WIP

yanncalec · May 19, 2024 · 5e9cce5 · 5e9cce5
1 parent 7dc15aa
commit 5e9cce5
Show file tree

Hide file tree

Showing 2 changed files with 62 additions and 21 deletions.
diff --git a/dpmhm/datasets/preprocessing.py b/dpmhm/datasets/preprocessing.py
@@ -17,7 +17,28 @@
 
 
 def get_mapping_supervised(labels:list, *, feature_field:str='feature', label_field:str='label') -> callable:
-    """Get a preprocessing mapping to transform a dataset to the format `(data,label)` for supervised training.
+    """Get a preprocessing mapping for supervised training.
+
+    This processing model performs the following transformations on a dataset:
+    - conversion of label from string to integer
+    - conversion from the format of channel first to channel last
+
+    After transformation, the dataset is in the format `(data,label)`.
+
+    Usage
+    -----
+    ```python
+    func = get_mapping_supervised(labels)
+    ds = ds.map(func, num_parallel_calls=tf.data.AUTOTUNE)
+    ds = ds.map(lambda x,y: (tf.ensure_shape(x, dimx), y),
+    ```
+
+    Note
+    ----
+    The `ensure_shape()` step seems necessary in Keras 3 to recover the dimension information lost after `.map()`. Otherwise the training will be failed.
+
+    See also:
+    https://github.com/tensorflow/tensorflow/issues/64177
     """
     label_layer = layers.StringLookup(
         # num_oov_indices=0,   # force zero-based integer
@@ -39,6 +60,8 @@ def nested_type_spec(sp:dict) -> dict:
     return tp
 
 
+# The following method doesn't work in Keras 3. Use `get_mapping_supervised()` instead.
+
 def keras_model_supervised(ds:Dataset, labels:list=None, normalize:bool=False, *, shape:tuple=None, feature_field:str='feature', label_field:str='label') -> models.Model:
     """Initialize a Keras preprocessing model for supervised training.
 
@@ -125,13 +148,9 @@ def keras_model_supervised(ds:Dataset, labels:list=None, normalize:bool=False, *
     feature_output = ops.transpose(feature_output, [0,2,3,1])
 
     # Restore the shape information
-    # if shape is not None:
-    #     feature_output = tf.reshape(feature_output, shape)
-    # label_output = tf.reshape(label_output, ())
     if shape is not None:
         feature_output = ops.reshape(feature_output, shape)
-    # label_output = ops.reshape(label_output, ())
-    # print(label_output)
+    label_output = ops.reshape(label_output, ())
 
     outputs = (feature_output, label_output)
 
@@ -150,4 +169,4 @@ def keras_model_supervised(ds:Dataset, labels:list=None, normalize:bool=False, *
     return model
 
 
-__all__ = ['nested_type_spec', 'keras_model_supervised']
+__all__ = ['nested_type_spec', 'get_mapping_supervised']
diff --git a/notebooks/models/supervised_vggish.ipynb b/notebooks/models/supervised_vggish.ipynb
@@ -600,7 +600,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": null,
+   "execution_count": 14,
    "id": "2f8354ee-7bf5-49c6-90bc-b5ee6ba6b604",
    "metadata": {},
    "outputs": [
@@ -611,7 +611,13 @@
       "Epoch 1/5\n",
       "\u001b[1m357/357\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m127s\u001b[0m 350ms/step - accuracy: 0.5214 - loss: 3.7552 - val_accuracy: 0.9101 - val_loss: 0.2769\n",
       "Epoch 2/5\n",
-      "\u001b[1m 72/357\u001b[0m \u001b[32m━━━━\u001b[0m\u001b[37m━━━━━━━━━━━━━━━━\u001b[0m \u001b[1m1:26\u001b[0m 305ms/step - accuracy: 0.9204 - loss: 0.2379"
+      "\u001b[1m357/357\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m123s\u001b[0m 346ms/step - accuracy: 0.9255 - loss: 0.2217 - val_accuracy: 0.9515 - val_loss: 0.1628\n",
+      "Epoch 3/5\n",
+      "\u001b[1m357/357\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m123s\u001b[0m 345ms/step - accuracy: 0.9483 - loss: 0.1590 - val_accuracy: 0.9727 - val_loss: 0.0853\n",
+      "Epoch 4/5\n",
+      "\u001b[1m357/357\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m122s\u001b[0m 341ms/step - accuracy: 0.9588 - loss: 0.1418 - val_accuracy: 0.9693 - val_loss: 0.0822\n",
+      "Epoch 5/5\n",
+      "\u001b[1m357/357\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m122s\u001b[0m 342ms/step - accuracy: 0.9665 - loss: 0.1127 - val_accuracy: 0.9834 - val_loss: 0.0552\n"
      ]
     }
    ],
@@ -626,22 +632,40 @@
   },
   {
    "cell_type": "code",
-   "execution_count": null,
-   "id": "fc947ba1",
+   "execution_count": 15,
+   "id": "06f79296-3a90-47e5-ba61-815507251d89",
    "metadata": {},
-   "outputs": [],
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\u001b[1m1630/1630\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m22s\u001b[0m 13ms/step - accuracy: 0.9777 - loss: 0.0619\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "[0.05482785403728485, 0.9822086095809937]"
+      ]
+     },
+     "execution_count": 15,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
    "source": [
-    "model.save(str(outdir / 'saved_model'))"
+    "model.evaluate(ds_test)"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": null,
-   "id": "06f79296-3a90-47e5-ba61-815507251d89",
+   "execution_count": 17,
+   "id": "fc947ba1",
    "metadata": {},
    "outputs": [],
    "source": [
-    "model.evaluate(ds_test)"
+    "model.save(str(outdir / 'saved_model.keras'))"
    ]
   },
   {
@@ -656,7 +680,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 19,
+   "execution_count": 18,
    "id": "10b97f80-0af9-4ee6-a943-1ae9658be30e",
    "metadata": {},
    "outputs": [],
@@ -672,7 +696,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 20,
+   "execution_count": null,
    "id": "4fc06497-9ce1-4ca9-aa9d-9f68500eea2a",
    "metadata": {},
    "outputs": [
@@ -681,9 +705,7 @@
      "output_type": "stream",
      "text": [
       "Epoch 1/2\n",
-      "713/713 [==============================] - 454s 633ms/step - loss: 0.1665 - accuracy: 0.9604 - val_loss: 0.2006 - val_accuracy: 0.9521\n",
-      "Epoch 2/2\n",
-      "713/713 [==============================] - 459s 643ms/step - loss: 0.0477 - accuracy: 0.9876 - val_loss: 0.0283 - val_accuracy: 0.9914\n"
+      "\u001b[1m174/357\u001b[0m \u001b[32m━━━━━━━━━\u001b[0m\u001b[37m━━━━━━━━━━━\u001b[0m \u001b[1m2:46\u001b[0m 911ms/step - accuracy: 0.9695 - loss: 0.1080"
      ]
     }
    ],