File size: 14,451 Bytes
43b7e92
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
# coding=utf-8
# Copyright 2024 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 tempfile
import unittest

import numpy as np

from diffusers import (
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    OnnxStableDiffusionPipeline,
    PNDMScheduler,
)
from diffusers.utils.testing_utils import is_onnx_available, nightly, require_onnxruntime, require_torch_gpu

from ..test_pipelines_onnx_common import OnnxPipelineTesterMixin


if is_onnx_available():
    import onnxruntime as ort


class OnnxStableDiffusionPipelineFastTests(OnnxPipelineTesterMixin, unittest.TestCase):
    hub_checkpoint = "hf-internal-testing/tiny-random-OnnxStableDiffusionPipeline"

    def get_dummy_inputs(self, seed=0):
        generator = np.random.RandomState(seed)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 7.5,
            "output_type": "np",
        }
        return inputs

    def test_pipeline_default_ddim(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.65072, 0.58492, 0.48219, 0.55521, 0.53180, 0.55939, 0.50697, 0.39800, 0.46455])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_pipeline_pndm(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config, skip_prk_steps=True)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.65863, 0.59425, 0.49326, 0.56313, 0.53875, 0.56627, 0.51065, 0.39777, 0.46330])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_pipeline_lms(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.53755, 0.60786, 0.47402, 0.49488, 0.51869, 0.49819, 0.47985, 0.38957, 0.44279])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_pipeline_euler(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.53755, 0.60786, 0.47402, 0.49488, 0.51869, 0.49819, 0.47985, 0.38957, 0.44279])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_pipeline_euler_ancestral(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.53817, 0.60812, 0.47384, 0.49530, 0.51894, 0.49814, 0.47984, 0.38958, 0.44271])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_pipeline_dpm_multistep(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.53895, 0.60808, 0.47933, 0.49608, 0.51886, 0.49950, 0.48053, 0.38957, 0.44200])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_prompt_embeds(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        inputs["prompt"] = 3 * [inputs["prompt"]]

        # forward
        output = pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        inputs = self.get_dummy_inputs()
        prompt = 3 * [inputs.pop("prompt")]

        text_inputs = pipe.tokenizer(
            prompt,
            padding="max_length",
            max_length=pipe.tokenizer.model_max_length,
            truncation=True,
            return_tensors="np",
        )
        text_inputs = text_inputs["input_ids"]

        prompt_embeds = pipe.text_encoder(input_ids=text_inputs.astype(np.int32))[0]

        inputs["prompt_embeds"] = prompt_embeds

        # forward
        output = pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

    def test_stable_diffusion_negative_prompt_embeds(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs()
        negative_prompt = 3 * ["this is a negative prompt"]
        inputs["negative_prompt"] = negative_prompt
        inputs["prompt"] = 3 * [inputs["prompt"]]

        # forward
        output = pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        inputs = self.get_dummy_inputs()
        prompt = 3 * [inputs.pop("prompt")]

        embeds = []
        for p in [prompt, negative_prompt]:
            text_inputs = pipe.tokenizer(
                p,
                padding="max_length",
                max_length=pipe.tokenizer.model_max_length,
                truncation=True,
                return_tensors="np",
            )
            text_inputs = text_inputs["input_ids"]

            embeds.append(pipe.text_encoder(input_ids=text_inputs.astype(np.int32))[0])

        inputs["prompt_embeds"], inputs["negative_prompt_embeds"] = embeds

        # forward
        output = pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4


@nightly
@require_onnxruntime
@require_torch_gpu
class OnnxStableDiffusionPipelineIntegrationTests(unittest.TestCase):
    @property
    def gpu_provider(self):
        return (
            "CUDAExecutionProvider",
            {
                "gpu_mem_limit": "15000000000",  # 15GB
                "arena_extend_strategy": "kSameAsRequested",
            },
        )

    @property
    def gpu_options(self):
        options = ort.SessionOptions()
        options.enable_mem_pattern = False
        return options

    def test_inference_default_pndm(self):
        # using the PNDM scheduler by default
        sd_pipe = OnnxStableDiffusionPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4",
            revision="onnx",
            safety_checker=None,
            feature_extractor=None,
            provider=self.gpu_provider,
            sess_options=self.gpu_options,
        )
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        np.random.seed(0)
        output = sd_pipe([prompt], guidance_scale=6.0, num_inference_steps=10, output_type="np")
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.0452, 0.0390, 0.0087, 0.0350, 0.0617, 0.0364, 0.0544, 0.0523, 0.0720])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3

    def test_inference_ddim(self):
        ddim_scheduler = DDIMScheduler.from_pretrained(
            "runwayml/stable-diffusion-v1-5", subfolder="scheduler", revision="onnx"
        )
        sd_pipe = OnnxStableDiffusionPipeline.from_pretrained(
            "runwayml/stable-diffusion-v1-5",
            revision="onnx",
            scheduler=ddim_scheduler,
            safety_checker=None,
            feature_extractor=None,
            provider=self.gpu_provider,
            sess_options=self.gpu_options,
        )
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "open neural network exchange"
        generator = np.random.RandomState(0)
        output = sd_pipe([prompt], guidance_scale=7.5, num_inference_steps=10, generator=generator, output_type="np")
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.2867, 0.1974, 0.1481, 0.7294, 0.7251, 0.6667, 0.4194, 0.5642, 0.6486])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3

    def test_inference_k_lms(self):
        lms_scheduler = LMSDiscreteScheduler.from_pretrained(
            "runwayml/stable-diffusion-v1-5", subfolder="scheduler", revision="onnx"
        )
        sd_pipe = OnnxStableDiffusionPipeline.from_pretrained(
            "runwayml/stable-diffusion-v1-5",
            revision="onnx",
            scheduler=lms_scheduler,
            safety_checker=None,
            feature_extractor=None,
            provider=self.gpu_provider,
            sess_options=self.gpu_options,
        )
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "open neural network exchange"
        generator = np.random.RandomState(0)
        output = sd_pipe([prompt], guidance_scale=7.5, num_inference_steps=10, generator=generator, output_type="np")
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.2306, 0.1959, 0.1593, 0.6549, 0.6394, 0.5408, 0.5065, 0.6010, 0.6161])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3

    def test_intermediate_state(self):
        number_of_steps = 0

        def test_callback_fn(step: int, timestep: int, latents: np.ndarray) -> None:
            test_callback_fn.has_been_called = True
            nonlocal number_of_steps
            number_of_steps += 1
            if step == 0:
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array(
                    [-0.6772, -0.3835, -1.2456, 0.1905, -1.0974, 0.6967, -1.9353, 0.0178, 1.0167]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-3
            elif step == 5:
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array(
                    [-0.3351, 0.2241, -0.1837, -0.2325, -0.6577, 0.3393, -0.0241, 0.5899, 1.3875]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-3

        test_callback_fn.has_been_called = False

        pipe = OnnxStableDiffusionPipeline.from_pretrained(
            "runwayml/stable-diffusion-v1-5",
            revision="onnx",
            safety_checker=None,
            feature_extractor=None,
            provider=self.gpu_provider,
            sess_options=self.gpu_options,
        )
        pipe.set_progress_bar_config(disable=None)

        prompt = "Andromeda galaxy in a bottle"

        generator = np.random.RandomState(0)
        pipe(
            prompt=prompt,
            num_inference_steps=5,
            guidance_scale=7.5,
            generator=generator,
            callback=test_callback_fn,
            callback_steps=1,
        )
        assert test_callback_fn.has_been_called
        assert number_of_steps == 6

    def test_stable_diffusion_no_safety_checker(self):
        pipe = OnnxStableDiffusionPipeline.from_pretrained(
            "runwayml/stable-diffusion-v1-5",
            revision="onnx",
            safety_checker=None,
            feature_extractor=None,
            provider=self.gpu_provider,
            sess_options=self.gpu_options,
        )
        assert isinstance(pipe, OnnxStableDiffusionPipeline)
        assert pipe.safety_checker is None

        image = pipe("example prompt", num_inference_steps=2).images[0]
        assert image is not None

        # check that there's no error when saving a pipeline with one of the models being None
        with tempfile.TemporaryDirectory() as tmpdirname:
            pipe.save_pretrained(tmpdirname)
            pipe = OnnxStableDiffusionPipeline.from_pretrained(tmpdirname)

        # sanity check that the pipeline still works
        assert pipe.safety_checker is None
        image = pipe("example prompt", num_inference_steps=2).images[0]
        assert image is not None