MythEclipse/zeavis-edu
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

a

Browse Source
This commit is contained in:
MythEclipse
2026-06-12 15:24:49 +00:00
parent 731b5728f7
commit 2e925a4188
2 changed files with 16 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Machine_Learning/notebook.ipynb
+4 -109
View File
File diff suppressed because one or more lines are too long
Machine_Learning/save_model.py
+12 -6
View File
@@ -11,7 +11,7 @@ log = logging.getLogger(__name__)
def cbam_block(x, ratio=8, name="cbam"): def cbam_block(x, ratio=8, name="cbam"):
"""Convolutional Block Attention Module — lightweight foreground attention.""" """Convolutional Block Attention Module — lightweight foreground attention."""
channels = tf.shape(x)[-1] channels = x.shape[-1]
# Channel attention # Channel attention
avg_pool = layers.GlobalAveragePooling2D()(x) avg_pool = layers.GlobalAveragePooling2D()(x)
@@ -34,17 +34,23 @@ def cbam_block(x, ratio=8, name="cbam"):
return x return x
def build_clean_model(num_classes, img_size=(224, 224)): def build_clean_model(num_classes, target_size=(224, 224)):
"""Build the production architecture: CBAM + lightweight head, outputting raw logits.""" """Build the production architecture: CBAM + lightweight head, outputting raw logits.
Mirrors the notebook's build_model() exactly so set_weights() maps correctly.
"""
base_model = tf.keras.applications.EfficientNetV2B0( base_model = tf.keras.applications.EfficientNetV2B0(
input_shape=img_size + (3,), input_shape=target_size + (3,),
include_top=False, include_top=False,
weights=None, weights=None,
) )
base_model.trainable = False base_model.trainable = False
inputs = tf.keras.Input(shape=img_size + (3,), name="input") inputs = tf.keras.Input(shape=(None, None, 3), name="input")
x = base_model(inputs, training=False) x = layers.Resizing(target_size[0], target_size[1], interpolation="bilinear",
name="resize_input")(inputs)
x = layers.GaussianNoise(0.05, name="gauss_noise")(x)
x = base_model(x, training=False)
# CBAM attention — focus on leaf regions, ignore background # CBAM attention — focus on leaf regions, ignore background
x = cbam_block(x, ratio=8, name="cbam") x = cbam_block(x, ratio=8, name="cbam")
x = layers.GlobalAveragePooling2D(name="gap")(x) x = layers.GlobalAveragePooling2D(name="gap")(x)