Parameter Adjustmens and Ensemble Model Implementation

audio_toolset/audio_augmentation.py

@@ -0,0 +1,58 @@
+import numpy as np
+
+
+class NoiseInjection(object):
+
+    def __init__(self, noise_factor: float, sigma=0.5, mu=0.5):
+        assert noise_factor > 0, f'max_shift_ratio has to be greater then 0, but was: {noise_factor}.'
+        self.mu = mu
+        self.sigma = sigma
+        self.noise_factor = noise_factor
+
+    def __call__(self, x: np.ndarray):
+        noise = np.random.normal(loc=self.mu, scale=self.sigma, size=x.shape)
+        augmented_data = x + self.noise_factor * noise
+        # Cast back to same data type
+        augmented_data = augmented_data.astype(x.dtype)
+        return augmented_data
+
+
+class LoudnessManipulator(object):
+
+    def __init__(self, max_factor: float):
+        assert 1 > max_factor > 0, f'max_shift_ratio has to be between [0,1], but was: {max_factor}.'
+
+        self.max_factor = max_factor
+
+    def __call__(self, x: np.ndarray):
+        augmented_data = x + x * (np.random.random() * self.max_factor)
+        # Cast back to same data type
+        augmented_data = augmented_data.astype(x.dtype)
+        return augmented_data
+
+
+class ShiftTime(object):
+
+    valid_shifts = ['right', 'left', 'any']
+
+    def __init__(self, max_shift_ratio: float, shift_direction: str = 'any'):
+        assert 1 > max_shift_ratio > 0, f'max_shift_ratio has to be between [0,1], but was: {max_shift_ratio}.'
+        assert shift_direction.lower() in self.valid_shifts, f'shift_direction has to be one of: {self.valid_shifts}'
+        self.max_shift_ratio = max_shift_ratio
+        self.shift_direction = shift_direction.lower()
+
+    def __call__(self, x: np.ndarray):
+        shift = np.random.randint(max(int(self.max_shift_ratio * x.shape[-1]), 1))
+        if self.shift_direction == 'right':
+            shift = -1 * shift
+        elif self.shift_direction == 'any':
+            direction = np.random.choice([1, -1], 1)
+            shift = direction * shift
+        augmented_data = np.roll(x, shift)
+        # Set to silence for heading/ tailing
+        shift = int(shift)
+        if shift > 0:
+            augmented_data[:shift] = 0
+        else:
+            augmented_data[shift:] = 0
+        return augmented_data

audio_toolset/audio_io.py

@@ -1,5 +1,4 @@
 import librosa
-import torch
 from scipy.signal import butter, lfilter
 
 import numpy as np
@@ -36,31 +35,34 @@ class MFCC(object):
 
 class NormalizeLocal(object):
     def __init__(self):
-        self.cache: torch.Tensor
+        self.cache: np.ndarray
         pass
 
-    def __call__(self, x: torch.Tensor):
+    def __call__(self, x: np.ndarray):
         mean = x.mean()
-        std = x.std()
+        std = x.std() + 0.0001
 
-        x = x.__sub__(mean).__div__(std)
-        x[torch.isnan(x)] = 0
-        x[torch.isinf(x)] = 0
+        # Pytorch Version:
+        # x = x.__sub__(mean).__div__(std)
+        # Numpy Version
+        x = (x - mean) / std
+        x[np.isnan(x)] = 0
+        x[np.isinf(x)] = 0
         return x
 
 
 class NormalizeMelband(object):
     def __init__(self):
-        self.cache: torch.Tensor
+        self.cache: np.ndarray
         pass
 
-    def __call__(self, x: torch.Tensor):
+    def __call__(self, x: np.ndarray):
         mean = x.mean(-1).unsqueeze(-1)
         std = x.std(-1).unsqueeze(-1)
 
         x = x.__sub__(mean).__div__(std)
-        x[torch.isnan(x)] = 0
-        x[torch.isinf(x)] = 0
+        x[np.isnan(x)] = 0
+        x[np.isinf(x)] = 0
         return x
 
 

modules/utils.py

@@ -24,17 +24,6 @@ class F_x(object):
 
 class ShapeMixin:
 
-    @property
-    def shape(self):
-        x = torch.randn(self.in_shape).unsqueeze(0)
-        output: torch.Tensor = self(x)
-        return output.shape[1:]
-
-
-# Utility - Modules
-###################
-class Flatten(nn.Module):
-
     @property
     def shape(self):
         try:
@@ -45,6 +34,11 @@ class Flatten(nn.Module):
             print(e)
             return -1
 
+
+# Utility - Modules
+###################
+class Flatten(ShapeMixin, nn.Module):
+
     def __init__(self, in_shape, to=-1):
         assert isinstance(to, int) or isinstance(to, tuple)
         super(Flatten, self).__init__()
@@ -172,29 +166,6 @@ class LightningBaseModule(pl.LightningModule, ABC):
         self.apply(weight_initializer)
 
 
-class BaseModuleMixin_Dataloaders(ABC):
-
-    # Dataloaders
-    # ================================================================================
-    # Train Dataloader
-    def train_dataloader(self):
-        return DataLoader(dataset=self.dataset.train_dataset, shuffle=True,
-                          batch_size=self.params.batch_size,
-                          num_workers=self.params.worker)
-
-    # Test Dataloader
-    def test_dataloader(self):
-        return DataLoader(dataset=self.dataset.test_dataset, shuffle=True,
-                          batch_size=self.params.batch_size,
-                          num_workers=self.params.worker)
-
-    # Validation Dataloader
-    def val_dataloader(self):
-        return DataLoader(dataset=self.dataset.val_dataset, shuffle=True,
-                          batch_size=self.params.batch_size,
-                          num_workers=self.params.worker)
-
-
 class FilterLayer(nn.Module):
 
     def __init__(self):
@@ -253,7 +224,7 @@ class HorizontalSplitter(nn.Module):
         self.in_shape = in_shape
 
         self.channel, self.height, self.width = self.in_shape
-        self.new_height = (self.height // self.n) + 1 if self.height % self.n != 0 else 0
+        self.new_height = (self.height // self.n) + (1 if self.height % self.n != 0 else 0)
 
         self.shape = (self.channel, self.new_height, self.width)
         self.autopad = AutoPadToShape(self.shape)

utils/config.py

@@ -7,8 +7,7 @@ from argparse import Namespace, ArgumentParser
 from collections import defaultdict
 from configparser import ConfigParser
 from pathlib import Path
-
-from ml_lib.utils.model_io import ModelParameters
+import hashlib
 
 
 def is_jsonable(x):
@@ -22,6 +21,30 @@ def is_jsonable(x):
 
 class Config(ConfigParser, ABC):
 
+    @property
+    def name(self):
+        short_name = "".join(c for c in self.model.type if c.isupper())
+        return f'{short_name}_{self.fingerprint}'
+
+    @property
+    def version(self):
+        return f'version_{self.main.seed}'
+
+    @property
+    def exp_path(self):
+        return Path(self.train.outpath) / self.model.type / self.name
+
+    @property
+    def fingerprint(self):
+        h = hashlib.md5()
+        params = deepcopy(self.as_dict)
+        del params['model']['type']
+        del params['data']['worker']
+        del params['main']
+        h.update(str(params).encode())
+        fingerprint = h.hexdigest()
+        return fingerprint
+
     @property
     def _model_weight_init(self):
         mod = __import__('torch.nn.init', fromlist=[self.model.weight_init])
@@ -33,8 +56,8 @@ class Config(ConfigParser, ABC):
         This is function is supposed to return a dict, which holds a mapping from string model names to model classes
 
         Example:
-        from models.binary_classifier import BinaryClassifier
-        return dict(BinaryClassifier=BinaryClassifier,
+        from models.binary_classifier import ConvClassifier
+        return dict(ConvClassifier=ConvClassifier,
                     )
         :return:
         """
@@ -46,8 +69,7 @@ class Config(ConfigParser, ABC):
         try:
             return self._model_map[self.model.type]
         except KeyError:
-            raise KeyError(rf'The model alias you provided ("{self.get("model", "type")}") does not exist! \n'
-                           f'Try one of these:\n{list(self._model_map.keys())}')
+            raise KeyError(rf'The model alias you provided ("{self.get("model", "type")}") does not exist! Try one of these: {list(self._model_map.keys())}')
 
     # TODO: Do this programmatically; This did not work:
     # Initialize Default Sections as Property
@@ -83,6 +105,7 @@ class Config(ConfigParser, ABC):
         params.update(self.train.__dict__)
         assert all(key not in list(params.keys()) for key in self.data.__dict__)
         params.update(self.data.__dict__)
+        params.update(exp_path=str(self.exp_path), exp_fingerprint=str(self.fingerprint))
         return params
 
     @property
@@ -134,7 +157,6 @@ class Config(ConfigParser, ABC):
         new_config.read_dict(sorted_dict)
         return new_config
 
-
     def build_model(self):
         return self.model_class(self.model_paramters)
 

utils/logging.py

@@ -25,7 +25,7 @@ class Logger(LightningLoggerBase, ABC):
 
     @property
     def name(self):
-        return self.config.model.type
+        return self.config.name
 
     @property
     def project_name(self):
@@ -37,7 +37,11 @@ class Logger(LightningLoggerBase, ABC):
 
     @property
     def outpath(self):
-        raise NotImplementedError
+        return Path(self.config.train.outpath) / self.config.model.type
+
+    @property
+    def exp_path(self):
+        return Path(self.outpath) / self.name
 
     def __init__(self, config: Config):
         """
@@ -58,10 +62,12 @@ class Logger(LightningLoggerBase, ABC):
         self._testtube_kwargs = dict(save_dir=self.outpath, version=self.version, name=self.name)
         self._neptune_kwargs = dict(offline_mode=self.debug,
                                     api_key=self.config.project.neptune_key,
+                                    experiment_name=self.name,
                                     project_name=self.project_name,
                                     upload_source_files=list())
         self.neptunelogger = NeptuneLogger(**self._neptune_kwargs)
         self.testtubelogger = TestTubeLogger(**self._testtube_kwargs)
+        self.log_config_as_ini()
 
     def log_hyperparams(self, params):
         self.neptunelogger.log_hyperparams(params)
@@ -80,6 +86,10 @@ class Logger(LightningLoggerBase, ABC):
     def log_config_as_ini(self):
         self.config.write(self.log_dir / 'config.ini')
 
+    def log_text(self, name, text, step_nb=0, **kwargs):
+        # TODO Implement Offline variant.
+        self.neptunelogger.log_text(name, text, step_nb)
+
     def log_metric(self, metric_name, metric_value, **kwargs):
         self.testtubelogger.log_metrics(dict(metric_name=metric_value))
         self.neptunelogger.log_metric(metric_name, metric_value, **kwargs)
@@ -97,7 +107,6 @@ class Logger(LightningLoggerBase, ABC):
     def finalize(self, status):
         self.testtubelogger.finalize(status)
         self.neptunelogger.finalize(status)
-        self.log_config_as_ini()
 
     def __enter__(self):
         return self

utils/transforms.py

@@ -0,0 +1,8 @@
+from torchvision.transforms import ToTensor as TorchvisionToTensor
+
+
+class ToTensor(TorchvisionToTensor):
+
+    def __call__(self, pic):
+        tensor = super(ToTensor, self).__call__(pic).float()
+        return tensor