DL: Fix num_class parsing from model architecture
JIRA: MADLIB-1472
get_num_classes function did not work in certain models that end
with activation layers. The regresion was caused by the changes
from the multi-io commit.
diff --git a/src/ports/postgres/modules/deep_learning/madlib_keras.sql_in b/src/ports/postgres/modules/deep_learning/madlib_keras.sql_in
index 429c0f0..05edc0e 100644
--- a/src/ports/postgres/modules/deep_learning/madlib_keras.sql_in
+++ b/src/ports/postgres/modules/deep_learning/madlib_keras.sql_in
@@ -84,7 +84,7 @@
kernel version (1.14). Using a newer or older version may or may not work as intended.
MADlib's deep learning methods are designed to use the TensorFlow package and its built in Keras
-functions. To ensure consistency, please use tensorflow.keras objects (models, layers, etc.)
+functions. To ensure consistency, please use tensorflow.keras objects (models, layers, etc.)
instead of importing Keras and using its objects.
@note CUDA GPU memory cannot be released until the process holding it is terminated.
@@ -165,15 +165,15 @@
@note
- Custom loss functions and custom metrics can be used as defined in
<a href="group__grp__custom__function.html">Define Custom Functions.</a>
- List the custom function name and provide the name of the table where the
+ List the custom function name and provide the name of the table where the
serialized Python objects reside using the parameter 'object_table' below.
- The following loss function is
not supported: <em>sparse_categorical_crossentropy</em>.
The following metrics are not
supported: <em>sparse_categorical_accuracy, sparse_top_k_categorical_accuracy</em>.
- - The Keras accuracy parameter <em>top_k_categorical_accuracy</em> returns top 5 accuracy by
+ - The Keras accuracy parameter <em>top_k_categorical_accuracy</em> returns top 5 accuracy by
default. If you want a different top k value, use the helper function
- <a href="group__grp__custom__function.html#top_k_function">Top k Accuracy Function</a>
+ <a href="group__grp__custom__function.html#top_k_function">Top k Accuracy Function</a>
to create a custom
Python function to compute the top k accuracy that you want.
@@ -609,10 +609,10 @@
<DD>TEXT. Column with independent variables in the test table.
If a 'normalizing_const' is specified when preprocessing the
training dataset, this same normalization will be applied to
- the independent variables used in predict. In the case that there
- are multiple independent variables,
+ the independent variables used in predict. In the case that there
+ are multiple independent variables,
representing a multi-input neural network,
- put the columns as a comma
+ put the columns as a comma
separated list, e.g., 'indep_var1, indep_var2, indep_var3' in the same
way as was done in the preprocessor step for the training data.
</DD>
@@ -695,7 +695,8 @@
pred_type,
use_gpus,
class_values,
- normalizing_const
+ normalizing_const,
+ dependent_count
)
</pre>
@@ -805,6 +806,11 @@
array by. For example, you would use 255 for this value if the image data is
in the form 0-255.
</DD>
+
+ <DT>dependent_count (optional)</DT>
+ <DD>INTEGER, default: 1.
+ The number of dependent variables in the model.
+ </DD>
</DL>
diff --git a/src/ports/postgres/modules/deep_learning/model_arch_info.py_in b/src/ports/postgres/modules/deep_learning/model_arch_info.py_in
index 9c28c43..0081e58 100644
--- a/src/ports/postgres/modules/deep_learning/model_arch_info.py_in
+++ b/src/ports/postgres/modules/deep_learning/model_arch_info.py_in
@@ -66,12 +66,15 @@
arch_layers = _get_layers(model_arch)
num_classes = []
- layer_count = len(arch_layers) - 1
- for i in range(multi_dep_count):
- if 'units' in arch_layers[layer_count-i]['config']:
- num_classes.append(arch_layers[layer_count-i]['config']['units'])
-
+ i = len(arch_layers) - 1
+ dep_counter = 0
+ while i >= 0 and dep_counter < multi_dep_count:
+ if 'units' in arch_layers[i]['config']:
+ num_classes.append(arch_layers[i]['config']['units'])
+ dep_counter +=1
+ i -= 1
if num_classes:
+ num_classes.reverse()
return num_classes
plpy.error('Unable to get number of classes from model architecture.')
diff --git a/src/ports/postgres/modules/deep_learning/test/unit_tests/test_madlib_keras.py_in b/src/ports/postgres/modules/deep_learning/test/unit_tests/test_madlib_keras.py_in
index 0c91eb4..e46efd7 100644
--- a/src/ports/postgres/modules/deep_learning/test/unit_tests/test_madlib_keras.py_in
+++ b/src/ports/postgres/modules/deep_learning/test/unit_tests/test_madlib_keras.py_in
@@ -1410,7 +1410,7 @@
self.model.add(Dense(1599))
with self.assertRaises(plpy.PLPYException) as error:
self.subject.validate_class_values(
- self.module_name, [range(1599), range(1598)], 'prob', self.model.to_json())
+ self.module_name, [range(1599)], 'prob', self.model.to_json())
self.assertIn('1600', str(error.exception))
def test_validate_class_values_valid_class_values_prob(self):
@@ -1487,6 +1487,70 @@
obj = self.subject._validate_gpu_config(self.module_name, 'foo', [1,0,0,1])
self.assertIn('does not have gpu', str(error.exception).lower())
+ def test_validate_class_values_last_layer_not_dense(self):
+ num_classes = 3
+ model = Sequential()
+ model.add(Conv2D(2, kernel_size=(1, 1), activation='relu',
+ input_shape=(1,1,1,), padding='same'))
+ model.add(Dense(num_classes))
+ model.add(Activation('relu'))
+ model.add(Activation('softmax'))
+
+ self.subject.validate_class_values(
+ self.module_name, [range(num_classes)], 'prob', model.to_json())
+
+ def test_validate_class_values_last_layer_not_dense_multiio(self):
+ num_classes = 3
+ model = Sequential()
+ model.add(Conv2D(2, kernel_size=(1, 1), activation='relu',
+ input_shape=(1,1,1,), padding='same'))
+ model.add(Dense(num_classes))
+ model.add(Dense(num_classes))
+ model.add(Activation('relu'))
+ model.add(Activation('softmax'))
+
+ self.subject.validate_class_values(
+ self.module_name, [range(num_classes), range(num_classes)], 'prob', model.to_json())
+
+ def test_validate_class_values_mismatch(self):
+ expected_error_regex = ".*do not match.*architecture"
+ num_classes = 3
+
+ # only one dense layer but len(dep_var) = 2
+ model = Sequential()
+ model.add(Conv2D(2, kernel_size=(1, 1), activation='relu',
+ input_shape=(1,1,1,), padding='same'))
+ model.add(Dense(num_classes))
+ model.add(Activation('relu'))
+ with self.assertRaisesRegexp(plpy.PLPYException, expected_error_regex):
+ self.subject.validate_class_values(
+ self.module_name, [range(num_classes), range(num_classes)], 'prob', model.to_json())
+
+ # two dense layers
+ model = Sequential()
+ model.add(Conv2D(2, kernel_size=(1, 1), activation='relu',
+ input_shape=(1,1,1,), padding='same'))
+ model.add(Dense(2))
+ model.add(Dense(num_classes))
+ with self.assertRaisesRegexp(plpy.PLPYException, expected_error_regex):
+ self.subject.validate_class_values(
+ self.module_name, [range(num_classes), range(num_classes)], 'prob', model.to_json())
+
+ def test_validate_class_values_no_units(self):
+ expected_error_regex = ".*Unable.*classes.*architecture"
+ num_classes = 3
+ #model arch is missing a dense layer
+ model = Sequential()
+ model.add(Activation('relu'))
+ model.add(Conv2D(2, kernel_size=(1, 1), activation='relu',
+ input_shape=(1,1,1,), padding='same'))
+ with self.assertRaisesRegexp(plpy.PLPYException, expected_error_regex):
+ self.subject.validate_class_values(
+ self.module_name, [range(num_classes)], 'prob', model.to_json())
+ with self.assertRaisesRegexp(plpy.PLPYException, expected_error_regex):
+ self.subject.validate_class_values(
+ self.module_name, [range(num_classes), range(num_classes)], 'prob', model.to_json())
+
class MadlibSerializerTestCase(unittest.TestCase):
def setUp(self):
