declare-lab/tango2

1# coding=utf-82# Copyright 2023 HuggingFace Inc.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15 16import gc17import unittest18 19import torch20from parameterized import parameterized21 22from diffusers import AutoencoderKL23from diffusers.utils import floats_tensor, load_hf_numpy, require_torch_gpu, slow, torch_all_close, torch_device24 25from ..test_modeling_common import ModelTesterMixin26 27 28torch.backends.cuda.matmul.allow_tf32 = False29 30 31class AutoencoderKLTests(ModelTesterMixin, unittest.TestCase):32    model_class = AutoencoderKL33 34    @property35    def dummy_input(self):36        batch_size = 437        num_channels = 338        sizes = (32, 32)39 40        image = floats_tensor((batch_size, num_channels) + sizes).to(torch_device)41 42        return {"sample": image}43 44    @property45    def input_shape(self):46        return (3, 32, 32)47 48    @property49    def output_shape(self):50        return (3, 32, 32)51 52    def prepare_init_args_and_inputs_for_common(self):53        init_dict = {54            "block_out_channels": [32, 64],55            "in_channels": 3,56            "out_channels": 3,57            "down_block_types": ["DownEncoderBlock2D", "DownEncoderBlock2D"],58            "up_block_types": ["UpDecoderBlock2D", "UpDecoderBlock2D"],59            "latent_channels": 4,60        }61        inputs_dict = self.dummy_input62        return init_dict, inputs_dict63 64    def test_forward_signature(self):65        pass66 67    def test_training(self):68        pass69 70    @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")71    def test_gradient_checkpointing(self):72        # enable deterministic behavior for gradient checkpointing73        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()74        model = self.model_class(**init_dict)75        model.to(torch_device)76 77        assert not model.is_gradient_checkpointing and model.training78 79        out = model(**inputs_dict).sample80        # run the backwards pass on the model. For backwards pass, for simplicity purpose,81        # we won't calculate the loss and rather backprop on out.sum()82        model.zero_grad()83 84        labels = torch.randn_like(out)85        loss = (out - labels).mean()86        loss.backward()87 88        # re-instantiate the model now enabling gradient checkpointing89        model_2 = self.model_class(**init_dict)90        # clone model91        model_2.load_state_dict(model.state_dict())92        model_2.to(torch_device)93        model_2.enable_gradient_checkpointing()94 95        assert model_2.is_gradient_checkpointing and model_2.training96 97        out_2 = model_2(**inputs_dict).sample98        # run the backwards pass on the model. For backwards pass, for simplicity purpose,99        # we won't calculate the loss and rather backprop on out.sum()100        model_2.zero_grad()101        loss_2 = (out_2 - labels).mean()102        loss_2.backward()103 104        # compare the output and parameters gradients105        self.assertTrue((loss - loss_2).abs() < 1e-5)106        named_params = dict(model.named_parameters())107        named_params_2 = dict(model_2.named_parameters())108        for name, param in named_params.items():109            self.assertTrue(torch_all_close(param.grad.data, named_params_2[name].grad.data, atol=5e-5))110 111    def test_from_pretrained_hub(self):112        model, loading_info = AutoencoderKL.from_pretrained("fusing/autoencoder-kl-dummy", output_loading_info=True)113        self.assertIsNotNone(model)114        self.assertEqual(len(loading_info["missing_keys"]), 0)115 116        model.to(torch_device)117        image = model(**self.dummy_input)118 119        assert image is not None, "Make sure output is not None"120 121    def test_output_pretrained(self):122        model = AutoencoderKL.from_pretrained("fusing/autoencoder-kl-dummy")123        model = model.to(torch_device)124        model.eval()125 126        if torch_device == "mps":127            generator = torch.manual_seed(0)128        else:129            generator = torch.Generator(device=torch_device).manual_seed(0)130 131        image = torch.randn(132            1,133            model.config.in_channels,134            model.config.sample_size,135            model.config.sample_size,136            generator=torch.manual_seed(0),137        )138        image = image.to(torch_device)139        with torch.no_grad():140            output = model(image, sample_posterior=True, generator=generator).sample141 142        output_slice = output[0, -1, -3:, -3:].flatten().cpu()143 144        # Since the VAE Gaussian prior's generator is seeded on the appropriate device,145        # the expected output slices are not the same for CPU and GPU.146        if torch_device == "mps":147            expected_output_slice = torch.tensor(148                [149                    -4.0078e-01,150                    -3.8323e-04,151                    -1.2681e-01,152                    -1.1462e-01,153                    2.0095e-01,154                    1.0893e-01,155                    -8.8247e-02,156                    -3.0361e-01,157                    -9.8644e-03,158                ]159            )160        elif torch_device == "cpu":161            expected_output_slice = torch.tensor(162                [-0.1352, 0.0878, 0.0419, -0.0818, -0.1069, 0.0688, -0.1458, -0.4446, -0.0026]163            )164        else:165            expected_output_slice = torch.tensor(166                [-0.2421, 0.4642, 0.2507, -0.0438, 0.0682, 0.3160, -0.2018, -0.0727, 0.2485]167            )168 169        self.assertTrue(torch_all_close(output_slice, expected_output_slice, rtol=1e-2))170 171 172@slow173class AutoencoderKLIntegrationTests(unittest.TestCase):174    def get_file_format(self, seed, shape):175        return f"gaussian_noise_s={seed}_shape={'_'.join([str(s) for s in shape])}.npy"176 177    def tearDown(self):178        # clean up the VRAM after each test179        super().tearDown()180        gc.collect()181        torch.cuda.empty_cache()182 183    def get_sd_image(self, seed=0, shape=(4, 3, 512, 512), fp16=False):184        dtype = torch.float16 if fp16 else torch.float32185        image = torch.from_numpy(load_hf_numpy(self.get_file_format(seed, shape))).to(torch_device).to(dtype)186        return image187 188    def get_sd_vae_model(self, model_id="CompVis/stable-diffusion-v1-4", fp16=False):189        revision = "fp16" if fp16 else None190        torch_dtype = torch.float16 if fp16 else torch.float32191 192        model = AutoencoderKL.from_pretrained(193            model_id,194            subfolder="vae",195            torch_dtype=torch_dtype,196            revision=revision,197        )198        model.to(torch_device).eval()199 200        return model201 202    def get_generator(self, seed=0):203        if torch_device == "mps":204            return torch.manual_seed(seed)205        return torch.Generator(device=torch_device).manual_seed(seed)206 207    @parameterized.expand(208        [209            # fmt: off210            [33, [-0.1603, 0.9878, -0.0495, -0.0790, -0.2709, 0.8375, -0.2060, -0.0824], [-0.2395, 0.0098, 0.0102, -0.0709, -0.2840, -0.0274, -0.0718, -0.1824]],211            [47, [-0.2376, 0.1168, 0.1332, -0.4840, -0.2508, -0.0791, -0.0493, -0.4089], [0.0350, 0.0847, 0.0467, 0.0344, -0.0842, -0.0547, -0.0633, -0.1131]],212            # fmt: on213        ]214    )215    def test_stable_diffusion(self, seed, expected_slice, expected_slice_mps):216        model = self.get_sd_vae_model()217        image = self.get_sd_image(seed)218        generator = self.get_generator(seed)219 220        with torch.no_grad():221            sample = model(image, generator=generator, sample_posterior=True).sample222 223        assert sample.shape == image.shape224 225        output_slice = sample[-1, -2:, -2:, :2].flatten().float().cpu()226        expected_output_slice = torch.tensor(expected_slice_mps if torch_device == "mps" else expected_slice)227 228        assert torch_all_close(output_slice, expected_output_slice, atol=1e-3)229 230    @parameterized.expand(231        [232            # fmt: off233            [33, [-0.0513, 0.0289, 1.3799, 0.2166, -0.2573, -0.0871, 0.5103, -0.0999]],234            [47, [-0.4128, -0.1320, -0.3704, 0.1965, -0.4116, -0.2332, -0.3340, 0.2247]],235            # fmt: on236        ]237    )238    @require_torch_gpu239    def test_stable_diffusion_fp16(self, seed, expected_slice):240        model = self.get_sd_vae_model(fp16=True)241        image = self.get_sd_image(seed, fp16=True)242        generator = self.get_generator(seed)243 244        with torch.no_grad():245            sample = model(image, generator=generator, sample_posterior=True).sample246 247        assert sample.shape == image.shape248 249        output_slice = sample[-1, -2:, :2, -2:].flatten().float().cpu()250        expected_output_slice = torch.tensor(expected_slice)251 252        assert torch_all_close(output_slice, expected_output_slice, atol=1e-2)253 254    @parameterized.expand(255        [256            # fmt: off257            [33, [-0.1609, 0.9866, -0.0487, -0.0777, -0.2716, 0.8368, -0.2055, -0.0814], [-0.2395, 0.0098, 0.0102, -0.0709, -0.2840, -0.0274, -0.0718, -0.1824]],258            [47, [-0.2377, 0.1147, 0.1333, -0.4841, -0.2506, -0.0805, -0.0491, -0.4085], [0.0350, 0.0847, 0.0467, 0.0344, -0.0842, -0.0547, -0.0633, -0.1131]],259            # fmt: on260        ]261    )262    def test_stable_diffusion_mode(self, seed, expected_slice, expected_slice_mps):263        model = self.get_sd_vae_model()264        image = self.get_sd_image(seed)265 266        with torch.no_grad():267            sample = model(image).sample268 269        assert sample.shape == image.shape270 271        output_slice = sample[-1, -2:, -2:, :2].flatten().float().cpu()272        expected_output_slice = torch.tensor(expected_slice_mps if torch_device == "mps" else expected_slice)273 274        assert torch_all_close(output_slice, expected_output_slice, atol=1e-3)275 276    @parameterized.expand(277        [278            # fmt: off279            [13, [-0.2051, -0.1803, -0.2311, -0.2114, -0.3292, -0.3574, -0.2953, -0.3323]],280            [37, [-0.2632, -0.2625, -0.2199, -0.2741, -0.4539, -0.4990, -0.3720, -0.4925]],281            # fmt: on282        ]283    )284    @require_torch_gpu285    def test_stable_diffusion_decode(self, seed, expected_slice):286        model = self.get_sd_vae_model()287        encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64))288 289        with torch.no_grad():290            sample = model.decode(encoding).sample291 292        assert list(sample.shape) == [3, 3, 512, 512]293 294        output_slice = sample[-1, -2:, :2, -2:].flatten().cpu()295        expected_output_slice = torch.tensor(expected_slice)296 297        assert torch_all_close(output_slice, expected_output_slice, atol=1e-3)298 299    @parameterized.expand(300        [301            # fmt: off302            [27, [-0.0369, 0.0207, -0.0776, -0.0682, -0.1747, -0.1930, -0.1465, -0.2039]],303            [16, [-0.1628, -0.2134, -0.2747, -0.2642, -0.3774, -0.4404, -0.3687, -0.4277]],304            # fmt: on305        ]306    )307    @require_torch_gpu308    def test_stable_diffusion_decode_fp16(self, seed, expected_slice):309        model = self.get_sd_vae_model(fp16=True)310        encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64), fp16=True)311 312        with torch.no_grad():313            sample = model.decode(encoding).sample314 315        assert list(sample.shape) == [3, 3, 512, 512]316 317        output_slice = sample[-1, -2:, :2, -2:].flatten().float().cpu()318        expected_output_slice = torch.tensor(expected_slice)319 320        assert torch_all_close(output_slice, expected_output_slice, atol=5e-3)321 322    @parameterized.expand(323        [324            # fmt: off325            [33, [-0.3001, 0.0918, -2.6984, -3.9720, -3.2099, -5.0353, 1.7338, -0.2065, 3.4267]],326            [47, [-1.5030, -4.3871, -6.0355, -9.1157, -1.6661, -2.7853, 2.1607, -5.0823, 2.5633]],327            # fmt: on328        ]329    )330    def test_stable_diffusion_encode_sample(self, seed, expected_slice):331        model = self.get_sd_vae_model()332        image = self.get_sd_image(seed)333        generator = self.get_generator(seed)334 335        with torch.no_grad():336            dist = model.encode(image).latent_dist337            sample = dist.sample(generator=generator)338 339        assert list(sample.shape) == [image.shape[0], 4] + [i // 8 for i in image.shape[2:]]340 341        output_slice = sample[0, -1, -3:, -3:].flatten().cpu()342        expected_output_slice = torch.tensor(expected_slice)343 344        tolerance = 1e-3 if torch_device != "mps" else 1e-2345        assert torch_all_close(output_slice, expected_output_slice, atol=tolerance)346