Spaces:
Running
on
Zero
Running
on
Zero
# 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 unittest | |
import torch | |
from parameterized import parameterized | |
from diffusers import AutoencoderKL | |
from diffusers.utils import floats_tensor, load_hf_numpy, require_torch_gpu, slow, torch_all_close, torch_device | |
from ..test_modeling_common import ModelTesterMixin | |
torch.backends.cuda.matmul.allow_tf32 = False | |
class AutoencoderKLTests(ModelTesterMixin, unittest.TestCase): | |
model_class = AutoencoderKL | |
def dummy_input(self): | |
batch_size = 4 | |
num_channels = 3 | |
sizes = (32, 32) | |
image = floats_tensor((batch_size, num_channels) + sizes).to(torch_device) | |
return {"sample": image} | |
def input_shape(self): | |
return (3, 32, 32) | |
def output_shape(self): | |
return (3, 32, 32) | |
def prepare_init_args_and_inputs_for_common(self): | |
init_dict = { | |
"block_out_channels": [32, 64], | |
"in_channels": 3, | |
"out_channels": 3, | |
"down_block_types": ["DownEncoderBlock2D", "DownEncoderBlock2D"], | |
"up_block_types": ["UpDecoderBlock2D", "UpDecoderBlock2D"], | |
"latent_channels": 4, | |
} | |
inputs_dict = self.dummy_input | |
return init_dict, inputs_dict | |
def test_forward_signature(self): | |
pass | |
def test_training(self): | |
pass | |
def test_gradient_checkpointing(self): | |
# enable deterministic behavior for gradient checkpointing | |
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common() | |
model = self.model_class(**init_dict) | |
model.to(torch_device) | |
assert not model.is_gradient_checkpointing and model.training | |
out = model(**inputs_dict).sample | |
# run the backwards pass on the model. For backwards pass, for simplicity purpose, | |
# we won't calculate the loss and rather backprop on out.sum() | |
model.zero_grad() | |
labels = torch.randn_like(out) | |
loss = (out - labels).mean() | |
loss.backward() | |
# re-instantiate the model now enabling gradient checkpointing | |
model_2 = self.model_class(**init_dict) | |
# clone model | |
model_2.load_state_dict(model.state_dict()) | |
model_2.to(torch_device) | |
model_2.enable_gradient_checkpointing() | |
assert model_2.is_gradient_checkpointing and model_2.training | |
out_2 = model_2(**inputs_dict).sample | |
# run the backwards pass on the model. For backwards pass, for simplicity purpose, | |
# we won't calculate the loss and rather backprop on out.sum() | |
model_2.zero_grad() | |
loss_2 = (out_2 - labels).mean() | |
loss_2.backward() | |
# compare the output and parameters gradients | |
self.assertTrue((loss - loss_2).abs() < 1e-5) | |
named_params = dict(model.named_parameters()) | |
named_params_2 = dict(model_2.named_parameters()) | |
for name, param in named_params.items(): | |
self.assertTrue(torch_all_close(param.grad.data, named_params_2[name].grad.data, atol=5e-5)) | |
def test_from_pretrained_hub(self): | |
model, loading_info = AutoencoderKL.from_pretrained("fusing/autoencoder-kl-dummy", output_loading_info=True) | |
self.assertIsNotNone(model) | |
self.assertEqual(len(loading_info["missing_keys"]), 0) | |
model.to(torch_device) | |
image = model(**self.dummy_input) | |
assert image is not None, "Make sure output is not None" | |
def test_output_pretrained(self): | |
model = AutoencoderKL.from_pretrained("fusing/autoencoder-kl-dummy") | |
model = model.to(torch_device) | |
model.eval() | |
if torch_device == "mps": | |
generator = torch.manual_seed(0) | |
else: | |
generator = torch.Generator(device=torch_device).manual_seed(0) | |
image = torch.randn( | |
1, | |
model.config.in_channels, | |
model.config.sample_size, | |
model.config.sample_size, | |
generator=torch.manual_seed(0), | |
) | |
image = image.to(torch_device) | |
with torch.no_grad(): | |
output = model(image, sample_posterior=True, generator=generator).sample | |
output_slice = output[0, -1, -3:, -3:].flatten().cpu() | |
# Since the VAE Gaussian prior's generator is seeded on the appropriate device, | |
# the expected output slices are not the same for CPU and GPU. | |
if torch_device == "mps": | |
expected_output_slice = torch.tensor( | |
[ | |
-4.0078e-01, | |
-3.8323e-04, | |
-1.2681e-01, | |
-1.1462e-01, | |
2.0095e-01, | |
1.0893e-01, | |
-8.8247e-02, | |
-3.0361e-01, | |
-9.8644e-03, | |
] | |
) | |
elif torch_device == "cpu": | |
expected_output_slice = torch.tensor( | |
[-0.1352, 0.0878, 0.0419, -0.0818, -0.1069, 0.0688, -0.1458, -0.4446, -0.0026] | |
) | |
else: | |
expected_output_slice = torch.tensor( | |
[-0.2421, 0.4642, 0.2507, -0.0438, 0.0682, 0.3160, -0.2018, -0.0727, 0.2485] | |
) | |
self.assertTrue(torch_all_close(output_slice, expected_output_slice, rtol=1e-2)) | |
class AutoencoderKLIntegrationTests(unittest.TestCase): | |
def get_file_format(self, seed, shape): | |
return f"gaussian_noise_s={seed}_shape={'_'.join([str(s) for s in shape])}.npy" | |
def tearDown(self): | |
# clean up the VRAM after each test | |
super().tearDown() | |
gc.collect() | |
torch.cuda.empty_cache() | |
def get_sd_image(self, seed=0, shape=(4, 3, 512, 512), fp16=False): | |
dtype = torch.float16 if fp16 else torch.float32 | |
image = torch.from_numpy(load_hf_numpy(self.get_file_format(seed, shape))).to(torch_device).to(dtype) | |
return image | |
def get_sd_vae_model(self, model_id="CompVis/stable-diffusion-v1-4", fp16=False): | |
revision = "fp16" if fp16 else None | |
torch_dtype = torch.float16 if fp16 else torch.float32 | |
model = AutoencoderKL.from_pretrained( | |
model_id, | |
subfolder="vae", | |
torch_dtype=torch_dtype, | |
revision=revision, | |
) | |
model.to(torch_device).eval() | |
return model | |
def get_generator(self, seed=0): | |
if torch_device == "mps": | |
return torch.manual_seed(seed) | |
return torch.Generator(device=torch_device).manual_seed(seed) | |
def test_stable_diffusion(self, seed, expected_slice, expected_slice_mps): | |
model = self.get_sd_vae_model() | |
image = self.get_sd_image(seed) | |
generator = self.get_generator(seed) | |
with torch.no_grad(): | |
sample = model(image, generator=generator, sample_posterior=True).sample | |
assert sample.shape == image.shape | |
output_slice = sample[-1, -2:, -2:, :2].flatten().float().cpu() | |
expected_output_slice = torch.tensor(expected_slice_mps if torch_device == "mps" else expected_slice) | |
assert torch_all_close(output_slice, expected_output_slice, atol=1e-3) | |
def test_stable_diffusion_fp16(self, seed, expected_slice): | |
model = self.get_sd_vae_model(fp16=True) | |
image = self.get_sd_image(seed, fp16=True) | |
generator = self.get_generator(seed) | |
with torch.no_grad(): | |
sample = model(image, generator=generator, sample_posterior=True).sample | |
assert sample.shape == image.shape | |
output_slice = sample[-1, -2:, :2, -2:].flatten().float().cpu() | |
expected_output_slice = torch.tensor(expected_slice) | |
assert torch_all_close(output_slice, expected_output_slice, atol=1e-2) | |
def test_stable_diffusion_mode(self, seed, expected_slice, expected_slice_mps): | |
model = self.get_sd_vae_model() | |
image = self.get_sd_image(seed) | |
with torch.no_grad(): | |
sample = model(image).sample | |
assert sample.shape == image.shape | |
output_slice = sample[-1, -2:, -2:, :2].flatten().float().cpu() | |
expected_output_slice = torch.tensor(expected_slice_mps if torch_device == "mps" else expected_slice) | |
assert torch_all_close(output_slice, expected_output_slice, atol=1e-3) | |
def test_stable_diffusion_decode(self, seed, expected_slice): | |
model = self.get_sd_vae_model() | |
encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64)) | |
with torch.no_grad(): | |
sample = model.decode(encoding).sample | |
assert list(sample.shape) == [3, 3, 512, 512] | |
output_slice = sample[-1, -2:, :2, -2:].flatten().cpu() | |
expected_output_slice = torch.tensor(expected_slice) | |
assert torch_all_close(output_slice, expected_output_slice, atol=1e-3) | |
def test_stable_diffusion_decode_fp16(self, seed, expected_slice): | |
model = self.get_sd_vae_model(fp16=True) | |
encoding = self.get_sd_image(seed, shape=(3, 4, 64, 64), fp16=True) | |
with torch.no_grad(): | |
sample = model.decode(encoding).sample | |
assert list(sample.shape) == [3, 3, 512, 512] | |
output_slice = sample[-1, -2:, :2, -2:].flatten().float().cpu() | |
expected_output_slice = torch.tensor(expected_slice) | |
assert torch_all_close(output_slice, expected_output_slice, atol=5e-3) | |
def test_stable_diffusion_encode_sample(self, seed, expected_slice): | |
model = self.get_sd_vae_model() | |
image = self.get_sd_image(seed) | |
generator = self.get_generator(seed) | |
with torch.no_grad(): | |
dist = model.encode(image).latent_dist | |
sample = dist.sample(generator=generator) | |
assert list(sample.shape) == [image.shape[0], 4] + [i // 8 for i in image.shape[2:]] | |
output_slice = sample[0, -1, -3:, -3:].flatten().cpu() | |
expected_output_slice = torch.tensor(expected_slice) | |
tolerance = 1e-3 if torch_device != "mps" else 1e-2 | |
assert torch_all_close(output_slice, expected_output_slice, atol=tolerance) | |