diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..cdc8f46
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,126 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+pretrained/
+tb_history/
+logs/
+cls_runs/
+slurm_logs/
+
+#experiments/
+experiments/
+experiments_da/
+config[0-9]*
+*.png
+*.pth
+
+# Distribution / packaging
+.Python
+env/
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# PyInstaller
+# Usually these files are written by a python script from a template
+# before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# pyenv
+.python-version
+
+# celery beat schedule file
+celerybeat-schedule
+
+# SageMath parsed files
+*.sage.py
+
+# dotenv
+.env
+
+# virtualenv
+.venv
+venv/
+ENV/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+
+# input data, saved log, checkpoints
+data/
+input/
+saved/
+outputs/
+datasets/
+
+# editor, os cache directory
+.vscode/
+.idea/
+__MACOSX/
+
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000..eca2867
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Yassine Ouali
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..6f62f3b
--- /dev/null
+++ b/README.md
@@ -0,0 +1,183 @@
+
+
+## Semi-Supervised Semantic Segmentation with Cross-Consistecy Training (CCT)
+
+#### [Paper](https://arxiv.org/abs/2003.09005), [Project Page](https://yassouali.github.io/cct_page/)
+
+This repo contains the official implementation of CVPR 2020 paper: Semi-Supervised Semantic Segmentation with Cross-Consistecy Training, which
+adapts the traditional consistency training framework of semi-supervised learning for semantic segmentation, with an extension to weak-supervised
+learning and learning on multiple domains.
+
+
+
+### Highlights
+
+**(1) Consistency Training for semantic segmentation.** \
+We observe that for semantic segmentation, due to the dense nature of the task,
+the cluster assumption is more easily enforced over the hidden representations rather than the inputs.
+
+**(2) Cross-Consistecy Training.** \
+We propose CCT (Cross-Consistecy Training) for semi-supervised semantic segmentation, where we define
+a number of novel perturbations, and show the effectiveness of enforcing consistency over the encoder's outputs
+rather than the inputs.
+
+**(3) Using weak-labels and pixel-level labels from multiple domains.** \
+The proposed method is quite simple and flexible, and can easily be extended to use image-level labels and
+pixel-level labels from multiple-domains.
+
+
+
+### Requirements
+
+This repo was tested with Ubuntu 18.04.3 LTS, Python 3.7, PyTorch 1.1.0, and CUDA 10.0. But it should be runnable with recent PyTorch versions >=1.1.0.
+
+The required packages are `pytorch` and `torchvision`, together with `PIL` and `opencv` for data-preprocessing and `tqdm` for showing the training progress.
+With some additional modules like `dominate` to save the results in the form of HTML files. To setup the necessary modules, simply run:
+
+```bash
+pip install -r requirements.txt
+```
+
+### Dataset
+
+In this repo, we use **Pascal VOC**, to obtain it, first download the [original dataset](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar), after extracting the files we'll end up with `VOCtrainval_11-May-2012/VOCdevkit/VOC2012` containing the image sets, the XML annotation for both object detection and segmentation, and JPEG images.\
+The second step is to augment the dataset using the additionnal annotations provided by [Semantic Contours from Inverse Detectors](http://home.bharathh.info/pubs/pdfs/BharathICCV2011.pdf). Download the rest of the annotations [SegmentationClassAug](https://www.dropbox.com/s/oeu149j8qtbs1x0/SegmentationClassAug.zip?dl=0) and add them to the path `VOCtrainval_11-May-2012/VOCdevkit/VOC2012`, now we're set, for training use the path to `VOCtrainval_11-May-2012`.
+
+
+### Training
+
+To train a model, first download PASCAL VOC as detailed above, then set `data_dir` to the dataset path in the config file in `configs/config.json` and set the rest of the parameters, like the number of GPUs, cope size, data augmentation ... etc ,you can also change CCT hyperparameters if you wish, more details below. Then simply run:
+
+```bash
+python train.py --config configs/config.json
+```
+
+The log files and the `.pth` checkpoints will be saved in `saved\EXP_NAME`, to monitor the training using tensorboard, please run:
+
+```bash
+tensorboard --logdir saved
+```
+
+To resume training using a saved `.pth` model:
+
+```bash
+python train.py --config configs/config.json --resume saved/CCT/checkpoint.pth
+```
+
+**Results**: The results will be saved in `saved` as an html file, containing the validation results,
+and the name it will take is `experim_name` specified in `configs/config.json`.
+
+### Pseudo-labels
+
+If you want to use image level labels to train the auxiliary labels as explained in section 3.3 of the paper. First generate the pseudo-labels
+using the code in `pseudo_labels`:
+
+
+```bash
+cd pseudo_labels
+python run.py --voc12_root DATA_PATH
+```
+
+`DATA_PATH` must point to the folder containing `JPEGImages` in Pascal Voc dataset. The results will be
+saved in `pseudo_labels/result/pseudo_labels` as PNG files, the flag `use_weak_labels` needs to be set to True in the config file, and
+then we can train the model as detailed above.
+
+
+### Inference
+
+For inference, we need a pretrained model, the jpg images we'd like to segment and the config used in training (to load the correct model and other parameters),
+
+```bash
+python inference.py --config config.json --model best_model.pth --images images_folder
+```
+
+The predictions will be saved as `.png` images in `outputs\` is used, for Pacal VOC the default palette is:
+
+
+
+Here are the flags available for inference:
+
+```
+--images Folder containing the jpg images to segment.
+--model Path to the trained pth model.
+--config The config file used for training the model.
+```
+
+### Citation ✏️ 📄
+
+If you find this repo useful for your research, please consider citing the paper as follows:
+
+```
+@inproceedings{ouali2020semi,
+ title={Semi-Supervised Semantic Segmentation with Cross-Consistency Training},
+ author={Ouali, Yassine and Hudelot, C{\'e}line and Tami, Myriam},
+ year = {2020},
+ booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
+ note = {to appear},
+ pubstate = {published},
+ tppubtype = {inproceedings}
+}
+```
+
+For any questions, please contact Yassine Ouali (yassine.ouali@centralesupelec.fr).
+
+#### Config file details ⚙️
+
+Bellow we detail the CCT parameters that can be controlled in the config file `configs/config.json`, the rest of the parameters
+are self-explanatory.
+
+```javascript
+{
+ "name": "CCT",
+ "experim_name": "CCT", // The name the results will take (html and the folder in /saved)
+ "n_gpu": 1, // Number of GPUs
+ "n_labeled_examples": 1000, // Number of labeled examples (choices are 60, 100, 200,
+ // 300, 500, 800, 1000, 1464, and the splits are in dataloaders/voc_splits)
+ "diff_lrs": true,
+ "ramp_up": 0.1, // The unsupervised loss will be slowly scaled up in the first 10% of Training time
+ "unsupervised_w": 30, // Weighting of the unsupervised loss
+ "ignore_index": 255,
+ "lr_scheduler": "Poly",
+ "use_weak_labels": false, // If the pseudo-labels were generated, we can use them to train the aux. decoders
+ "weakly_loss_w": 0.4, // Weighting of the weakly-supervised loss
+ "pretrained": true,
+
+ "model":{
+ "supervised": true, // Supervised setting (training only on the labeled examples)
+ "semi": false, // Semi-supervised setting
+ "supervised_w": 1, // Weighting of the supervised loss
+
+ "sup_loss": "CE", // supervised loss, choices are CE and ab-CE = ["CE", "ABCE"]
+ "un_loss": "MSE", // unsupervised loss, choices are CE and KL-divergence = ["MSE", "KL"]
+
+ "softmax_temp": 1,
+ "aux_constraint": false, // Pair-wise loss (sup. mat.)
+ "aux_constraint_w": 1,
+ "confidence_masking": false, // Confidence masking (sup. mat.)
+ "confidence_th": 0.5,
+
+ "drop": 6, // Number of DropOut decoders
+ "drop_rate": 0.5, // Dropout probability
+ "spatial": true,
+
+ "cutout": 6, // Number of G-Cutout decoders
+ "erase": 0.4, // We drop 40% of the area
+
+ "vat": 2, // Number of I-VAT decoders
+ "xi": 1e-6, // VAT parameters
+ "eps": 2.0,
+
+ "context_masking": 2, // Number of Con-Msk decoders
+ "object_masking": 2, // Number of Obj-Msk decoders
+ "feature_drop": 6, // Number of F-Drop decoders
+
+ "feature_noise": 6, // Number of F-Noise decoders
+ "uniform_range": 0.3 // The range of the noise
+ },
+```
+
+#### Acknowledgements
+
+- Pseudo-labels generation is based on Jiwoon Ahn's implementation [irn](https://github.com/jiwoon-ahn/irn).
+- Code structure was based on [Pytorch-Template](https://github.com/victoresque/pytorch-template/blob/master/README.m)
+- ResNet backbone was downloaded from [torchcv](https://github.com/donnyyou/torchcv)
diff --git a/base/__init__.py b/base/__init__.py
new file mode 100644
index 0000000..f5c8e0b
--- /dev/null
+++ b/base/__init__.py
@@ -0,0 +1,6 @@
+from .base_dataloader import *
+from .base_dataset import *
+from .base_model import *
+from .base_trainer import *
+
+
diff --git a/base/base_dataloader.py b/base/base_dataloader.py
new file mode 100644
index 0000000..09b72b0
--- /dev/null
+++ b/base/base_dataloader.py
@@ -0,0 +1,48 @@
+import numpy as np
+from copy import deepcopy
+import torch
+from torch.utils.data import DataLoader
+from torch.utils.data.sampler import SubsetRandomSampler
+
+class BaseDataLoader(DataLoader):
+ def __init__(self, dataset, batch_size, shuffle, num_workers, val_split = 0.0):
+ self.shuffle = shuffle
+ self.dataset = dataset
+ self.nbr_examples = len(dataset)
+ if val_split:
+ self.train_sampler, self.val_sampler = self._split_sampler(val_split)
+ else:
+ self.train_sampler, self.val_sampler = None, None
+
+ self.init_kwargs = {
+ 'dataset': self.dataset,
+ 'batch_size': batch_size,
+ 'shuffle': self.shuffle,
+ 'num_workers': num_workers,
+ 'pin_memory': True
+ }
+ super(BaseDataLoader, self).__init__(sampler=self.train_sampler, **self.init_kwargs)
+
+ def _split_sampler(self, split):
+ if split == 0.0:
+ return None, None
+
+ self.shuffle = False
+
+ split_indx = int(self.nbr_examples * split)
+ np.random.seed(0)
+
+ indxs = np.arange(self.nbr_examples)
+ np.random.shuffle(indxs)
+ train_indxs = indxs[split_indx:]
+ val_indxs = indxs[:split_indx]
+ self.nbr_examples = len(train_indxs)
+
+ train_sampler = SubsetRandomSampler(train_indxs)
+ val_sampler = SubsetRandomSampler(val_indxs)
+ return train_sampler, val_sampler
+
+ def get_val_loader(self):
+ if self.val_sampler is None:
+ return None
+ return DataLoader(sampler=self.val_sampler, **self.init_kwargs)
diff --git a/base/base_dataset.py b/base/base_dataset.py
new file mode 100644
index 0000000..22e938b
--- /dev/null
+++ b/base/base_dataset.py
@@ -0,0 +1,188 @@
+import random, math
+import numpy as np
+import cv2
+import torch
+import torch.nn.functional as F
+from torch.utils.data import Dataset
+from PIL import Image
+from torchvision import transforms
+from scipy import ndimage
+from math import ceil
+
+class BaseDataSet(Dataset):
+ def __init__(self, data_dir, split, mean, std, ignore_index, base_size=None, augment=True, val=False,
+ jitter=False, use_weak_lables=False, weak_labels_output=None, crop_size=None, scale=False, flip=False, rotate=False,
+ blur=False, return_id=False, n_labeled_examples=None):
+
+ self.root = data_dir
+ self.split = split
+ self.mean = mean
+ self.std = std
+ self.augment = augment
+ self.crop_size = crop_size
+ self.jitter = jitter
+ self.image_padding = (np.array(mean)*255.).tolist()
+ self.ignore_index = ignore_index
+ self.return_id = return_id
+ self.n_labeled_examples = n_labeled_examples
+ self.val = val
+
+ self.use_weak_lables = use_weak_lables
+ self.weak_labels_output = weak_labels_output
+
+ if self.augment:
+ self.base_size = base_size
+ self.scale = scale
+ self.flip = flip
+ self.rotate = rotate
+ self.blur = blur
+
+ self.jitter_tf = transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1)
+ self.to_tensor = transforms.ToTensor()
+ self.normalize = transforms.Normalize(mean, std)
+
+ self.files = []
+ self._set_files()
+
+ cv2.setNumThreads(0)
+
+ def _set_files(self):
+ raise NotImplementedError
+
+ def _load_data(self, index):
+ raise NotImplementedError
+
+ def _rotate(self, image, label):
+ # Rotate the image with an angle between -10 and 10
+ h, w, _ = image.shape
+ angle = random.randint(-10, 10)
+ center = (w / 2, h / 2)
+ rot_matrix = cv2.getRotationMatrix2D(center, angle, 1.0)
+ image = cv2.warpAffine(image, rot_matrix, (w, h), flags=cv2.INTER_CUBIC)#, borderMode=cv2.BORDER_REFLECT)
+ label = cv2.warpAffine(label, rot_matrix, (w, h), flags=cv2.INTER_NEAREST)#, borderMode=cv2.BORDER_REFLECT)
+ return image, label
+
+ def _crop(self, image, label):
+ # Padding to return the correct crop size
+ if (isinstance(self.crop_size, list) or isinstance(self.crop_size, tuple)) and len(self.crop_size) == 2:
+ crop_h, crop_w = self.crop_size
+ elif isinstance(self.crop_size, int):
+ crop_h, crop_w = self.crop_size, self.crop_size
+ else:
+ raise ValueError
+
+ h, w, _ = image.shape
+ pad_h = max(crop_h - h, 0)
+ pad_w = max(crop_w - w, 0)
+ pad_kwargs = {
+ "top": 0,
+ "bottom": pad_h,
+ "left": 0,
+ "right": pad_w,
+ "borderType": cv2.BORDER_CONSTANT,}
+ if pad_h > 0 or pad_w > 0:
+ image = cv2.copyMakeBorder(image, value=self.image_padding, **pad_kwargs)
+ label = cv2.copyMakeBorder(label, value=self.ignore_index, **pad_kwargs)
+
+ # Cropping
+ h, w, _ = image.shape
+ start_h = random.randint(0, h - crop_h)
+ start_w = random.randint(0, w - crop_w)
+ end_h = start_h + crop_h
+ end_w = start_w + crop_w
+ image = image[start_h:end_h, start_w:end_w]
+ label = label[start_h:end_h, start_w:end_w]
+ return image, label
+
+ def _blur(self, image, label):
+ # Gaussian Blud (sigma between 0 and 1.5)
+ sigma = random.random() * 1.5
+ ksize = int(3.3 * sigma)
+ ksize = ksize + 1 if ksize % 2 == 0 else ksize
+ image = cv2.GaussianBlur(image, (ksize, ksize), sigmaX=sigma, sigmaY=sigma, borderType=cv2.BORDER_REFLECT_101)
+ return image, label
+
+ def _flip(self, image, label):
+ # Random H flip
+ if random.random() > 0.5:
+ image = np.fliplr(image).copy()
+ label = np.fliplr(label).copy()
+ return image, label
+
+ def _resize(self, image, label, bigger_side_to_base_size=True):
+ if isinstance(self.base_size, int):
+ h, w, _ = image.shape
+ if self.scale:
+ longside = random.randint(int(self.base_size*0.5), int(self.base_size*2.0))
+ #longside = random.randint(int(self.base_size*0.5), int(self.base_size*1))
+ else:
+ longside = self.base_size
+
+ if bigger_side_to_base_size:
+ h, w = (longside, int(1.0 * longside * w / h + 0.5)) if h > w else (int(1.0 * longside * h / w + 0.5), longside)
+ else:
+ h, w = (longside, int(1.0 * longside * w / h + 0.5)) if h < w else (int(1.0 * longside * h / w + 0.5), longside)
+ image = np.asarray(Image.fromarray(np.uint8(image)).resize((w, h), Image.BICUBIC))
+ label = cv2.resize(label, (w, h), interpolation=cv2.INTER_NEAREST)
+ return image, label
+
+ elif (isinstance(self.base_size, list) or isinstance(self.base_size, tuple)) and len(self.base_size) == 2:
+ h, w, _ = image.shape
+ if self.scale:
+ scale = random.random() * 1.5 + 0.5 # Scaling between [0.5, 2]
+ h, w = int(self.base_size[0] * scale), int(self.base_size[1] * scale)
+ else:
+ h, w = self.base_size
+ image = np.asarray(Image.fromarray(np.uint8(image)).resize((w, h), Image.BICUBIC))
+ label = cv2.resize(label, (w, h), interpolation=cv2.INTER_NEAREST)
+ return image, label
+
+ else:
+ raise ValueError
+
+ def _val_augmentation(self, image, label):
+ if self.base_size is not None:
+ image, label = self._resize(image, label)
+ image = self.normalize(self.to_tensor(Image.fromarray(np.uint8(image))))
+ return image, label
+
+ image = self.normalize(self.to_tensor(Image.fromarray(np.uint8(image))))
+ return image, label
+
+ def _augmentation(self, image, label):
+ h, w, _ = image.shape
+
+ if self.base_size is not None:
+ image, label = self._resize(image, label)
+
+ if self.crop_size is not None:
+ image, label = self._crop(image, label)
+
+ if self.flip:
+ image, label = self._flip(image, label)
+
+ image = Image.fromarray(np.uint8(image))
+ image = self.jitter_tf(image) if self.jitter else image
+
+ return self.normalize(self.to_tensor(image)), label
+
+ def __len__(self):
+ return len(self.files)
+
+ def __getitem__(self, index):
+ image, label, image_id = self._load_data(index)
+ if self.val:
+ image, label = self._val_augmentation(image, label)
+ elif self.augment:
+ image, label = self._augmentation(image, label)
+
+ label = torch.from_numpy(np.array(label, dtype=np.int32)).long()
+ return image, label
+
+ def __repr__(self):
+ fmt_str = "Dataset: " + self.__class__.__name__ + "\n"
+ fmt_str += " # data: {}\n".format(self.__len__())
+ fmt_str += " Split: {}\n".format(self.split)
+ fmt_str += " Root: {}".format(self.root)
+ return fmt_str
+
diff --git a/base/base_model.py b/base/base_model.py
new file mode 100644
index 0000000..47a91bc
--- /dev/null
+++ b/base/base_model.py
@@ -0,0 +1,22 @@
+import logging
+import torch.nn as nn
+import numpy as np
+
+class BaseModel(nn.Module):
+ def __init__(self):
+ super(BaseModel, self).__init__()
+ self.logger = logging.getLogger(self.__class__.__name__)
+
+ def forward(self):
+ raise NotImplementedError
+
+ def summary(self):
+ model_parameters = filter(lambda p: p.requires_grad, self.parameters())
+ nbr_params = sum([np.prod(p.size()) for p in model_parameters])
+ self.logger.info(f'Nbr of trainable parameters: {nbr_params}')
+
+ def __str__(self):
+ model_parameters = filter(lambda p: p.requires_grad, self.parameters())
+ nbr_params = int(sum([np.prod(p.size()) for p in model_parameters]))
+ return f'\nNbr of trainable parameters: {nbr_params}'
+ #return super(BaseModel, self).__str__() + f'\nNbr of trainable parameters: {nbr_params}'
diff --git a/base/base_trainer.py b/base/base_trainer.py
new file mode 100644
index 0000000..9c61684
--- /dev/null
+++ b/base/base_trainer.py
@@ -0,0 +1,170 @@
+import os, json, math, logging, sys, datetime
+import torch
+from torch.utils import tensorboard
+from utils import helpers
+from utils import logger
+import utils.lr_scheduler
+from utils.htmlwriter import HTML
+
+def get_instance(module, name, config, *args):
+ return getattr(module, config[name]['type'])(*args, **config[name]['args'])
+
+class BaseTrainer:
+ def __init__(self, model, resume, config, iters_per_epoch, train_logger=None):
+ self.model = model
+ self.config = config
+
+ self.train_logger = train_logger
+ self.logger = logging.getLogger(self.__class__.__name__)
+ self.do_validation = self.config['trainer']['val']
+ self.start_epoch = 1
+ self.improved = False
+
+ # SETTING THE DEVICE
+ self.device, availble_gpus = self._get_available_devices(self.config['n_gpu'])
+ self.model = torch.nn.DataParallel(self.model, device_ids=availble_gpus)
+ self.model.to(self.device)
+
+ # CONFIGS
+ cfg_trainer = self.config['trainer']
+ self.epochs = cfg_trainer['epochs']
+ self.save_period = cfg_trainer['save_period']
+
+ # OPTIMIZER
+ trainable_params = [{'params': filter(lambda p:p.requires_grad, self.model.module.get_other_params())},
+ {'params': filter(lambda p:p.requires_grad, self.model.module.get_backbone_params()),
+ 'lr': config['optimizer']['args']['lr'] / 10}]
+
+ self.optimizer = get_instance(torch.optim, 'optimizer', config, trainable_params)
+ model_params = sum([i.shape.numel() for i in list(model.parameters())])
+ opt_params = sum([i.shape.numel() for j in self.optimizer.param_groups for i in j['params']])
+ assert opt_params == model_params, 'some params are missing in the opt'
+
+ self.lr_scheduler = getattr(utils.lr_scheduler, config['lr_scheduler'])(optimizer=self.optimizer, num_epochs=self.epochs,
+ iters_per_epoch=iters_per_epoch)
+
+ # MONITORING
+ self.monitor = cfg_trainer.get('monitor', 'off')
+ if self.monitor == 'off':
+ self.mnt_mode = 'off'
+ self.mnt_best = 0
+ else:
+ self.mnt_mode, self.mnt_metric = self.monitor.split()
+ assert self.mnt_mode in ['min', 'max']
+ self.mnt_best = -math.inf if self.mnt_mode == 'max' else math.inf
+ self.early_stoping = cfg_trainer.get('early_stop', math.inf)
+
+ # CHECKPOINTS & TENSOBOARD
+ date_time = datetime.datetime.now().strftime('%m-%d_%H-%M')
+ run_name = config['experim_name']
+ self.checkpoint_dir = os.path.join(cfg_trainer['save_dir'], run_name)
+ helpers.dir_exists(self.checkpoint_dir)
+ config_save_path = os.path.join(self.checkpoint_dir, 'config.json')
+ with open(config_save_path, 'w') as handle:
+ json.dump(self.config, handle, indent=4, sort_keys=True)
+
+ writer_dir = os.path.join(cfg_trainer['log_dir'], run_name)
+ self.writer = tensorboard.SummaryWriter(writer_dir)
+ self.html_results = HTML(web_dir=config['trainer']['save_dir'], exp_name=config['experim_name'],
+ save_name=config['experim_name'], config=config, resume=resume)
+
+ if resume: self._resume_checkpoint(resume)
+
+ def _get_available_devices(self, n_gpu):
+ sys_gpu = torch.cuda.device_count()
+ if sys_gpu == 0:
+ self.logger.warning('No GPUs detected, using the CPU')
+ n_gpu = 0
+ elif n_gpu > sys_gpu:
+ self.logger.warning(f'Nbr of GPU requested is {n_gpu} but only {sys_gpu} are available')
+ n_gpu = sys_gpu
+
+ device = torch.device('cuda:0' if n_gpu > 0 else 'cpu')
+ self.logger.info(f'Detected GPUs: {sys_gpu} Requested: {n_gpu}')
+ available_gpus = list(range(n_gpu))
+ return device, available_gpus
+
+
+
+ def train(self):
+ for epoch in range(self.start_epoch, self.epochs+1):
+ results = self._train_epoch(epoch)
+ if self.do_validation and epoch % self.config['trainer']['val_per_epochs'] == 0:
+ results = self._valid_epoch(epoch)
+ self.logger.info('\n\n')
+ for k, v in results.items():
+ self.logger.info(f' {str(k):15s}: {v}')
+
+ if self.train_logger is not None:
+ log = {'epoch' : epoch, **results}
+ self.train_logger.add_entry(log)
+
+ # CHECKING IF THIS IS THE BEST MODEL (ONLY FOR VAL)
+ if self.mnt_mode != 'off' and epoch % self.config['trainer']['val_per_epochs'] == 0:
+ try:
+ if self.mnt_mode == 'min': self.improved = (log[self.mnt_metric] < self.mnt_best)
+ else: self.improved = (log[self.mnt_metric] > self.mnt_best)
+ except KeyError:
+ self.logger.warning(f'The metrics being tracked ({self.mnt_metric}) has not been calculated. Training stops.')
+ break
+
+ if self.improved:
+ self.mnt_best = log[self.mnt_metric]
+ self.not_improved_count = 0
+ else:
+ self.not_improved_count += 1
+
+ if self.not_improved_count > self.early_stoping:
+ self.logger.info(f'\nPerformance didn\'t improve for {self.early_stoping} epochs')
+ self.logger.warning('Training Stoped')
+ break
+
+ # SAVE CHECKPOINT
+ if epoch % self.save_period == 0:
+ self._save_checkpoint(epoch, save_best=self.improved)
+ self.html_results.save()
+
+
+ def _save_checkpoint(self, epoch, save_best=False):
+ state = {
+ 'arch': type(self.model).__name__,
+ 'epoch': epoch,
+ 'state_dict': self.model.state_dict(),
+ 'monitor_best': self.mnt_best,
+ 'config': self.config
+ }
+
+ filename = os.path.join(self.checkpoint_dir, f'checkpoint.pth')
+ self.logger.info(f'\nSaving a checkpoint: {filename} ...')
+ torch.save(state, filename)
+
+ if save_best:
+ filename = os.path.join(self.checkpoint_dir, f'best_model.pth')
+ torch.save(state, filename)
+ self.logger.info("Saving current best: best_model.pth")
+
+ def _resume_checkpoint(self, resume_path):
+ self.logger.info(f'Loading checkpoint : {resume_path}')
+ checkpoint = torch.load(resume_path)
+ self.start_epoch = checkpoint['epoch'] + 1
+ self.mnt_best = checkpoint['monitor_best']
+ self.not_improved_count = 0
+
+ try:
+ self.model.load_state_dict(checkpoint['state_dict'])
+ except Exception as e:
+ print(f'Error when loading: {e}')
+ self.model.load_state_dict(checkpoint['state_dict'], strict=False)
+
+ if "logger" in checkpoint.keys():
+ self.train_logger = checkpoint['logger']
+ self.logger.info(f'Checkpoint <