Merge in some of the TF2 stuff

examples/deep_weeds.py

@@ -2,10 +2,10 @@
 Train an image classifier on the deep weeds dataset
 """
 
-from miso.training.parameters import MisoConfig
+from miso.training.parameters import MisoParameters
 from miso.training.trainer import train_image_classification_model
 
-tp = MisoConfig()
+tp = MisoParameters()
 
 # -----------------------------------------------------------------------------
 # Dataset
@@ -38,7 +38,7 @@
 # - resnet[18,34,50]
 # - vgg[16,19]
 # - efficientnetB[0-7]
-tp.cnn.type = "resnet50_tl"
+tp.cnn.id = "resnet50_tl"
 # Input image shape, set to None to use default size ([128, 128, 1] for custom, [224, 224, 3] for others)
 tp.cnn.img_shape = [224, 224, 3]
 # Input image colour space [greyscale/rgb]

examples/endless_forams.py

@@ -1,10 +1,10 @@
 import os
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
-from miso.training.parameters import MisoConfig
+from miso.training.parameters import MisoParameters
 from miso.training.trainer import train_image_classification_model
 
-tp = MisoConfig()
+tp = MisoParameters()
 
 # -----------------------------------------------------------------------------
 # Name
@@ -44,8 +44,8 @@
 # - resnet[18,34,50]
 # - vgg[16,19]
 # - efficientnetB[0-7]
-tp.cnn.type = r"base_cyclic"
-tp.cnn.type = r"resnet50_tl"
+tp.cnn.id = r"base_cyclic"
+tp.cnn.id = r"resnet50_tl"
 # Input image shape, set to None to use default size ([128, 128, 1] for custom, [224, 224, 3] for others)
 tp.cnn.img_shape = [128, 128, 1]
 tp.cnn.img_shape = [224, 224, 3]

examples/olzo_mini.py

@@ -7,13 +7,10 @@
 
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
-from miso.training.parameters import MisoConfig
+from miso.training.parameters import MisoParameters
 from miso.training.trainer import train_image_classification_model
 
-tp = MisoConfig()
-
-import tensorflow as tf
-print(tf.config.list_physical_devices('GPU'))
+tp = MisoParameters()
 
 # -----------------------------------------------------------------------------
 # Dataset
@@ -47,7 +44,7 @@
 # - resnet[18,34,50]
 # - vgg[16,19]
 # - efficientnetB[0-7]
-tp.cnn.type = "base_cyclic"
+tp.cnn.id = "convnexttiny"
 # Input image shape, set to None to use default size ([128, 128, 1] for custom, [224, 224, 3] for others)
 tp.cnn.img_shape = [128, 128, 1]
 # Input image colour space [greyscale/rgb]

miso/__main__.py

@@ -16,7 +16,7 @@
 
     tp = MisoParameters()
     tp.source = args.input
-    tp.output_dir = args.output
+    tp.save_dir = args.output
     tp.cnn_type = args.type
     tp.filters = args.filters
     tp.min_count = args.min_count

miso/deploy/model_info.py

@@ -5,7 +5,7 @@
 import numpy as np
 from collections import OrderedDict
 
-from miso.training.parameters import MisoConfig
+from miso.training.parameters import MisoParameters
 
 
 class ModelInfo:
@@ -15,7 +15,7 @@ def __init__(self,
                  type: str,
                  date: datetime.datetime,
                  protobuf: str,
-                 params: MisoConfig,
+                 params: MisoParameters,
                  inputs: OrderedDict,
                  outputs: OrderedDict,
                  data_source_name: str,
@@ -78,9 +78,9 @@ def to_xml(self):
         ET.SubElement(root, "date").text = "{0:%Y-%m-%d_%H%M%S}".format(self.date)
         ET.SubElement(root, "protobuf").text = self.protobuf
 
-        parent_node = ET.SubElement(root, "params")
-        for key, value in self.params.asdict().items():
-            ET.SubElement(parent_node, key).text = str(value)
+        # parent_node = ET.SubElement(root, "params")
+        # for key, value in self.params.asdict().items():
+        #     ET.SubElement(parent_node, key).text = str(value)
 
         parent_node = ET.SubElement(root, "inputs")
         for name, tensor in self.inputs.items():

miso/models/factory.py

@@ -1,21 +1,35 @@
 import math
 
+from keras import Model
 from keras.optimizers import SGD, Adam
 
 from miso.models.keras_models import head, tail, KERAS_MODEL_PARAMETERS
 from miso.models.base_cyclic import *
 from miso.models.resnet_cyclic import *
-from miso.training.parameters import MisoConfig
+from miso.training.parameters import MisoParameters
 
 try:
     from tensorflow.keras.applications.efficientnet import *
 except ImportError:
     pass
 
 
-def generate(tp: MisoConfig):
+def create_optimizer(tp: MisoParameters):
+    if tp.optimizer.name == "sgd":
+        opt = SGD(learning_rate=tp.optimizer.learning_rate,
+                  decay=tp.optimizer.decay,
+                  momentum=tp.optimizer.momentum,
+                  nesterov=tp.optimizer.nesterov)
+    elif tp.optimizer.name == "adam":
+        opt = Adam(learning_rate=tp.optimizer.learning_rate, decay=tp.optimizer.decay)
+    else:
+        raise ValueError(f"The optimizer {tp.optimizer.name} is not supported, valid optimizers are: sgd, adam")
+    return opt
+
+
+def generate(tp: MisoParameters):
     # Base Cyclic - custom network created at CEREGE specifically for foraminifera by adding cyclic layers
-    if tp.cnn.type.startswith("base_cyclic"):
+    if tp.cnn.id.startswith("base_cyclic"):
         model = base_cyclic(input_shape=tp.cnn.img_shape,
                             nb_classes=tp.dataset.num_classes,
                             filters=tp.cnn.filters,
@@ -29,7 +43,7 @@ def generate(tp: MisoConfig):
         model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
 
     # ResNet Cyclic - custom network created at CEREGE specifically for foraminifera by adding cyclic layers
-    elif tp.cnn.type.startswith("resnet_cyclic"):
+    elif tp.cnn.id.startswith("resnet_cyclic"):
         blocks = int(math.log2(tp.cnn.img_shape[0]) - 2)
         blocks -= 1  # Resnet has one block to start with already
         resnet_params = ResnetModelParameters('resnet_cyclic',
@@ -40,21 +54,21 @@ def generate(tp: MisoConfig):
                                     use_cyclic=True,
                                     global_pooling=tp.cnn.global_pooling)
         model = ResNetCyclic(resnet_params, tp.cnn.img_shape, None, True, tp.dataset.num_classes)
-        opt = Adam(lr=0.001)
+        opt = create_optimizer(tp)
         model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
 
     # Standard networks in keras applications (with options for transfer learning or cyclic gain
     else:
         # Legacy support where transfer learning, cyclic and gain were part of the cnn type
-        parts = tp.cnn.type.split("_")
+        parts = tp.cnn.id.split("_")
         if len(parts) > 0:
-            tp.cnn.type = parts[0]
+            tp.cnn.id = parts[0]
             if "cyclic" in parts:
                 tp.cnn.use_cyclic = True
-            if "cyclicgain" in parts:
-                tp.cnn.use_cyclic_gain = True
-        if tp.cnn.type in KERAS_MODEL_PARAMETERS.keys():
-            model_head = head(tp.cnn.type,
+                if "gain" in parts:
+                    tp.cnn.use_cyclic_gain = True
+        if tp.cnn.id in KERAS_MODEL_PARAMETERS.keys():
+            model_head = head(tp.cnn.id,
                               use_cyclic=tp.cnn.use_cyclic,
                               use_gain=tp.cnn.use_cyclic_gain,
                               input_shape=tp.cnn.img_shape,
@@ -65,24 +79,24 @@ def generate(tp: MisoConfig):
             model = combine_head_and_tail(model_head, model_tail)
         else:
             raise ValueError(
-                "The CNN type {} is not supported, valid CNNs are {}".format(tp.cnn.type, KERAS_MODEL_PARAMETERS.keys()))
+                "The CNN type {} is not supported, valid CNNs are {}".format(tp.cnn.id, KERAS_MODEL_PARAMETERS.keys()))
         # opt = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
-        opt = Adam(lr=0.001)
+        opt = create_optimizer(tp)
         model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
     return model
 
 
-def generate_tl(tp: MisoConfig):
+def generate_tl(tp: MisoParameters):
     # Legacy support where transfer learning, cyclic and gain were part of the cnn type
-    parts = tp.cnn.type.split("_")
+    parts = tp.cnn.id.split("_")
     if len(parts) > 0:
-        tp.cnn.type = parts[0]
+        tp.cnn.id = parts[0]
         if "cyclic" in parts:
             tp.cnn.use_cyclic = True
-        if "cyclicgain" in parts:
-            tp.cnn.use_cyclic_gain = True
-    if tp.cnn.type in KERAS_MODEL_PARAMETERS.keys():
-        model_head = head(tp.cnn.type,
+            if "gain" in parts:
+                tp.cnn.use_cyclic_gain = True
+    if tp.cnn.id in KERAS_MODEL_PARAMETERS.keys():
+        model_head = head(tp.cnn.id,
                           use_cyclic=tp.cnn.use_cyclic,
                           use_gain=tp.cnn.use_cyclic_gain,
                           input_shape=tp.cnn.img_shape,
@@ -92,8 +106,8 @@ def generate_tl(tp: MisoConfig):
         model_tail = tail(tp.dataset.num_classes, [model_head.layers[-1].output.shape[-1], ])
     else:
         raise ValueError(
-            "The CNN type {} is not supported, valid CNNs are {}".format(tp.cnn.type, KERAS_MODEL_PARAMETERS.keys()))
-    opt = Adam(lr=0.001)
+            "The CNN type {} is not supported, valid CNNs are {}".format(tp.cnn.id, KERAS_MODEL_PARAMETERS.keys()))
+    opt = create_optimizer(tp)
     model_tail.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
     return model_head, model_tail
 
@@ -102,7 +116,12 @@ def combine_head_and_tail(model_head, model_tail):
     return Model(inputs=model_head.input, outputs=model_tail.call(model_head.output))
 
 
-def generate_vector_from_model(model, cnn_type):
-    vector_tensor = model.get_layer(index=-2).get_output_at(1)
-    vector_model = Model(model.inputs, vector_tensor)
+def generate_vector_from_model(model, tp):
+    try:
+        vector_model = Model(model.inputs, model.get_layer(index=-2).get_output_at(1))
+        return vector_model
+    except:
+        pass
+
+    vector_model = Model(model.inputs, model.get_layer(index=-2).get_output_at(0))
     return vector_model

miso/models/keras_models.py

@@ -133,5 +133,9 @@ def tail_vector(num_classes, input_shape, dropout=(0.5, 0.5)):
     'efficientnetb5': KerasModelParameters(ka.efficientnet.EfficientNetB5, no_prepro, [456, 456, 3]),
     'efficientnetb6': KerasModelParameters(ka.efficientnet.EfficientNetB6, no_prepro, [528, 528, 3]),
     'efficientnetb7': KerasModelParameters(ka.efficientnet.EfficientNetB7, no_prepro, [600, 600, 3]),
-    'convnexttiny': KerasModelParameters(ka.convnext.ConvNeXtTiny, no_prepro, [600, 600, 3]),
+    'convnexttiny': KerasModelParameters(ka.convnext.ConvNeXtTiny, no_prepro, [224, 224, 3]),
+    'convnextsmall': KerasModelParameters(ka.convnext.ConvNeXtSmall, no_prepro, [224, 224, 3]),
+    'convnextbase': KerasModelParameters(ka.convnext.ConvNeXtBase, no_prepro, [224, 224, 3]),
+    'convnextlarge': KerasModelParameters(ka.convnext.ConvNeXtLarge, no_prepro, [224, 224, 3]),
+    'convnextxlarge': KerasModelParameters(ka.convnext.ConvNeXtXLarge, no_prepro, [224, 224, 3]),
 }

miso/training/parameters.py

@@ -17,7 +17,7 @@ def default_field(obj):
     return field(default_factory=lambda: copy.copy(obj))
 
 @dataclass
-class BaseConfig(object):
+class BaseParameters(object):
     class Meta:
         ordered = True
 
@@ -42,9 +42,9 @@ def load(cls, path: Union[Path, str]):
 
 
 @dataclass
-class ModelConfig(BaseConfig):
+class ModelParameters(BaseParameters):
     # Common values
-    type: str = "base_cyclic"
+    id: str = "base_cyclic"
     img_shape: List[int] = field(default=(128, 128, 1))
     img_type: str = "greyscale"
 
@@ -63,7 +63,7 @@ class ModelConfig(BaseConfig):
 
 
 @dataclass
-class TrainingConfig(BaseConfig):
+class TrainingParameters(BaseParameters):
     batch_size: int = 64
     max_epochs: int = 10000
     alr_epochs: int = 10
@@ -77,7 +77,16 @@ class TrainingConfig(BaseConfig):
 
 
 @dataclass
-class DatasetConfig(BaseConfig):
+class OptimizerParameters(BaseParameters):
+    name: str = "adam"
+    learning_rate: float = 0.001
+    momentum: float = 0.9
+    decay: float = 0.0
+    nesterov: bool = False
+
+
+@dataclass
+class DatasetParameters(BaseParameters):
     num_classes: Optional[int] = None
     source: Optional[str] = None
     min_count: Optional[int] = 10
@@ -89,7 +98,7 @@ class DatasetConfig(BaseConfig):
 
 
 @dataclass
-class AugmentationConfig(BaseConfig):
+class AugmentationParameters(BaseParameters):
     rotation: List[float] = field(default=(0, 360))
     gain: Optional[List[float]] = field(default=(0.8, 1, 1.2))
     gamma: Optional[List[float]] = field(default=(0.5, 1, 2))
@@ -101,31 +110,32 @@ class AugmentationConfig(BaseConfig):
 
 
 @dataclass
-class OutputConfig(BaseConfig):
-    output_dir: str = None
+class OutputParameters(BaseParameters):
+    save_dir: str = None
     save_model: bool = True
     save_mislabeled: bool = False
 
 
 @dataclass
-class MisoConfig(BaseConfig):
+class MisoParameters(BaseParameters):
     name: str = ""
     description: str = ""
-    cnn: ModelConfig = default_field(ModelConfig())
-    dataset: DatasetConfig = default_field(DatasetConfig())
-    training: TrainingConfig = default_field(TrainingConfig())
-    augmentation: AugmentationConfig = default_field(AugmentationConfig())
-    output: OutputConfig = default_field(OutputConfig())
+    cnn: ModelParameters = default_field(ModelParameters())
+    dataset: DatasetParameters = default_field(DatasetParameters())
+    training: TrainingParameters = default_field(TrainingParameters())
+    augmentation: AugmentationParameters = default_field(AugmentationParameters())
+    output: OutputParameters = default_field(OutputParameters())
+    optimizer: OptimizerParameters = default_field(OptimizerParameters())
 
     def sanitise(self):
         if self.name == "":
-            self.name = self.dataset.source[:64] + "_" + self.cnn.type
+            self.name = self.dataset.source[:64] + "_" + self.cnn.id
             self.name = re.sub('[^A-Za-z0-9]+', '-', self.name)
         if self.cnn.img_shape is None:
-            if self.cnn.type.endswith("_tl"):
-                shape = KERAS_MODEL_PARAMETERS[self.cnn.type.split('_')[0]].default_input_shape
+            if self.cnn.id.endswith("_tl"):
+                shape = KERAS_MODEL_PARAMETERS[self.cnn.id.split('_')[0]].default_input_shape
             else:
-                if self.cnn.type.startswith("base_cyclic") or self.cnn.type.startswith("resnet_cyclic"):
+                if self.cnn.id.startswith("base_cyclic") or self.cnn.id.startswith("resnet_cyclic"):
                     shape = [128, 128, 3]
                 else:
                     shape = [224, 224, 3]
@@ -136,7 +146,7 @@ def sanitise(self):
                     shape[2] = 1
                     self.augmentation.orig_img_shape[2] = 3
             self.cnn.img_shape = shape
-        elif self.cnn.type.startswith("base_cyclic") or self.cnn.type.startswith("resnet_cyclic"):
+        elif self.cnn.id.startswith("base_cyclic") or self.cnn.id.startswith("resnet_cyclic"):
             pass
         else:
             self.cnn.img_shape[2] = 3
@@ -150,7 +160,7 @@ def get_default_shape(cnn_type):
 
 
 if __name__ == "__main__":
-    m = MisoConfig()
+    m = MisoParameters()
     print(m.dumps())