Fix #1, #2

src/cli.py

@@ -6,30 +6,34 @@ def cli():
     import argparse
     parser = argparse.ArgumentParser(description='CLI for speech recognition functionality.')
     parser.add_argument('action', metavar='ACTION', type=str, nargs=1,
-                        choices=['train', 'test', 'record', 'continuous'],
+                        choices=['isolated_train', 'isolated_test', 'record', 'continuous_train'],
                         help='Action to perform.')
-    parser.add_argument("-d", "--model-directory", default='models-4gaussians-em',
-                        help="Directory which the trained models are stored, or test models are used.")
+    parser.add_argument("-i", "--input", default='models-continuous-4gaussians-em-realign/',
+                        help="Directory which the input models are stored for continuous training, or test models are"
+                             "used. \nIgnored for isolated model training")
+    parser.add_argument("-o", "--output", default='models-continuous-4gaussians-em-realign/',
+                        help="Directory which the trained models are stored")
+    # TODO: use -g and -e for continuous training
     parser.add_argument("-g", "--gmm", help="Use GMM-HMM as the model.", default=False, action='store_true')
     parser.add_argument("-e", "--em", help="Use EM algorithm to train models.", default=False, action='store_true')
     args = parser.parse_args()
 
-    if args.action[0] == 'train':
+    if args.action[0] == 'isolated_train':
         print('training...')
         # data folder location
         folder = os.path.join(data_path, 'train')
         for digit in digit_names:
             filenames = [os.path.join(folder, f) for f in os.listdir(folder) if
                          re.match('[A-Z]+_' + digit + '[AB].wav', f)]
 
-            train(filenames, args.model_directory, digit, n_segs=5, use_gmm=args.gmm, use_em=args.em)
+            train(filenames, args.output, digit, n_segs=5, use_gmm=args.gmm, use_em=args.em)
 
-    if args.action[0] == 'test':
+    if args.action[0] == 'isolated_test':
         print('testing...')
         # get all the models from pickle files
         models = []
         for digit in digit_names:
-            file = open(os.path.join(args.model_directory, digit + '.pkl'), 'rb')
+            file = open(os.path.join(args.input, digit + '.pkl'), 'rb')
             models.append(pickle.load(file))
             file.close()
         # get file patterns for each digit
@@ -46,7 +50,7 @@ def cli():
         # get all the models from pickle files
         models = []
         for digit in digit_names:
-            file = open(args.model_directory + digit + '.pkl', 'rb')
+            file = open(args.input + digit + '.pkl', 'rb')
             models.append(pickle.load(file))
             file.close()
 
@@ -61,8 +65,8 @@ def cli():
                 best_model = i
         print(best_model)
 
-    if args.action[0] == 'continuous':
-        aurora_continuous_train()
+    if args.action[0] == 'continuous_train':
+        aurora_continuous_train(args.input, args.output)
 
 
 if __name__ == "__main__":

src/sr/core.py

@@ -54,9 +54,9 @@ def make_HMM(filenames, n_segs, use_gmm, use_em):
     return model
 
 
-def train(filenames, model_folder, model_name, n_segs, use_gmm, use_em):
+def train(filenames, output_path, model_name, n_segs, use_gmm, use_em):
     models = make_HMM(filenames, n_segs, use_gmm=use_gmm, use_em=use_em)
-    file = open(os.path.join(model_folder, model_name + '.pkl'), 'wb')
+    file = open(os.path.join(output_path, model_name + '.pkl'), 'wb')
     pickle.dump(models, file)
 
 
@@ -94,14 +94,12 @@ def test(models, folder, file_patterns):
     return n_passed / n_tests
 
 
-def aurora_continuous_train():
+def aurora_continuous_train(input_path, output_path):
     models = []
     hmm_index = 0
     # for digit in digit_names:
     for digit in range(11):
-        # TODO: use command line argument for input model path
-        file = open('models-continuous-4gaussians-em-realign/' + str(digit) + '.pkl', 'rb')
-        # file = open('models-4gaussians-em/' + str(digit) + '.pkl', 'rb')
+        file = open(input_path + str(digit) + '.pkl', 'rb')
 
         model: HMM = pickle.load(file)
         # set value of hmm_state.parent to the index of the hmm it belongs to
@@ -140,4 +138,4 @@ def aurora_continuous_train():
         pickle.dump(data, f)
         f.close()
 
-    continuous_train(data, models, labels)
+    continuous_train(data, models, labels, output_path)

src/sr/recognition/continuous_speech.py

@@ -5,7 +5,7 @@
 from .hmm import HMM
 import os
 import pickle
-from typing import List
+from typing import List, AnyStr
 import copy
 import numpy as np
 
@@ -53,7 +53,8 @@ def build_state_sequences(hmms: List[HMM], label_matrix: List[List[int]]):
     return seq, trans, end_state_indices[-1]
 
 
-def continuous_train(data: List[np.ndarray], models: List[HMM], label_seqs: List[List[int]], n_gaussians: int = 4,
+def continuous_train(data: List[np.ndarray], models: List[HMM], label_seqs: List[List[int]], output_path: AnyStr,
+                     n_gaussians: int = 4,
                      n_segments: int = 5,
                      max_iteration: int = 1000):
     # remember old models
@@ -163,9 +164,8 @@ def continuous_train(data: List[np.ndarray], models: List[HMM], label_seqs: List
                 new_models[mi].transitions[si, si] = -np.log(1 - p_jump)
 
         # save new models to files
-        # TODO: use command line argument for output model path
         for i in range(len(new_models)):
-            file = open(os.path.join('models-continuous-4gaussians-em-realign', str(i) + '.pkl'), 'wb')
+            file = open(os.path.join(output_path, str(i) + '.pkl'), 'wb')
             pickle.dump(new_models[i], file)
             file.close()
         # check if converged