Post-Training Quantization and Weights Compression of DeepFloyd IF model with NNCFยถ

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS.

Github

The goal of this tutorial is to demonstrate how to speed up the model by applying 8-bit post-training quantization and weights compression from NNCF (Neural Network Compression Framework) and infer optimized model via OpenVINOโ„ข Toolkit.

NOTE: you should run 238-deep-floyd-if-convert notebook first to generate OpenVINO IR model that is used for optimization.

The optimization process contains the following steps: 1. Compress weights of the converted OpenVINO text encoder from notebook with NNCF. 2. Quantize the converted stage_1 and stage_2 U-Nets from notebook with NNCF. 2. Check the model result using the same input data from the notebook. 3. Compare model size of converted and optimized models. 4. Compare performance of converted and optimized models.

Table of contents:

Prerequisitesยถ

%pip install -q datasets "nncf>=2.6.0"
import nncf
import torch
import openvino as ov

from diffusers import DiffusionPipeline
from diffusers.utils.pil_utils import pt_to_pil
from pathlib import Path
from typing import Any, List

from utils import TextEncoder, UnetFirstStage, UnetSecondStage

checkpoint_variant = 'fp16'
model_dtype = torch.float32
RANDOM_SEED = 42
N_DIFFUSION_STEPS = 50
UNET_2_STEPS = 20

core = ov.Core()
INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino
MODEL_DIR = Path('./models')
TEXT_ENCODER_IR_PATH = MODEL_DIR / "encoder_ir.xml"
UNET_I_IR_PATH = MODEL_DIR / "unet_ir_I.xml"
UNET_II_IR_PATH = MODEL_DIR / "unet_ir_II.xml"

if not (TEXT_ENCODER_IR_PATH.exists() and UNET_I_IR_PATH.exists() and UNET_II_IR_PATH.exists()):
    raise RuntimeError('This notebook should be run after 238-deep-floyd-if notebook')
import ipywidgets as widgets

device = widgets.Dropdown(
    options=core.available_devices + ["AUTO"],
    value='AUTO',
    description='Device:',
    disabled=False,
)

device
Dropdown(description='Device:', index=2, options=('CPU', 'GPU', 'AUTO'), value='AUTO')

Compress weightsยถ

Text encoder model consumes ~22 GB of disk space. To avoid running out of memory, we suggest using 8-bit weights compression instead of quantization. An optimized model will show less speed up than a quantized model, but this will significantly reduce the model footprint.

%%time

text_encoder = core.read_model(TEXT_ENCODER_IR_PATH)
text_encoder_optimized = nncf.compress_weights(text_encoder)

TEXT_ENCODER_INT8_IR_PATH = Path("_optimized.".join(TEXT_ENCODER_IR_PATH.as_posix().split(".")))
ov.save_model(text_encoder_optimized, TEXT_ENCODER_INT8_IR_PATH)
2023-10-30 08:36:34.384792: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable TF_ENABLE_ONEDNN_OPTS=0.
2023-10-30 08:36:34.423283: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-10-30 08:36:35.184200: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
CPU times: user 3min 16s, sys: 58 s, total: 4min 14s
Wall time: 4min 12s

Quantizeยถ

Prepare datasetยถ

DeepFloyd IF consists of a U-Net model for first and second stages. First stage U-Net generates 64x64 px image based on text prompt, second stage U-Net generates a 256x256 px image based on image from previous step. We use a portion of train LAION2B dataset from Hugging Face as calibration data. LAION2B is the English subset of the LAION5B dataset, contains over 2 billion objects.

import numpy as np
from datasets import load_dataset

np.random.seed(RANDOM_SEED)

def get_negative_prompt():
    negative_prompts = [
        "amateur", "blurred", "deformed", "disfigured", "disgusting", "jpeg artifacts", "low contrast",
        "low quality", "low saturation", "mangled", "morbid", "mutilated", "mutation",
        "out of frame", "out of frame", "ugly", "uncentered", "underexposed", "unreal",
    ]
    num_elements = np.random.randint(2, 6)
    random_elements = np.random.choice(negative_prompts, num_elements)
    return [" ".join(random_elements)]

def prepare_calibration_data(dataloader, stage_1):
    """
    This function prepares calibration data from a dataloader for a specified number of initialization steps.
    It iterates over the dataloader, fetching batches and storing the relevant data.
    """
    data = []
    for batch in dataloader:
        prompt = batch["TEXT"]
        negative_prompt = get_negative_prompt()
        prompt_embeds, negative_embeds = stage_1.encode_prompt(prompt, negative_prompt=negative_prompt)
        data.append((prompt_embeds, negative_embeds))
    return data


def prepare_dataset(stage_1, opt_init_steps=300):
    """
    Prepares a text dataset for quantization.
    """
    dataset = load_dataset("laion/laion2B-en", streaming=True, split="train")
    train_dataset = dataset.shuffle(seed=RANDOM_SEED, buffer_size=1000).take(opt_init_steps)
    dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=1)
    calibration_data = prepare_calibration_data(dataloader, stage_1)
    return calibration_data
%%time

generator = torch.manual_seed(RANDOM_SEED)
opt_init_steps = 300
selection_prob = 0.5
prompts_number = np.ceil(opt_init_steps // (min(N_DIFFUSION_STEPS, UNET_2_STEPS) * selection_prob))

stage_1 = DiffusionPipeline.from_pretrained(
    "DeepFloyd/IF-I-M-v1.0",
    variant=checkpoint_variant,
    torch_dtype=model_dtype
)
encoded_prompts = prepare_dataset(stage_1, int(prompts_number))
safety_checker/model.safetensors not found

A mixture of fp16 and non-fp16 filenames will be loaded.
Loaded fp16 filenames:
[text_encoder/pytorch_model.fp16-00001-of-00002.bin, unet/diffusion_pytorch_model.fp16.bin, text_encoder/pytorch_model.fp16-00002-of-00002.bin]
Loaded non-fp16 filenames:
[safety_checker/pytorch_model.bin, watermarker/diffusion_pytorch_model.bin
If this behavior is not expected, please check your folder structure.
Cannot initialize model with low cpu memory usage because accelerate was not found in the environment. Defaulting to low_cpu_mem_usage=False. It is strongly recommended to install accelerate for faster and less memory-intense model loading. You can do so with:
`
pip install accelerate
`
.
Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]
You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the legacy (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set legacy=False. This should only be set if you understand what it means, and thouroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
/home/ea/work/ov_venv/lib/python3.8/site-packages/torch/cuda/__init__.py:138: UserWarning: CUDA initialization: The NVIDIA driver on your system is too old (found version 11080). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver. (Triggered internally at ../c10/cuda/CUDAFunctions.cpp:108.)
  return torch._C._cuda_getDeviceCount() > 0
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
Downloading readme:   0%|          | 0.00/56.0 [00:00<?, ?B/s]
Resolving data files:   0%|          | 0/128 [00:00<?, ?it/s]
CPU times: user 18min 16s, sys: 1min 2s, total: 19min 18s
Wall time: 2min 5s

To collect intermediate model inputs for calibration we should customize CompiledModel.

class CompiledModelDecorator(ov.CompiledModel):
    def __init__(self, compiled_model, prob: float, data_cache: List[Any] = []):
        super().__init__(compiled_model)
        self.data_cache = data_cache
        self.prob = np.clip(prob, 0, 1)

    def __call__(self, *args, **kwargs):
        if np.random.rand() >= self.prob:
            self.data_cache.append(*args)
        return super().__call__(*args, **kwargs)
stage_1.unet = UnetFirstStage(
    UNET_I_IR_PATH,
    stage_1.unet.config,
    dtype=model_dtype,
    device=device.value
)
stage_1.set_progress_bar_config(disable=True)

stage_1_data_cache = []
stage_1.unet.unet_openvino = CompiledModelDecorator(stage_1.unet.unet_openvino, prob=selection_prob, data_cache=stage_1_data_cache)

generator = torch.manual_seed(RANDOM_SEED)
stage_2_inputs = []  # to speed up dataset preparation for stage 2 U-Net we can collect several images below
for data in encoded_prompts:
    prompt_embeds, negative_embeds = data
    image = stage_1(prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds,
                    generator=generator, output_type="pt", num_inference_steps=N_DIFFUSION_STEPS).images
    stage_2_inputs.append((image, prompt_embeds, negative_embeds))

    if len(stage_1_data_cache) >= opt_init_steps:
        break

Quantize first stage U-Netยถ

%%time

ov_model = core.read_model(UNET_I_IR_PATH)
stage_1_calibration_dataset = nncf.Dataset(stage_1_data_cache, lambda x: x)

quantized_model = nncf.quantize(
    model=ov_model,
    calibration_dataset=stage_1_calibration_dataset,
    model_type=nncf.ModelType.TRANSFORMER,
    advanced_parameters=nncf.AdvancedQuantizationParameters(smooth_quant_alpha=0.25)
)

UNET_I_INT8_PATH = "_optimized.".join(UNET_I_IR_PATH.as_posix().split("."))
ov.save_model(quantized_model, UNET_I_INT8_PATH)
Statistics collection: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 300/300 [01:35<00:00,  3.14it/s]
Applying Smooth Quant: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 73/73 [00:04<00:00, 17.55it/s]
Statistics collection: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 300/300 [05:44<00:00,  1.15s/it]
Applying Fast Bias correction: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 268/268 [00:35<00:00,  7.50it/s]
CPU times: user 1h 8min 46s, sys: 1min 22s, total: 1h 10min 8s
Wall time: 9min 46s
%%time

from tqdm.notebook import tqdm

start = len(stage_2_inputs)
for i, data in tqdm(enumerate(encoded_prompts[start:])):
    prompt_embeds, negative_embeds = data
    image = stage_1(prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds,
                    generator=generator, output_type="pt", num_inference_steps=N_DIFFUSION_STEPS).images
    stage_2_inputs.append((image, prompt_embeds, negative_embeds))
0it [00:00, ?it/s]
CPU times: user 1h 17min 46s, sys: 44.9 s, total: 1h 18min 31s
Wall time: 4min 46s
%%time

generator = torch.manual_seed(RANDOM_SEED)
opt_init_steps = 300

stage_2 = DiffusionPipeline.from_pretrained(
    "DeepFloyd/IF-II-M-v1.0",
    text_encoder=None,
    variant=checkpoint_variant,
    torch_dtype=model_dtype
)
stage_2.set_progress_bar_config(disable=True)

stage_2.unet = UnetSecondStage(
    UNET_II_IR_PATH,
    stage_2.unet.config,
    dtype=model_dtype,
    device=device.value
)
stage_2_data_cache = []
stage_2.unet.unet_openvino = CompiledModelDecorator(stage_2.unet.unet_openvino, prob=selection_prob, data_cache=stage_2_data_cache)

for data in tqdm(stage_2_inputs):
    image, prompt_embeds, negative_embeds = data
    image = stage_2(image=image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds,
                    generator=generator, output_type="pt", num_inference_steps=UNET_2_STEPS).images

    if len(stage_2_data_cache) >= opt_init_steps:
        break
A mixture of fp16 and non-fp16 filenames will be loaded.
Loaded fp16 filenames:
[text_encoder/model.fp16-00001-of-00002.safetensors, unet/diffusion_pytorch_model.fp16.safetensors, text_encoder/model.fp16-00002-of-00002.safetensors, safety_checker/model.fp16.safetensors]
Loaded non-fp16 filenames:
[watermarker/diffusion_pytorch_model.safetensors
If this behavior is not expected, please check your folder structure.
Cannot initialize model with low cpu memory usage because accelerate was not found in the environment. Defaulting to low_cpu_mem_usage=False. It is strongly recommended to install accelerate for faster and less memory-intense model loading. You can do so with:
`
pip install accelerate
`
.
Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]
CPU times: user 6h 28min 3s, sys: 2min 11s, total: 6h 30min 15s
Wall time: 24min 32s

Quantize second stage U-Netยถ

%%time

ov_model = core.read_model(UNET_II_IR_PATH)

calibration_dataset = nncf.Dataset(stage_2_data_cache, lambda x: x)
quantized_model = nncf.quantize(
    model=ov_model,
    calibration_dataset=calibration_dataset,
    model_type=nncf.ModelType.TRANSFORMER,
)

UNET_II_INT8_PATH = "_optimized.".join(UNET_II_IR_PATH.as_posix().split("."))
ov.save_model(quantized_model, UNET_II_INT8_PATH)
Statistics collection: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 300/300 [12:02<00:00,  2.41s/it]
Applying Smooth Quant: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 54/54 [00:03<00:00, 15.80it/s]
Statistics collection: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 300/300 [34:51<00:00,  6.97s/it]
Applying Fast Bias correction: 100%|โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ| 245/245 [00:39<00:00,  6.17it/s]
CPU times: user 7h 57min 5s, sys: 6min 43s, total: 8h 3min 49s
Wall time: 49min 24s

Run optimized OpenVINO modelยถ

Let us check predictions with the optimized OpenVINO DeepFloyd IF model result using the same input data from the 1st notebook.

prompt = 'ultra close color photo portrait of rainbow owl with deer horns in the woods'
negative_prompt = 'blurred unreal uncentered occluded'
%%time

stage_1 = DiffusionPipeline.from_pretrained(
    "DeepFloyd/IF-I-M-v1.0",
    variant=checkpoint_variant,
    torch_dtype=model_dtype
)

# Initialize the First Stage U-Net wrapper class
stage_1.unet = UnetFirstStage(
    UNET_I_INT8_PATH,
    stage_1.unet.config,
    dtype=model_dtype,
    device=device.value
)

stage_1.text_encoder = TextEncoder(TEXT_ENCODER_INT8_IR_PATH, dtype=model_dtype, device=device.value)
print('The model has been loaded')

# Generate text embeddings
prompt_embeds, negative_embeds = stage_1.encode_prompt(prompt, negative_prompt=negative_prompt)

# Fix PRNG seed
generator = torch.manual_seed(RANDOM_SEED)

# Inference
image = stage_1(prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds,
                generator=generator, output_type="pt", num_inference_steps=N_DIFFUSION_STEPS).images

# Show the image
pt_to_pil(image)[0]
safety_checker/model.safetensors not found

A mixture of fp16 and non-fp16 filenames will be loaded.
Loaded fp16 filenames:
[text_encoder/pytorch_model.fp16-00001-of-00002.bin, unet/diffusion_pytorch_model.fp16.bin, text_encoder/pytorch_model.fp16-00002-of-00002.bin]
Loaded non-fp16 filenames:
[safety_checker/pytorch_model.bin, watermarker/diffusion_pytorch_model.bin
If this behavior is not expected, please check your folder structure.
Cannot initialize model with low cpu memory usage because accelerate was not found in the environment. Defaulting to low_cpu_mem_usage=False. It is strongly recommended to install accelerate for faster and less memory-intense model loading. You can do so with:
`
pip install accelerate
`
.
Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
The model has been loaded
0%|          | 0/50 [00:00<?, ?it/s]
CPU times: user 3min 39s, sys: 21 s, total: 4min
Wall time: 58.7 s
../_images/238-deep-floyd-if-optimize-with-output_23_6.png
%%time

stage_2 = DiffusionPipeline.from_pretrained(
    "DeepFloyd/IF-II-M-v1.0",
    text_encoder=None,
    variant=checkpoint_variant,
    torch_dtype=model_dtype
)

# Initialize the Second Stage U-Net wrapper class
stage_2.unet = UnetSecondStage(
    UNET_II_INT8_PATH,
    stage_2.unet.config,
    dtype=model_dtype,
    device=device.value
)
print('The model has been loaded')

image = stage_2(
    image=image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds,
    generator=generator, output_type="pt", num_inference_steps=UNET_2_STEPS).images

# Show the image
pil_image = pt_to_pil(image)[0]
pil_image
A mixture of fp16 and non-fp16 filenames will be loaded.
Loaded fp16 filenames:
[text_encoder/model.fp16-00001-of-00002.safetensors, unet/diffusion_pytorch_model.fp16.safetensors, text_encoder/model.fp16-00002-of-00002.safetensors, safety_checker/model.fp16.safetensors]
Loaded non-fp16 filenames:
[watermarker/diffusion_pytorch_model.safetensors
If this behavior is not expected, please check your folder structure.
Cannot initialize model with low cpu memory usage because accelerate was not found in the environment. Defaulting to low_cpu_mem_usage=False. It is strongly recommended to install accelerate for faster and less memory-intense model loading. You can do so with:
`
pip install accelerate
`
.
Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]
The model has been loaded
0%|          | 0/20 [00:00<?, ?it/s]
CPU times: user 6min 20s, sys: 6.78 s, total: 6min 27s
Wall time: 32.1 s
../_images/238-deep-floyd-if-optimize-with-output_24_5.png
import cv2
import numpy as np
from utils import convert_result_to_image, download_omz_model

# 1032: 4x superresolution, 1033: 3x superresolution
model_name = 'single-image-super-resolution-1032'
download_omz_model(model_name, MODEL_DIR)

sr_model_xml_path = MODEL_DIR / f'{model_name}.xml'
model = core.read_model(model=sr_model_xml_path)
model.reshape({
    0: [1, 3, 256, 256],
    1: [1, 3, 1024, 1024]
})
compiled_sr_model = core.compile_model(model=model, device_name=device.value)

original_image = np.array(pil_image)
bicubic_image = cv2.resize(
    src=original_image, dsize=(1024, 1024), interpolation=cv2.INTER_CUBIC
)

# Reshape the images from (H,W,C) to (N,C,H,W) as expected by the model.
input_image_original = np.expand_dims(original_image.transpose(2, 0, 1), axis=0)
input_image_bicubic = np.expand_dims(bicubic_image.transpose(2, 0, 1), axis=0)

# Model Inference
result = compiled_sr_model(
    [input_image_original, input_image_bicubic]
)[compiled_sr_model.output(0)]

img = convert_result_to_image(result)
img
single-image-super-resolution-1032 already downloaded to models
../_images/238-deep-floyd-if-optimize-with-output_25_1.png

Note

Accuracy of quantized models can generally be improved by increasing calibration dataset size. For U-Net models, you can collect a more diverse dataset by using a smaller selection_prob value, but this will increase the dataset collection time.

Letโ€™s calculate the compression rate of the optimized IRs file size relative to the FP16 OpenVINO models file size

def calculate_compression_rate(ov_model_path):
    fp16_ir_model_size = Path(ov_model_path).with_suffix(".bin").stat().st_size / 1024 / 1024
    int8_model_path = "_optimized.".join(ov_model_path.as_posix().split("."))
    quantized_model_size = Path(int8_model_path).with_suffix(".bin").stat().st_size / 1024 / 1024
    print(f'{ov_model_path.as_posix().split(".")[0]}')
    print(f"    * FP16 IR model size: {fp16_ir_model_size:.2f} MB")
    print(f"    * INT8 model size: {quantized_model_size:.2f} MB")
    print(f"    * Model compression rate: {fp16_ir_model_size / quantized_model_size:.3f}")
for model_path in [TEXT_ENCODER_IR_PATH, UNET_I_IR_PATH, UNET_II_IR_PATH]:
    calculate_compression_rate(model_path)
models/encoder_ir
    * FP16 IR model size: 22006.77 MB
    * INT8 model size: 4546.70 MB
    * Model compression rate: 4.840
models/unet_ir_I
    * FP16 IR model size: 1417.56 MB
    * INT8 model size: 355.16 MB
    * Model compression rate: 3.991
models/unet_ir_II
    * FP16 IR model size: 1758.82 MB
    * INT8 model size: 440.49 MB
    * Model compression rate: 3.993

To measure the inference performance of OpenVINO FP16 and INT8 models, use Benchmark Tool.

NOTE: For more accurate performance, run benchmark_app in a terminal/command prompt after closing other applications. Run benchmark_app --help to see an overview of all command-line options.

import re

def get_fps(benchmark_output: str):
    parsed_output = [line for line in benchmark_output if 'Throughput:' in line]
    fps = re.findall(r"\d+\.\d+", parsed_output[0])[0]
    return fps

Text encoder

benchmark_output = !benchmark_app -m $TEXT_ENCODER_IR_PATH -d $device.value -api async
original_fps = get_fps(benchmark_output)
print(f"FP16 Text Encoder Throughput: {original_fps} FPS")

benchmark_output = !benchmark_app -m $TEXT_ENCODER_INT8_IR_PATH -d $device.value -api async
optimized_fps = get_fps(benchmark_output)
print(f"INT8 Text Encoder Throughput: {optimized_fps} FPS")
print(f"Text encoder speed up: {float(optimized_fps) / float(original_fps)}")
FP16 Text Encoder Throughput: 0.99 FPS
INT8 Text Encoder Throughput: 2.47 FPS
Text encoder speed up: 2.4949494949494953

First stage UNet

benchmark_output = !benchmark_app -m $UNET_I_IR_PATH -d $device.value -api async
original_fps = get_fps(benchmark_output)
print(f"FP16 1 stage U-Net Throughput: {original_fps} FPS")

benchmark_output = !benchmark_app -m $UNET_I_INT8_PATH -d $device.value -api async
optimized_fps = get_fps(benchmark_output)
print(f"INT8 1 stage U-Net Throughput: {optimized_fps} FPS")
print(f"1 stage U-Net speed up: {float(optimized_fps) / float(original_fps)}")
FP16 1 stage U-Net Throughput: 4.65 FPS
INT8 1 stage U-Net Throughput: 12.06 FPS
1 stage U-Net speed up: 2.593548387096774

Second stage UNet

benchmark_output = !benchmark_app -m $UNET_II_IR_PATH -d $device.value -api async
original_fps = get_fps(benchmark_output)
print(f"FP16 2 stage U-Net Throughput: {original_fps} FPS")

benchmark_output = !benchmark_app -m $UNET_II_INT8_PATH -d $device.value -api async
optimized_fps = get_fps(benchmark_output)
print(f"INT8 2 stage U-Net Throughput: {optimized_fps} FPS")
print(f"2 stage U-Net speed up: {float(optimized_fps) / float(original_fps)}")
FP16 2 stage U-Net Throughput: 0.28 FPS
INT8 2 stage U-Net Throughput: 0.92 FPS
2 stage U-Net speed up: 3.2857142857142856