EVOLUTION-MANAGER

Edit File: ColorizeTrainingStableLargeBatch.ipynb

{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "## Stable Model Training (Large Batch/Limited GPU Memory Support)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## IMPORTANT: Training has -not- been verified by myself for this notebook ~jantic" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### NOTES: \n", "* This is \"NoGAN\" based training, described in the DeOldify readme.\n", "* This model prioritizes stable and reliable renderings. It does particularly well on portraits and landscapes. It's not as colorful as the artistic model." ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import os\n", "os.environ['CUDA_VISIBLE_DEVICES']='0' " ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import fastai\n", "from fastai import *\n", "from fastai.vision import *\n", "from fastai.callbacks.tensorboard import *\n", "from fastai.vision.gan import *\n", "from deoldify.generators import *\n", "from deoldify.critics import *\n", "from deoldify.dataset import *\n", "from deoldify.loss import *\n", "from deoldify.save import *\n", "from PIL import Image, ImageDraw, ImageFont\n", "from PIL import ImageFile" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Setup" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Activate Large Model Support for PyTorch\n", "This will allow us to fit the model within a GPU with smaller memory capacity (e.g. GTX 1070 8Gb)." ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Large Model Support (LMS) is a feature provided in IBM Watson Machine Learning Community Edition (WML-CE) PyTorch V1.1.0 that allows the successful training of deep learning models that would otherwise exhaust GPU memory and abort with “out-of-memory” errors. LMS manages this oversubscription of GPU memory by temporarily swapping tensors to host memory when they are not needed. One or more elements of a deep learning model can lead to GPU memory exhaustion.\n", "\n", "Requires the use of IBM WML-CE (Available here: https://www.ibm.com/support/knowledgecenter/en/SS5SF7_1.6.1/welcome/welcome.html)\n", "\n", "Further Reading on PyTorch with Large Model Support: https://www.ibm.com/support/knowledgecenter/en/SS5SF7_1.6.1/navigation/wmlce_getstarted_pytorch.html" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import shutil" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Set limit of GPU used before swapping to tensors to host memory\n", "max_gpu_mem = 7\n", "\n", "def gb_to_bytes(gb):\n", " return gb*1024*1024*1024\n", "\n", "# Enable PyTorch LMS\n", "torch.cuda.set.enabled_lms(True)\n", "# Set LMS limit\n", "torch.cuda.set_limit_lms(gb_to_bytes(max_gpu_mem))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Check LMS is enabled\n", "torch.cuda.get_enabled_lms()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Check LMS Limit has been set\n", "torch.cuda.get_limit_lms()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ " " ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Path to Training Data\n", "path = Path('data/imagenet/ILSVRC/Data/CLS-LOC')\n", "path_hr = path\n", "\n", "# Path to Black and White images\n", "path_bandw = Path('/training/DeOldify')\n", "path_lr = path_bandw/'bandw'\n", "\n", "# Name of Model\n", "proj_id = 'StableModel'\n", "\n", "# Name of Generator\n", "gen_name = proj_id + '_gen'\n", "pre_gen_name = gen_name + '_0'\n", "\n", "# Name of Critic\n", "crit_name = proj_id + '_crit'\n", "\n", "# Name of Generated Images folder, located within the Black and White folder\n", "name_gen = proj_id + '_image_gen'\n", "path_gen = path/name_gen\n", "\n", "# Path to tensorboard data\n", "TENSORBOARD_PATH = Path('data/tensorboard/' + proj_id)\n", "\n", "nf_factor = 2\n", "pct_start = 1e-8\n", "\n", "# Number of workers for DataLoader\n", "num_works = 2" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "def get_data(bs:int, sz:int, keep_pct:float):\n", " return get_colorize_data(sz=sz, bs=bs, crappy_path=path_lr, good_path=path_hr, \n", " random_seed=None, keep_pct=keep_pct, num_workers=num_works)\n", "\n", "def get_crit_data(classes, bs, sz):\n", " src = ImageList.from_folder(path, include=classes, recurse=True).split_by_rand_pct(0.1, seed=42)\n", " ll = src.label_from_folder(classes=classes)\n", " data = (ll.transform(get_transforms(max_zoom=2.), size=sz)\n", " .databunch(bs=bs).normalize(imagenet_stats))\n", " return data\n", "\n", "def create_training_images(fn,i):\n", " dest = path_lr/fn.relative_to(path_hr)\n", " dest.parent.mkdir(parents=True, exist_ok=True)\n", " img = PIL.Image.open(fn).convert('LA').convert('RGB')\n", " img.save(dest) \n", " \n", "def save_preds(dl):\n", " i=0\n", " names = dl.dataset.items\n", " \n", " for b in dl:\n", " preds = learn_gen.pred_batch(batch=b, reconstruct=True)\n", " for o in preds:\n", " o.save(path_gen/names[i].name)\n", " i += 1\n", " \n", "def save_gen_images():\n", " if path_gen.exists(): shutil.rmtree(path_gen)\n", " path_gen.mkdir(exist_ok=True)\n", " data_gen = get_data(bs=bs, sz=sz, keep_pct=0.085)\n", " save_preds(data_gen.fix_dl)\n", " PIL.Image.open(path_gen.ls()[0])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Create black and white training images" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Only runs if the directory isn't already created." ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "if not path_lr.exists():\n", " il = ImageList.from_folder(path_hr)\n", " parallel(create_training_images, il.items)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Pre-train generator" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### NOTE\n", "Most of the training takes place here in pretraining for NoGAN. The goal here is to take the generator as far as possible with conventional training, as that is much easier to control and obtain glitch-free results compared to GAN training." ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 64px" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "bs=88 # This can be increased if using PyTorch LMS, training could be slower.\n", "sz=64\n", "keep_pct=1.0" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "data_gen = get_data(bs=bs, sz=sz, keep_pct=keep_pct)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen = gen_learner_wide(data=data_gen, gen_loss=FeatureLoss(), nf_factor=nf_factor)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.callback_fns.append(partial(ImageGenTensorboardWriter, base_dir=TENSORBOARD_PATH, name='GenPre'))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.fit_one_cycle(1, pct_start=0.8, max_lr=slice(1e-3))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.save(pre_gen_name)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.unfreeze()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.fit_one_cycle(1, pct_start=pct_start, max_lr=slice(3e-7, 3e-4))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.save(pre_gen_name)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 128px" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "bs=40 # This can be increased if using PyTorch LMS, training could be slower.\n", "sz=128\n", "keep_pct=1.0" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.data = get_data(sz=sz, bs=bs, keep_pct=keep_pct)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.unfreeze()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.fit_one_cycle(1, pct_start=pct_start, max_lr=slice(1e-7,1e-4))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.save(pre_gen_name)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 192px" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "bs=16 # This can be increased if using PyTorch LMS, training could be slower.\n", "sz=192\n", "keep_pct=0.50" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.data = get_data(sz=sz, bs=bs, keep_pct=keep_pct)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.unfreeze()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.fit_one_cycle(1, pct_start=pct_start, max_lr=slice(5e-8,5e-5))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.save(pre_gen_name)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 256px" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "bs=8 # This can be increased if using PyTorch LMS, training could be slower.\n", "sz=256\n", "keep_pct=0.50" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.data = get_data(sz=sz, bs=bs, keep_pct=keep_pct)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.unfreeze()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.fit_one_cycle(1, pct_start=pct_start, max_lr=slice(5e-8,5e-5))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen.save(pre_gen_name)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Repeatable GAN Cycle" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### NOTE\n", "Best results so far have been based on repeating the cycle below a few times (about 5-8?), until diminishing returns are hit (no improvement in image quality). Each time you repeat the cycle, you want to increment that old_checkpoint_num by 1 so that new check points don't overwrite the old. " ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "old_checkpoint_num = 0\n", "checkpoint_num = old_checkpoint_num + 1\n", "gen_old_checkpoint_name = gen_name + '_' + str(old_checkpoint_num)\n", "gen_new_checkpoint_name = gen_name + '_' + str(checkpoint_num)\n", "crit_old_checkpoint_name = crit_name + '_' + str(old_checkpoint_num)\n", "crit_new_checkpoint_name= crit_name + '_' + str(checkpoint_num)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Save Generated Images" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "bs=8\n", "sz=256" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen = gen_learner_wide(data=data_gen, gen_loss=FeatureLoss(), nf_factor=nf_factor).load(gen_old_checkpoint_name, with_opt=False)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "save_gen_images()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Pretrain Critic" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "##### Only need full pretraining of critic when starting from scratch. Otherwise, just finetune!" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "if old_checkpoint_num == 0:\n", " bs=64\n", " sz=128\n", " learn_gen=None\n", " gc.collect()\n", " data_crit = get_crit_data([name_gen, 'test'], bs=bs, sz=sz)\n", " data_crit.show_batch(rows=3, ds_type=DatasetType.Train, imgsize=3)\n", " learn_critic = colorize_crit_learner(data=data_crit, nf=256)\n", " learn_critic.callback_fns.append(partial(LearnerTensorboardWriter, base_dir=TENSORBOARD_PATH, name='CriticPre'))\n", " learn_critic.fit_one_cycle(6, 1e-3)\n", " learn_critic.save(crit_old_checkpoint_name)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "bs=8\n", "sz=256" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "data_crit = get_crit_data([name_gen, 'test'], bs=bs, sz=sz)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "data_crit.show_batch(rows=3, ds_type=DatasetType.Train, imgsize=3)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_critic = colorize_crit_learner(data=data_crit, nf=256).load(crit_old_checkpoint_name, with_opt=False)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_critic.callback_fns.append(partial(LearnerTensorboardWriter, base_dir=TENSORBOARD_PATH, name='CriticPre'))" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_critic.fit_one_cycle(4, 1e-4)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_critic.save(crit_new_checkpoint_name)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### GAN" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_crit=None\n", "learn_gen=None\n", "gc.collect()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "lr=2e-5\n", "sz=256\n", "bs=5" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "data_crit = get_crit_data([name_gen, 'test'], bs=bs, sz=sz)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_crit = colorize_crit_learner(data=data_crit, nf=256).load(crit_new_checkpoint_name, with_opt=False)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn_gen = gen_learner_wide(data=data_gen, gen_loss=FeatureLoss(), nf_factor=nf_factor).load(gen_old_checkpoint_name, with_opt=False)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "switcher = partial(AdaptiveGANSwitcher, critic_thresh=0.65)\n", "learn = GANLearner.from_learners(learn_gen, learn_crit, weights_gen=(1.0,1.5), show_img=False, switcher=switcher,\n", " opt_func=partial(optim.Adam, betas=(0.,0.9)), wd=1e-3)\n", "learn.callback_fns.append(partial(GANDiscriminativeLR, mult_lr=5.))\n", "learn.callback_fns.append(partial(GANTensorboardWriter, base_dir=TENSORBOARD_PATH, name='GanLearner', visual_iters=100))\n", "learn.callback_fns.append(partial(GANSaveCallback, learn_gen=learn_gen, filename=gen_new_checkpoint_name, save_iters=100))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "#### Instructions: \n", "Find the checkpoint just before where glitches start to be introduced. This is all very new so you may need to play around with just how far you go here with keep_pct." ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "learn.data = get_data(sz=sz, bs=bs, keep_pct=0.03)\n", "learn_gen.freeze_to(-1)\n", "learn.fit(1,lr)" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.7.0" } }, "nbformat": 4, "nbformat_minor": 2 }