Realtime translation (mostly a cdec wrapper for now)

7 files changed, 200 insertions, 38 deletions

diff --git a/realtime/README.md b/realtime/README.md
new file mode 100644
index 00000000..b37dddc8
--- /dev/null
+++ b/realtime/README.md
@@ -0,0 +1 @@
+More to come.
diff --git a/realtime/realtime.py b/realtime/realtime.py
new file mode 100755
index 00000000..a6b3ed52
--- /dev/null
+++ b/realtime/realtime.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python
+import argparse
+import os
+import logging
+import cdec.configobj
+import cdec.sa
+import shutil
+import sys
+import subprocess
+import tempfile
+import time
+
+from rt import ForceAligner
+from rt import CdecDecoder
+
+class RealtimeDecoder:
+
+    def __init__(self, configdir, tmpdir='/tmp'):
+
+        # Temporary work dir
+        self.tmp = tempfile.mkdtemp(dir=tmpdir)
+        logging.info('Using temp dir {}'.format(self.tmp))
+
+        # Word aligner
+        fwd_params = os.path.join(configdir, 'a.fwd_params')
+        fwd_err = os.path.join(configdir, 'a.fwd_err')
+        rev_params = os.path.join(configdir, 'a.rev_params')
+        rev_err = os.path.join(configdir, 'a.rev_err')
+        self.aligner = ForceAligner(fwd_params, fwd_err, rev_params, rev_err)
+
+        # Grammar extractor
+        sa_config = os.path.join(configdir, 'sa.ini')
+        self.extractor = cdec.sa.GrammarExtractor(sa_config)
+
+        # Decoder
+        decoder_config = os.path.join(configdir, 'cdec.ini')
+        decoder_weights = os.path.join(configdir, 'weights.final')
+        self.decoder = CdecDecoder(decoder_config, decoder_weights)
+
+    def close(self):
+        logging.info('Closing processes')
+        self.aligner.close()
+        self.decoder.close()
+        logging.info('Deleting {}'.format(self.tmp))
+        shutil.rmtree(self.tmp)
+
+    def grammar(self, sentence):
+        grammar_file = tempfile.mkstemp(dir=self.tmp)[1]
+        with open(grammar_file, 'w') as output:
+            for rule in self.extractor.grammar(sentence):
+                output.write(str(rule) + '\n')
+        return grammar_file
+        
+    def decode(self, sentence):
+        grammar_file = self.grammar(sentence)
+        start_time = time.time()
+        hyp = self.decoder.decode(sentence, grammar_file)
+        stop_time = time.time()
+        logging.info('Translation time: {} seconds'.format(stop_time - start_time))
+        os.remove(grammar_file)
+        return hyp
+
+    def learn(self, source, target):
+        pass
+
+def main():
+
+    parser = argparse.ArgumentParser(description='Real-time adaptive translation with cdec.')
+    parser.add_argument('-c', '--config', required=True, help='Config directory (see README.md)')
+    parser.add_argument('-T', '--temp', help='Temp directory (default /tmp)', default='/tmp')
+    parser.add_argument('-v', '--verbose', help='Info to stderr', action='store_true')
+    args = parser.parse_args()
+
+    if not args.config:
+        parser.error('specify a configuration directory')
+
+    if args.verbose:
+        logging.basicConfig(level=logging.INFO)
+
+    rtd = RealtimeDecoder(args.config)
+
+    try:
+        for line in sys.stdin:
+            input = [f.strip() for f in line.split('|||')]
+            if len(input) == 1:
+                hyp = rtd.decode(input[0])
+                sys.stdout.write('{}\n'.format(hyp))
+            elif len(input) == 2:
+                rtd.learn(*input)
+
+    # Clean exit on ctrl+c
+    except KeyboardInterrupt:
+        logging.info('Caught KeyboardInterrupt, exiting')
+
+    # Cleanup
+    rtd.close()
+
+
+def mkconfig():
+    pass
+
+if __name__ == '__main__':
+    main()
diff --git a/realtime/rt/__init__.py b/realtime/rt/__init__.py
new file mode 100644
index 00000000..7a1aeda7
--- /dev/null
+++ b/realtime/rt/__init__.py
@@ -0,0 +1,2 @@
+from aligner import *
+from decoder import *
diff --git a/realtime/rt/aligner.py b/realtime/rt/aligner.py
new file mode 100644
index 00000000..d94dbda0
--- /dev/null
+++ b/realtime/rt/aligner.py
@@ -0,0 +1,50 @@
+import os
+import sys
+import subprocess
+
+import util
+
+class ForceAligner:
+
+    def __init__(self, fwd_params, fwd_err, rev_params, rev_err):
+
+        cdec_root = os.path.dirname(os.path.dirname(os.path.dirname(__file__)))
+        fast_align = os.path.join(cdec_root, 'word-aligner', 'fast_align')
+        atools = os.path.join(cdec_root, 'utils', 'atools')
+
+        (fwd_T, fwd_m) = self.read_err(fwd_err)
+        (rev_T, rev_m) = self.read_err(rev_err)
+
+        fwd_cmd = [fast_align, '-i', '-', '-d', '-T', fwd_T, '-m', fwd_m, '-f', fwd_params]
+        rev_cmd = [fast_align, '-i', '-', '-d', '-T', rev_T, '-m', rev_m, '-f', rev_params, '-r']
+        tools_cmd = [atools, '-i', '-', '-j', '-', '-c', 'grow-diag-final-and']
+
+        self.fwd_align = util.popen_io(fwd_cmd)
+        self.rev_align = util.popen_io(rev_cmd)
+        self.tools = util.popen_io(tools_cmd)
+
+    def align(self, line):
+        self.fwd_align.stdin.write('{}\n'.format(line))
+        self.rev_align.stdin.write('{}\n'.format(line))
+        # f words ||| e words ||| links ||| score
+        fwd_line = self.fwd_align.stdout.readline().split('|||')[2].strip()
+        rev_line = self.rev_align.stdout.readline().split('|||')[2].strip()
+        self.tools.stdin.write('{}\n'.format(fwd_line))
+        self.tools.stdin.write('{}\n'.format(rev_line))
+        return self.tools.stdout.readline().strip()
+ 
+    def close(self):
+        self.fwd_align.stdin.close()
+        self.rev_align.stdin.close()
+        self.tools.stdin.close()
+
+    def read_err(self, err):
+        (T, m) = ('', '')
+        for line in open(err):
+            # expected target length = source length * N
+            if 'expected target length' in line:
+                m = line.split()[-1]
+            # final tension: N
+            elif 'final tension' in line:
+                T = line.split()[-1]
+        return (T, m)
diff --git a/realtime/rt/decoder.py b/realtime/rt/decoder.py
new file mode 100644
index 00000000..f4fea0e2
--- /dev/null
+++ b/realtime/rt/decoder.py
@@ -0,0 +1,23 @@
+import os
+import subprocess
+
+import util
+
+class Decoder:
+
+    def close(self):
+        self.decoder.stdin.close()
+
+class CdecDecoder(Decoder):
+    
+    def __init__(self, config, weights):
+        cdec_root = os.path.dirname(os.path.dirname(os.path.dirname(__file__)))
+        decoder = os.path.join(cdec_root, 'decoder', 'cdec')
+        decoder_cmd = [decoder, '-c', config, '-w', weights]
+        self.decoder = util.popen_io(decoder_cmd)
+
+    def decode(self, sentence, grammar):
+        input = '<seg grammar="{g}">{s}</seg>\n'.format(i=id, s=sentence, g=grammar)
+        self.decoder.stdin.write(input)
+        return self.decoder.stdout.readline().strip()
+
diff --git a/realtime/rt/util.py b/realtime/rt/util.py
new file mode 100644
index 00000000..7f877161
--- /dev/null
+++ b/realtime/rt/util.py
@@ -0,0 +1,13 @@
+import subprocess
+import threading
+
+def popen_io(cmd):
+    p = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    consume_stream(p.stderr)
+    return p
+
+def consume_stream(stream):
+    def consume(s):
+        for _ in s:
+            pass
+    threading.Thread(target=consume, args=(stream,)).start()
diff --git a/word-aligner/force_align.py b/word-aligner/force_align.py
index a0c1aad7..ad6d95fa 100755
--- a/word-aligner/force_align.py
+++ b/word-aligner/force_align.py
@@ -2,7 +2,6 @@
 
 import os
 import sys
-import subprocess
 
 def main():
 
@@ -15,45 +14,16 @@ def main():
         sys.stderr.write('  {} fwd_params fwd_err rev_params rev_err <in.f-e >out.f-e.gdfa\n'.format(sys.argv[0]))
         sys.exit(2)
 
-    (f_p, f_err, r_p, r_err) = sys.argv[1:]
-    
-    (f_T, f_m) = find_Tm(f_err)
-    (r_T, r_m) = find_Tm(r_err)
-    
-    fast_align = os.path.join(os.path.dirname(__file__), 'fast_align')
-    f_cmd = [fast_align, '-i', '-', '-d', '-T', f_T, '-m', f_m, '-f', f_p]
-    r_cmd = [fast_align, '-i', '-', '-d', '-T', r_T, '-m', r_m, '-f', r_p, '-r']
-
-    atools = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'utils', 'atools')
-    tools_cmd = [atools, '-i', '-', '-j', '-', '-c', 'grow-diag-final-and']
-
-    sys.stderr.write('running: {}\n'.format(' '.join(f_cmd)))
-    sys.stderr.write('running: {}\n'.format(' '.join(r_cmd)))
-    sys.stderr.write('running: {}\n'.format(' '.join(tools_cmd)))
-
-    f_a = subprocess.Popen(f_cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
-    r_a = subprocess.Popen(r_cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
-    tools = subprocess.Popen(tools_cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+    # Hook into realtime
+    sys.path.append(os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), 'realtime'))
+    from rt import ForceAligner
 
+    aligner = ForceAligner(*sys.argv[1:])
+    
     for line in sys.stdin:
-        f_a.stdin.write(line)
-        r_a.stdin.write(line)
-        # f words ||| e words ||| links ||| score
-        f_line = f_a.stdout.readline().split('|||')[2].strip()
-        r_line = r_a.stdout.readline().split('|||')[2].strip()
-        tools.stdin.write('{}\n'.format(f_line))
-        tools.stdin.write('{}\n'.format(r_line))
-        sys.stdout.write(tools.stdout.readline())
-
-def find_Tm(err):
-    (T, m) = ('', '')
-    for line in open(err):
-        # expected target length = source length * N
-        if 'expected target length' in line:
-            m = line.split()[-1]
-        elif 'final tension' in line:
-            T = line.split()[-1]
-    return (T, m)
+        sys.stdout.write('{}\n'.format(aligner.align(line.strip())))
 
+    aligner.close()
+    
 if __name__ == '__main__':
     main()