test_modeling_common.py 16.5 KB
Newer Older
1
# coding=utf-8
Patrick von Platen's avatar
Patrick von Platen committed
2
# Copyright 2023 HuggingFace Inc.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
#
# 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 inspect
import tempfile
18
import unittest
19
import unittest.mock as mock
20
from typing import Dict, List, Tuple
21
22

import numpy as np
23
import requests_mock
24
import torch
25
from requests.exceptions import HTTPError
26

27
from diffusers.models import ModelMixin, UNet2DConditionModel
28
from diffusers.training_utils import EMAModel
29
from diffusers.utils import torch_device
30
31


32
class ModelUtilsTest(unittest.TestCase):
33
34
35
36
37
38
39
    def tearDown(self):
        super().tearDown()

        import diffusers

        diffusers.utils.import_utils._safetensors_available = True

40
41
42
43
44
45
46
    def test_accelerate_loading_error_message(self):
        with self.assertRaises(ValueError) as error_context:
            UNet2DConditionModel.from_pretrained("hf-internal-testing/stable-diffusion-broken", subfolder="unet")

        # make sure that error message states what keys are missing
        assert "conv_out.bias" in str(error_context.exception)

47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
    def test_cached_files_are_used_when_no_internet(self):
        # A mock response for an HTTP head request to emulate server down
        response_mock = mock.Mock()
        response_mock.status_code = 500
        response_mock.headers = {}
        response_mock.raise_for_status.side_effect = HTTPError
        response_mock.json.return_value = {}

        # Download this model to make sure it's in the cache.
        orig_model = UNet2DConditionModel.from_pretrained(
            "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet"
        )

        # Under the mock environment we get a 500 error when trying to reach the model.
        with mock.patch("requests.request", return_value=response_mock):
            # Download this model to make sure it's in the cache.
            model = UNet2DConditionModel.from_pretrained(
                "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet", local_files_only=True
            )

        for p1, p2 in zip(orig_model.parameters(), model.parameters()):
            if p1.data.ne(p2.data).sum() > 0:
                assert False, "Parameters not the same!"

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
    def test_one_request_upon_cached(self):
        # TODO: For some reason this test fails on MPS where no HEAD call is made.
        if torch_device == "mps":
            return

        import diffusers

        diffusers.utils.import_utils._safetensors_available = False

        with tempfile.TemporaryDirectory() as tmpdirname:
            with requests_mock.mock(real_http=True) as m:
                UNet2DConditionModel.from_pretrained(
                    "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet", cache_dir=tmpdirname
                )

            download_requests = [r.method for r in m.request_history]
            assert download_requests.count("HEAD") == 2, "2 HEAD requests one for config, one for model"
            assert download_requests.count("GET") == 2, "2 GET requests one for config, one for model"

            with requests_mock.mock(real_http=True) as m:
                UNet2DConditionModel.from_pretrained(
                    "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet", cache_dir=tmpdirname
                )

            cache_requests = [r.method for r in m.request_history]
            assert (
                "HEAD" == cache_requests[0] and len(cache_requests) == 1
            ), "We should call only `model_info` to check for _commit hash and `send_telemetry`"

        diffusers.utils.import_utils._safetensors_available = True

102

103
class ModelTesterMixin:
104
    def test_from_save_pretrained(self):
105
106
107
108
109
110
111
112
113
114
115
116
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with tempfile.TemporaryDirectory() as tmpdirname:
            model.save_pretrained(tmpdirname)
            new_model = self.model_class.from_pretrained(tmpdirname)
            new_model.to(torch_device)

        with torch.no_grad():
117
118
119
120
121
            # Warmup pass when using mps (see #372)
            if torch_device == "mps" and isinstance(model, ModelMixin):
                _ = model(**self.dummy_input)
                _ = new_model(**self.dummy_input)

122
123
            image = model(**inputs_dict)
            if isinstance(image, dict):
124
                image = image.sample
125
126
127
128

            new_image = new_model(**inputs_dict)

            if isinstance(new_image, dict):
129
                new_image = new_image.sample
130
131
132

        max_diff = (image - new_image).abs().sum().item()
        self.assertLessEqual(max_diff, 5e-5, "Models give different forward passes")
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

    def test_from_save_pretrained_variant(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with tempfile.TemporaryDirectory() as tmpdirname:
            model.save_pretrained(tmpdirname, variant="fp16")
            new_model = self.model_class.from_pretrained(tmpdirname, variant="fp16")

            # non-variant cannot be loaded
            with self.assertRaises(OSError) as error_context:
                self.model_class.from_pretrained(tmpdirname)

            # make sure that error message states what keys are missing
            assert "Error no file named diffusion_pytorch_model.bin found in directory" in str(error_context.exception)

            new_model.to(torch_device)

        with torch.no_grad():
            # Warmup pass when using mps (see #372)
            if torch_device == "mps" and isinstance(model, ModelMixin):
                _ = model(**self.dummy_input)
                _ = new_model(**self.dummy_input)

            image = model(**inputs_dict)
            if isinstance(image, dict):
                image = image.sample

            new_image = new_model(**inputs_dict)

            if isinstance(new_image, dict):
                new_image = new_image.sample

        max_diff = (image - new_image).abs().sum().item()
        self.assertLessEqual(max_diff, 5e-5, "Models give different forward passes")
171

172
173
174
175
176
177
178
179
180
181
182
183
184
    def test_from_save_pretrained_dtype(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        for dtype in [torch.float32, torch.float16, torch.bfloat16]:
            if torch_device == "mps" and dtype == torch.bfloat16:
                continue
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.to(dtype)
                model.save_pretrained(tmpdirname)
185
                new_model = self.model_class.from_pretrained(tmpdirname, low_cpu_mem_usage=True, torch_dtype=dtype)
186
                assert new_model.dtype == dtype
187
                new_model = self.model_class.from_pretrained(tmpdirname, low_cpu_mem_usage=False, torch_dtype=dtype)
188
189
                assert new_model.dtype == dtype

190
191
192
193
194
    def test_determinism(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()
195

196
        with torch.no_grad():
197
198
199
200
            # Warmup pass when using mps (see #372)
            if torch_device == "mps" and isinstance(model, ModelMixin):
                model(**self.dummy_input)

201
202
            first = model(**inputs_dict)
            if isinstance(first, dict):
203
                first = first.sample
204
205
206

            second = model(**inputs_dict)
            if isinstance(second, dict):
207
                second = second.sample
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225

        out_1 = first.cpu().numpy()
        out_2 = second.cpu().numpy()
        out_1 = out_1[~np.isnan(out_1)]
        out_2 = out_2[~np.isnan(out_2)]
        max_diff = np.amax(np.abs(out_1 - out_2))
        self.assertLessEqual(max_diff, 1e-5)

    def test_output(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)

            if isinstance(output, dict):
226
                output = output.sample
227
228
229
230
231

        self.assertIsNotNone(output)
        expected_shape = inputs_dict["sample"].shape
        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")

232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
    def test_forward_with_norm_groups(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        init_dict["norm_num_groups"] = 16
        init_dict["block_out_channels"] = (16, 32)

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)

            if isinstance(output, dict):
                output = output.sample

        self.assertIsNotNone(output)
        expected_shape = inputs_dict["sample"].shape
        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")

252
253
254
255
256
257
258
259
260
261
262
    def test_forward_signature(self):
        init_dict, _ = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        signature = inspect.signature(model.forward)
        # signature.parameters is an OrderedDict => so arg_names order is deterministic
        arg_names = [*signature.parameters.keys()]

        expected_arg_names = ["sample", "timestep"]
        self.assertListEqual(arg_names[:2], expected_arg_names)

263
    def test_model_from_pretrained(self):
264
265
266
267
268
269
270
271
272
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        # test if the model can be loaded from the config
        # and has all the expected shape
        with tempfile.TemporaryDirectory() as tmpdirname:
273
274
            model.save_pretrained(tmpdirname)
            new_model = self.model_class.from_pretrained(tmpdirname)
275
276
277
            new_model.to(torch_device)
            new_model.eval()

278
        # check if all parameters shape are the same
279
280
281
282
283
284
285
286
287
        for param_name in model.state_dict().keys():
            param_1 = model.state_dict()[param_name]
            param_2 = new_model.state_dict()[param_name]
            self.assertEqual(param_1.shape, param_2.shape)

        with torch.no_grad():
            output_1 = model(**inputs_dict)

            if isinstance(output_1, dict):
288
                output_1 = output_1.sample
289
290
291
292

            output_2 = new_model(**inputs_dict)

            if isinstance(output_2, dict):
293
                output_2 = output_2.sample
294
295
296

        self.assertEqual(output_1.shape, output_2.shape)

297
    @unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
298
299
300
301
302
303
304
305
306
    def test_training(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.train()
        output = model(**inputs_dict)

        if isinstance(output, dict):
307
            output = output.sample
308
309
310
311
312

        noise = torch.randn((inputs_dict["sample"].shape[0],) + self.output_shape).to(torch_device)
        loss = torch.nn.functional.mse_loss(output, noise)
        loss.backward()

313
    @unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
314
315
316
317
318
319
    def test_ema_training(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.train()
320
        ema_model = EMAModel(model.parameters())
321
322
323
324

        output = model(**inputs_dict)

        if isinstance(output, dict):
325
            output = output.sample
326
327
328
329

        noise = torch.randn((inputs_dict["sample"].shape[0],) + self.output_shape).to(torch_device)
        loss = torch.nn.functional.mse_loss(output, noise)
        loss.backward()
330
        ema_model.step(model.parameters())
331

332
    def test_outputs_equivalence(self):
333
        def set_nan_tensor_to_zero(t):
334
335
336
337
338
            # Temporary fallback until `aten::_index_put_impl_` is implemented in mps
            # Track progress in https://github.com/pytorch/pytorch/issues/77764
            device = t.device
            if device.type == "mps":
                t = t.to("cpu")
339
            t[t != t] = 0
340
            return t.to(device)
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

        def recursive_check(tuple_object, dict_object):
            if isinstance(tuple_object, (List, Tuple)):
                for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object.values()):
                    recursive_check(tuple_iterable_value, dict_iterable_value)
            elif isinstance(tuple_object, Dict):
                for tuple_iterable_value, dict_iterable_value in zip(tuple_object.values(), dict_object.values()):
                    recursive_check(tuple_iterable_value, dict_iterable_value)
            elif tuple_object is None:
                return
            else:
                self.assertTrue(
                    torch.allclose(
                        set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                    ),
                    msg=(
                        "Tuple and dict output are not equal. Difference:"
                        f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
                        f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
                        f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
                    ),
                )

        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

370
371
372
373
374
375
376
        with torch.no_grad():
            # Warmup pass when using mps (see #372)
            if torch_device == "mps" and isinstance(model, ModelMixin):
                model(**self.dummy_input)

            outputs_dict = model(**inputs_dict)
            outputs_tuple = model(**inputs_dict, return_dict=False)
377
378

        recursive_check(outputs_tuple, outputs_dict)
379

Anton Lozhkov's avatar
Anton Lozhkov committed
380
    @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
    def test_enable_disable_gradient_checkpointing(self):
        if not self.model_class._supports_gradient_checkpointing:
            return  # Skip test if model does not support gradient checkpointing

        init_dict, _ = self.prepare_init_args_and_inputs_for_common()

        # at init model should have gradient checkpointing disabled
        model = self.model_class(**init_dict)
        self.assertFalse(model.is_gradient_checkpointing)

        # check enable works
        model.enable_gradient_checkpointing()
        self.assertTrue(model.is_gradient_checkpointing)

        # check disable works
        model.disable_gradient_checkpointing()
        self.assertFalse(model.is_gradient_checkpointing)
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417

    def test_deprecated_kwargs(self):
        has_kwarg_in_model_class = "kwargs" in inspect.signature(self.model_class.__init__).parameters
        has_deprecated_kwarg = len(self.model_class._deprecated_kwargs) > 0

        if has_kwarg_in_model_class and not has_deprecated_kwarg:
            raise ValueError(
                f"{self.model_class} has `**kwargs` in its __init__ method but has not defined any deprecated kwargs"
                " under the `_deprecated_kwargs` class attribute. Make sure to either remove `**kwargs` if there are"
                " no deprecated arguments or add the deprecated argument with `_deprecated_kwargs ="
                " [<deprecated_argument>]`"
            )

        if not has_kwarg_in_model_class and has_deprecated_kwarg:
            raise ValueError(
                f"{self.model_class} doesn't have `**kwargs` in its __init__ method but has defined deprecated kwargs"
                " under the `_deprecated_kwargs` class attribute. Make sure to either add the `**kwargs` argument to"
                f" {self.model_class}.__init__ if there are deprecated arguments or remove the deprecated argument"
                " from `_deprecated_kwargs = [<deprecated_argument>]`"
            )