test_text_to_image.py 14.6 KB
Newer Older
1
#!/usr/bin/env python
2
# coding=utf-8
3
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
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
#
# 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 logging
import os
import shutil
import sys
import tempfile

from diffusers import DiffusionPipeline, UNet2DConditionModel  # noqa: E402


sys.path.append("..")
from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402


logging.basicConfig(level=logging.DEBUG)

logger = logging.getLogger()
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)


class TextToImage(ExamplesTestsAccelerate):
    def test_text_to_image(self):
        with tempfile.TemporaryDirectory() as tmpdir:
            test_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path hf-internal-testing/tiny-stable-diffusion-pipe
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
                --max_train_steps 2
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                """.split()

            run_command(self._launch_args + test_args)
            # save_pretrained smoke test
            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))

    def test_text_to_image_checkpointing(self):
        pretrained_model_name_or_path = "hf-internal-testing/tiny-stable-diffusion-pipe"
        prompt = "a prompt"

        with tempfile.TemporaryDirectory() as tmpdir:
            # Run training script with checkpointing
68
            # max_train_steps == 4, checkpointing_steps == 2
69
70
71
72
73
74
75
76
77
78
79
            # Should create checkpoints at steps 2, 4

            initial_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
80
                --max_train_steps 4
81
82
83
84
85
86
87
88
89
90
91
92
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                --checkpointing_steps=2
                --seed=0
                """.split()

            run_command(self._launch_args + initial_run_args)

            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
93
            pipe(prompt, num_inference_steps=1)
94
95
96
97
98
99
100
101
102
103

            # check checkpoint directories exist
            self.assertEqual(
                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
                {"checkpoint-2", "checkpoint-4"},
            )

            # check can run an intermediate checkpoint
            unet = UNet2DConditionModel.from_pretrained(tmpdir, subfolder="checkpoint-2/unet")
            pipe = DiffusionPipeline.from_pretrained(pretrained_model_name_or_path, unet=unet, safety_checker=None)
104
            pipe(prompt, num_inference_steps=1)
105
106
107
108

            # Remove checkpoint 2 so that we can check only later checkpoints exist after resuming
            shutil.rmtree(os.path.join(tmpdir, "checkpoint-2"))

109
            # Run training script for 2 total steps resuming from checkpoint 4
110
111
112
113
114
115
116
117
118
119

            resume_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
120
                --max_train_steps 2
121
122
123
124
125
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
126
                --checkpointing_steps=1
127
128
129
130
131
132
133
134
                --resume_from_checkpoint=checkpoint-4
                --seed=0
                """.split()

            run_command(self._launch_args + resume_run_args)

            # check can run new fully trained pipeline
            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
135
            pipe(prompt, num_inference_steps=1)
136

137
138
            # no checkpoint-2 -> check old checkpoints do not exist
            # check new checkpoints exist
139
140
            self.assertEqual(
                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
141
                {"checkpoint-4", "checkpoint-5"},
142
143
144
145
146
147
148
149
            )

    def test_text_to_image_checkpointing_use_ema(self):
        pretrained_model_name_or_path = "hf-internal-testing/tiny-stable-diffusion-pipe"
        prompt = "a prompt"

        with tempfile.TemporaryDirectory() as tmpdir:
            # Run training script with checkpointing
150
            # max_train_steps == 4, checkpointing_steps == 2
151
152
153
154
155
156
157
158
159
160
161
            # Should create checkpoints at steps 2, 4

            initial_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
162
                --max_train_steps 4
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                --checkpointing_steps=2
                --use_ema
                --seed=0
                """.split()

            run_command(self._launch_args + initial_run_args)

            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
            pipe(prompt, num_inference_steps=2)

            # check checkpoint directories exist
            self.assertEqual(
                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
                {"checkpoint-2", "checkpoint-4"},
            )

            # check can run an intermediate checkpoint
            unet = UNet2DConditionModel.from_pretrained(tmpdir, subfolder="checkpoint-2/unet")
            pipe = DiffusionPipeline.from_pretrained(pretrained_model_name_or_path, unet=unet, safety_checker=None)
187
            pipe(prompt, num_inference_steps=1)
188
189
190
191

            # Remove checkpoint 2 so that we can check only later checkpoints exist after resuming
            shutil.rmtree(os.path.join(tmpdir, "checkpoint-2"))

192
            # Run training script for 2 total steps resuming from checkpoint 4
193
194
195
196
197
198
199
200
201
202

            resume_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
203
                --max_train_steps 2
204
205
206
207
208
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
209
                --checkpointing_steps=1
210
211
212
213
214
215
216
217
218
                --resume_from_checkpoint=checkpoint-4
                --use_ema
                --seed=0
                """.split()

            run_command(self._launch_args + resume_run_args)

            # check can run new fully trained pipeline
            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
219
            pipe(prompt, num_inference_steps=1)
220

221
222
            # no checkpoint-2 -> check old checkpoints do not exist
            # check new checkpoints exist
223
224
            self.assertEqual(
                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
225
                {"checkpoint-4", "checkpoint-5"},
226
227
228
229
230
231
232
233
            )

    def test_text_to_image_checkpointing_checkpoints_total_limit(self):
        pretrained_model_name_or_path = "hf-internal-testing/tiny-stable-diffusion-pipe"
        prompt = "a prompt"

        with tempfile.TemporaryDirectory() as tmpdir:
            # Run training script with checkpointing
234
            # max_train_steps == 6, checkpointing_steps == 2, checkpoints_total_limit == 2
235
236
237
238
239
240
241
242
243
244
245
246
            # Should create checkpoints at steps 2, 4, 6
            # with checkpoint at step 2 deleted

            initial_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
247
                --max_train_steps 6
248
249
250
251
252
253
254
255
256
257
258
259
260
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                --checkpointing_steps=2
                --checkpoints_total_limit=2
                --seed=0
                """.split()

            run_command(self._launch_args + initial_run_args)

            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
261
            pipe(prompt, num_inference_steps=1)
262
263

            # check checkpoint directories exist
264
265
            # checkpoint-2 should have been deleted
            self.assertEqual({x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-4", "checkpoint-6"})
266
267
268
269
270
271
272

    def test_text_to_image_checkpointing_checkpoints_total_limit_removes_multiple_checkpoints(self):
        pretrained_model_name_or_path = "hf-internal-testing/tiny-stable-diffusion-pipe"
        prompt = "a prompt"

        with tempfile.TemporaryDirectory() as tmpdir:
            # Run training script with checkpointing
273
274
            # max_train_steps == 4, checkpointing_steps == 2
            # Should create checkpoints at steps 2, 4
275
276
277
278
279
280
281
282
283
284

            initial_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
285
                --max_train_steps 4
286
287
288
289
290
291
292
293
294
295
296
297
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                --checkpointing_steps=2
                --seed=0
                """.split()

            run_command(self._launch_args + initial_run_args)

            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
298
            pipe(prompt, num_inference_steps=1)
299
300
301
302

            # check checkpoint directories exist
            self.assertEqual(
                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
303
                {"checkpoint-2", "checkpoint-4"},
304
305
            )

306
            # resume and we should try to checkpoint at 6, where we'll have to remove
307
308
309
310
311
312
313
314
315
316
317
            # checkpoint-2 and checkpoint-4 instead of just a single previous checkpoint

            resume_run_args = f"""
                examples/text_to_image/train_text_to_image.py
                --pretrained_model_name_or_path {pretrained_model_name_or_path}
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
318
                --max_train_steps 8
319
320
321
322
323
324
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                --checkpointing_steps=2
325
326
                --resume_from_checkpoint=checkpoint-4
                --checkpoints_total_limit=2
327
328
329
330
331
332
                --seed=0
                """.split()

            run_command(self._launch_args + resume_run_args)

            pipe = DiffusionPipeline.from_pretrained(tmpdir, safety_checker=None)
333
            pipe(prompt, num_inference_steps=1)
334
335
336
337

            # check checkpoint directories exist
            self.assertEqual(
                {x for x in os.listdir(tmpdir) if "checkpoint" in x},
338
                {"checkpoint-6", "checkpoint-8"},
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
            )


class TextToImageSDXL(ExamplesTestsAccelerate):
    def test_text_to_image_sdxl(self):
        with tempfile.TemporaryDirectory() as tmpdir:
            test_args = f"""
                examples/text_to_image/train_text_to_image_sdxl.py
                --pretrained_model_name_or_path hf-internal-testing/tiny-stable-diffusion-xl-pipe
                --dataset_name hf-internal-testing/dummy_image_text_data
                --resolution 64
                --center_crop
                --random_flip
                --train_batch_size 1
                --gradient_accumulation_steps 1
                --max_train_steps 2
                --learning_rate 5.0e-04
                --scale_lr
                --lr_scheduler constant
                --lr_warmup_steps 0
                --output_dir {tmpdir}
                """.split()

            run_command(self._launch_args + test_args)
            # save_pretrained smoke test
            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "unet", "diffusion_pytorch_model.safetensors")))
            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "scheduler", "scheduler_config.json")))