Add params and output modality type

Author: christinadionysio

src/main/python/systemds/scuro/representations/bert.py

@@ -29,8 +29,10 @@
 
 
 class Bert(UnimodalRepresentation):
-    def __init__(self, output_file=None):
-        super().__init__("Bert")
+    def __init__(self, model_name="bert", output_file=None):
+        parameters = {"model_name": "bert"}
+        self.model_name = model_name
+        super().__init__("Bert", ModalityType.EMBEDDING, parameters)
 
         self.output_file = output_file
 

src/main/python/systemds/scuro/representations/bow.py

@@ -27,8 +27,9 @@
 
 
 class BoW(UnimodalRepresentation):
-    def __init__(self, ngram_range, min_df, output_file=None):
-        super().__init__("BoW")
+    def __init__(self, ngram_range=2, min_df=2, output_file=None):
+        parameters = {"ngram_range": [ngram_range], "min_df": [min_df]}
+        super().__init__("BoW", ModalityType.EMBEDDING, parameters)
         self.ngram_range = ngram_range
         self.min_df = min_df
         self.output_file = output_file

src/main/python/systemds/scuro/representations/fusion.py

@@ -27,12 +27,12 @@
 
 
 class Fusion(Representation):
-    def __init__(self, name):
+    def __init__(self, name, parameters=None):
         """
         Parent class for different multimodal fusion types
         :param name: Name of the fusion type
         """
-        super().__init__(name)
+        super().__init__(name, parameters)
 
     def transform(self, modalities: List[Modality]):
         """

src/main/python/systemds/scuro/representations/glove.py

@@ -39,7 +39,7 @@ def load_glove_embeddings(file_path):
 
 class GloVe(UnimodalRepresentation):
     def __init__(self, glove_path, output_file=None):
-        super().__init__("GloVe")
+        super().__init__("GloVe", ModalityType.TEXT)
         self.glove_path = glove_path
         self.output_file = output_file
 

src/main/python/systemds/scuro/representations/mel_spectrogram.py

@@ -28,8 +28,16 @@
 
 
 class MelSpectrogram(UnimodalRepresentation):
-    def __init__(self):
-        super().__init__("MelSpectrogram")
+    def __init__(self, n_mels=128, hop_length=512, n_fft=2048):
+        parameters = {
+            "n_mels": [20, 32, 64, 128],
+            "hop_length": [256, 512, 1024, 2048],
+            "n_fft": [1024, 2048, 4096],
+        }
+        super().__init__("MelSpectrogram", ModalityType.TIMESERIES, parameters)
+        self.n_mels = n_mels
+        self.hop_length = hop_length
+        self.n_fft = n_fft
 
     def transform(self, modality):
         transformed_modality = TransformedModality(

src/main/python/systemds/scuro/representations/representation.py

@@ -18,8 +18,24 @@
 # under the License.
 #
 # -------------------------------------------------------------
+from abc import abstractmethod
 
 
 class Representation:
-    def __init__(self, name):
+    def __init__(self, name, parameters):
         self.name = name
+        self._parameters = parameters
+
+    @property
+    def parameters(self):
+        return self._parameters
+
+    def get_current_parameters(self):
+        current_params = {}
+        for parameter in self.parameters.keys():
+            current_params[parameter] = getattr(self, parameter)
+        return current_params
+
+    def set_parameters(self, parameters):
+        for parameter in parameters:
+            setattr(self, parameter, parameters[parameter])

src/main/python/systemds/scuro/representations/resnet.py

@@ -38,7 +38,9 @@
 
 class ResNet(UnimodalRepresentation):
     def __init__(self, layer="avgpool", model_name="ResNet18", output_file=None):
-        super().__init__("ResNet")
+        self.model_name = model_name
+        parameters = self._get_parameters()
+        super().__init__("ResNet", ModalityType.TIMESERIES, parameters) # TODO: TIMESERIES only for videos - images would be handled as EMBEDDIGN
 
         self.output_file = output_file
         self.layer_name = layer
@@ -82,6 +84,25 @@ def model(self, model):
         else:
             raise NotImplementedError
 
+    def _get_parameters(self, high_level=True):
+        parameters = {"model_name": [], "layer_name": []}
+        for m in ["ResNet18", "ResNet34", "ResNet50", "ResNet101", "ResNet152"]:
+            parameters["model_name"].append(m)
+
+        if high_level:
+            parameters["layer_name"] = [
+                "conv1",
+                "layer1",
+                "layer2",
+                "layer3",
+                "layer4",
+                "avgpool",
+            ]
+        else:
+            for name, layer in self.model.named_modules():
+                parameters["layer_name"].append(name)
+        return parameters
+
     def transform(self, modality):
 
         t = transforms.Compose(

src/main/python/systemds/scuro/representations/tfidf.py

@@ -27,8 +27,9 @@
 
 
 class TfIdf(UnimodalRepresentation):
-    def __init__(self, min_df, output_file=None):
-        super().__init__("TF-IDF")
+    def __init__(self, min_df=2, output_file=None):
+        parameters = {"min_df": [min_df]}
+        super().__init__("TF-IDF", ModalityType.EMBEDDING, parameters)
         self.min_df = min_df
         self.output_file = output_file
 

src/main/python/systemds/scuro/representations/unimodal.py

@@ -18,17 +18,25 @@
 # under the License.
 #
 # -------------------------------------------------------------
+import abc
+
 from systemds.scuro.representations.representation import Representation
 
 
 class UnimodalRepresentation(Representation):
-    def __init__(self, name):
+    def __init__(self, name: str, output_modality_type, parameters=None):
         """
         Parent class for all unimodal representation types
         :param name: name of the representation
+        :param parameters: parameters of the representation; name of the parameter and
+        possible parameter values
         """
-        super().__init__(name)
+        super().__init__(name, parameters)
+        self.output_modality_type = output_modality_type
+        if parameters is None:
+            parameters = {}
 
+    @abc.abstractmethod
     def transform(self, data):
         raise f"Not implemented for {self.name}"
 

src/main/python/systemds/scuro/representations/word2vec.py

@@ -40,8 +40,13 @@ def get_embedding(sentence, model):
 
 
 class W2V(UnimodalRepresentation):
-    def __init__(self, vector_size, min_count, window, output_file=None):
-        super().__init__("Word2Vec")
+    def __init__(self, vector_size=3, min_count=2, window=2, output_file=None):
+        parameters = {
+            "vector_size": [vector_size],
+            "min_count": [min_count],
+            "window": [window],
+        }  # TODO
+        super().__init__("Word2Vec", ModalityType.EMBEDDING, parameters)
         self.vector_size = vector_size
         self.min_count = min_count
         self.window = window