Cleanup code

.travis.yml

@@ -68,10 +68,8 @@ script:
   - cd ..
   # run stateful tests
   - cd test/stateful_test
-  - echo "compiling stateful recurrent tests"
   - $CXX -I../../include -std=c++14 -O3 stateful_recurrent_tests.cpp -o stateful_recurrent_tests_cpp
   - mkdir models
-  - echo "running stateful recurrent tests"
   - python3 stateful_recurrent_tests.py
   - cd ../..
   # release possibly new version to conan

test/stateful_test/readme.md

@@ -1,20 +1,20 @@
-## Stand-Alone Stateful RNN Tests
+# Stand-Alone Stateful RNN Tests
 
 This is a set of tests that is not part of the frugally-deep unit tests.  These test the functionality of the stateful property for RNNs (e.g., LSTMs, GRUs, and bidirectional variants).  This requires multiple calls to `model.stateful_predict()` to test and that is why it is not covered by the standard unit tests which only verify test vectors for a single call to `model.predict()` or `model.stateful_predict()`.
 
 To execute this test, just run the following commands:
 
 ```bash
 mkdir models
-g++ -I../../include -std=c++14 -O3 stateful_recurrent_tests.cpp -o stateful_recurrent_tests_cppts
+g++ -I../../include -std=c++14 -O3 stateful_recurrent_tests.cpp -o stateful_recurrent_tests_cpp
 python3 stateful_recurrent_tests.py
 ```
 
 This does the following:
 
 1. Runs a series of RNN models in Keras and stores the models and some test vectors in the `models` subdirectory.
 2. Runs the `../../keras_export/convert_model.py ` script to convert all of the generated Keras models. The resulting json files also are written to the `models` subdirectory.
-3.  Compiles and runs `stateful_recurrent_tests.cpp` which loads each of the json model files and generates test vectors using frugrally-deep that correspond to those in step 1.
+3. Compiles and runs `stateful_recurrent_tests.cpp` which loads each of the json model files and generates test vectors using frugrally-deep that correspond to those in step 1.
 4. Compares the test vectors from Keras and frugally-deep to report PASS/FAIL on the tests.
 
 The main Python script also creates the models directory as needed and clears it at the start of its execution.

test/stateful_test/stateful_recurrent_tests.py

@@ -1,14 +1,12 @@
-import h5py
 import keras
 import numpy as np
 from keras.layers import Input, Dense, GRU, LSTM, Bidirectional
 import tensorflow as tf
-
-tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
 import os
 import sys
 import errno
-import shutil
+
+tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
 
 __author__ = "Keith Chugg"
 __copyright__ = "Copyright 2019, Keith Chugg"
@@ -26,7 +24,7 @@ def get_trained_model(x_train, y_train, layer_name, n_recurrent_units, bidi):
         REC_LAYER = LSTM
     else:
         REC_LAYER = GRU
-    ######  Define/Build/Train Training Model
+    #  Define/Build/Train Training Model
     training_in_shape = x_train.shape[1:]
     training_in = Input(shape=training_in_shape)
     if bidi:
@@ -45,7 +43,7 @@ def get_trained_model(x_train, y_train, layer_name, n_recurrent_units, bidi):
 
 def get_test_model(n_recurrent_units, sequence_length, feature_dim, layer_name, stateful, initialize_states, weights,
                    bidi):
-    features_in = Input(batch_shape=(1, sequence_length, feature_dim))  ## stateful ==> needs batch_shape specified
+    features_in = Input(batch_shape=(1, sequence_length, feature_dim))  # stateful ==> needs batch_shape specified
     if layer_name == 'LSTM':
         REC_LAYER = LSTM
     else:
@@ -82,10 +80,9 @@ def get_test_model(n_recurrent_units, sequence_length, feature_dim, layer_name,
             state_h_in = Input(batch_shape=(1, n_recurrent_units))
             state_c_in = Input(batch_shape=(1, n_recurrent_units))
             if layer_name == 'LSTM':
-                recurrent_out = REC_LAYER(n_recurrent_units, return_sequences=True, stateful=stateful)(features_in,
-                                                                                                       initial_state=[
-                                                                                                           state_h_in,
-                                                                                                           state_c_in])
+                recurrent_out = REC_LAYER(n_recurrent_units,
+                                          return_sequences=True,
+                                          stateful=stateful)(features_in, initial_state=[state_h_in, state_c_in])
                 pred = Dense(1)(recurrent_out)
                 test_model = keras.Model(inputs=[features_in, state_h_in, state_c_in], outputs=pred)
             else:
@@ -132,7 +129,7 @@ def eval_test_model(baseline_out, test_model, x_in, layer_name, bidi, stateful,
             # msg += f'{layer_name}; Sequence {s}; stateful {stateful}; Initialzied state: {states_initialized}; State Reset: {state_reset}\n'
             if VERBOSE:
                 print(msg)
-            pred_in = x_in[s].reshape(x_inf[1:].shape)
+            pred_in = x_in[s].reshape(x_in[1:].shape)
             if not states_initialized:  # no initial state
                 pred_seq = test_model.predict(pred_in)
             else:
@@ -162,85 +159,92 @@ def eval_test_model(baseline_out, test_model, x_in, layer_name, bidi, stateful,
     return results
 
 
-print("Starting stateful recurrent tests", flush=True)
-
-##### generate toy data
-train_seq_length = 4
-feature_dim = 1
-num_seqs = 8
-
-x_train = np.random.normal(0, 1, num_seqs * train_seq_length * feature_dim)
-x_train = x_train.reshape(num_seqs, train_seq_length, feature_dim)
-y_train = np.random.normal(0, 1, num_seqs * train_seq_length)
-y_train = y_train.reshape(num_seqs, train_seq_length, 1)
-
-n_recurrent_units = 2
-all_results = np.zeros(0, dtype=np.float32)
-
-# hand-generate the input data to make it easy to input into C++ code by hand.
-# this hard-codes the train_seq_length, feature_dim vars for this purpose
-x_inf = np.asarray([2.1, -1.2, 3.14, 1.2, 1, 3, -2, 10], dtype=np.float32)  # simple
-x_inf = x_inf.reshape((2, train_seq_length, 1))
-
-initial_states = np.asarray([1.1, -2.1, 2.7, 3.1, -2.5, 3.0, -2.0, -10.0], dtype=np.float32)
-initial_states = initial_states.reshape((4, 1, 2))
-
-model_file_names = []
-
-for bidi in [False, True]:
-    for layer_name in ['GRU', 'LSTM']:
-        ### train with no initial state, no statefulness
-        training_model, trained_weights = get_trained_model(x_train, y_train, layer_name, n_recurrent_units, bidi)
-        y_inf = training_model.predict(x_inf)
-        for stateful in [False, True]:
-            for initialize_states in [False, True]:
-                ### evaluate the model
-                test_model, model_fname = get_test_model(n_recurrent_units, train_seq_length, feature_dim, layer_name,
-                                                         stateful, initialize_states, trained_weights, bidi)
-                result = eval_test_model(y_inf, test_model, x_inf, layer_name, bidi, stateful, initialize_states,
-                                         initial_states)
-                all_results = np.append(all_results, result)
-                model_file_names.append(model_fname)
-
-if VERBOSE:
-    print('\n\n')
-    print(all_results)
-    print(model_file_names)
-
-all_results.tofile('models/python_results.npy')
-
-for h5_fname in model_file_names:
-    json_fname = h5_fname.replace('.h5', '.json')
-    cmd = 'python3 ../../keras_export/convert_model.py ' + h5_fname + ' ' + json_fname
-    os.system(cmd)
-
-os.system('./stateful_recurrent_tests_cpp')
-
-frugally_deep_results = np.fromfile('models/fd_results.bin', dtype=np.float32)
-
-num_test_models = len(model_file_names)
-test_sequences_per_model = 4
-frugally_deep_results = frugally_deep_results.reshape((num_test_models, test_sequences_per_model * train_seq_length))
-all_results = all_results.reshape((num_test_models, test_sequences_per_model * train_seq_length))
-all_results = all_results.astype(np.float32)
-
-all_tests_passed = True
-for i, model_fname in enumerate(model_file_names):
-    test_name = os.path.basename(model_fname)
-    test_name = test_name.split('.h5')[0]
-    print('Test ', i + 1, ' ', test_name, ': ')
-    diff = np.abs(all_results[i] - frugally_deep_results[i])
-    max_delta = np.max(diff)
+def main():
+    print("Starting stateful recurrent tests")
+
+    # generate toy data
+    train_seq_length = 4
+    feature_dim = 1
+    num_seqs = 8
+
+    x_train = np.random.normal(0, 1, num_seqs * train_seq_length * feature_dim)
+    x_train = x_train.reshape(num_seqs, train_seq_length, feature_dim)
+    y_train = np.random.normal(0, 1, num_seqs * train_seq_length)
+    y_train = y_train.reshape(num_seqs, train_seq_length, 1)
+
+    n_recurrent_units = 2
+    all_results = np.zeros(0, dtype=np.float32)
+
+    # hand-generate the input data to make it easy to input into C++ code by hand.
+    # this hard-codes the train_seq_length, feature_dim vars for this purpose
+    x_inf = np.asarray([2.1, -1.2, 3.14, 1.2, 1, 3, -2, 10], dtype=np.float32)  # simple
+    x_inf = x_inf.reshape((2, train_seq_length, 1))
+
+    initial_states = np.asarray([1.1, -2.1, 2.7, 3.1, -2.5, 3.0, -2.0, -10.0], dtype=np.float32)
+    initial_states = initial_states.reshape((4, 1, 2))
+
+    model_file_names = []
+
+    for bidi in [False, True]:
+        for layer_name in ['GRU', 'LSTM']:
+            # train with no initial state, no statefulness
+            training_model, trained_weights = get_trained_model(x_train, y_train, layer_name, n_recurrent_units, bidi)
+            y_inf = training_model.predict(x_inf)
+            for stateful in [False, True]:
+                for initialize_states in [False, True]:
+                    # evaluate the model
+                    test_model, model_fname = get_test_model(n_recurrent_units, train_seq_length, feature_dim,
+                                                             layer_name,
+                                                             stateful, initialize_states, trained_weights, bidi)
+                    result = eval_test_model(y_inf, test_model, x_inf, layer_name, bidi, stateful, initialize_states,
+                                             initial_states)
+                    all_results = np.append(all_results, result)
+                    model_file_names.append(model_fname)
+
     if VERBOSE:
-        print('Max delta: ', "{:4.3e}".format(max_delta))
-    if max_delta < TEST_EPSILON:
-        print('PASSED')
-    else:
-        print('*********  FAILED !!!!!!!!!!!!\n\n')
-        print('Keras: ', all_results[i], '\n')
-        print('Frugally-deep: ', frugally_deep_results[i], '\n')
-        all_tests_passed = False
-
-if not all_tests_passed:
-    sys.exit(errno.EIO)
-print('\n\nPassed all stateful tests')
+        print('\n\n')
+        print(all_results)
+        print(model_file_names)
+
+    all_results.tofile('models/python_results.npy')
+
+    for h5_fname in model_file_names:
+        json_fname = h5_fname.replace('.h5', '.json')
+        cmd = 'python3 ../../keras_export/convert_model.py ' + h5_fname + ' ' + json_fname
+        os.system(cmd)
+
+    os.system('./stateful_recurrent_tests_cpp')
+
+    frugally_deep_results = np.fromfile('models/fd_results.bin', dtype=np.float32)
+
+    num_test_models = len(model_file_names)
+    test_sequences_per_model = 4
+    frugally_deep_results = frugally_deep_results.reshape(
+        (num_test_models, test_sequences_per_model * train_seq_length))
+    all_results = all_results.reshape((num_test_models, test_sequences_per_model * train_seq_length))
+    all_results = all_results.astype(np.float32)
+
+    all_tests_passed = True
+    for i, model_fname in enumerate(model_file_names):
+        test_name = os.path.basename(model_fname)
+        test_name = test_name.split('.h5')[0]
+        print('Test ', i + 1, ' ', test_name, ': ')
+        diff = np.abs(all_results[i] - frugally_deep_results[i])
+        max_delta = np.max(diff)
+        if VERBOSE:
+            print('Max delta: ', "{:4.3e}".format(max_delta))
+        if max_delta < TEST_EPSILON:
+            print('PASSED')
+        else:
+            print('********* FAILED !!!!!!!!!!!!\n\n')
+            print('Keras: ', all_results[i], '\n')
+            print('Frugally-deep: ', frugally_deep_results[i], '\n')
+            all_tests_passed = False
+
+    if not all_tests_passed:
+        sys.exit(errno.EIO)
+    print('\n\nPassed all stateful tests')
+
+
+if __name__ == "__main__":
+    main()