add changes to load wlasl100 dataset

config.py

@@ -1,3 +1,4 @@
-MIN_INPUT_HEIGHT = 60
-MAX_INPUT_HEIGHT = 60
+AUTSL_INPUT_HEIGHT = 60
+MEJIAPEREZ_INPUT_HEIGHT = 64
+WLASL100_INPUT_HEIGHT = 64
 RANDOM_SEED = 0

dataset.py

@@ -1,6 +1,6 @@
 from enum import IntEnum
 import tensorflow as tf
-from config import MAX_INPUT_HEIGHT, MIN_INPUT_HEIGHT
+from config import AUTSL_INPUT_HEIGHT, MEJIAPEREZ_INPUT_HEIGHT, WLASL100_INPUT_HEIGHT
 from data_augmentation import RandomFlip, RandomScale, RandomShift, RandomRotation, RandomSpeed
 from preprocessing import Center, CenterAtFirstFrame2D, FillBlueWithAngle, PadIfLessThan, RemoveZ, ResizeIfMoreThan, TranslationScaleInvariant
 import tensorflow_datasets as tfds
@@ -51,15 +51,15 @@ class LayerType(IntEnum):
     },
     'train_resize': {
         'type': LayerType.Normalization,
-        'layer': ResizeIfMoreThan(frames=MIN_INPUT_HEIGHT)
+        'layer': ResizeIfMoreThan(frames=100)
     },
     'test_resize': {
         'type': LayerType.Normalization,
-        'layer': ResizeIfMoreThan(frames=MAX_INPUT_HEIGHT)
+        'layer': ResizeIfMoreThan(frames=100)
     },
     'pad': {
         'type': LayerType.Normalization,
-        'layer': PadIfLessThan(frames=MIN_INPUT_HEIGHT)
+        'layer': PadIfLessThan(frames=100)
     },
     'angle': {
         'type': LayerType.Data,
@@ -205,18 +205,34 @@ def label_to_one_hot(item):
         # obtain characteristics of the dataset
         num_train_examples = ds["train"].cardinality()
         num_val_examples = ds["validation"].cardinality()
-        if "test" in ds.keys:
+        if "test" in ds.keys():
             ds["test"] = ds["test"].map(label_to_one_hot)
             num_test_examples = ds["test"].cardinality()
         else:
             num_test_examples = 0
         num_total_examples = num_train_examples + num_val_examples + num_test_examples
+
+        if name == "autsl_tssi":
+            input_height = AUTSL_INPUT_HEIGHT
+        elif name == "mejiaperez_tssi":
+            input_height = MEJIAPEREZ_INPUT_HEIGHT
+        elif name == "wlasl100_tssi":
+            input_height = WLASL100_INPUT_HEIGHT
+        else:
+            raise Exception("Dataset " + name + " not found.")
+
+
+        LayerDict["train_resize"]["layer"] = ResizeIfMoreThan(frames=input_height)
+        LayerDict["test_resize"]["layer"] = ResizeIfMoreThan(frames=input_height)
+        LayerDict["pad"]["layer"] = PadIfLessThan(frames=input_height)
 
         self.ds = ds
+        self.name = name
         self.num_train_examples = num_train_examples
         self.num_val_examples = num_val_examples
         self.num_test_examples = num_test_examples
         self.num_total_examples = num_total_examples
+        self.input_height = input_height
         self.input_width = info.features['pose'].shape[1]
         self.num_classes = info.features['label'].num_classes
 
@@ -239,7 +255,7 @@ def train_map_fn(x, y):
             batch = RemoveZ()(batch)
             batch = preprocessing_pipeline(batch, training=True)
             x = tf.ensure_shape(
-                batch[0], [MIN_INPUT_HEIGHT, self.input_width, 2])
+                batch[0], [self.input_height, self.input_width, 3])
             return x, y
 
         train_ds = self.ds["train"]

datasets/download_datasets.sh

@@ -1,9 +1,9 @@
-wget -O autsl_tssi.zip https://storage.googleapis.com/slr_tfds_datasets/autsl_tssi.zip
-wget -O autsl_tssi.zip https://storage.googleapis.com/slr_tfds_datasets/wlasl100_tssi.zip
-wget -O autsl_tssi.zip https://storage.googleapis.com/slr_tfds_datasets/mejiaperez_tssi.zip
-unzip autsl_tssi.zip
+# wget -O autsl_tssi.zip https://storage.googleapis.com/slr_tfds_datasets/autsl_tssi.zip
+wget -O wlasl100_tssi.zip https://storage.googleapis.com/slr_tfds_datasets/wlasl100_tssi.zip
+# wget -O mejiaperez_tssi.zip https://storage.googleapis.com/slr_tfds_datasets/mejiaperez_tssi.zip
+# unzip autsl_tssi.zip
 unzip wlasl100_tssi.zip
-unzip mejiaperez_tssi.zip
-rm autsl_tssi.zip
+# unzip mejiaperez_tssi.zip
+# rm autsl_tssi.zip
 rm wlasl100_tssi.zip
-rm mejiaperez_tssi.zip
+# rm mejiaperez_tssi.zip

model.py

@@ -1,19 +1,17 @@
 import tensorflow as tf
-from efficient_net_b0 import EfficientNetB0
 from loss import select_loss
 from tensorflow.keras.metrics import TopKCategoricalAccuracy
 from tensorflow.keras.layers import Dropout, Dense
 from tensorflow.keras import Input
 from tensorflow.keras.models import Model
 from tensorflow.keras.applications.densenet import DenseNet121
-# from tensorflow.keras.applications.efficientnet import EfficientNetB0
 
 
 def build_densenet121_model(input_shape=[None, 135, 2],
                             dropout=0,
                             optimizer=None,
                             pretraining=False,
-                            use_loss="crossentroypy",
+                            use_loss="crossentropy",
                             num_classes=None):
     # setup backbone
     weights = 'imagenet' if pretraining else None
@@ -41,37 +39,3 @@ def build_densenet121_model(input_shape=[None, 135, 2],
     model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
 
     return model
-
-
-def build_efficientnet_model(input_shape=[None, 128, 3],
-                             dropout=0,
-                             optimizer=None,
-                             pretraining=True,
-                             use_loss="crossentropy",
-                             num_classes=100):
-    # setup backbone
-    weights = "imagenet" if pretraining else None
-    backbone = EfficientNetB0(input_shape=input_shape,
-                              weights=weights,
-                              include_top=False,
-                              pooling="avg")
-
-    # setup model
-    inputs = Input(shape=input_shape)
-    x = backbone(inputs)
-    x = Dropout(dropout)(x)
-    predictions = Dense(num_classes, activation='softmax')(x)
-    model = Model(inputs=inputs, outputs=predictions)
-
-    # setup the metrics
-    metrics = [
-        TopKCategoricalAccuracy(k=1, name='top_1', dtype=tf.float32)
-    ]
-
-    # setup the model
-    loss = select_loss(use_loss)
-
-    # compile the model
-    model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
-
-    return model