File size: 9,968 Bytes
f1069cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding=utf-8
# Copyright 2023 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 math
import unittest

import torch

from diffusers import UNet2DModel
from diffusers.utils import floats_tensor, logging, slow, torch_all_close, torch_device

from ..test_modeling_common import ModelTesterMixin


logger = logging.get_logger(__name__)
torch.backends.cuda.matmul.allow_tf32 = False


class Unet2DModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = UNet2DModel

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor([10]).to(torch_device)

        return {"sample": noise, "timestep": time_step}

    @property
    def input_shape(self):
        return (3, 32, 32)

    @property
    def output_shape(self):
        return (3, 32, 32)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "block_out_channels": (32, 64),
            "down_block_types": ("DownBlock2D", "AttnDownBlock2D"),
            "up_block_types": ("AttnUpBlock2D", "UpBlock2D"),
            "attention_head_dim": None,
            "out_channels": 3,
            "in_channels": 3,
            "layers_per_block": 2,
            "sample_size": 32,
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict


class UNetLDMModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = UNet2DModel

    @property
    def dummy_input(self):
        batch_size = 4
        num_channels = 4
        sizes = (32, 32)

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor([10]).to(torch_device)

        return {"sample": noise, "timestep": time_step}

    @property
    def input_shape(self):
        return (4, 32, 32)

    @property
    def output_shape(self):
        return (4, 32, 32)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "sample_size": 32,
            "in_channels": 4,
            "out_channels": 4,
            "layers_per_block": 2,
            "block_out_channels": (32, 64),
            "attention_head_dim": 32,
            "down_block_types": ("DownBlock2D", "DownBlock2D"),
            "up_block_types": ("UpBlock2D", "UpBlock2D"),
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict

    def test_from_pretrained_hub(self):
        model, loading_info = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)

        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        image = model(**self.dummy_input).sample

        assert image is not None, "Make sure output is not None"

    @unittest.skipIf(torch_device != "cuda", "This test is supposed to run on GPU")
    def test_from_pretrained_accelerate(self):
        model, _ = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)
        model.to(torch_device)
        image = model(**self.dummy_input).sample

        assert image is not None, "Make sure output is not None"

    @unittest.skipIf(torch_device != "cuda", "This test is supposed to run on GPU")
    def test_from_pretrained_accelerate_wont_change_results(self):
        # by defautl model loading will use accelerate as `low_cpu_mem_usage=True`
        model_accelerate, _ = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)
        model_accelerate.to(torch_device)
        model_accelerate.eval()

        noise = torch.randn(
            1,
            model_accelerate.config.in_channels,
            model_accelerate.config.sample_size,
            model_accelerate.config.sample_size,
            generator=torch.manual_seed(0),
        )
        noise = noise.to(torch_device)
        time_step = torch.tensor([10] * noise.shape[0]).to(torch_device)

        arr_accelerate = model_accelerate(noise, time_step)["sample"]

        # two models don't need to stay in the device at the same time
        del model_accelerate
        torch.cuda.empty_cache()
        gc.collect()

        model_normal_load, _ = UNet2DModel.from_pretrained(
            "fusing/unet-ldm-dummy-update", output_loading_info=True, low_cpu_mem_usage=False
        )
        model_normal_load.to(torch_device)
        model_normal_load.eval()
        arr_normal_load = model_normal_load(noise, time_step)["sample"]

        assert torch_all_close(arr_accelerate, arr_normal_load, rtol=1e-3)

    def test_output_pretrained(self):
        model = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update")
        model.eval()
        model.to(torch_device)

        noise = torch.randn(
            1,
            model.config.in_channels,
            model.config.sample_size,
            model.config.sample_size,
            generator=torch.manual_seed(0),
        )
        noise = noise.to(torch_device)
        time_step = torch.tensor([10] * noise.shape[0]).to(torch_device)

        with torch.no_grad():
            output = model(noise, time_step).sample

        output_slice = output[0, -1, -3:, -3:].flatten().cpu()
        # fmt: off
        expected_output_slice = torch.tensor([-13.3258, -20.1100, -15.9873, -17.6617, -23.0596, -17.9419, -13.3675, -16.1889, -12.3800])
        # fmt: on

        self.assertTrue(torch_all_close(output_slice, expected_output_slice, rtol=1e-3))


class NCSNppModelTests(ModelTesterMixin, unittest.TestCase):
    model_class = UNet2DModel

    @property
    def dummy_input(self, sizes=(32, 32)):
        batch_size = 4
        num_channels = 3

        noise = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [10]).to(dtype=torch.int32, device=torch_device)

        return {"sample": noise, "timestep": time_step}

    @property
    def input_shape(self):
        return (3, 32, 32)

    @property
    def output_shape(self):
        return (3, 32, 32)

    def prepare_init_args_and_inputs_for_common(self):
        init_dict = {
            "block_out_channels": [32, 64, 64, 64],
            "in_channels": 3,
            "layers_per_block": 1,
            "out_channels": 3,
            "time_embedding_type": "fourier",
            "norm_eps": 1e-6,
            "mid_block_scale_factor": math.sqrt(2.0),
            "norm_num_groups": None,
            "down_block_types": [
                "SkipDownBlock2D",
                "AttnSkipDownBlock2D",
                "SkipDownBlock2D",
                "SkipDownBlock2D",
            ],
            "up_block_types": [
                "SkipUpBlock2D",
                "SkipUpBlock2D",
                "AttnSkipUpBlock2D",
                "SkipUpBlock2D",
            ],
        }
        inputs_dict = self.dummy_input
        return init_dict, inputs_dict

    @slow
    def test_from_pretrained_hub(self):
        model, loading_info = UNet2DModel.from_pretrained("google/ncsnpp-celebahq-256", output_loading_info=True)
        self.assertIsNotNone(model)
        self.assertEqual(len(loading_info["missing_keys"]), 0)

        model.to(torch_device)
        inputs = self.dummy_input
        noise = floats_tensor((4, 3) + (256, 256)).to(torch_device)
        inputs["sample"] = noise
        image = model(**inputs)

        assert image is not None, "Make sure output is not None"

    @slow
    def test_output_pretrained_ve_mid(self):
        model = UNet2DModel.from_pretrained("google/ncsnpp-celebahq-256")
        model.to(torch_device)

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        batch_size = 4
        num_channels = 3
        sizes = (256, 256)

        noise = torch.ones((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [1e-4]).to(torch_device)

        with torch.no_grad():
            output = model(noise, time_step).sample

        output_slice = output[0, -3:, -3:, -1].flatten().cpu()
        # fmt: off
        expected_output_slice = torch.tensor([-4836.2231, -6487.1387, -3816.7969, -7964.9253, -10966.2842, -20043.6016, 8137.0571, 2340.3499, 544.6114])
        # fmt: on

        self.assertTrue(torch_all_close(output_slice, expected_output_slice, rtol=1e-2))

    def test_output_pretrained_ve_large(self):
        model = UNet2DModel.from_pretrained("fusing/ncsnpp-ffhq-ve-dummy-update")
        model.to(torch_device)

        torch.manual_seed(0)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(0)

        batch_size = 4
        num_channels = 3
        sizes = (32, 32)

        noise = torch.ones((batch_size, num_channels) + sizes).to(torch_device)
        time_step = torch.tensor(batch_size * [1e-4]).to(torch_device)

        with torch.no_grad():
            output = model(noise, time_step).sample

        output_slice = output[0, -3:, -3:, -1].flatten().cpu()
        # fmt: off
        expected_output_slice = torch.tensor([-0.0325, -0.0900, -0.0869, -0.0332, -0.0725, -0.0270, -0.0101, 0.0227, 0.0256])
        # fmt: on

        self.assertTrue(torch_all_close(output_slice, expected_output_slice, rtol=1e-2))

    def test_forward_with_norm_groups(self):
        # not required for this model
        pass