Version 2.2.5

README.md

@@ -64,7 +64,7 @@ AudioSet (ADS) and Primate Vocalize Corpus (PVC) cannot be installed automatical
 [comment]: <> (TODO : For PVC install !)
 
 ## Usage
-The main scripts available are in directory ```standalone``` :
+The main scripts available are in ```standalone``` directory :
 ```
 standalone
 ├── deep_co_training.py
@@ -82,27 +82,49 @@ The code use Hydra for parsing args. The syntax of setting an argument is "name=
 
 Example 1 : MixMatch on ESC10
 ```bash
-python mixmatch.py dataset=esc10
+python mixmatch.py data=esc10
 ```
 
 Example 2 : Supervised+Weak on GSC
 ```bash
-python supervised.py dataset=gsc expt.augm_train=weak bsize=256 epochs=300
+python supervised.py data=gsc expt.augm_train=weak bsize=256 epochs=300
 ```
 
 Example 3 : FixMatch+MixUp on UBS8K
 ```bash
-python fixmatch.py dataset=ubs8K dataset.root="../data/UBS8K" expt=fixmatch_mixup bsize_s=128 bsize_u=128 epochs=300
+python fixmatch.py data=ubs8K expt=fixmatch_mixup bsize_s=128 bsize_u=128 epochs=300
 ```
-(note: default folds used are in "config/dataset/ubs8k.yaml")
+(note: default folds used for UBS8K are in "config/data/ubs8k.yaml")
+
+Example 4 : ReMixMatch on CIFAR-10
+```bash
+python remixmatch.py data=cifar10 model.n_input_channels=3
+```
+
+## List of main arguments
+
+| Name | Description | Values | Default |
+| --- | --- | --- | --- |
+| data | Dataset used | ads, cifar10, esc10, fsd50k, gsc, pvc, ubs8k | esc10 |
+| expt | Training method (experiment) used | *(depends of the python script, see the filenames in config/expt/ folder)* | *(depends of the python script)* |
+| model | Pytorch model to use | mobilenetv1, mobilenetv2, vgg, wideresnet28 | wideresnet28 |
+| optim | Optimizer used | adam, sgd | adam |
+| sched | Learning rate scheduler | cosine, softcosine, none | softcosine |
+| epochs | Number of training epochs | int | 1 |
+| bsize | Batch size in SUP methods | int | 60 |
+| ratio | Ratio of the training data used in SUP methods | float in [0, 1] | 1.0 |
+| bsize_s | Batch size of supervised part in SSL methods | int | 30 |
+| bsize_u | Batch size of unsupervised part in SSL methods | int | 30 |
+| ratio_s | Ratio of the supervised training data used in SSL methods | float in [0, 1] | 0.1 |
+| ratio_u | Ratio of the unsupervised training data used in SSL methods | float in [0, 1] | 0.9 |
+
 
 ## SSLH Package overview
 ```
 sslh
 ├── callbacks
 ├── datamodules
-│     ├── fully_supervised
-│     ├── partial_supervised
+│     ├── supervised
 │     └── semi_supervised
 ├── datasets
 ├── expt
@@ -129,7 +151,7 @@ This repository has been created by Etienne Labbé (Labbeti on Github).
 
 It contains also some code from the following authors :
 - Léo Cancès (leocances on github)
-  - For AudioSet, ESC10, GSC, PVC and UBS8K datasets.
+  - For AudioSet, ESC10, GSC, PVC and UBS8K datasets base code.
 - Qiuqiang Kong (qiuqiangkong on Github)
   - For MobileNetV1 & V2 model implementation from [PANN](https://github.com/qiuqiangkong/audioset_tagging_cnn).
 
@@ -161,5 +183,6 @@ It contains also some code from the following authors :
 | _s | Supervised |
 | sched | Scheduler |
 | SSL | Semi-Supervised Learning |
+| SUP | Supervised Learning |
 | _u | Unsupervised |
 | UBS8K | UrbanSound8K dataset |

config/data/gsc.yaml

@@ -16,4 +16,4 @@ download: true
 transform:
   n_mels: 64
   hop_length: 512
-  n_fft: 2048
+  n_fft: 2048

config/sched/cosine.yaml

@@ -1,5 +1,6 @@
 # @package _group_
 
+# math.cos(7.0 / 16.0 * math.pi * min(step / n_steps, 1.0))
 name: "CosineScheduler"
 acronym: "cosine"
 on_epoch: ${data.log_on_epoch}

config/sched/softcosine.yaml

@@ -1,5 +1,6 @@
 # @package _group_
 
+# 0.5 * (1.0 + math.cos(math.pi * min(step / n_steps, 1.0)))
 name: "SoftCosineScheduler"
 acronym: "softcosine"
 on_epoch: ${data.log_on_epoch}

post_setup.sh

@@ -6,7 +6,6 @@ for dir in logs boards
 do
   for dataset in ADS CIFAR10 ESC10 FSD50K GSC PVC UBS8K
   do
-    echo "Building directory '$root/$dir/$dataset' ..."
     mkdir -p $root/$dir/$dataset
   done
 done

sslh/__init__.py

@@ -6,4 +6,4 @@
 __email__ = 'etienne.labbe31@gmail.com'
 __maintainer__ = 'Etienne Labbé'
 __status__ = 'Released'
-__version__ = '2.2.4'
+__version__ = '2.2.5'

sslh/datasets/esc50.py

@@ -73,7 +73,6 @@ def __init__(
 		download: bool = False,
 		transform: Module = None
 	) -> None:
-
 		super().__init__()
 
 		self.root = root
@@ -160,8 +159,10 @@ def download(self) -> None:
 			print('Dataset already downloaded and verified.')
 
 		else:
+			if not os.path.isdir(self.root):
+				os.makedirs(self.root)
+
 			archive_path = os.path.join(self.root, FOLDER_IN_ARCHIVE + '.zip')
-
 			download_url(self.url, self.root)
 			extract_archive(archive_path, self.root)
 

sslh/datasets/ubs8k.py

@@ -9,23 +9,24 @@
 
 
 class URBANSOUND8K(Dataset):
-	N_CLASSES = 10
+	ROOT_DNAME: str = 'UrbanSound8K'
+	N_CLASSES: int = 10
 
-	def __init__(self, root: str, folds: List[int], resample_sr: int = 22050):
+	def __init__(self, root: str, folds: List[int], resample_sr: int = 22050) -> None:
 		"""
 			UBS8K dataset.
 
 			:param root: The directory path to the dataset root.
 			:param folds: The folds to use.
-			:param resample_sr:
+			:param resample_sr: The resample rate of the waveform.
 		"""
 		super().__init__()
 		self.root = root
 		self.folds = list(folds)
 		self.resample_sr = resample_sr
 
 		self.meta = self._load_metadata()
-		self.wav_dir = os.path.join(root, 'UrbanSound8K', 'audio')
+		self.wav_dir = os.path.join(root, self.ROOT_DNAME, 'audio')
 
 	def __getitem__(self, idx) -> Tuple[Tensor, int]:
 		filename = self.meta['filename'][idx]
@@ -45,10 +46,10 @@ def __len__(self) -> int:
 		return len(self.meta['filename'])
 
 	def _load_metadata(self) -> Dict[str, list]:
-		csv_path = os.path.join(self.root, 'UrbanSound8K', 'metadata', 'UrbanSound8K.csv')
+		csv_path = os.path.join(self.root, self.ROOT_DNAME, 'metadata', 'UrbanSound8K.csv')
 
-		with open(csv_path) as f:
-			lines = f.read().splitlines()
+		with open(csv_path) as file:
+			lines = file.read().splitlines()
 			lines = lines[1:]  # remove the header
 
 		info = {'filename': [], 'fold': [], 'target': []}
@@ -69,7 +70,8 @@ def _to_mono(self, waveform: Tensor) -> Tensor:
 		if len(waveform.shape) == 2:
 			if waveform.shape[0] == 1:
 				return waveform
-			return waveform.mean(dim=0)
+			else:
+				return waveform.mean(dim=0)
 		else:
 			raise ValueError(
 				f'waveform tensor should be of shape (channels, time). currently is of shape {waveform.shape}'

sslh/transforms/cifar10.py

@@ -1,4 +1,6 @@
 
+import torch
+
 from torchvision.transforms import ToTensor, Normalize
 from typing import Callable, List, Optional
 
@@ -33,7 +35,7 @@ def get_transform_cifar10(
 
 
 def get_target_transform_cifar10(smooth: Optional[float] = None) -> Callable:
-	return OneHot(N_CLASSES, smooth)
+	return OneHot(N_CLASSES, smooth, dtype=torch.float)
 
 
 def get_self_transform_cifar10(**kwargs) -> Callable:

sslh/utils/average_pred.py

@@ -4,19 +4,19 @@
 
 
 class AveragePred:
-	def __init__(self, history: int = 128):
+	def __init__(self, history: int = 128) -> None:
 		self.history = history
 		self.preds = []
 		self.cur_idx = 0
 
-	def reset(self):
+	def reset(self) -> None:
 		"""
 			Reset the history.
 		"""
 		self.preds = []
 		self.cur_idx = 0
 
-	def add_pred(self, pred: Tensor):
+	def add_pred(self, pred: Tensor) -> None:
 		"""
 			Add a batch of predictions of shape (bsize, n_classes) to the history for computing the classes distributions.
 		"""