File size: 12,801 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
import random
import unittest

import numpy as np
import torch
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer

import diffusers
from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    MotionAdapter,
    PIAPipeline,
    UNet2DConditionModel,
    UNetMotionModel,
)
from diffusers.utils import is_xformers_available, logging
from diffusers.utils.testing_utils import floats_tensor, torch_device

from ..test_pipelines_common import IPAdapterTesterMixin, PipelineFromPipeTesterMixin, PipelineTesterMixin


def to_np(tensor):
    if isinstance(tensor, torch.Tensor):
        tensor = tensor.detach().cpu().numpy()

    return tensor


class PIAPipelineFastTests(IPAdapterTesterMixin, PipelineTesterMixin, PipelineFromPipeTesterMixin, unittest.TestCase):
    pipeline_class = PIAPipeline
    params = frozenset(
        [
            "prompt",
            "height",
            "width",
            "guidance_scale",
            "negative_prompt",
            "prompt_embeds",
            "negative_prompt_embeds",
            "cross_attention_kwargs",
        ]
    )
    batch_params = frozenset(["prompt", "image", "generator"])
    required_optional_params = frozenset(
        [
            "num_inference_steps",
            "generator",
            "latents",
            "return_dict",
            "callback_on_step_end",
            "callback_on_step_end_tensor_inputs",
        ]
    )

    def get_dummy_components(self):
        torch.manual_seed(0)
        unet = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("CrossAttnDownBlock2D", "DownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
            norm_num_groups=2,
        )
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="linear",
            clip_sample=False,
        )
        torch.manual_seed(0)
        vae = 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,
        )
        torch.manual_seed(0)
        text_encoder_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,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
        motion_adapter = MotionAdapter(
            block_out_channels=(32, 64),
            motion_layers_per_block=2,
            motion_norm_num_groups=2,
            motion_num_attention_heads=4,
            conv_in_channels=9,
        )

        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "motion_adapter": motion_adapter,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "feature_extractor": None,
            "image_encoder": None,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)

        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        inputs = {
            "image": image,
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 7.5,
            "output_type": "pt",
        }
        return inputs

    def test_motion_unet_loading(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        assert isinstance(pipe.unet, UNetMotionModel)

    def test_ip_adapter_single(self):
        expected_pipe_slice = None

        if torch_device == "cpu":
            expected_pipe_slice = np.array(
                [
                    0.5609,
                    0.5756,
                    0.4830,
                    0.4420,
                    0.4547,
                    0.5129,
                    0.3779,
                    0.4042,
                    0.3772,
                    0.4450,
                    0.5710,
                    0.5536,
                    0.4835,
                    0.4308,
                    0.5578,
                    0.5578,
                    0.4395,
                    0.5440,
                    0.6051,
                    0.4651,
                    0.6258,
                    0.5662,
                    0.3988,
                    0.5108,
                    0.4153,
                    0.3993,
                    0.4803,
                ]
            )
        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)

    def test_dict_tuple_outputs_equivalent(self):
        expected_slice = None
        if torch_device == "cpu":
            expected_slice = np.array([0.3740, 0.4284, 0.4038, 0.5417, 0.4405, 0.5521, 0.4273, 0.4124, 0.4997])
        return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)

    @unittest.skip("Attention slicing is not enabled in this pipeline")
    def test_attention_slicing_forward_pass(self):
        pass

    def test_inference_batch_single_identical(
        self,
        batch_size=2,
        expected_max_diff=1e-4,
        additional_params_copy_to_batched_inputs=["num_inference_steps"],
    ):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        for components in pipe.components.values():
            if hasattr(components, "set_default_attn_processor"):
                components.set_default_attn_processor()

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        inputs = self.get_dummy_inputs(torch_device)
        # Reset generator in case it is has been used in self.get_dummy_inputs
        inputs["generator"] = self.get_generator(0)

        logger = logging.get_logger(pipe.__module__)
        logger.setLevel(level=diffusers.logging.FATAL)

        # batchify inputs
        batched_inputs = {}
        batched_inputs.update(inputs)

        for name in self.batch_params:
            if name not in inputs:
                continue

            value = inputs[name]
            if name == "prompt":
                len_prompt = len(value)
                batched_inputs[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
                batched_inputs[name][-1] = 100 * "very long"

            else:
                batched_inputs[name] = batch_size * [value]

        if "generator" in inputs:
            batched_inputs["generator"] = [self.get_generator(i) for i in range(batch_size)]

        if "batch_size" in inputs:
            batched_inputs["batch_size"] = batch_size

        for arg in additional_params_copy_to_batched_inputs:
            batched_inputs[arg] = inputs[arg]

        output = pipe(**inputs)
        output_batch = pipe(**batched_inputs)

        assert output_batch[0].shape[0] == batch_size

        max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
        assert max_diff < expected_max_diff

    @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
    def test_to_device(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.set_progress_bar_config(disable=None)

        pipe.to("cpu")
        # pipeline creates a new motion UNet under the hood. So we need to check the device from pipe.components
        model_devices = [
            component.device.type for component in pipe.components.values() if hasattr(component, "device")
        ]
        self.assertTrue(all(device == "cpu" for device in model_devices))

        output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
        self.assertTrue(np.isnan(output_cpu).sum() == 0)

        pipe.to("cuda")
        model_devices = [
            component.device.type for component in pipe.components.values() if hasattr(component, "device")
        ]
        self.assertTrue(all(device == "cuda" for device in model_devices))

        output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
        self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)

    def test_to_dtype(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.set_progress_bar_config(disable=None)

        # pipeline creates a new motion UNet under the hood. So we need to check the dtype from pipe.components
        model_dtypes = [component.dtype for component in pipe.components.values() if hasattr(component, "dtype")]
        self.assertTrue(all(dtype == torch.float32 for dtype in model_dtypes))

        pipe.to(dtype=torch.float16)
        model_dtypes = [component.dtype for component in pipe.components.values() if hasattr(component, "dtype")]
        self.assertTrue(all(dtype == torch.float16 for dtype in model_dtypes))

    def test_prompt_embeds(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.set_progress_bar_config(disable=None)
        pipe.to(torch_device)

        inputs = self.get_dummy_inputs(torch_device)
        inputs.pop("prompt")
        inputs["prompt_embeds"] = torch.randn((1, 4, 32), device=torch_device)
        pipe(**inputs)

    def test_free_init(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.set_progress_bar_config(disable=None)
        pipe.to(torch_device)

        inputs_normal = self.get_dummy_inputs(torch_device)
        frames_normal = pipe(**inputs_normal).frames[0]

        pipe.enable_free_init(
            num_iters=2,
            use_fast_sampling=True,
            method="butterworth",
            order=4,
            spatial_stop_frequency=0.25,
            temporal_stop_frequency=0.25,
        )
        inputs_enable_free_init = self.get_dummy_inputs(torch_device)
        frames_enable_free_init = pipe(**inputs_enable_free_init).frames[0]

        pipe.disable_free_init()
        inputs_disable_free_init = self.get_dummy_inputs(torch_device)
        frames_disable_free_init = pipe(**inputs_disable_free_init).frames[0]

        sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_init)).sum()
        max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_init)).max()
        self.assertGreater(
            sum_enabled, 1e1, "Enabling of FreeInit should lead to results different from the default pipeline results"
        )
        self.assertLess(
            max_diff_disabled,
            1e-4,
            "Disabling of FreeInit should lead to results similar to the default pipeline results",
        )

    @unittest.skipIf(
        torch_device != "cuda" or not is_xformers_available(),
        reason="XFormers attention is only available with CUDA and `xformers` installed",
    )
    def test_xformers_attention_forwardGenerator_pass(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        for component in pipe.components.values():
            if hasattr(component, "set_default_attn_processor"):
                component.set_default_attn_processor()
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output_without_offload = pipe(**inputs).frames[0]
        output_without_offload = (
            output_without_offload.cpu() if torch.is_tensor(output_without_offload) else output_without_offload
        )

        pipe.enable_xformers_memory_efficient_attention()
        inputs = self.get_dummy_inputs(torch_device)
        output_with_offload = pipe(**inputs).frames[0]
        output_with_offload = (
            output_with_offload.cpu() if torch.is_tensor(output_with_offload) else output_without_offload
        )

        max_diff = np.abs(to_np(output_with_offload) - to_np(output_without_offload)).max()
        self.assertLess(max_diff, 1e-4, "XFormers attention should not affect the inference results")