lots of efforts to accelerate training

.gitignore

@@ -3,7 +3,7 @@ src
 __pycache__/
 *.py[cod]
 *$py.class
-
+outputs/
 .vscode
 
 # C extensions

README.md

@@ -1,146 +1,51 @@
-<p align = 'center'>
-<em><b>Fast Forward Computer Vision</b>: train models at a fraction of the cost with accelerated data loading!</em>
-</p>
-<img src='assets/logo.svg' width='100%'/>
+# Fast Forward Computer Vision for Pretraining
+
 <p align = 'center'>
 <!-- <br /> -->
 [<a href="#install-with-anaconda">install</a>]
-[<a href="#quickstart">quickstart</a>]
-[<a href="#features">features</a>]
+[<a href="#features">new features</a>]
 [<a href="https://docs.ffcv.io">docs</a>]
-[<a href="https://join.slack.com/t/ffcv-workspace/shared_invite/zt-11olgvyfl-dfFerPxlm6WtmlgdMuw_2A">support slack</a>]
-[<a href="https://ffcv.io">homepage</a>]
 [<a href="https://arxiv.org/abs/2306.12517">paper</a>]
-<br>
-Maintainers:
-<a href="https://twitter.com/gpoleclerc">Guillaume Leclerc</a>,
-<a href="https://twitter.com/andrew_ilyas">Andrew Ilyas</a> and
-<a href="https://twitter.com/logan_engstrom">Logan Engstrom</a>
 </p>
 
-`ffcv` is a drop-in data loading system that dramatically increases data throughput in model training:
-
-- [Train an ImageNet model](#prepackaged-computer-vision-benchmarks)
-on one GPU in 35 minutes (98¢/model on AWS)
-- [Train a CIFAR-10 model](https://docs.ffcv.io/ffcv_examples/cifar10.html)
-on one GPU in 36 seconds (2¢/model on AWS)
-- Train a `$YOUR_DATASET` model `$REALLY_FAST` (for `$WAY_LESS`)
-
-Keep your training algorithm the same, just replace the data loader! Look at these speedups:
-
-<img src="assets/headline.svg" width='830px'/>
-
-`ffcv` also comes prepacked with [fast, simple code](https://github.com/libffcv/imagenet-example) for [standard vision benchmarks]((https://docs.ffcv.io/benchmarks.html)):
-
-<img src="docs/_static/perf_scatterplot.svg" width='830px'/>
+This library is derived from [FFCV](https://github.com/libffcv/ffcv) to optimize the memory usage and accelerate data loading. 
 
 ## Installation
-### Linux
+### Running Environment
 ```
-conda create -y -n ffcv python=3.9 cupy pkg-config libjpeg-turbo opencv pytorch torchvision cudatoolkit=11.3 numba -c pytorch -c conda-forge
+conda create -y -n ffcv "python>=3.9" cupy pkg-config "libjpeg-turbo>=3.0.0" opencv numba -c conda-forge
 conda activate ffcv
-pip install ffcv
-```
-Troubleshooting note 1: if the above commands result in a package conflict error, try running ``conda config --env --set channel_priority flexible`` in the environment and rerunning the installation command.
-
-Troubleshooting note 2: on some systems (but rarely), you'll need to add the ``compilers`` package to the first command above.
-
-Troubleshooting note 3: courtesy of @kschuerholt, here is a [Dockerfile](https://github.com/kschuerholt/pytorch_cuda_opencv_ffcv_docker) that may help with conda-free installation
-
-### Windows
-* Install <a href="https://opencv.org/releases/">opencv4</a>
-  * Add `..../opencv/build/x64/vc15/bin` to PATH environment variable
-* Install <a href="https://sourceforge.net/projects/libjpeg-turbo/files/">libjpeg-turbo</a>, download libjpeg-turbo-x.x.x-vc64.exe, not gcc64
-  * Add `..../libjpeg-turbo64/bin` to PATH environment variable
-* Install <a href="https://www.sourceware.org/pthreads-win32/">pthread</a>, download last release.zip
-  * After unzip, rename Pre-build.2 folder to pthread
-  * Open `pthread/include/pthread.h`, and add the code below to the top of the file.  
-  ```cpp
-  #define HAVE_STRUCT_TIMESPEC
-  ```
-  * Add `..../pthread/dll` to PATH environment variable
-* Install <a href="https://docs.cupy.dev/en/stable/install.html#installing-cupy">cupy</a> depending on your CUDA Toolkit version.
-* `pip install ffcv`
-
-## Citation
-If you use FFCV, please cite it as:
-
-```
-@inproceedings{leclerc2023ffcv,
-    author = {Guillaume Leclerc and Andrew Ilyas and Logan Engstrom and Sung Min Park and Hadi Salman and Aleksander Madry},
-    title = {{FFCV}: Accelerating Training by Removing Data Bottlenecks},
-    year = {2023},
-    booktitle = {Computer Vision and Pattern Recognition (CVPR)},
-    note = {\url{https://github.com/libffcv/ffcv/}. commit xxxxxxx}
-}
-```
-(Make sure to replace xxxxxxx above with the hash of the commit used!)
-
-## Quickstart
-Accelerate <a href="#features">*any*</a> learning system with `ffcv`.
-First,
-convert your dataset into `ffcv` format (`ffcv` converts both indexed PyTorch datasets and
-<a href="https://github.com/webdataset/webdataset">WebDatasets</a>):
-```python
-from ffcv.writer import DatasetWriter
-from ffcv.fields import RGBImageField, IntField
-
-# Your dataset (`torch.utils.data.Dataset`) of (image, label) pairs
-my_dataset = make_my_dataset()
-write_path = '/output/path/for/converted/ds.beton'
-
-# Pass a type for each data field
-writer = DatasetWriter(write_path, {
-    # Tune options to optimize dataset size, throughput at train-time
-    'image': RGBImageField(max_resolution=256),
-    'label': IntField()
-})
-
-# Write dataset
-writer.from_indexed_dataset(my_dataset)
+conda install pytorch-cuda=11.3 torchvision  -c pytorch -c nvidia
+pip install .
 ```
-Then replace your old loader with the `ffcv` loader at train time (in PyTorch,
-no other changes required!):
-```python
-from ffcv.loader import Loader, OrderOption
-from ffcv.transforms import ToTensor, ToDevice, ToTorchImage, Cutout
-from ffcv.fields.decoders import IntDecoder, RandomResizedCropRGBImageDecoder
-
-# Random resized crop
-decoder = RandomResizedCropRGBImageDecoder((224, 224))
-
-# Data decoding and augmentation
-image_pipeline = [decoder, Cutout(), ToTensor(), ToTorchImage(), ToDevice(0)]
-label_pipeline = [IntDecoder(), ToTensor(), ToDevice(0)]
-
-# Pipeline for each data field
-pipelines = {
-    'image': image_pipeline,
-    'label': label_pipeline
-}
-
-# Replaces PyTorch data loader (`torch.utils.data.Dataloader`)
-loader = Loader(write_path, batch_size=bs, num_workers=num_workers,
-                order=OrderOption.RANDOM, pipelines=pipelines)
-
-# rest of training / validation proceeds identically
-for epoch in range(epochs):
-    ...
-```
-[See here](https://docs.ffcv.io/basics.html) for a more detailed guide to deploying `ffcv` for your dataset.
 
 ## Prepackaged Computer Vision Benchmarks
 From gridding to benchmarking to fast research iteration, there are many reasons
 to want faster model training. Below we present premade codebases for training
 on ImageNet and CIFAR, including both (a) extensible codebases and (b)
 numerous premade training configurations.
 
+## Make Dataset
+We provide a script to make the dataset `examples/write_dataset.py`, which provides three mode:
+- `jpg`: The script will compress all the images to jpg format.
+- `png`: The script will compress all the images to png format. This format is too slow.
+- `raw`: The script will not compress the images.
+- `smart`: The script will compress the images larger than the `threshold`.
+- `proportion`: The script will compress a random subset of the data with size specified by the `compress_probability` argument.
+
+```
+python examples/write_dataset.py --cfg.write_mode=smart --cfg.threshold=206432  --cfg.jpeg_quality=90  \
+    --cfg.num_workers=40 --cfg.max_resolution=500 \
+    --cfg.data_dir=$IMAGENET_DIR/train \
+    --cfg.write_path=$write_path 
+```
 ### ImageNet
 We provide a self-contained script for training ImageNet <it>fast</it>.
 Above we plot the training time versus
 accuracy frontier, and the dataloading speeds, for 1-GPU ResNet-18 and 8-GPU
 ResNet-50 alongside a few baselines.
 
+TODO:
 
 | Link to Config                                                                                                                         |   top_1 |   top_5 |   # Epochs |   Time (mins) | Architecture   | Setup    |
 |:---------------------------------------------------------------------------------------------------------------------------------------|--------:|--------:|-----------:|--------------:|:---------------|:---------|
@@ -167,69 +72,36 @@ potential to raise the accuracy even further). You can find the training script
 <a href="https://github.com/libffcv/ffcv/tree/main/examples/cifar">here</a>.
 
 ## Features
-<img src='docs/_static/clippy-transparent-2.png' width='100%'/>
-
-Computer vision or not, FFCV can help make training faster in a variety of
-resource-constrained settings!
-Our <a href="https://docs.ffcv.io/performance_guide.html">performance guide</a>
-has a more detailed account of the ways in which FFCV can adapt to different
-performance bottlenecks.
-
-
-- **Plug-and-play with any existing training code**: Rather than changing
-  aspects of model training itself, FFCV focuses on removing *data bottlenecks*,
-  which turn out to be a problem everywhere from neural network training to
-  linear regression. This means that:
-
-    - FFCV can be introduced into any existing training code in just a few
-      lines of code (e.g., just swapping out the data loader and optionally the
-      augmentation pipeline);
-    - You don't have to change the model itself to make it faster (e.g., feel
-      free to analyze models *without* CutMix, Dropout, momentum scheduling, etc.);
-    - FFCV can speed up a lot more beyond just neural network training---in
-      fact, the more data-bottlenecked the application (e.g., linear regression,
-      bulk inference, etc.), the faster FFCV will make it!
-
-  See our [Getting started](https://docs.ffcv.io/basics.html) guide,
-  [Example walkthroughs](https://docs.ffcv.io/examples.html), and
-  [Code examples](https://github.com/libffcv/ffcv/tree/main/examples)
-  to see how easy it is to get started!
-- **Fast data processing without the pain**: FFCV automatically handles data
-  reading, pre-fetching, caching, and transfer between devices in an extremely
-  efficiently way, so that users don't have to think about it.
-- **Automatically fused-and-compiled data processing**: By either using
-  [pre-written](https://docs.ffcv.io/api/transforms.html) FFCV transformations
-  or
-  [easily writing custom ones](https://docs.ffcv.io/ffcv_examples/custom_transforms.html),
-  users can
-  take advantage of FFCV's compilation and pipelining abilities, which will
-  automatically fuse and compile simple Python augmentations to machine code
-  using [Numba](https://numba.pydata.org), and schedule them asynchronously to avoid
-  loading delays.
-- **Load data fast from RAM, SSD, or networked disk**: FFCV exposes
-  user-friendly options that can be adjusted based on the resources
-  available. For example, if a dataset fits into memory, FFCV can cache it
-  at the OS level and ensure that multiple concurrent processes all get fast
-  data access. Otherwise, FFCV can use fast process-level caching and will
-  optimize data loading to minimize the underlying number of disk reads. See
-  [The Bottleneck Doctor](https://docs.ffcv.io/bottleneck_doctor.html)
-  guide for more information.
-- **Training multiple models per GPU**: Thanks to fully asynchronous
-  thread-based data loading, you can now interleave training multiple models on
-  the same GPU efficiently, without any data-loading overhead. See
-  [this guide](https://docs.ffcv.io/parameter_tuning.html) for more info.
-- **Dedicated tools for image handling**: All the features above work are
-  equally applicable to all sorts of machine learning models, but FFCV also
-  offers some vision-specific features, such as fast JPEG encoding and decoding,
-  storing datasets as mixtures of raw and compressed images to trade off I/O
-  overhead and compute overhead, etc. See the
-  [Working with images](https://docs.ffcv.io/working_with_images.html) guide for
-  more information.
-
-# Contributors
-
-- [Guillaume Leclerc](https://github.com/GuillaumeLeclerc)
-- [Logan Engstrom](http://loganengstrom.com/)
-- [Andrew Ilyas](http://andrewilyas.com/)
-- [Sam Park](http://sungminpark.com/)
-- [Hadi Salman](http://hadisalman.com/)
+
+Compared to the original FFCV, this library has the following new features:
+
+- **crop decode**: RandomCrop and CenterCrop are now implemented to decode the crop region, which can save memory and accelerate decoding.
+
+- **cache strategy**: There is a potential issue that the OS cache will be swapped out. We use `FFCV_DEFAULT_CACHE_PROCESS` to control the cache process. The choices for the cache process are:
+  - `0`: os cache
+  - `1`: process cache
+  - `2`: Shared Memory 
+
+- **lossless compression**: PNG is supported for lossless compression. We use `RGBImageField(mode='png')` to enable the lossless compression.
+
+- **few memory**: We optimize the memory usage and accelerate data loading.
+
+Comparison of throughput:
+
+| img\_size    |     112 |     160 |     192 |   224   |         |         |         |         |    512 |
+|--------------|--------:|--------:|--------:|:-------:|--------:|--------:|--------:|--------:|-------:|
+| batch\_size  |     512 |     512 |     512 |     128 |     256 |   512   |         |         |    512 |
+| num\_workers |      10 |      10 |      10 |      10 |      10 |       5 |      10 |      20 |     10 |
+| loader       |         |         |         |         |         |         |         |         |        |
+| ours         | 23024.0 | 19396.5 | 16503.6 | 16536.1 | 16338.5 | 12369.7 | 14521.4 | 14854.6 | 4260.3 |
+| ffcv         | 16853.2 | 13906.3 | 13598.4 | 12192.7 | 11960.2 |  9112.7 | 12539.4 | 12601.8 | 3577.8 |
+
+Comparison of memory usage:
+| img\_size    |  112 |  160 |  192 | 224 |      |      |      |      |  512 |
+|--------------|-----:|-----:|-----:|:---:|-----:|-----:|-----:|-----:|-----:|
+| batch\_size  |  512 |  512 |  512 | 128 |  256 |  512 |      |      |  512 |
+| num\_workers |   10 |   10 |   10 |  10 |   10 |    5 |   10 |   20 |   10 |
+| loader       |      |      |      |     |      |      |      |      |      |
+| ours         |  9.0 |  9.8 | 11.4 | 5.8 |  7.7 | 11.4 | 11.4 | 11.4 | 34.0 |
+| ffcv         | 13.4 | 14.8 | 17.7 | 7.6 | 11.0 | 17.7 | 17.7 | 17.7 | 56.6 |
+