pytorch-finetuner
Setup
$ git clone https://github.com/knjcode/pytorch-finetuner
$ cd pytorch-finetuner
$ pip install -r requirements.txt
Example usage
(Prerequisites) Arrange images into their respective directories
A training data directory (images/train), validation data directory (images/valid), and test data directory (images/test) should containing one subdirectory per image class.
For example, arrange training, validation, and test data as follows.
images/
train/
airplanes/
airplane001.jpg
airplane002.jpg
...
watch/
watch001.jpg
watch002.jpg
...
valid/
airplanes/
airplane101.jpg
airplane102.jpg
...
watch/
watch101.jpg
watch102.jpg
...
test/
airplanes/
airplane201.jpg
airplane202.jpg
...
watch/
watch201.jpg
watch202.jpg
...
Prepare example images using bundled shell script
Download Caltech 101 dataset, and split part of it into the example_images directory.
$ util/caltech101_prepare.sh
example_images/trainis train set of 60 images for each classesexample_images/validis validation set of 20 images for each classesexample_imags/testis test set of 20 images for each classes
$ util/counter.sh example_images/train
example_images/train contains 10 directories
Faces 60
Leopards 60
Motorbikes 60
airplanes 60
bonsai 60
car_side 60
chandelier 60
hawksbill 60
ketch 60
watch 60
With this dataset you can immediately try fine-tuning with pytorch-finetuner.
$ ./train.py example_images --model resnet50 --epochs 30 --lr-step-epochs 10,20
train
$ ./train.py example_images --epochs 5
$ ./train.py example_images --epochs 5
Running script with args: Namespace(base_lr=0.0125, batch_size=128, cosine_annealing_eta_min=1e-05, cosine_annealing_mult=2, cosine_annealing_t_max=None, cuda=True, cutout=False, cutout_holes=None, cutout_length=None, data='example_images', disp_batches=1, epochs=3, image_dump=False, input_size=224, jitter_brightness=0.1, jitter_contrast=0.1, jitter_hue=0.05, jitter_saturation=0.1, log_dir='logs', lr_factor=0.1, lr_patience=None, lr_step_epochs=[30, 60, 90], mixup=False, mixup_alpha=None, model='resnet18', model_dir='model', momentum=0.9, no_cuda=False, optimizer='sgd', prefix='20181226184319', random_erasing=False, random_erasing_p=None, random_erasing_r1=None, random_erasing_r2=None, random_erasing_sh=None, random_erasing_sl=None, random_horizontal_flip=0.5, random_resized_crop_ratio=[0.75, 1.3333333333333333], random_resized_crop_scale=[0.08, 1.0], random_rotate_degree=3.0, random_vertical_flip=0.0, resume=None, rgb_mean=[0.485, 0.456, 0.406], rgb_std=[0.229, 0.224, 0.225], ricap=False, ricap_beta=None, ricap_with_line=False, save_best_and_last=False, save_best_only=False, scale_size=256, scratch=False, seed=None, start_epoch=0, warmup_epochs=5, wd=0.0001, workers=22)
scale_size: 256 input_size: 224
rgb_mean: [0.0, 0.0, 0.0]
rgb_std: [1.0, 1.0, 1.0]
number of train dataset: 600
number of validation dataset: 200
number of classes: 10
=> using pre-trained model 'resnet18'
=> using GPU
=> using optimizer: sgd
=> using MultiStepLR scheduler
=> model and logs prefix: 20181226184319
=> log dir: logs
=> model dir: model
=> tensorboardX log dir: logs/20181226184319-tensorboardX
Epoch[0] Batch[1] [128/600 (20%)] speed: 15.22 samples/sec accuracy: 0.0781250000 loss: 2.5856597424
Epoch[0] Batch[2] [256/600 (40%)] speed: 30.00 samples/sec accuracy: 0.0820312500 loss: 2.5729613304
Epoch[0] Batch[3] [384/600 (60%)] speed: 44.36 samples/sec accuracy: 0.0885416642 loss: 2.5417649746
Epoch[0] Batch[4] [512/600 (80%)] speed: 58.31 samples/sec accuracy: 0.0996093750 loss: 2.5194582939
Epoch[0] Batch[5] [600/600 (100%)] speed: 60.46 samples/sec accuracy: 0.1183238626 loss: 2.4788246155
Epoch[0] Train-accuracy: 0.11832386255264282
Epoch[0] Train-loss: 2.4788246154785156
Epoch[0] learning-rate: 0.0020833333333333333
Epoch[0] Validation-accuracy: 0.2777777910232544
Epoch[0] Validation-loss: 2.0763139724731445
Epoch[0] Time cost: 13.761763095855713 [sec]
=> Saved checkpoint to "model/20181226184319-resnet18-0001.model"
=> Saved checkpoint to "model/20181226184319-resnet18-best.model"
Epoch[1] Batch[1] [128/600 (20%)] speed: 42.31 samples/sec accuracy: 0.2578125000 loss: 2.2109849453
Epoch[1] Batch[2] [256/600 (40%)] speed: 81.27 samples/sec accuracy: 0.3007812500 loss: 2.1127164364
...
(snip)
...
Epoch[4] Batch[1] [128/600 (20%)] speed: 42.00 samples/sec accuracy: 0.9531250000 loss: 0.2027403414
Epoch[4] Batch[2] [256/600 (40%)] speed: 80.71 samples/sec accuracy: 0.9648437500 loss: 0.1552993804
Epoch[4] Batch[3] [384/600 (60%)] speed: 116.50 samples/sec accuracy: 0.9661458135 loss: 0.1483899951
Epoch[4] Batch[4] [512/600 (80%)] speed: 149.72 samples/sec accuracy: 0.9726562500 loss: 0.1362725645
Epoch[4] Batch[5] [600/600 (100%)] speed: 170.88 samples/sec accuracy: 0.9690340757 loss: 0.1381170452
Epoch[4] Train-accuracy: 0.9690340757369995
Epoch[4] Train-loss: 0.13811704516410828
Epoch[4] learning-rate: 0.010416666666666668
Epoch[4] Validation-accuracy: 1.0
Epoch[4] Validation-loss: 0.016005855053663254
Epoch[4] Time cost: 5.863810300827026 [sec]
=> Saved checkpoint to "model/20181226184319-resnet18-0005.model"
=> Saved checkpoint to "model/20181226184319-resnet18-best.model"
test
$ ./test.py example_images -m model/20181226184319-resnet18-best.model
$ ./test.py example_images -m model/20181226184319-resnet18-best.model
Running script with args: Namespace(batch_size=128, cuda=True, data='example_images', input_size=None, log_dir='logs', model='model/20181226184319-resnet18-best.model', no_cuda=False, num_classes=None, prefix='20190117043959', print_cr=False, rgb_mean='0,0,0', rgb_std='1,1,1', scale_size=None, seed=None, topk=3, tta=False, tta_custom_seven_crop=False, tta_custom_six_crop=False, tta_custom_ten_crop=False, tta_custom_twenty_crop=False, tta_ten_crop=False, workers=22)
=> loading saved checkpoint 'model/20181226184319-resnet18-best.model'
scale_size: 256 input_size: 224
rgb_mean: [0.0, 0.0, 0.0]
rgb_std: [1.0, 1.0, 1.0]
number of test dataset: 200
number of classes: 10
Test: 100%|██████████| 2/2 [00:06<00:00, 4.37s/it, loss=0.0231, accuracy=99.6]
=> Saved test results to "logs/20190117043959-test-results.log"
=> Saved classification report to "logs/20190117043959-test-classification_report.log"
model: model/20181226184319-resnet18-best.model
Test-loss: 0.023076653480529785
Test-accuracy: 0.995 (199/200)
=> Saved test log to "logs/20190117043959-test.log"
Calculate RGB mean and std of dataset
$ ./train.py example_images/train --calc-rgb-mean-and-std --batch-size 600
=> Calculate rgb mean and std (dir: example_images/train images: 600 batch-size: 600)
Calc rgb mean/std: 100%|████████████████████████| 1/1 [00:04<00:00, 4.19s/it]
=> processed: 600 images
=> calculated rgb mean: [0.5068701 0.50441307 0.47790593]
=> calculated rgb std: [0.28773245 0.27445307 0.29044855]
Please use following command options when train and test:
--rgb-mean 0.507,0.504,0.478 --rgb-std 0.288,0.274,0.290
With Data Augmentation
Random Rotation (Enabled by default)
./train.py example_images --random-rotate-degree 3.0
Random Resized Crop (Enabled by default)
./train.py example_images \
--random-resized-crop-scale 0.08,1.0 \
--random-resized-crop-ratio 0.75,1.3333333333333333
Random Horizontal Flip (Enabled by default)
./train.py example_images --random-horizontal-flip 0.5
Random Vertical Flip
./train.py example_images --random-vertical-flip 0.5
Color Jitter (Enabled by default)
./train.py example_images \
--jitter-brightness 0.10 \
--jitter-contrast 0.10 \
--jitter-saturation 0.10 \
--jitter-hue 0.05
Normalize
./train.py example_images \
--rgb-mean 0.485,0.456,0.406 \
--rgb-std 0.229,0.224,0.225
Cutout, Random Erasing, mixup, RICAP, ICAP, CutMix
# Cutout
$ ./train.py example_images --cutout
# Random Erasing
$ ./train.py example_images --random-erasing
# mixup
$ ./train.py example_images --mixup
# RICAP
$ ./train.py example_images --ricap
# ICAP
$ ./train.py example_images --icap
# CutMix
$ ./train.py example_images --cutmix
# mixup + Cutout
$ ./train.py example_images --mixup --cutout
# mixup + Random Erasing
$ ./train.py example_images --mixup --random-erasing
# RICAP + Cutout
$ ./train.py example_images --ricap --cutout
# RICAP + Random Erasing
$ ./train.py example_images --ricap --random-erasing
- Cutout: Improved Regularization of Convolutional Neural Networks with Cutout
- Random Erasing: Random Erasing Data Augmentation
- mixup: mixup: Beyond Empirical Risk Minimization
- RICAP: Data Augmentation using Random Image Cropping and Patching for Deep CNNs
- CutMix: CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features
Usage
usage: train.py [-h] [--model ARCH] [--from-scratch] [--epochs EPOCHS]
[--batch-size BATCH_SIZE] [--val-batch-size VAL_BATCH_SIZE]
[-j WORKERS] [--prefix PREFIX] [--log-dir LOG_DIR]
[--model-dir MODEL_DIR] [--resume MODEL]
[--start-epoch START_EPOCH] [--disp-batches DISP_BATCHES]
[--save-best-only] [--save-best-and-last] [--drop-last]
[--base-lr BASE_LR] [--lr-factor LR_FACTOR]
[--lr-step-epochs LR_STEP_EPOCHS] [--lr-patience LR_PATIENCE]
[--cosine-annealing-t-max COSINE_ANNEALING_T_MAX]
[--cosine-annealing-mult COSINE_ANNEALING_MULT]
[--cosine-annealing-eta-min COSINE_ANNEALING_ETA_MIN]
[--final-lr FINAL_LR] [--optimizer OPTIMIZER]
[--momentum MOMENTUM] [--wd WD]
[--warmup-epochs WARMUP_EPOCHS] [--scale-size SCALE_SIZE]
[--input-size INPUT_SIZE] [--rgb-mean RGB_MEAN]
[--rgb-std RGB_STD]
[--random-resized-crop-scale RANDOM_RESIZED_CROP_SCALE]
[--random-resized-crop-ratio RANDOM_RESIZED_CROP_RATIO]
[--random-horizontal-flip RANDOM_HORIZONTAL_FLIP]
[--random-vertical-flip RANDOM_VERTICAL_FLIP]
[--jitter-brightness JITTER_BRIGHTNESS]
[--jitter-contrast JITTER_CONTRAST]
[--jitter-saturation JITTER_SATURATION]
[--jitter-hue JITTER_HUE]
[--random-rotate-degree RANDOM_ROTATE_DEGREE]
[--interpolation {BILINEAR,BICUBIC,NEAREST}] [--image-dump]
[--calc-rgb-mean-and-std] [--no-cuda] [--seed SEED]
[--warm_restart_next WARM_RESTART_NEXT]
[--warm_restart_current WARM_RESTART_CURRENT] [--cutout]
[--cutout-holes CUTOUT_HOLES] [--cutout-length CUTOUT_LENGTH]
[--random-erasing] [--random-erasing-p RANDOM_ERASING_P]
[--random-erasing-sl RANDOM_ERASING_SL]
[--random-erasing-sh RANDOM_ERASING_SH]
[--random-erasing-r1 RANDOM_ERASING_R1]
[--random-erasing-r2 RANDOM_ERASING_R2] [--mixup]
[--mixup-alpha MIXUP_ALPHA] [--ricap]
[--ricap-beta RICAP_BETA] [--ricap-with-line] [--icap]
[--icap-beta ICAP_BETA] [--cutmix] [--cutmix-beta CUTMIX_BETA]
[--cutmix-prob CUTMIX_PROB]
DIR
train
positional arguments:
DIR path to dataset (train and validation)
optional arguments:
-h, --help show this help message and exit
--model ARCH, -m ARCH
specify model architecture (default: resnet18)
--from-scratch do not use pre-trained weights (default: False)
--epochs EPOCHS number of total epochs to run (default: 30)
--batch-size BATCH_SIZE, -b BATCH_SIZE
the batch size (default: 128)
--val-batch-size VAL_BATCH_SIZE
the validation batch size (default: 256)
-j WORKERS, --workers WORKERS
number of data loading workers (default: 80% of the
number of cores)
--prefix PREFIX prefix of model and logs (default: auto)
--log-dir LOG_DIR log directory (default: logs)
--model-dir MODEL_DIR
model saving dir (default: model)
--resume MODEL path to saved model (default: None)
--start-epoch START_EPOCH
manual epoch number (default: 0)
--disp-batches DISP_BATCHES
show progress for every n batches (default: auto)
--save-best-only save only the latest best model according to the
validation accuracy (default: False)
--save-best-and-last save last and latest best model according to the
validation accuracy (default: False)
--drop-last drop the last incomplete batch, if the dataset size is
not divisible by the batch size. (default: False)
--base-lr BASE_LR initial learning rate (default: 0.001)
--lr-factor LR_FACTOR
the ratio to reduce lr on each step (default: 0.1)
--lr-step-epochs LR_STEP_EPOCHS
the epochs to reduce the lr (default: 10,20)
--lr-patience LR_PATIENCE
enable ReduceLROnPlateau lr scheduler with specified
patience (default: None)
--cosine-annealing-t-max COSINE_ANNEALING_T_MAX
enable CosineAnnealinigLR scheduler with specified
T_max (default: None)
--cosine-annealing-mult COSINE_ANNEALING_MULT
T_mult of CosineAnnealingLR scheduler
--cosine-annealing-eta-min COSINE_ANNEALING_ETA_MIN
Minimum learning rate of CosineannealingLR scheduler
--final-lr FINAL_LR final_lr of AdaBound optimizer
--optimizer OPTIMIZER
the optimizer type (default: sgd)
--momentum MOMENTUM momentum (default: 0.9)
--wd WD weight decay (default: 1e-04)
--warmup-epochs WARMUP_EPOCHS
number of warmup epochs (default: 5)
--scale-size SCALE_SIZE
scale size (default: auto)
--input-size INPUT_SIZE
input size (default: auto)
--rgb-mean RGB_MEAN RGB mean (default: 0,0,0)
--rgb-std RGB_STD RGB std (default: 1,1,1)
--random-resized-crop-scale RANDOM_RESIZED_CROP_SCALE
range of size of the origin size cropped (default:
0.08,1.0)
--random-resized-crop-ratio RANDOM_RESIZED_CROP_RATIO
range of aspect ratio of the origin aspect ratio
cropped (defaullt: 0.75,1.3333333333333333)
--random-horizontal-flip RANDOM_HORIZONTAL_FLIP
probability of the image being flipped (default: 0.5)
--random-vertical-flip RANDOM_VERTICAL_FLIP
probability of the image being flipped (default: 0.0)
--jitter-brightness JITTER_BRIGHTNESS
jitter brightness of data augmentation (default: 0.10)
--jitter-contrast JITTER_CONTRAST
jitter contrast of data augmentation (default: 0.10)
--jitter-saturation JITTER_SATURATION
jitter saturation of data augmentation (default: 0.10)
--jitter-hue JITTER_HUE
jitter hue of data augmentation (default: 0.05)
--random-rotate-degree RANDOM_ROTATE_DEGREE
rotate degree of data augmentation (default: 3.0)
--interpolation {BILINEAR,BICUBIC,NEAREST}
interpolation. (default: BILINEAR)
--image-dump dump batch images and exit (default: False)
--calc-rgb-mean-and-std
calculate rgb mean and std of train images and exit
(default: False)
--no-cuda disables CUDA training (default: False)
--seed SEED seed for initializing training. (default: None)
--warm_restart_next WARM_RESTART_NEXT
next warm restart epoch (default: None
--warm_restart_current WARM_RESTART_CURRENT
current warm restart epoch (default: None)
--cutout apply cutout (default: False)
--cutout-holes CUTOUT_HOLES
number of holes to cut out from image (default: 1)
--cutout-length CUTOUT_LENGTH
length of the holes (default: 64)
--random-erasing apply random erasing (default: False)
--random-erasing-p RANDOM_ERASING_P
random erasing p (default: 0.5)
--random-erasing-sl RANDOM_ERASING_SL
random erasing sl (default: 0.02)
--random-erasing-sh RANDOM_ERASING_SH
random erasing sh (default: 0.4)
--random-erasing-r1 RANDOM_ERASING_R1
random erasing r1 (default: 0.3)
--random-erasing-r2 RANDOM_ERASING_R2
random erasing r2 (default: 3.3333333333333335)
--mixup apply mixup (default: Falsse)
--mixup-alpha MIXUP_ALPHA
mixup alpha (default: 0.2)
--ricap apply RICAP (default: False)
--ricap-beta RICAP_BETA
RICAP beta (default: 0.3)
--ricap-with-line RICAP with boundary line (default: False)
--icap apply ICAP (default: False)
--icap-beta ICAP_BETA
ICAP beta (default: 0.3)
--cutmix apply CutMix (default: False)
--cutmix-beta CUTMIX_BETA
CutMix beta (default: 1.0)
--cutmix-prob CUTMIX_PROB
CutMix probability (default: 1.0)