segmodels-keras 0.1.0.dev5__tar.gz → 0.1.0.dev6__tar.gz

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: segmodels_keras
-Version: 0.1.0.dev5
+Version: 0.1.0.dev6
 Summary: Image segmentation models with pre-trained backbones with Keras.
 Home-page: https://github.com/orthoseg/segmodels_keras
 Author: Pieter Roggemans

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/segmodels_keras/_version.py

@@ -1,2 +1,2 @@
-__versioninfo__ = (0, 1, 0, "dev5")
+__versioninfo__ = (0, 1, 0, "dev6")
 __version__ = ".".join(map(str, __versioninfo__))

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/segmodels_keras/models/fpn.py

@@ -108,6 +108,7 @@ def build_fpn(
     aggregation="sum",
     dropout=None,
     weights_notop=None,
+    freeze_notop=False,
 ):
     input_ = backbone.input
     x = backbone.output
@@ -177,6 +178,8 @@ def build_fpn(
     if weights_notop is not None:
         model_notop = models.Model(input_, x)
         model_notop.load_weights(weights_notop)
+        if freeze_notop:
+            freeze_model(model_notop)
         x = model_notop.output
 
     # model head (define number of output classes)
@@ -208,6 +211,7 @@ def FPN(
     activation="softmax",
     weights=None,
     weights_notop=None,
+    freeze_notop=False,
     encoder_weights="imagenet",
     encoder_freeze=False,
     encoder_features="default",
@@ -230,6 +234,8 @@ def FPN(
         weights: optional, path to model weights to be loaded.
         weights_notop: optional, path to model weights without top (without segmentation
             head) to be loaded.
+        freeze_notop: if ``True``, set all layers of the model except the top as
+            non-trainable.
         activation: name of one of ``keras.activations`` for last model layer
             (e.g. ``sigmoid``, ``softmax``, ``linear``).
         encoder_weights: one of ``None`` (random initialization), ``imagenet``
@@ -276,6 +282,7 @@ def FPN(
         classes=classes,
         aggregation=pyramid_aggregation,
         weights_notop=weights_notop,
+        freeze_notop=freeze_notop,
     )
 
     # lock encoder weights for fine-tuning

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/segmodels_keras/models/linknet.py

@@ -142,6 +142,7 @@ def build_linknet(
     activation="sigmoid",
     use_batchnorm=True,
     weights_notop=None,
+    freeze_notop=False,
 ):
     input_ = backbone.input
     x = backbone.output
@@ -174,6 +175,8 @@ def build_linknet(
     if weights_notop is not None:
         model_notop = models.Model(input_, x)
         model_notop.load_weights(weights_notop)
+        if freeze_notop:
+            freeze_model(model_notop)
         x = model_notop.output
 
     # model head (define number of output classes)
@@ -204,6 +207,7 @@ def Linknet(
     activation="sigmoid",
     weights=None,
     weights_notop=None,
+    freeze_notop=False,
     encoder_weights="imagenet",
     encoder_freeze=False,
     encoder_features="default",
@@ -230,6 +234,8 @@ def Linknet(
         weights: optional, path to model weights to be loaded.
         weights_notop: optional, path to model weights without top (without segmentation
             head) to be loaded.
+        freeze_notop: if ``True``, set all layers of the model except the top as
+            non-trainable.
         encoder_weights: one of ``None`` (random initialization), ``imagenet``
             (pre-training on ImageNet).
         encoder_freeze: if ``True`` set all layers of encoder (backbone model) as
@@ -285,6 +291,7 @@ def Linknet(
         n_upsample_blocks=len(decoder_filters),
         use_batchnorm=decoder_use_batchnorm,
         weights_notop=weights_notop,
+        freeze_notop=freeze_notop,
     )
 
     # lock encoder weights for fine-tuning

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/segmodels_keras/models/pspnet.py

@@ -125,6 +125,7 @@ def build_psp(
     activation="softmax",
     dropout=None,
     weights_notop=None,
+    freeze_notop=False,
 ):
     input_ = backbone.input
     x = (
@@ -152,6 +153,8 @@ def build_psp(
     if weights_notop is not None:
         model_notop = models.Model(input_, x)
         model_notop.load_weights(weights_notop)
+        if freeze_notop:
+            freeze_model(model_notop)
         x = model_notop.output
 
     # model head
@@ -185,6 +188,7 @@ def PSPNet(
     activation="softmax",
     weights=None,
     weights_notop=None,
+    freeze_notop=False,
     encoder_weights="imagenet",
     encoder_freeze=False,
     downsample_factor=8,
@@ -208,6 +212,8 @@ def PSPNet(
         weights: optional, path to model weights to be loaded.
         weights_notop: optional, path to model weights without top (without segmentation
             head) to be loaded.
+        freeze_notop: if ``True``, set all layers of the model except the top as
+            non-trainable.
         encoder_weights: one of ``None`` (random initialization), ``imagenet``
             (pre-training on ImageNet).
         encoder_freeze: if ``True`` set all layers of encoder (backbone model) as
@@ -261,6 +267,7 @@ def PSPNet(
         activation=activation,
         dropout=psp_dropout,
         weights_notop=weights_notop,
+        freeze_notop=freeze_notop,
     )
 
     # lock encoder weights for fine-tuning

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/segmodels_keras/models/unet.py

@@ -114,6 +114,7 @@ def build_unet(
     activation="sigmoid",
     use_batchnorm=True,
     weights_notop=None,
+    freeze_notop=False,
 ):
     input_ = backbone.input
     x = backbone.output
@@ -146,6 +147,8 @@ def build_unet(
     if weights_notop is not None:
         model_notop = models.Model(input_, x)
         model_notop.load_weights(weights_notop)
+        if freeze_notop:
+            freeze_model(model_notop)
         x = model_notop.output
 
     # model head (define number of output classes)
@@ -177,6 +180,7 @@ def Unet(
     activation="sigmoid",
     weights=None,
     weights_notop=None,
+    freeze_notop=False,
     encoder_weights="imagenet",
     encoder_freeze=False,
     encoder_features="default",
@@ -201,6 +205,8 @@ def Unet(
         weights: optional, path to model weights to be loaded.
         weights_notop: optional, path to model weights without top (without segmentation
             head) to be loaded.
+        freeze_notop: if ``True``, set all layers of the model except the top as
+            non-trainable.
         encoder_weights: one of ``None`` (random initialization), ``imagenet``
             (pre-training on ImageNet).
         encoder_freeze: if ``True`` set all layers of encoder (backbone model) as
@@ -256,6 +262,7 @@ def Unet(
         n_upsample_blocks=len(decoder_filters),
         use_batchnorm=decoder_use_batchnorm,
         weights_notop=weights_notop,
+        freeze_notop=freeze_notop,
     )
 
     # lock encoder weights for fine-tuning

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/segmodels_keras.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: segmodels_keras
-Version: 0.1.0.dev5
+Version: 0.1.0.dev6
 Summary: Image segmentation models with pre-trained backbones with Keras.
 Home-page: https://github.com/orthoseg/segmodels_keras
 Author: Pieter Roggemans

{segmodels_keras-0.1.0.dev5 → segmodels_keras-0.1.0.dev6}/tests/test_utils.py

@@ -127,13 +127,13 @@ def test_save_model_weights_notop(tmp_path, decoder):
     # Test loading the model weights without the top layers again
     assert path.exists()
     if decoder == "pspnet":
-        model = sm.PSPNet(backbone_name, weights_notop=path)
+        model = sm.PSPNet(backbone_name, weights_notop=path, freeze_notop=True)
     elif decoder == "fpn":
-        model = sm.FPN(backbone_name, weights_notop=path)
+        model = sm.FPN(backbone_name, weights_notop=path, freeze_notop=True)
     elif decoder == "linknet":
-        model = sm.Linknet(backbone_name, weights_notop=path)
+        model = sm.Linknet(backbone_name, weights_notop=path, freeze_notop=True)
     elif decoder == "unet":
-        model = sm.Unet(backbone_name, weights_notop=path)
+        model = sm.Unet(backbone_name, weights_notop=path, freeze_notop=True)
     else:
         raise ValueError(f"Incorrect decoder {decoder}")