Merge pull request #109 from mjq2020/dev

Optimizing Dataset Sampling

Author: mjq2020

configs/accelerometer/3axes_accelerometer_62.5Hz_1s_classify.py

@@ -4,20 +4,19 @@
 
 num_classes = 3
 num_axes = 3
-frequency = 62.5
-window = 1000
+window_size = 30
+stride = 20
 
 model = dict(
     type='AccelerometerClassifier',
     backbone=dict(
         type='AxesNet',
         num_axes=num_axes,
-        frequency=frequency,
-        window=window,
+        window_size=window_size,
         num_classes=num_classes,
     ),
     head=dict(
-        type='edgelab.ClsHead',
+        type='edgelab.AxesClsHead',
         loss=dict(type='mmcls.CrossEntropyLoss', loss_weight=1.0),
         topk=(1, 5),
     ),
@@ -29,15 +28,15 @@
 batch_size = 1
 workers = 1
 
-shape = num_classes * int(62.5 * 1000 / 1000)
+shape = [1, num_axes * window_size]
 
 train_pipeline = [
-    dict(type='edgelab.LoadSensorFromFile'),
+    # dict(type='edgelab.LoadSensorFromFile'),
     dict(type='edgelab.PackSensorInputs'),
 ]
 
 test_pipeline = [
-    dict(type='edgelab.LoadSensorFromFile'),
+    # dict(type='edgelab.LoadSensorFromFile'),
     dict(type='edgelab.PackSensorInputs'),
 ]
 
@@ -49,6 +48,8 @@
         data_root=data_root,
         data_prefix='training',
         ann_file='info.labels',
+        window_size=window_size,
+        stride=stride,
         pipeline=train_pipeline,
     ),
     sampler=dict(type='DefaultSampler', shuffle=True),
@@ -61,6 +62,8 @@
     dataset=dict(
         type=dataset_type,
         data_root=data_root,
+        window_size=window_size,
+        stride=stride,
         data_prefix='testing',
         ann_file='info.labels',
         pipeline=test_pipeline,

edgelab/datasets/sensordataset.py

@@ -1,7 +1,10 @@
 import json
 import os
-from typing import Optional, Union
+import os.path as osp
+from typing import List, Optional, Union
 
+import cbor
+import numpy as np
 from mmcls.datasets import CustomDataset
 
 from edgelab.registry import DATASETS
@@ -17,6 +20,11 @@ def __init__(
         metainfo: Optional[dict] = None,
         data_root: str = '',
         data_prefix: Union[str, dict] = '',
+        window_size: int = 80,
+        stride: int = 30,
+        retention: float = 0.8,
+        # source: str = 'EI',
+        flatten: bool = True,
         multi_label: bool = False,
         **kwargs,
     ):
@@ -26,6 +34,12 @@ def __init__(
         self.data_root = data_root
         self.ann_file = ann_file
         self.data_prefix = data_prefix
+        self.window_size = window_size
+        self.stride = stride
+        self.retention = retention
+        self.flatten = flatten
+
+        self.data_dir = osp.join(self.data_root, self.data_prefix)
 
         self.info_lables = json.load(open(os.path.join(self.data_root, self.data_prefix, self.ann_file)))
 
@@ -58,7 +72,6 @@ def _find_samples(self):
                     gt_label = j
                     break
             samples.append((filename, gt_label))
-        print(samples)
         return samples
 
     def load_data_list(self):
@@ -75,12 +88,56 @@ def load_data_list(self):
 
         data_list = []
         for filename, gt_label in samples:
-            img_path = os.path.join(self.img_prefix, filename)
-            info = {'file_path': img_path, 'gt_label': int(gt_label)}
-            data_list.append(info)
+            ann_path = os.path.join(self.data_dir, filename)
+            data_list.extend(
+                [{'data': np.asanyarray([data]), 'gt_label': int(gt_label)} for data in self.read_split_data(ann_path)]
+            )
 
         return data_list
 
+    def read_split_data(self, file_path: str) -> List:
+        if file_path.lower().endswith('.cbor'):
+            with open(file_path, 'rb') as f:
+                data = cbor.loads(f.read())
+        elif file_path.lower().endswith('.json'):
+            with open(file_path, 'r') as f:
+                data = json.load(f)
+
+        values = np.asanyarray(data['payload']['values'])
+
+        result = []
+        values_len = len(values)
+        if values_len <= self.window_size:
+            result.append(self.pad_data(values, self.window_size).transpose(0, 1).reshape(-1))
+        else:
+            indexes = range(0, values_len, self.stride)
+            for i in indexes:
+                if (values_len - i + 1) < self.window_size or i == indexes[-1]:
+                    if self.retention * self.window_size < (values_len - i + 1):
+                        data = self.pad_data(values[i:], self.window_size)
+                    else:
+                        continue
+                else:
+                    end = i + self.window_size
+                    if end >= values_len:
+                        if self.retention * self.window_size < (values_len - i + 1):
+                            data = self.pad_data(values[i:], self.window_size)
+                        else:
+                            continue
+                    else:
+                        data = values[i:end]
+                if self.flatten:
+                    data = data.transpose(0, 1).reshape(-1)
+                result.append(data)
+        return result
+
+    def pad_data(self, data: np.asanyarray, total_len: int, mode='constant', pad_val=0) -> np.array:
+        pad_len = total_len - len(data)
+        front = pad_len // 2
+        arfter = pad_len - front
+        data = np.pad(data, ((front, arfter), (0, 0)), mode=mode, constant_values=pad_val)
+        return data
+
     def is_valid_file(self, filename: str) -> bool:
         """Check if a file is a valid sample."""
         return True

edgelab/models/backbones/AxesNet.py

@@ -6,12 +6,10 @@
 
 @MODELS.register_module()
 class AxesNet(nn.Module):
-    def __init__(
-        self, num_axes=3, frequency=62.5, window=1000, num_classes=-1  # axes number  # sample frequency  # window size
-    ):
+    def __init__(self, num_axes=3, window_size=80, num_classes=-1):  # axes number  # sample frequency  # window size
         super().__init__()
         self.num_classes = num_classes
-        self.intput_feature = num_axes * int(frequency * window / 1000)
+        self.intput_feature = num_axes * window_size
         liner_feature = self.liner_feature_fit()
         self.fc1 = nn.Linear(in_features=self.intput_feature, out_features=liner_feature, bias=True)
         self.fc2 = nn.Linear(in_features=liner_feature, out_features=liner_feature, bias=True)
@@ -23,6 +21,7 @@ def liner_feature_fit(self):
         return (int(self.intput_feature / 1024) + 1) * 256
 
     def forward(self, x):
+        x = x[0] if isinstance(x, list) else x
         x = F.relu(self.fc1(x))
         x = F.relu(self.fc2(x))
 

requirements/base.txt

@@ -1,5 +1,8 @@
 # common
 albumentations>=1.3.0
+
+# sensor
+cbor
 numpy>=1.23.5
 
 # vision