File size: 4,058 Bytes
3a25a0a |
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 |
# coding=utf-8
# Copyright 2024 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 diffusers import (
AutoencoderKL,
)
from diffusers.utils.testing_utils import (
enable_full_determinism,
load_hf_numpy,
numpy_cosine_similarity_distance,
require_torch_gpu,
slow,
torch_device,
)
enable_full_determinism()
@slow
@require_torch_gpu
class AutoencoderKLSingleFileTests(unittest.TestCase):
model_class = AutoencoderKL
ckpt_path = (
"https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors"
)
repo_id = "stabilityai/sd-vae-ft-mse"
main_input_name = "sample"
base_precision = 1e-2
def setUp(self):
super().setUp()
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
def get_file_format(self, seed, shape):
return f"gaussian_noise_s={seed}_shape={'_'.join([str(s) for s in shape])}.npy"
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 test_single_file_inference_same_as_pretrained(self):
model_1 = self.model_class.from_pretrained(self.repo_id).to(torch_device)
model_2 = self.model_class.from_single_file(self.ckpt_path, config=self.repo_id).to(torch_device)
image = self.get_sd_image(33)
generator = torch.Generator(torch_device)
with torch.no_grad():
sample_1 = model_1(image, generator=generator.manual_seed(0)).sample
sample_2 = model_2(image, generator=generator.manual_seed(0)).sample
assert sample_1.shape == sample_2.shape
output_slice_1 = sample_1.flatten().float().cpu()
output_slice_2 = sample_2.flatten().float().cpu()
assert numpy_cosine_similarity_distance(output_slice_1, output_slice_2) < 1e-4
def test_single_file_components(self):
model = self.model_class.from_pretrained(self.repo_id)
model_single_file = self.model_class.from_single_file(self.ckpt_path, config=self.repo_id)
PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
for param_name, param_value in model_single_file.config.items():
if param_name in PARAMS_TO_IGNORE:
continue
assert (
model.config[param_name] == param_value
), f"{param_name} differs between pretrained loading and single file loading"
def test_single_file_arguments(self):
model_default = self.model_class.from_single_file(self.ckpt_path, config=self.repo_id)
assert model_default.config.scaling_factor == 0.18215
assert model_default.config.sample_size == 256
assert model_default.dtype == torch.float32
scaling_factor = 2.0
sample_size = 512
torch_dtype = torch.float16
model = self.model_class.from_single_file(
self.ckpt_path,
config=self.repo_id,
sample_size=sample_size,
scaling_factor=scaling_factor,
torch_dtype=torch_dtype,
)
assert model.config.scaling_factor == scaling_factor
assert model.config.sample_size == sample_size
assert model.dtype == torch_dtype
|