="np",
+        )
+        image = output.images[0]
+
+        assert image.shape == (512, 512, 3)
+        assert np.abs(expected_image - image).max() < 1e-3
+
+    def test_stable_diffusion_inpaint_pipeline_fp16(self):
+        init_image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+            "/sd2-inpaint/init_image.png"
+        )
+        mask_image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint/mask.png"
+        )
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint"
+            "/yellow_cat_sitting_on_a_park_bench_fp16.npy"
+        )
+
+        model_id = "stabilityai/stable-diffusion-2-inpainting"
+        pipe = StableDiffusionInpaintPipeline.from_pretrained(
+            model_id,
+            revision="fp16",
+            torch_dtype=torch.float16,
+            safety_checker=None,
+        )
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        output = pipe(
+            prompt=prompt,
+            image=init_image,
+            mask_image=mask_image,
+            generator=generator,
+            output_type="np",
+        )
+        image = output.images[0]
+
+        assert image.shape == (512, 512, 3)
+        assert np.abs(expected_image - image).max() < 5e-1
+
+    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
+        torch.cuda.empty_cache()
+        torch.cuda.reset_max_memory_allocated()
+        torch.cuda.reset_peak_memory_stats()
+
+        init_image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+            "/sd2-inpaint/init_image.png"
+        )
+        mask_image = load_image(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-inpaint/mask.png"
+        )
+
+        model_id = "stabilityai/stable-diffusion-2-inpainting"
+        pndm = PNDMScheduler.from_pretrained(model_id, subfolder="scheduler")
+        pipe = StableDiffusionInpaintPipeline.from_pretrained(
+            model_id,
+            safety_checker=None,
+            scheduler=pndm,
+            device_map="auto",
+            revision="fp16",
+            torch_dtype=torch.float16,
+        )
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing(1)
+        pipe.enable_sequential_cpu_offload()
+
+        prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+        _ = pipe(
+            prompt=prompt,
+            image=init_image,
+            mask_image=mask_image,
+            generator=generator,
+            num_inference_steps=5,
+            output_type="np",
+        )
+
+        mem_bytes = torch.cuda.max_memory_allocated()
+        # make sure that less than 2.65 GB is allocated
+        assert mem_bytes < 2.65 * 10**9