diff --git a/condor/eval.sbatch b/condor/eval.sbatch
new file mode 100644
index 0000000000000000000000000000000000000000..5e2882807a0a42ede6f765200926135a5c092a0b
--- /dev/null
+++ b/condor/eval.sbatch
@@ -0,0 +1,14 @@
+#!/bin/bash
+#SBATCH -p big
+#SBATCH --time=24:00:00
+#SBATCH --nodes=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=20
+#SBATCH --mem=400G
+#SBATCH --job-name=eval
+#SBATCH --output=slurm.out
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+
+vitrun --nproc_per_node=8 ${@}
\ No newline at end of file
diff --git a/condor/eval_run.submit b/condor/eval_run.submit
index 8be080e45f32e42086db63a9f621b66a73136f6b..802d850f8ccadaac83f419834ad6115333193fd8 100644
--- a/condor/eval_run.submit
+++ b/condor/eval_run.submit
@@ -40,8 +40,8 @@ Rank = CUDAClockMhz * CUDAComputeUnits
request_GPUs = $(gpus)
# this needs to be specified for the AI@Surrey cluster if requesting a GPU
+GPUMem = 10000
-request_CPUs = 12
-request_memory = 80G
+request_CPUs = 60
+request_memory = 300G
#This job will complete in less than 1 hour
+JobRunTime = 20
diff --git a/pil_img.png b/pil_img.png
new file mode 100644
index 0000000000000000000000000000000000000000..37e0865aaa6a0ce9e546c1da08d4501679db3afe
Binary files /dev/null and b/pil_img.png differ
diff --git a/test_dl.ipynb b/test_dl.ipynb
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/vitookit/datasets/ffcv_transform.py b/vitookit/datasets/ffcv_transform.py
index 036b5933a48720ea48be09229cb40f9f801a6363..bc93642b350feea76691e44d51f72bbdebd23a89 100644
--- a/vitookit/datasets/ffcv_transform.py
+++ b/vitookit/datasets/ffcv_transform.py
@@ -35,6 +35,45 @@ from torch import nn
IMAGENET_MEAN = np.array([0.485, 0.456, 0.406]) * 255
IMAGENET_STD = np.array([0.229, 0.224, 0.225]) * 255
+
+@gin.configurable
+def SimplePipeline(img_size=224,scale=(0.2,1), ratio=(3.0/4.0, 4.0/3.0),blur=False):
+ image_pipeline = [
+ RandomResizedCropRGBImageDecoder((img_size, img_size), scale=scale,ratio=ratio),
+ RandomHorizontalFlip(),
+ NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float16),
+ ToTensor(),
+ ToDevice(torch.device('cuda')),
+ ToTorchImage(),
+ ]
+ if blur:
+ image_pipeline.append(transforms.GaussianBlur(3))
+ label_pipeline = [IntDecoder(), ToTensor(),ToDevice(torch.device('cuda'))]
+ # Pipeline for each data field
+ pipelines = {
+ 'image': image_pipeline,
+ 'label': label_pipeline
+ }
+ return pipelines
+
+@gin.configurable
+def ValPipeline(img_size=224,ratio= 224/256):
+ image_pipeline = [
+ CenterCropRGBImageDecoder((img_size, img_size), ratio),
+ NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float16),
+ ToTensor(),
+ ToDevice(torch.device('cuda')),
+ ToTorchImage(),
+ ]
+ label_pipeline = [IntDecoder(), ToTensor(),ToDevice(torch.device('cuda')),View(-1)]
+ # Pipeline for each data field
+ pipelines = {
+ 'image': image_pipeline,
+ 'label': label_pipeline
+ }
+ return pipelines
+
+
@njit(parallel=False, fastmath=True, inline="always")
def apply_cj(
im,
@@ -363,19 +402,14 @@ def convolution(image: np.ndarray, kernel: list | tuple, output: np.ndarray) ->
pad_y = (m_k - 1) // 2
pad_x = (n_k - 1) // 2
- sub_matrix = image[:m_k, :n_k]
- # center region
+ output.fill(0)
for i in range(y_strides):
for j in range(x_strides):
for c in range(c_i): # looping over the all channels
- sub_matrix = image[i:i+m_k, j:j+n_k,c]
- output[i+pad_y,j+pad_x, c] = np.sum(sub_matrix * kernel)
- # no filter for the padding region
- output[:pad_y] = image[:pad_y]
- output[-pad_y:] = image[-pad_y:]
- output[:, :pad_x] = image[:, :pad_x]
- output[:, -pad_x:] = image[:, -pad_x:]
-
+ for k_i in range(m_k):
+ for k_j in range(n_k):
+ output[i+pad_y,j+pad_x,c] += image[i+k_i,j+k_j,c] * kernel[k_i,k_j]
+
return output
@@ -409,20 +443,36 @@ class ThreeAugmentation(Operation):
+ 0.1140 * images[i, ..., 2:3]
)
else:
- sigma = np.random.uniform(radius_min, radius_max)
- kernel = generate_gaussian_filter(sigma,filter_shape=(5, 5))
+ sigma = np.random.uniform(radius_min, radius_max)
+ kernel_width = int(3 * sigma)
+ if kernel_width % 2 == 0:
+ kernel_width = kernel_width + 1 # make sure kernel width only sth 3,5,7 etc
+ kernel = generate_gaussian_filter(sigma,filter_shape=(kernel_width, kernel_width))
convolution(images[i], kernel, dst[i])
return dst
- randchoice.is_parallel = True
+ randchoice.parallel = True
return randchoice
def declare_state_and_memory(self, previous_state: State) -> Tuple[State, Optional[AllocationQuery]]:
# No updates to state or extra memory necessary!
mem_alloc = AllocationQuery(previous_state.shape,dtype=previous_state.dtype)
- return previous_state, mem_alloc
+ return replace(previous_state,jit_mode=True), mem_alloc
+
+class RandomChoice(nn.Module):
+ """Apply single transformation randomly picked from a list. This transform does not support torchscript."""
+
+ def __init__(self, transforms, p=None):
+ super().__init__()
+ self.transforms = transforms
+ self.p = p
+ def __call__(self, *args):
+ t = random.choices(self.transforms, weights=self.p)[0]
+ return t(*args)
+ def __repr__(self) -> str:
+ return f"{super().__repr__()}(p={self.p})"
@gin.configurable
def ThreeAugmentPipeline(img_size=224,scale=(0.08,1), color_jitter=None):
@@ -434,14 +484,20 @@ def ThreeAugmentPipeline(img_size=224,scale=(0.08,1), color_jitter=None):
[ RandomResizedCropRGBImageDecoder((img_size, img_size), scale=scale,),
RandomHorizontalFlip(),]+
# second_tfl
- [ ThreeAugmentation(),] +
+ # [ ThreeAugmentation(),] +
( [RandomColorJitter(jitter_prob=1, brightness=color_jitter, contrast=color_jitter, saturation=color_jitter, hue=0, seed=None)] if color_jitter else []) +
# final_tfl
[
- NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float16),
+ NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float32),
ToTensor(),
# ToDevice(torch.device('cuda')),
ToTorchImage(),
+ # ThreeAugmentation()
+ RandomChoice([
+ tfms.RandomSolarize(0,1),
+ tfms.RandomGrayscale(1),
+ tfms.GaussianBlur(7,sigma=(0.1,2)),
+ ])
])
label_pipeline = [IntDecoder(), ToTensor(),View(-1)]
# Pipeline for each data field
diff --git a/vitookit/datasets/transform.py b/vitookit/datasets/transform.py
index 24f877c932987a09d0cdb08036bc7f785d16e4ad..5d017f75cf4b5ca98de5782c834a9706a391ed9f 100644
--- a/vitookit/datasets/transform.py
+++ b/vitookit/datasets/transform.py
@@ -6,48 +6,6 @@ from torchvision import transforms
import torch
import gin
-from ffcv.transforms import ToTensor, ToDevice, ToTorchImage, NormalizeImage,RandomHorizontalFlip,View
-from ffcv.fields.decoders import IntDecoder, RandomResizedCropRGBImageDecoder, SimpleRGBImageDecoder, CenterCropRGBImageDecoder
-
-IMAGENET_MEAN = np.array([0.485, 0.456, 0.406]) * 255
-IMAGENET_STD = np.array([0.229, 0.224, 0.225]) * 255
-
-@gin.configurable
-def SimplePipeline(img_size=224,scale=(0.2,1), ratio=(3.0/4.0, 4.0/3.0),blur=False):
- image_pipeline = [
- RandomResizedCropRGBImageDecoder((img_size, img_size), scale=scale,ratio=ratio),
- RandomHorizontalFlip(),
- NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float16),
- ToTensor(),
- ToDevice(torch.device('cuda')),
- ToTorchImage(),
- ]
- if blur:
- image_pipeline.append(transforms.GaussianBlur(3))
- label_pipeline = [IntDecoder(), ToTensor(),ToDevice(torch.device('cuda'))]
- # Pipeline for each data field
- pipelines = {
- 'image': image_pipeline,
- 'label': label_pipeline
- }
- return pipelines
-
-@gin.configurable
-def ValPipeline(img_size=224,ratio= 224/256):
- image_pipeline = [
- CenterCropRGBImageDecoder((img_size, img_size), ratio),
- NormalizeImage(IMAGENET_MEAN, IMAGENET_STD, np.float16),
- ToTensor(),
- ToDevice(torch.device('cuda')),
- ToTorchImage(),
- ]
- label_pipeline = [IntDecoder(), ToTensor(),ToDevice(torch.device('cuda')),View(-1)]
- # Pipeline for each data field
- pipelines = {
- 'image': image_pipeline,
- 'label': label_pipeline
- }
- return pipelines
class PermutePatch(object):
"""
diff --git a/vitookit/evaluation/eval_cls_ffcv.py b/vitookit/evaluation/eval_cls_ffcv.py
index 70a79f0645280dda6fffc0507d2f83c60317ed7b..e9d399e444dcfdcb9183162b93e563f2906ae6d4 100644
--- a/vitookit/evaluation/eval_cls_ffcv.py
+++ b/vitookit/evaluation/eval_cls_ffcv.py
@@ -25,25 +25,22 @@ import math
import sys
import copy
import scipy.io as scio
-from vitookit.datasets.ffcv_transform import ThreeAugmentPipeline
-from vitookit.datasets.transform import ValPipeline, three_augmentation
+from vitookit.datasets.ffcv_transform import *
from vitookit.utils.helper import *
from vitookit.utils import misc
from vitookit.models.build_model import build_model
-from vitookit.datasets import build_dataset
import wandb
from pathlib import Path
from typing import Iterable, Optional
from torch.nn import functional as F
-from timm.models import create_model
+
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from timm.scheduler import create_scheduler
from timm.optim import create_optimizer
from timm.utils import NativeScaler, get_state_dict, ModelEma, accuracy
-from timm.data import Mixup, create_transform
-from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from timm.data import Mixup
from timm.layers import trunc_normal_
from ffcv import Loader
diff --git a/vitookit/evaluation/eval_linear_ffcv.py b/vitookit/evaluation/eval_linear_ffcv.py
index 143b6a8f7fa58120aa56fad70d9265dbdeb781bf..bbdb9f8e41afb6a674d807bdb2bfdf6f8a7419cf 100644
--- a/vitookit/evaluation/eval_linear_ffcv.py
+++ b/vitookit/evaluation/eval_linear_ffcv.py
@@ -19,7 +19,7 @@ import os
import time
from pathlib import Path
-import torch
+
import torch.backends.cudnn as cudnn
import wandb
from vitookit.models.build_model import build_model
@@ -30,7 +30,7 @@ from vitookit.utils.helper import aug_parse, load_pretrained_weights, restart_fr
from timm.models.layers import trunc_normal_
from vitookit.utils.lars import LARS
-from vitookit.datasets.transform import SimplePipeline, ValPipeline
+from vitookit.datasets.ffcv_transform import SimplePipeline, ValPipeline
from ffcv import Loader
from ffcv.loader import OrderOption
@@ -126,7 +126,7 @@ def main(args):
print('job dir: {}'.format(os.path.dirname(os.path.realpath(__file__))))
print("{}".format(args).replace(', ', ',\n'))
-
+ import torch
device = torch.device(args.device)
# fix the seed for reproducibility
diff --git a/vitookit/utils/helper.py b/vitookit/utils/helper.py
index 69c4fdf8f88cf259d2838d166a955ece526a7a4a..6eb89d34caed54147076ad52817fc6968338f512 100644
--- a/vitookit/utils/helper.py
+++ b/vitookit/utils/helper.py
@@ -24,7 +24,7 @@ def bool_flag(s):
def aug_parse(parser: argparse.ArgumentParser):
import yaml
parser.add_argument('--no_resume',default=False,action='store_true',help="")
- parser.add_argument('--cfgs', nargs='+',
+ parser.add_argument('--cfgs', nargs='+', default=[],
help='<Required> Config files *.gin.', required=False)
parser.add_argument('--gin', nargs='+',
help='Overrides config values. e.g. --gin "section.option=value"')