f

model_training_nnn_tpu/dataset_tf.py

@@ -633,12 +633,12 @@ class DataAugmentationTF:
         reshaped_inputs = tf.expand_dims(inputs, axis=2)
 
         # Execute depthwise convolution
-        # This is a single, efficient operation replacing the original Python for-loop
+        # Use VALID padding to ensure output length <= input length
         smoothed = tf.nn.depthwise_conv2d(
             reshaped_inputs,
             kernel,
             strides=[1, 1, 1, 1],
-            padding='SAME'
+            padding='VALID'  # Changed from 'SAME' to 'VALID' to prevent length increase
         )
 
         # Remove the dummy width dimension to restore original shape
@@ -963,15 +963,15 @@ def analyze_dataset_shapes(dataset_tf: BrainToTextDatasetTF,
         'n_features': dataset_tf.feature_dim
     }
 
-    # 5. 添加适当的安全边际 - 现在分析已包含数据增强效果
+    # 5. 添加适当的安全边际 - Gaussian平滑现在使用VALID padding，不会增加长度
     if transform_args:
-        # 已应用数据增强分析：只需要很小的边际应对随机性
+        # 已应用数据增强分析：VALID padding确保长度不增加，只需小边际
         if sample_size == -1:
-            safety_margin = 1.05  # 5% buffer for augmentation randomness (full data)
-            margin_reason = "small buffer for augmentation randomness after full analysis"
+            safety_margin = 1.05  # 5% buffer for minor augmentation effects (full data)
+            margin_reason = "small buffer after full analysis with length-preserving augmentation"
         else:
-            safety_margin = 1.15  # 15% buffer for augmentation randomness + sampling
-            margin_reason = f"buffer for augmentation randomness + sampling {sample_size} trials"
+            safety_margin = 1.15  # 15% buffer for augmentation + sampling
+            margin_reason = f"buffer for augmentation + sampling {sample_size} trials"
     else:
         # 未应用数据增强：使用原始逻辑
         if sample_size == -1: