Split test_model_recurrent into recurrent + extended

The catch-all get_test_model_recurrent had drifted to include layers
that aren't recurrent at all (DepthwiseConv1D, EinsumDense, Adaptive*Pooling,
GroupedQueryAttention, RMS/GroupNormalization, Discretization, Random* image
augmentation). Extract those into a new get_test_model_extended driving a
parallel test_model_extended_test target so each test name reflects what it
covers.

test_model_recurrent now only contains genuinely-recurrent layers: LSTM, GRU,
SimpleRNN, Bidirectional, Masking, RNN with cells, ConvLSTM 1D/2D/3D.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Dobiasd

keras_export/generate_test_models.py

@@ -826,13 +826,12 @@ def get_test_model_sequential() -> Model:
 
 
 def get_test_model_recurrent() -> Model:
-    """Returns a test model exercising LSTM, GRU, SimpleRNN, Bidirectional, and other newer layers."""
+    """Returns a test model exercising recurrent layers (LSTM, GRU, SimpleRNN,
+    Bidirectional, RNN with cells, ConvLSTM, Masking)."""
     seq_len = 5
     n_features = 4
 
     inputs = Input(shape=(seq_len, n_features))
-    inputs_2d = Input(shape=(7, 8, 4))
-    inputs_3d = Input(shape=(5, 6, 7, 4))
 
     lstm_seq = LSTM(6, return_sequences=True)(inputs)
     lstm_last = LSTM(7)(lstm_seq)
@@ -857,6 +856,79 @@ def get_test_model_recurrent() -> Model:
     bidi_gru_ave = Bidirectional(GRU(4), merge_mode='ave')(inputs)
     bidi_rnn_concat = Bidirectional(SimpleRNN(3, return_sequences=True))(inputs)
 
+    masked = Masking(mask_value=0.0)(inputs)
+
+    rnn_lstm = RNN(LSTMCell(4))(inputs)
+    rnn_gru = RNN(GRUCell(5), return_sequences=True)(inputs)
+    rnn_simple = RNN(SimpleRNNCell(3))(inputs)
+    rnn_stacked = RNN(StackedRNNCells([LSTMCell(6), GRUCell(4), SimpleRNNCell(3)]))(inputs)
+    rnn_stacked_seq = RNN(StackedRNNCells([LSTMCell(5), GRUCell(3)]),
+        return_sequences=True)(inputs)
+
+    inputs_clstm1d = Input(shape=(4, 6, 3))
+    inputs_clstm2d = Input(shape=(3, 5, 5, 3))
+    inputs_clstm3d = Input(shape=(2, 3, 4, 4, 3))
+    clstm1d = ConvLSTM1D(filters=2, kernel_size=3, padding='same')(inputs_clstm1d)
+    clstm1d_valid = ConvLSTM1D(filters=2, kernel_size=2, padding='valid',
+        return_sequences=True)(inputs_clstm1d)
+    clstm1d_dilated = ConvLSTM1D(filters=2, kernel_size=2, padding='same',
+        dilation_rate=2)(inputs_clstm1d)
+    clstm2d = ConvLSTM2D(filters=2, kernel_size=(3, 3), padding='same',
+        return_sequences=True)(inputs_clstm2d)
+    clstm2d_valid_no_bias = ConvLSTM2D(filters=3, kernel_size=(2, 2), padding='valid',
+        use_bias=False, activation='relu')(inputs_clstm2d)
+    # Two ConvLSTM2Ds chained: validates state handoff between successive temporal layers.
+    clstm2d_chained = ConvLSTM2D(filters=2, kernel_size=(2, 2), padding='same')(
+        ConvLSTM2D(filters=3, kernel_size=(3, 3), padding='same',
+            return_sequences=True)(inputs_clstm2d))
+    clstm3d = ConvLSTM3D(filters=2, kernel_size=(2, 2, 2), padding='same')(inputs_clstm3d)
+    clstm3d_valid = ConvLSTM3D(filters=2, kernel_size=(1, 2, 2), padding='valid',
+        return_sequences=True)(inputs_clstm3d)
+
+    outputs = [
+        lstm_last, lstm_no_bias, lstm_relu, lstm_no_unit_forget,
+        lstm_state_out, lstm_state_h, lstm_state_c,
+        gru_last, gru_no_bias,
+        rnn_last, rnn_relu_no_bias,
+        bidi_lstm_last_concat,
+        bidi_lstm_sum,
+        bidi_gru_mul,
+        bidi_gru_ave,
+        bidi_rnn_concat,
+        masked,
+        rnn_lstm, rnn_gru, rnn_simple,
+        rnn_stacked, rnn_stacked_seq,
+        clstm1d, clstm1d_valid, clstm1d_dilated,
+        clstm2d, clstm2d_valid_no_bias, clstm2d_chained,
+        clstm3d, clstm3d_valid,
+    ]
+
+    model = Model(inputs=[inputs, inputs_clstm1d, inputs_clstm2d, inputs_clstm3d],
+        outputs=outputs, name='test_model_recurrent')
+    model.compile(loss='mse', optimizer='adam')
+
+    training_data_size = 2
+    data_in = generate_input_data(training_data_size,
+        [(seq_len, n_features), (4, 6, 3), (3, 5, 5, 3), (2, 3, 4, 4, 3)])
+    initial_data_out = model.predict(data_in)
+    data_out = generate_output_data(training_data_size, initial_data_out)
+    model.fit(data_in, data_out, epochs=1)
+    return model
+
+
+def get_test_model_extended() -> Model:
+    """Returns a test model exercising recently-added non-recurrent layers
+    (DepthwiseConv1D, EinsumDense, RMSNormalization, GroupNormalization,
+    AdaptiveAvg/MaxPooling, GroupedQueryAttention, Discretization, and a
+    handful of training-only Random* augmentation passthroughs)."""
+    seq_len = 5
+    n_features = 4
+
+    inputs = Input(shape=(seq_len, n_features))
+    inputs_2d = Input(shape=(7, 8, 4))
+    inputs_3d = Input(shape=(5, 6, 7, 4))
+    inputs_kv = Input(shape=(8, n_features))
+
     dwc1d_same = DepthwiseConv1D(kernel_size=3, padding='same')(inputs)
     dwc1d_valid = DepthwiseConv1D(kernel_size=2, padding='valid', strides=2)(inputs)
     dwc1d_no_bias = DepthwiseConv1D(kernel_size=3, padding='same', use_bias=False)(inputs)
@@ -893,53 +965,15 @@ def get_test_model_recurrent() -> Model:
     gqa_eq_heads = GroupQueryAttention(head_dim=4, num_query_heads=4, num_key_value_heads=4)(
         inputs, inputs, inputs)
     # distinct sequence lengths for query vs key/value.
-    inputs_kv = Input(shape=(8, n_features))
     gqa_kv_diff = GroupQueryAttention(head_dim=3, num_query_heads=4, num_key_value_heads=2)(
         inputs, inputs_kv, inputs_kv)
 
     discretized = Discretization(bin_boundaries=[-0.5, 0.0, 0.5, 1.0])(inputs)
-    masked = Masking(mask_value=0.0)(inputs)
     rand_brightness = RandomBrightness(factor=0.1)(inputs_2d)
     rand_flip = RandomFlip(mode='horizontal')(inputs_2d)
     rand_crop = RandomCrop(height=7, width=8)(inputs_2d)
 
-    rnn_lstm = RNN(LSTMCell(4))(inputs)
-    rnn_gru = RNN(GRUCell(5), return_sequences=True)(inputs)
-    rnn_simple = RNN(SimpleRNNCell(3))(inputs)
-    rnn_stacked = RNN(StackedRNNCells([LSTMCell(6), GRUCell(4), SimpleRNNCell(3)]))(inputs)
-    rnn_stacked_seq = RNN(StackedRNNCells([LSTMCell(5), GRUCell(3)]),
-        return_sequences=True)(inputs)
-
-    inputs_clstm1d = Input(shape=(4, 6, 3))
-    inputs_clstm2d = Input(shape=(3, 5, 5, 3))
-    inputs_clstm3d = Input(shape=(2, 3, 4, 4, 3))
-    clstm1d = ConvLSTM1D(filters=2, kernel_size=3, padding='same')(inputs_clstm1d)
-    clstm1d_valid = ConvLSTM1D(filters=2, kernel_size=2, padding='valid',
-        return_sequences=True)(inputs_clstm1d)
-    clstm1d_dilated = ConvLSTM1D(filters=2, kernel_size=2, padding='same',
-        dilation_rate=2)(inputs_clstm1d)
-    clstm2d = ConvLSTM2D(filters=2, kernel_size=(3, 3), padding='same',
-        return_sequences=True)(inputs_clstm2d)
-    clstm2d_valid_no_bias = ConvLSTM2D(filters=3, kernel_size=(2, 2), padding='valid',
-        use_bias=False, activation='relu')(inputs_clstm2d)
-    # Two ConvLSTM2Ds chained: validates state handoff between successive temporal layers.
-    clstm2d_chained = ConvLSTM2D(filters=2, kernel_size=(2, 2), padding='same')(
-        ConvLSTM2D(filters=3, kernel_size=(3, 3), padding='same',
-            return_sequences=True)(inputs_clstm2d))
-    clstm3d = ConvLSTM3D(filters=2, kernel_size=(2, 2, 2), padding='same')(inputs_clstm3d)
-    clstm3d_valid = ConvLSTM3D(filters=2, kernel_size=(1, 2, 2), padding='valid',
-        return_sequences=True)(inputs_clstm3d)
-
     outputs = [
-        lstm_last, lstm_no_bias, lstm_relu, lstm_no_unit_forget,
-        lstm_state_out, lstm_state_h, lstm_state_c,
-        gru_last, gru_no_bias,
-        rnn_last, rnn_relu_no_bias,
-        bidi_lstm_last_concat,
-        bidi_lstm_sum,
-        bidi_gru_mul,
-        bidi_gru_ave,
-        bidi_rnn_concat,
         dwc1d_same, dwc1d_valid, dwc1d_no_bias,
         rms, rms_eps,
         gn, gn_no_scale, gn_no_center, gn_one, gn_per_channel,
@@ -948,24 +982,17 @@ def get_test_model_recurrent() -> Model:
         ap2d_a, ap2d_m, ap2d_uneven,
         ap3d_a, ap3d_m, ap3d_uneven,
         gqa, gqa_gated, gqa_eq_heads, gqa_kv_diff,
-        rnn_lstm, rnn_gru, rnn_simple,
-        rnn_stacked, rnn_stacked_seq,
-        clstm1d, clstm1d_valid, clstm1d_dilated,
-        clstm2d, clstm2d_valid_no_bias, clstm2d_chained,
-        clstm3d, clstm3d_valid,
-        discretized, masked,
+        discretized,
         rand_brightness, rand_flip, rand_crop,
     ]
 
-    model = Model(inputs=[inputs, inputs_2d, inputs_3d,
-                          inputs_clstm1d, inputs_clstm2d, inputs_clstm3d, inputs_kv],
-        outputs=outputs, name='test_model_recurrent')
+    model = Model(inputs=[inputs, inputs_2d, inputs_3d, inputs_kv],
+        outputs=outputs, name='test_model_extended')
     model.compile(loss='mse', optimizer='adam')
 
     training_data_size = 2
     data_in = generate_input_data(training_data_size,
-        [(seq_len, n_features), (7, 8, 4), (5, 6, 7, 4),
-         (4, 6, 3), (3, 5, 5, 3), (2, 3, 4, 4, 3), (8, n_features)])
+        [(seq_len, n_features), (7, 8, 4), (5, 6, 7, 4), (8, n_features)])
     initial_data_out = model.predict(data_in)
     data_out = generate_output_data(training_data_size, initial_data_out)
     model.fit(data_in, data_out, epochs=1)
@@ -988,6 +1015,7 @@ def main() -> None:
             'autoencoder': get_test_model_autoencoder,
             'sequential': get_test_model_sequential,
             'recurrent': get_test_model_recurrent,
+            'extended': get_test_model_extended,
         }
 
         if not model_name in get_model_functions:

test/CMakeLists.txt

@@ -28,6 +28,10 @@ add_custom_command ( OUTPUT test_model_recurrent.keras
                      COMMAND bash -c "${Python3_EXECUTABLE} ${FDEEP_TOP_DIR}/keras_export/generate_test_models.py recurrent test_model_recurrent.keras"
                      WORKING_DIRECTORY ${CMAKE_BINARY_DIR}/)
 
+add_custom_command ( OUTPUT test_model_extended.keras
+                     COMMAND bash -c "${Python3_EXECUTABLE} ${FDEEP_TOP_DIR}/keras_export/generate_test_models.py extended test_model_extended.keras"
+                     WORKING_DIRECTORY ${CMAKE_BINARY_DIR}/)
+
 add_custom_command ( OUTPUT readme_example_model.keras
                      COMMAND bash -c "${Python3_EXECUTABLE} ${FDEEP_TOP_DIR}/test/readme_example_generate.py"
                      WORKING_DIRECTORY ${CMAKE_BINARY_DIR}/)
@@ -62,6 +66,11 @@ add_custom_command ( OUTPUT test_model_recurrent.json
                      COMMAND bash -c "${Python3_EXECUTABLE} ${FDEEP_TOP_DIR}/keras_export/convert_model.py test_model_recurrent.keras test_model_recurrent.json"
                      WORKING_DIRECTORY ${CMAKE_BINARY_DIR}/)
 
+add_custom_command ( OUTPUT test_model_extended.json
+                     DEPENDS test_model_extended.keras
+                     COMMAND bash -c "${Python3_EXECUTABLE} ${FDEEP_TOP_DIR}/keras_export/convert_model.py test_model_extended.keras test_model_extended.json"
+                     WORKING_DIRECTORY ${CMAKE_BINARY_DIR}/)
+
 add_custom_command ( OUTPUT readme_example_model.json
                      DEPENDS readme_example_model.keras
                      COMMAND bash -c "${Python3_EXECUTABLE} ${FDEEP_TOP_DIR}/keras_export/convert_model.py readme_example_model.keras readme_example_model.json"
@@ -84,6 +93,7 @@ _add_test(test_model_variable_test test_model_variable.json)
 _add_test(test_model_autoencoder_test test_model_autoencoder.json)
 _add_test(test_model_sequential_test test_model_sequential.json)
 _add_test(test_model_recurrent_test test_model_recurrent.json)
+_add_test(test_model_extended_test test_model_extended.json)
 _add_test(readme_example_main readme_example_model.json)
 
 add_custom_target(unittest
@@ -93,6 +103,7 @@ add_custom_target(unittest
   COMMAND test_model_autoencoder_test
   COMMAND test_model_sequential_test
   COMMAND test_model_recurrent_test
+  COMMAND test_model_extended_test
   COMMAND readme_example_main
 
   COMMENT "Running unittests\n\n"

test/test_model_extended_test.cpp

@@ -0,0 +1,21 @@
+// Copyright 2016, Tobias Hermann.
+// https://github.com/Dobiasd/frugally-deep
+// Distributed under the MIT License.
+// (See accompanying LICENSE file or at
+//  https://opensource.org/licenses/MIT)
+
+#define DOCTEST_CONFIG_IMPLEMENT_WITH_MAIN
+#include "doctest/doctest.h"
+#define FDEEP_FLOAT_TYPE double
+#include <fdeep/fdeep.hpp>
+
+TEST_CASE("test_model_extended_test, load_model")
+{
+    const auto model = fdeep::load_model("../test_model_extended.json",
+        true, fdeep::cout_logger, static_cast<fdeep::float_type>(0.00001));
+    const auto multi_inputs = fplus::generate<std::vector<fdeep::tensors>>(
+        [&]() -> fdeep::tensors { return model.generate_dummy_inputs(); },
+        10);
+    model.predict_multi(multi_inputs, false);
+    model.predict_multi(multi_inputs, true);
+}