Trainola inserts into the destination stream now

This makes it easier to use the filter stuff from other code, but it's
also turning it into more of a spaghetti nightmare.  Might not be
worth continuing down this path.

Makefile

@@ -8,22 +8,30 @@ else
 	@echo "Try 'make install'"
 endif
 
-test: test_cleanup
+test: test_trainola
+
+test_trainola:
+	-nilmtool -u http://bucket/nilmdb remove -s min -e max \
+		/sharon/prep-a-matches
+	nilmtools/trainola.py "$$(cat extras/trainola-test-param.js)"
 
 test_cleanup:
-	src/cleanup.py -e extras/cleanup.cfg
-	src/cleanup.py extras/cleanup.cfg
+	nilmtools/cleanup.py -e extras/cleanup.cfg
+	nilmtools/cleanup.py extras/cleanup.cfg
 
 test_insert:
-	@make install >/dev/null
-	src/insert.py --file --dry-run  /test/foo </dev/null
+	nilmtools/insert.py --file --dry-run  /test/foo </dev/null
 
 test_copy:
-	@make install >/dev/null
-	src/copy_wildcard.py -U "http://nilmdb.com/bucket/" -D /lees*
+	nilmtools/copy_wildcard.py -U "http://nilmdb.com/bucket/" -D /lees*
 
-test_prep:
-	@make install >/dev/null
+/tmp/raw.dat:
+	octave --eval 'fs = 8000;' \
+	--eval 't = (0:fs*10)*2*pi*60/fs;' \
+	--eval 'raw = transpose([sin(t); 0.3*sin(3*t)+sin(t)]);' \
+	--eval 'save("-ascii","/tmp/raw.dat","raw");'
+
+test_prep: /tmp/raw.dat
 	-nilmtool destroy -R /test/raw
 	-nilmtool destroy -R /test/sinefit
 	-nilmtool destroy -R /test/prep
@@ -31,8 +39,8 @@ test_prep:
 	nilmtool create /test/sinefit float32_3
 	nilmtool create /test/prep float32_8
 	nilmtool insert -s '@0' -t -r 8000 /test/raw /tmp/raw.dat
-	src/sinefit.py -c 1 /test/raw /test/sinefit
-	src/prep.py -c 2 /test/raw /test/sinefit /test/prep
+	nilmtools/sinefit.py -a 0.5 -c 1 /test/raw /test/sinefit
+	nilmtools/prep.py -c 2 /test/raw /test/sinefit /test/prep
 	nilmtool extract -s min -e max /test/prep | head -20
 
 test_decimate:
@@ -40,8 +48,8 @@ test_decimate:
 	-@nilmtool destroy /lees-compressor/no-leak/raw/16 || true
 	-@nilmtool create /lees-compressor/no-leak/raw/4 float32_18 || true
 	-@nilmtool create /lees-compressor/no-leak/raw/16 float32_18 || true
-	time python src/decimate.py -s '2013-02-04 18:10:00' -e '2013-02-04 18:11:00' /lees-compressor/no-leak/raw/1 /lees-compressor/no-leak/raw/4
-	python src/decimate.py -s '2013-02-04 18:10:00' -e '2013-02-04 18:11:00' /lees-compressor/no-leak/raw/4 /lees-compressor/no-leak/raw/16
+	time python nilmtools/decimate.py -s '2013-02-04 18:10:00' -e '2013-02-04 18:11:00' /lees-compressor/no-leak/raw/1 /lees-compressor/no-leak/raw/4
+	python nilmtools/decimate.py -s '2013-02-04 18:10:00' -e '2013-02-04 18:11:00' /lees-compressor/no-leak/raw/4 /lees-compressor/no-leak/raw/16
 
 version:
 	python setup.py version
@@ -63,4 +71,4 @@ clean::
 gitclean::
 	git clean -dXf
 
-.PHONY: all version dist sdist install clean gitclean
+.PHONY: all version dist sdist install clean gitclean test

README.txt

@@ -5,10 +5,10 @@ by Jim Paris <jim@jtan.com>
 Prerequisites:
 
   # Runtime and build environments
-  sudo apt-get install python2.7 python2.7-dev python-setuptools
-  sudo apt-get install python-numpy python-scipy python-matplotlib
+  sudo apt-get install python2.7 python2.7-dev python-setuptools python-pip
+  sudo apt-get install python-numpy python-scipy
 
-  nilmdb (1.5.0+)
+  nilmdb (1.8.1+)
 
 Install:
 

extras/trainola-test-param.js

@@ -0,0 +1,31 @@
+{ "url": "http://bucket.mit.edu/nilmdb",
+  "dest_stream": "/sharon/prep-a-matches",
+  "stream": "/sharon/prep-a",
+  "start": 1366111383280463,
+  "end": 1366126163457797,
+  "columns": [ { "name": "P1", "index": 0 },
+               { "name": "Q1", "index": 1 },
+               { "name": "P3", "index": 2 } ],
+  "exemplars": [
+      { "name": "Boiler Pump ON",
+        "url": "http://bucket.mit.edu/nilmdb",
+        "stream": "/sharon/prep-a",
+        "start": 1366260494269078,
+        "end": 1366260608185031,
+        "dest_column": 0,
+        "columns": [ { "name": "P1", "index": 0 },
+                     { "name": "Q1", "index": 1 }
+                   ]
+      },
+      { "name": "Boiler Pump OFF",
+        "url": "http://bucket.mit.edu/nilmdb",
+        "stream": "/sharon/prep-a",
+        "start": 1366260864215764,
+        "end": 1366260870882998,
+        "dest_column": 1,
+        "columns": [ { "name": "P1", "index": 0 },
+                     { "name": "Q1", "index": 1 }
+                   ]
+      }
+  ]
+}

nilmtools/_version.py

@@ -181,7 +181,7 @@ def versions_from_parentdir(parentdir_prefix, versionfile_source, verbose=False)
 
 tag_prefix = "nilmtools-"
 parentdir_prefix = "nilmtools-"
-versionfile_source = "src/_version.py"
+versionfile_source = "nilmtools/_version.py"
 
 def get_versions(default={"version": "unknown", "full": ""}, verbose=False):
     variables = { "refnames": git_refnames, "full": git_full }

nilmtools/cleanup.py

@@ -238,12 +238,15 @@ def main(argv = None):
                        timestamp_to_seconds(total)))
             continue
         printf("  removing data before %s\n", timestamp_to_human(remove_before))
-        if args.yes:
-            client.stream_remove(path, None, remove_before)
-        for ap in streams[path].also_clean_paths:
-            printf("  also removing from %s\n", ap)
+        # Clean in reverse order.  Since we only use the primary stream and not
+        # the decimated streams to figure out which data to remove, removing
+        # the primary stream last means that we might recover more nicely if
+        # we are interrupted and restarted.
+        clean_paths = list(reversed(streams[path].also_clean_paths)) + [ path ]
+        for p in clean_paths:
+            printf("  removing from %s\n", p)
             if args.yes:
-                client.stream_remove(ap, None, remove_before)
+                client.stream_remove(p, None, remove_before)
 
     # All done
     if not args.yes:

nilmtools/decimate_auto.py

@@ -4,15 +4,19 @@ import nilmtools.filter
 import nilmtools.decimate
 import nilmdb.client
 import argparse
+import fnmatch
 
 def main(argv = None):
     parser = argparse.ArgumentParser(
         formatter_class = argparse.RawDescriptionHelpFormatter,
-        version = "1.0",
+        version = nilmtools.__version__,
         description = """\
     Automatically create multiple decimations from a single source
     stream, continuing until the last decimated level contains fewer
     than 500 points total.
+
+    Wildcards and multiple paths are accepted.  Decimated paths are
+    ignored when matching wildcards.
     """)
     parser.add_argument("-u", "--url", action="store",
                         default="http://localhost/nilmdb/",
@@ -23,20 +27,36 @@ def main(argv = None):
                         default = False,
                         help="Force metadata changes if the dest "
                         "doesn't match")
-    parser.add_argument("path", action="store",
+    parser.add_argument("path", action="store", nargs='+',
                         help='Path of base stream')
     args = parser.parse_args(argv)
 
     # Pull out info about the base stream
     client = nilmdb.client.Client(args.url)
 
-    info = nilmtools.filter.get_stream_info(client, args.path)
-    if not info:
-        raise Exception("path " + args.path + " not found")
+    # Find list of paths to process
+    streams = [ unicode(s[0]) for s in client.stream_list() ]
+    streams = [ s for s in streams if "~decim-" not in s ]
+    paths = []
+    for path in args.path:
+        new = fnmatch.filter(streams, unicode(path))
+        if not new:
+            print "error: no stream matched path:", path
+            raise SystemExit(1)
+        paths.extend(new)
 
-    meta = client.stream_get_metadata(args.path)
+    for path in paths:
+        do_decimation(client, args, path)
+
+def do_decimation(client, args, path):
+    print "Decimating", path
+    info = nilmtools.filter.get_stream_info(client, path)
+    if not info:
+        raise Exception("path " + path + " not found")
+
+    meta = client.stream_get_metadata(path)
     if "decimate_source" in meta:
-        print "Stream", args.path, "was decimated from", meta["decimate_source"]
+        print "Stream", path, "was decimated from", meta["decimate_source"]
         print "You need to pass the base stream instead"
         raise SystemExit(1)
 
@@ -53,7 +73,7 @@ def main(argv = None):
         if info.rows <= 500:
             break
         factor *= args.factor
-        new_path = "%s~decim-%d" % (args.path, factor)
+        new_path = "%s~decim-%d" % (path, factor)
 
         # Create the stream if needed
         new_info = nilmtools.filter.get_stream_info(client, new_path)
@@ -72,5 +92,7 @@ def main(argv = None):
         # Update info using the newly decimated stream
         info = nilmtools.filter.get_stream_info(client, new_path)
 
+    return
+
 if __name__ == "__main__":
     main()

nilmtools/filter.py

@@ -19,6 +19,10 @@ import re
 import argparse
 import numpy as np
 import cStringIO
+import functools
+
+class ArgumentError(Exception):
+    pass
 
 class MissingDestination(Exception):
     def __init__(self, args, src, dest):
@@ -65,9 +69,73 @@ def get_stream_info(client, path):
         return None
     return StreamInfo(client.geturl(), streams[0])
 
+# Filter processing for a single interval of data.
+def process_numpy_interval(interval, extractor, inserter, warn_rows,
+                           function, args = None):
+    """For the given 'interval' of data, extract data, process it
+    through 'function', and insert the result.
+
+    'extractor' should be a function like NumpyClient.stream_extract_numpy
+    but with the the interval 'start' and 'end' as the only parameters,
+    e.g.:
+       extractor = functools.partial(NumpyClient.stream_extract_numpy,
+                                     src_path, layout = l, maxrows = m)
+
+    'inserter' should be a function like NumpyClient.stream_insert_context
+    but with the interval 'start' and 'end' as the only parameters, e.g.:
+       inserter = functools.partial(NumpyClient.stream_insert_context,
+                                    dest_path)
+
+    If 'warn_rows' is not None, print a warning to stdout when the
+    number of unprocessed rows exceeds this amount.
+
+    See process_numpy for details on 'function' and 'args'.
+    """
+    if args is None:
+        args = []
+
+    with inserter(interval.start, interval.end) as insert_ctx:
+        insert_func = insert_ctx.insert
+        old_array = np.array([])
+        for new_array in extractor(interval.start, interval.end):
+            # If we still had old data left, combine it
+            if old_array.shape[0] != 0:
+                array = np.vstack((old_array, new_array))
+            else:
+                array = new_array
+
+            # Pass the data to the user provided function
+            processed = function(array, interval, args, insert_func, False)
+
+            # Send any pending data that the user function inserted
+            insert_ctx.send()
+
+            # Save the unprocessed parts
+            if processed >= 0:
+                old_array = array[processed:]
+            else:
+                raise Exception(
+                    sprintf("%s return value %s must be >= 0",
+                            str(function), str(processed)))
+
+            # Warn if there's too much data remaining
+            if warn_rows is not None and old_array.shape[0] > warn_rows:
+                printf("warning: %d unprocessed rows in buffer\n",
+                       old_array.shape[0])
+
+        # Last call for this contiguous interval
+        if old_array.shape[0] != 0:
+            processed = function(old_array, interval, args,
+                                 insert_func, True)
+            if processed != old_array.shape[0]:
+                # Truncate the interval we're inserting at the first
+                # unprocessed data point.  This ensures that
+                # we'll not miss any data when we run again later.
+                insert_ctx.update_end(old_array[processed][0])
+
 class Filter(object):
 
-    def __init__(self):
+    def __init__(self, parser_description = None):
         self._parser = None
         self._client_src = None
         self._client_dest = None
@@ -78,6 +146,9 @@ class Filter(object):
         self.end = None
         self.interhost = False
         self.force_metadata = False
+        if parser_description is not None:
+            self.setup_parser(parser_description)
+            self.parse_args()
 
     @property
     def client_src(self):
@@ -131,63 +202,52 @@ class Filter(object):
         self._parser = parser
         return parser
 
-    def interval_string(self, interval):
-        return sprintf("[ %s -> %s ]",
-                       timestamp_to_human(interval.start),
-                       timestamp_to_human(interval.end))
-
-    def parse_args(self, argv = None):
-        args = self._parser.parse_args(argv)
-
-        if args.dest_url is None:
-            args.dest_url = args.url
-        if args.url != args.dest_url:
+    def set_args(self, url, dest_url, srcpath, destpath, start, end,
+                 parsed_args = None, quiet = True):
+        """Set arguments directly from parameters"""
+        if dest_url is None:
+            dest_url = url
+        if url != dest_url:
             self.interhost = True
 
-        self._client_src = Client(args.url)
-        self._client_dest = Client(args.dest_url)
+        self._client_src = Client(url)
+        self._client_dest = Client(dest_url)
 
-        if (not self.interhost) and (args.srcpath == args.destpath):
-            self._parser.error("source and destination path must be different")
+        if (not self.interhost) and (srcpath == destpath):
+            raise ArgumentError("source and destination path must be different")
 
-        # Open and print info about the streams
-        self.src = get_stream_info(self._client_src, args.srcpath)
+        # Open the streams
+        self.src = get_stream_info(self._client_src, srcpath)
         if not self.src:
-            self._parser.error("source path " + args.srcpath + " not found")
+            raise ArgumentError("source path " + srcpath + " not found")
 
-        self.dest = get_stream_info(self._client_dest, args.destpath)
+        self.dest = get_stream_info(self._client_dest, destpath)
         if not self.dest:
-            raise MissingDestination(args, self.src,
-                                     StreamInfo(args.dest_url, [args.destpath]))
+            raise MissingDestination(parsed_args, self.src,
+                                     StreamInfo(dest_url, [destpath]))
 
-        print "Source:", self.src.string(self.interhost)
-        print "  Dest:", self.dest.string(self.interhost)
+        self.start = start
+        self.end = end
 
-        if args.dry_run:
-            for interval in self.intervals():
-                print self.interval_string(interval)
-            raise SystemExit(0)
+        # Print info
+        if not quiet:
+            print "Source:", self.src.string(self.interhost)
+            print "  Dest:", self.dest.string(self.interhost)
+
+    def parse_args(self, argv = None):
+        """Parse arguments from a command line"""
+        args = self._parser.parse_args(argv)
+
+        self.set_args(args.url, args.dest_url, args.srcpath, args.destpath,
+                      args.start, args.end, quiet = False, parsed_args = args)
 
         self.force_metadata = args.force_metadata
-
-        self.start = args.start
-        self.end = args.end
-
+        if args.dry_run:
+            for interval in self.intervals():
+                print interval.human_string()
+            raise SystemExit(0)
         return args
 
-    def _optimize_int(self, it):
-        """Join and yield adjacent intervals from the iterator 'it'"""
-        saved_int = None
-        for interval in it:
-            if saved_int is not None:
-                if saved_int.end == interval.start:
-                    interval.start = saved_int.start
-                else:
-                    yield saved_int
-            saved_int = interval
-        if saved_int is not None:
-            yield saved_int
-
     def intervals(self):
         """Generate all the intervals that this filter should process"""
         self._using_client = True
@@ -214,12 +274,13 @@ class Filter(object):
                               self.src.path, diffpath = self.dest.path,
                               start = self.start, end = self.end) )
         # Optimize intervals: join intervals that are adjacent
-        for interval in self._optimize_int(intervals):
+        for interval in nilmdb.utils.interval.optimize(intervals):
             yield interval
         self._using_client = False
 
     # Misc helpers
-    def arg_time(self, toparse):
+    @staticmethod
+    def arg_time(toparse):
         """Parse a time string argument"""
         try:
             return nilmdb.utils.time.parse_time(toparse)
@@ -233,8 +294,14 @@ class Filter(object):
         metadata = self._client_dest.stream_get_metadata(self.dest.path)
         if not self.force_metadata:
             for key in data:
-                wanted = str(data[key])
+                wanted = data[key]
+                if not isinstance(wanted, basestring):
+                    wanted = str(wanted)
                 val = metadata.get(key, wanted)
+                # Force UTF-8 encoding for comparison and display
+                wanted = wanted.encode('utf-8')
+                val = val.encode('utf-8')
+                key = key.encode('utf-8')
                 if val != wanted and self.dest.rows > 0:
                     m =  "Metadata in destination stream:\n"
                     m += "  %s = %s\n" % (key, val)
@@ -250,13 +317,16 @@ class Filter(object):
 
     # The main filter processing method.
     def process_numpy(self, function, args = None, rows = 100000):
-        """For all intervals that exist in self.src but don't exist in
-        self.dest, call 'function' with a Numpy array corresponding to
-        the data.  The data is converted to a Numpy array in chunks of
-        'rows' rows at a time.
+        """Calls process_numpy_interval for each interval that currently
+        exists in self.src, but doesn't exist in self.dest.  It will
+        process the data in chunks as follows:
+
+        For each chunk of data, call 'function' with a Numpy array
+        corresponding to the data.  The data is converted to a Numpy
+        array in chunks of 'rows' rows at a time.
 
         'function' should be defined as:
-           def function(data, interval, args, insert_func, final)
+        # def function(data, interval, args, insert_func, final)
 
         'data': array of data to process -- may be empty
 
@@ -275,51 +345,23 @@ class Filter(object):
         Return value of 'function' is the number of data rows processed.
         Unprocessed data will be provided again in a subsequent call
         (unless 'final' is True).
+
+        If unprocessed data remains after 'final' is True, the interval
+        being inserted will be ended at the timestamp of the first
+        unprocessed data point.
         """
-        if args is None:
-            args = []
         extractor = NumpyClient(self.src.url).stream_extract_numpy
         inserter = NumpyClient(self.dest.url).stream_insert_numpy_context
 
-        for interval in self.intervals():
-            print "Processing", self.interval_string(interval)
-            with inserter(self.dest.path,
-                          interval.start, interval.end) as insert_ctx:
-                insert_function = insert_ctx.insert
-                old_array = np.array([])
-                for new_array in extractor(self.src.path,
-                                           interval.start, interval.end,
+        extractor_func = functools.partial(extractor, self.src.path,
                                            layout = self.src.layout,
-                                           maxrows = rows):
-                    # If we still had old data left, combine it
-                    if old_array.shape[0] != 0:
-                        array = np.vstack((old_array, new_array))
-                    else:
-                        array = new_array
+                                           maxrows = rows)
+        inserter_func = functools.partial(inserter, self.dest.path)
 
-                    # Pass it to the process function
-                    processed = function(array, interval, args,
-                                         insert_function, False)
-
-                    # Send any pending data
-                    insert_ctx.send()
-
-                    # Save the unprocessed parts
-                    if processed >= 0:
-                        old_array = array[processed:]
-                    else:
-                        raise Exception(
-                            sprintf("%s return value %s must be >= 0",
-                                    str(function), str(processed)))
-
-                    # Warn if there's too much data remaining
-                    if old_array.shape[0] > 3 * rows:
-                        printf("warning: %d unprocessed rows in buffer\n",
-                               old_array.shape[0])
-
-                # Last call for this contiguous interval
-                if old_array.shape[0] != 0:
-                    function(old_array, interval, args, insert_function, True)
+        for interval in self.intervals():
+            print "Processing", interval.human_string()
+            process_numpy_interval(interval, extractor_func, inserter_func,
+                                   rows * 3, function, args)
 
 def main(argv = None):
     # This is just a dummy function; actual filters can use the other
@@ -328,7 +370,7 @@ def main(argv = None):
     parser = f.setup_parser()
     args = f.parse_args(argv)
     for i in f.intervals():
-        print "Generic filter: need to handle", f.interval_string(i)
+        print "Generic filter: need to handle", i.human_string()
 
 if __name__ == "__main__":
     main()

nilmtools/median.py

@@ -0,0 +1,43 @@
+#!/usr/bin/python
+import nilmtools.filter, scipy.signal
+
+def main(argv = None):
+    f = nilmtools.filter.Filter()
+    parser = f.setup_parser("Median Filter")
+    group = parser.add_argument_group("Median filter options")
+    group.add_argument("-z", "--size", action="store", type=int, default=25,
+                       help = "median filter size (default %(default)s)")
+    group.add_argument("-d", "--difference", action="store_true",
+                       help = "store difference rather than filtered values")
+
+    try:
+        args = f.parse_args(argv)
+    except nilmtools.filter.MissingDestination as e:
+        print "Source is %s (%s)" % (e.src.path, e.src.layout)
+        print "Destination %s doesn't exist" % (e.dest.path)
+        print "You could make it with a command like:"
+        print "  nilmtool -u %s create %s %s" % (e.dest.url,
+                                                 e.dest.path, e.src.layout)
+        raise SystemExit(1)
+
+    meta = f.client_src.stream_get_metadata(f.src.path)
+    f.check_dest_metadata({ "median_filter_source": f.src.path,
+                            "median_filter_size": args.size,
+                            "median_filter_difference": repr(args.difference) })
+
+    f.process_numpy(median_filter, args = (args.size, args.difference))
+
+def median_filter(data, interval, args, insert, final):
+    (size, diff) = args
+    (rows, cols) = data.shape
+    for i in range(cols - 1):
+        filtered = scipy.signal.medfilt(data[:, i+1], size)
+        if diff:
+            data[:, i+1] -= filtered
+        else:
+            data[:, i+1] = filtered
+    insert(data)
+    return rows
+
+if __name__ == "__main__":
+    main()

nilmtools/prep.py

@@ -3,6 +3,8 @@
 # Spectral envelope preprocessor.
 # Requires two streams as input: the original raw data, and sinefit data.
 
+from nilmdb.utils.printf import *
+from nilmdb.utils.time import timestamp_to_human
 import nilmtools.filter
 import nilmdb.client
 from numpy import *
@@ -77,7 +79,8 @@ def main(argv = None):
     # Check and set metadata in prep stream
     f.check_dest_metadata({ "prep_raw_source": f.src.path,
                             "prep_sinefit_source": sinefit.path,
-                            "prep_column": args.column })
+                            "prep_column": args.column,
+                            "prep_rotation": repr(rotation) })
 
     # Run the processing function on all data
     f.process_numpy(process, args = (client_sinefit, sinefit.path, args.column,
@@ -105,7 +108,6 @@ def process(data, interval, args, insert_function, final):
     # Pull out sinefit data for the entire time range of this block
     for sinefit_line in client.stream_extract(sinefit_path,
                                               data[0, 0], data[rows-1, 0]):
-
         def prep_period(t_min, t_max, rot):
             """
             Compute prep coefficients from time t_min to t_max, which
@@ -162,7 +164,15 @@ def process(data, interval, args, insert_function, final):
                 break
             processed = idx_max
 
-    print "Processed", processed, "of", rows, "rows"
+    # If we processed no data but there's lots in here, pretend we
+    # processed half of it.
+    if processed == 0 and rows > 10000:
+        processed = rows / 2
+        printf("%s: warning: no periods found; skipping %d rows\n",
+               timestamp_to_human(data[0][0]), processed)
+    else:
+        printf("%s: processed %d of %d rows\n",
+               timestamp_to_human(data[0][0]), processed, rows)
     return processed
 
 if __name__ == "__main__":

nilmtools/sinefit.py

@@ -0,0 +1,262 @@
+#!/usr/bin/python
+
+# Sine wave fitting.
+from nilmdb.utils.printf import *
+import nilmtools.filter
+import nilmdb.client
+from nilmdb.utils.time import (timestamp_to_human,
+                               timestamp_to_seconds,
+                               seconds_to_timestamp)
+
+from numpy import *
+from scipy import *
+#import pylab as p
+import operator
+import sys
+
+def main(argv = None):
+    f = nilmtools.filter.Filter()
+    parser = f.setup_parser("Sine wave fitting")
+    group = parser.add_argument_group("Sine fit options")
+    group.add_argument('-c', '--column', action='store', type=int,
+                       help='Column number (first data column is 1)')
+    group.add_argument('-f', '--frequency', action='store', type=float,
+                       default=60.0,
+                       help='Approximate frequency (default: %(default)s)')
+    group.add_argument('-m', '--min-freq', action='store', type=float,
+                       help='Minimum valid frequency '
+                       '(default: approximate frequency / 2))')
+    group.add_argument('-M', '--max-freq', action='store', type=float,
+                       help='Maximum valid frequency '
+                       '(default: approximate frequency * 2))')
+    group.add_argument('-a', '--min-amp', action='store', type=float,
+                       default=20.0,
+                       help='Minimum signal amplitude (default: %(default)s)')
+
+    # Parse arguments
+    try:
+        args = f.parse_args(argv)
+    except nilmtools.filter.MissingDestination as e:
+        rec = "float32_3"
+        print "Source is %s (%s)" % (e.src.path, e.src.layout)
+        print "Destination %s doesn't exist" % (e.dest.path)
+        print "You could make it with a command like:"
+        print "  nilmtool -u %s create %s %s" % (e.dest.url, e.dest.path, rec)
+        raise SystemExit(1)
+
+    if args.column is None or args.column < 1:
+        parser.error("need a column number >= 1")
+    if args.frequency < 0.1:
+        parser.error("frequency must be >= 0.1")
+    if args.min_freq is None:
+        args.min_freq = args.frequency / 2
+    if args.max_freq is None:
+        args.max_freq = args.frequency * 2
+    if (args.min_freq > args.max_freq or
+        args.min_freq > args.frequency or
+        args.max_freq < args.frequency):
+        parser.error("invalid min or max frequency")
+    if args.min_amp < 0:
+        parser.error("min amplitude must be >= 0")
+
+    f.check_dest_metadata({ "sinefit_source": f.src.path,
+                            "sinefit_column": args.column })
+    f.process_numpy(process, args = (args.column, args.frequency, args.min_amp,
+                                     args.min_freq, args.max_freq))
+
+class SuppressibleWarning(object):
+    def __init__(self, maxcount = 10, maxsuppress = 100):
+        self.maxcount = maxcount
+        self.maxsuppress = maxsuppress
+        self.count = 0
+        self.last_msg = ""
+
+    def _write(self, sec, msg):
+        if sec:
+            now = timestamp_to_human(seconds_to_timestamp(sec)) + ": "
+        else:
+            now = ""
+        sys.stderr.write(now + msg)
+
+    def warn(self, msg, seconds = None):
+        self.count += 1
+        if self.count <= self.maxcount:
+            self._write(seconds, msg)
+        if (self.count - self.maxcount) >= self.maxsuppress:
+            self.reset(seconds)
+
+    def reset(self, seconds = None):
+        if self.count > self.maxcount:
+            self._write(seconds, sprintf("(%d warnings suppressed)\n",
+                                         self.count - self.maxcount))
+        self.count = 0
+
+def process(data, interval, args, insert_function, final):
+    (column, f_expected, a_min, f_min, f_max) = args
+    rows = data.shape[0]
+
+    # Estimate sampling frequency from timestamps
+    fs = (rows-1) / (timestamp_to_seconds(data[-1][0]) -
+                     timestamp_to_seconds(data[0][0]))
+
+    # Pull out about 3.5 periods of data at once;
+    # we'll expect to match 3 zero crossings in each window
+    N = max(int(3.5 * fs / f_expected), 10)
+
+    # If we don't have enough data, don't bother processing it
+    if rows < N:
+        return 0
+
+    warn = SuppressibleWarning(3, 1000)
+
+    # Process overlapping windows
+    start = 0
+    num_zc = 0
+    last_inserted_timestamp = None
+    while start < (rows - N):
+        this = data[start:start+N, column]
+        t_min = timestamp_to_seconds(data[start, 0])
+        t_max = timestamp_to_seconds(data[start+N-1, 0])
+
+        # Do 4-parameter sine wave fit
+        (A, f0, phi, C) = sfit4(this, fs)
+
+        # Check bounds.  If frequency is too crazy, ignore this window
+        if f0 < f_min or f0 > f_max:
+            warn.warn(sprintf("frequency %s outside valid range %s - %s\n",
+                              str(f0), str(f_min), str(f_max)), t_min)
+            start += N
+            continue
+
+        # If amplitude is too low, results are probably just noise
+        if A < a_min:
+            warn.warn(sprintf("amplitude %s below minimum threshold %s\n",
+                              str(A), str(a_min)), t_min)
+            start += N
+            continue
+
+        #p.plot(arange(N), this)
+        #p.plot(arange(N), A * sin(f0/fs * 2 * pi * arange(N) + phi) + C, 'g')
+
+        # Period starts when the argument of sine is 0 degrees,
+        # so we're looking for sample number:
+        #     n = (0 - phi) / (f0/fs * 2 * pi)
+        zc_n = (0 - phi) / (f0 / fs * 2 * pi)
+        period_n = fs/f0
+
+        # Add periods to make N positive
+        while zc_n < 0:
+            zc_n += period_n
+
+        last_zc = None
+        # Mark the zero crossings until we're a half period away
+        # from the end of the window
+        while zc_n < (N - period_n/2):
+            #p.plot(zc_n, C, 'ro')
+            t = t_min + zc_n / fs
+            if (last_inserted_timestamp is None or
+                t > last_inserted_timestamp):
+                insert_function([[seconds_to_timestamp(t), f0, A, C]])
+                last_inserted_timestamp = t
+                warn.reset(t)
+            else:
+                warn.warn("timestamp overlap\n", t)
+            num_zc += 1
+            last_zc = zc_n
+            zc_n += period_n
+
+        # Advance the window one quarter period past the last marked
+        # zero crossing, or advance the window by half its size if we
+        # didn't mark any.
+        if last_zc is not None:
+            advance = min(last_zc + period_n/4, N)
+        else:
+            advance = N/2
+        #p.plot(advance, C, 'go')
+        #p.show()
+
+        start = int(round(start + advance))
+
+    # Return the number of rows we've processed
+    warn.reset(last_inserted_timestamp)
+    if last_inserted_timestamp:
+        now = timestamp_to_human(seconds_to_timestamp(
+            last_inserted_timestamp)) + ": "
+    else:
+        now = ""
+    printf("%sMarked %d zero-crossings in %d rows\n", now, num_zc, start)
+    return start
+
+def sfit4(data, fs):
+    """(A, f0, phi, C) = sfit4(data, fs)
+
+    Compute 4-parameter (unknown-frequency) least-squares fit to
+    sine-wave data, according to IEEE Std 1241-2010 Annex B
+
+    Input:
+      data  vector of input samples
+      fs    sampling rate (Hz)
+
+    Output:
+      Parameters [A, f0,  phi, C] to fit the equation
+        x[n] = A * sin(f0/fs * 2 * pi * n + phi) + C
+      where n is sample number.  Or, as a function of time:
+        x(t) = A * sin(f0 * 2 * pi * t + phi) + C
+
+    by Jim Paris
+    (Verified to match sfit4.m)
+    """
+    N = len(data)
+    t = linspace(0, (N-1) / float(fs), N)
+
+    ## Estimate frequency using FFT (step b)
+    Fc = fft(data)
+    F = abs(Fc)
+    F[0] = 0   # eliminate DC
+
+    # Find pair of spectral lines with largest amplitude:
+    # resulting values are in F(i) and F(i+1)
+    i = argmax(F[0:int(N/2)] + F[1:int(N/2+1)])
+
+    # Interpolate FFT to get a better result (from Markus [B37])
+    U1 = real(Fc[i])
+    U2 = real(Fc[i+1])
+    V1 = imag(Fc[i])
+    V2 = imag(Fc[i+1])
+    n = 2 * pi / N
+    ni1 = n * i
+    ni2 = n * (i+1)
+    K = ((V2-V1)*sin(ni1) + (U2-U1)*cos(ni1)) / (U2-U1)
+    Z1 = V1 * (K - cos(ni1)) / sin(ni1) + U1
+    Z2 = V2 * (K - cos(ni2)) / sin(ni2) + U2
+    i = arccos((Z2*cos(ni2) - Z1*cos(ni1)) / (Z2-Z1)) / n
+
+    # Convert to Hz
+    f0 = i * float(fs) / N
+
+    # Fit it.  We'll catch exceptions here and just returns zeros
+    # if something fails with the least squares fit, etc.
+    try:
+        # first guess for A0, B0 using 3-parameter fit (step c)
+        s = zeros(3)
+        w = 2*pi*f0
+
+        # Now iterate 7 times (step b, plus 6 iterations of step i)
+        for idx in range(7):
+            D = c_[cos(w*t), sin(w*t), ones(N),
+                  -s[0] * t * sin(w*t) + s[1] * t * cos(w*t) ] # eqn B.16
+            s = linalg.lstsq(D, data)[0] # eqn B.18
+            w = w + s[3]	# update frequency estimate
+
+        ## Extract results
+        A = sqrt(s[0]*s[0] + s[1]*s[1]) # eqn B.21
+        f0 = w / (2*pi)
+        phi = arctan2(s[0], s[1]) # eqn B.22 (flipped for sin instead of cos)
+        C = s[2]
+        return (A, f0, phi, C)
+    except Exception as e:
+        # something broke down, just return zeros
+        return (0, 0, 0, 0)
+
+if __name__ == "__main__":
+    main()

nilmtools/trainola.py

@@ -0,0 +1,279 @@
+#!/usr/bin/python
+
+from nilmdb.utils.printf import *
+import nilmdb.client
+import nilmtools.filter
+from nilmdb.utils.time import (timestamp_to_human,
+                               timestamp_to_seconds,
+                               seconds_to_timestamp)
+from nilmdb.utils.interval import Interval
+
+import numpy as np
+import scipy
+import scipy.signal
+from numpy.core.umath_tests import inner1d
+import nilmrun
+from collections import OrderedDict
+import sys
+import functools
+import collections
+
+class DataError(ValueError):
+    pass
+
+def build_column_mapping(colinfo, streaminfo):
+    """Given the 'columns' list from the JSON data, verify and
+    pull out a dictionary mapping for the column names/numbers."""
+    columns = OrderedDict()
+    for c in colinfo:
+        if (c['name'] in columns.keys() or
+            c['index'] in columns.values()):
+            raise DataError("duplicated columns")
+        if (c['index'] < 0 or c['index'] >= streaminfo.layout_count):
+            raise DataError("bad column number")
+        columns[c['name']] = c['index']
+    if not len(columns):
+        raise DataError("no columns")
+    return columns
+
+class Exemplar(object):
+    def __init__(self, exinfo, min_rows = 10, max_rows = 100000):
+        """Given a dictionary entry from the 'exemplars' input JSON,
+        verify the stream, columns, etc.  Then, fetch all the data
+        into self.data."""
+
+        self.name = exinfo['name']
+        self.url = exinfo['url']
+        self.stream = exinfo['stream']
+        self.start = exinfo['start']
+        self.end = exinfo['end']
+        self.dest_column = exinfo['dest_column']
+
+        # Get stream info
+        self.client = nilmdb.client.numpyclient.NumpyClient(self.url)
+        self.info = nilmtools.filter.get_stream_info(self.client, self.stream)
+
+        # Build up name => index mapping for the columns
+        self.columns = build_column_mapping(exinfo['columns'], self.info)
+
+        # Count points
+        self.count = self.client.stream_count(self.stream, self.start, self.end)
+
+        # Verify count
+        if self.count == 0:
+            raise DataError("No data in this exemplar!")
+        if self.count < min_rows:
+            raise DataError("Too few data points: " + str(self.count))
+        if self.count > max_rows:
+            raise DataError("Too many data points: " + str(self.count))
+
+        # Extract the data
+        datagen = self.client.stream_extract_numpy(self.stream,
+                                                   self.start, self.end,
+                                                   self.info.layout,
+                                                   maxrows = self.count)
+        self.data = list(datagen)[0]
+
+        # Discard timestamp
+        self.data = self.data[:,1:]
+
+        # Subtract the mean from each column
+        self.data = self.data - self.data.mean(axis=0)
+
+        # Get scale factors for each column by computing dot product
+        # of each column with itself.
+        self.scale = inner1d(self.data.T, self.data.T)
+
+        # Ensure a minimum (nonzero) scale and convert to list
+        self.scale = np.maximum(self.scale, [1e-9]).tolist()
+
+    def __str__(self):
+        return sprintf("\"%s\" %s [%s] %s rows",
+                       self.name, self.stream, ",".join(self.columns.keys()),
+                       self.count)
+
+def peak_detect(data, delta):
+    """Simple min/max peak detection algorithm, taken from my code
+    in the disagg.m from the 10-8-5 paper"""
+    mins = [];
+    maxs = [];
+    cur_min = (None, np.inf)
+    cur_max = (None, -np.inf)
+    lookformax = False
+    for (n, p) in enumerate(data):
+        if p > cur_max[1]:
+            cur_max = (n, p)
+        if p < cur_min[1]:
+            cur_min = (n, p)
+        if lookformax:
+            if p < (cur_max[1] - delta):
+                maxs.append(cur_max)
+                cur_min = (n, p)
+                lookformax = False
+        else:
+            if p > (cur_min[1] + delta):
+                mins.append(cur_min)
+                cur_max = (n, p)
+                lookformax = True
+    return (mins, maxs)
+
+def trainola_matcher(data, interval, args, insert_func, final_chunk):
+    """Perform cross-correlation match"""
+    ( src_columns, dest_count, exemplars ) = args
+    nrows = data.shape[0]
+
+    # We want at least 10% more points than the widest exemplar.
+    widest = max([ x.count for x in exemplars ])
+    if (widest * 1.1) > nrows:
+        return 0
+
+    # This is how many points we'll consider valid in the
+    # cross-correlation.
+    valid = nrows + 1 - widest
+    matches = collections.defaultdict(list)
+
+    # Try matching against each of the exemplars
+    for e in exemplars:
+        corrs = []
+
+        # Compute cross-correlation for each column
+        for col_name in e.columns:
+            a = data[:, src_columns[col_name] + 1]
+            b = e.data[:, e.columns[col_name]]
+            corr = scipy.signal.fftconvolve(a, np.flipud(b), 'valid')[0:valid]
+
+            # Scale by the norm of the exemplar
+            corr = corr / e.scale[e.columns[col_name]]
+            corrs.append(corr)
+
+        # Find the peaks using the column with the largest amplitude
+        biggest = e.scale.index(max(e.scale))
+        peaks_minmax = peak_detect(corrs[biggest], 0.1)
+        peaks = [ p[0] for p in peaks_minmax[1] ]
+
+        # Now look at every peak
+        for row in peaks:
+            # Correlation for each column must be close enough to 1.
+            for (corr, scale) in zip(corrs, e.scale):
+                # The accepted distance from 1 is based on the relative
+                # amplitude of the column.  Use a linear mapping:
+                # scale 1.0 -> distance 0.1
+                # scale 0.0 -> distance 1.0
+                distance = 1 - 0.9 * (scale / e.scale[biggest])
+                if abs(corr[row] - 1) > distance:
+                    # No match
+                    break
+            else:
+                # Successful match
+                matches[row].append(e)
+
+    # Insert matches into destination stream.
+    matched_rows = sorted(matches.keys())
+    out = np.zeros((len(matched_rows), dest_count + 1))
+
+    for n, row in enumerate(matched_rows):
+        # Fill timestamp
+        out[n][0] = data[row, 0]
+
+        # Mark matched exemplars
+        for exemplar in matches[row]:
+            out[n, exemplar.dest_column + 1] = 1.0
+
+    # Insert it
+    insert_func(out)
+
+    # Return how many rows we processed
+    return max(valid, 0)
+
+def trainola(conf):
+    print "Trainola", nilmtools.__version__
+
+    # Load main stream data
+    url = conf['url']
+    src_path = conf['stream']
+    dest_path = conf['dest_stream']
+    start = conf['start']
+    end = conf['end']
+
+    # Get info for the src and dest streams
+    src_client = nilmdb.client.numpyclient.NumpyClient(url)
+    src = nilmtools.filter.get_stream_info(src_client, src_path)
+    if not src:
+        raise DataError("source path '" + src_path + "' does not exist")
+    src_columns = build_column_mapping(conf['columns'], src)
+
+    dest_client = nilmdb.client.numpyclient.NumpyClient(url)
+    dest = nilmtools.filter.get_stream_info(dest_client, dest_path)
+    if not dest:
+        raise DataError("destination path '" + dest_path + "' does not exist")
+
+    printf("Source:\n")
+    printf("  %s [%s]\n", src.path, ",".join(src_columns.keys()))
+    printf("Destination:\n")
+    printf("  %s (%s columns)\n", dest.path, dest.layout_count)
+
+    # Pull in the exemplar data
+    exemplars = []
+    for n, exinfo in enumerate(conf['exemplars']):
+        printf("Loading exemplar %d:\n", n)
+        e = Exemplar(exinfo)
+        col = e.dest_column
+        if col < 0 or col >= dest.layout_count:
+            raise DataError(sprintf("bad destination column number %d\n" +
+                                    "dest stream only has 0 through %d",
+                                    col, dest.layout_count - 1))
+        printf("  %s, output column %d\n", str(e), col)
+        exemplars.append(e)
+    if len(exemplars) == 0:
+        raise DataError("missing exemplars")
+
+    # Verify that the exemplar columns are all represented in the main data
+    for n, ex in enumerate(exemplars):
+        for col in ex.columns:
+            if col not in src_columns:
+                raise DataError(sprintf("Exemplar %d column %s is not "
+                                        "available in source data", n, col))
+
+    # Figure out which intervals we should process
+    intervals = ( Interval(s, e) for (s, e) in
+                  src_client.stream_intervals(src_path,
+                                              diffpath = dest_path,
+                                              start = start, end = end) )
+    intervals = nilmdb.utils.interval.optimize(intervals)
+
+    # Do the processing
+    rows = 100000
+    extractor = functools.partial(src_client.stream_extract_numpy,
+                                  src.path, layout = src.layout, maxrows = rows)
+    inserter = functools.partial(dest_client.stream_insert_numpy_context,
+                                 dest.path)
+    for interval in intervals:
+        printf("Processing interval:\n")
+        printf("  %s\n", interval.human_string())
+        nilmtools.filter.process_numpy_interval(
+            interval, extractor, inserter, rows * 3,
+            trainola_matcher, (src_columns, dest.layout_count, exemplars))
+
+    return "done"
+
+def main(argv = None):
+    import simplejson as json
+    import sys
+
+    if argv is None:
+        argv = sys.argv[1:]
+    if len(argv) != 1:
+        raise DataError("need one argument, either a dictionary or JSON string")
+
+    try:
+        # Passed in a JSON string (e.g. on the command line)
+        conf = json.loads(argv[0])
+    except TypeError as e:
+        # Passed in the config dictionary (e.g. from NilmRun)
+        conf = argv[0]
+
+    return trainola(conf)
+
+if __name__ == "__main__":
+    main()
+

setup.py

@@ -30,7 +30,7 @@ except ImportError:
 # Versioneer manages version numbers from git tags.
 # https://github.com/warner/python-versioneer
 import versioneer
-versioneer.versionfile_source = 'src/_version.py'
+versioneer.versionfile_source = 'nilmtools/_version.py'
 versioneer.versionfile_build = 'nilmtools/_version.py'
 versioneer.tag_prefix = 'nilmtools-'
 versioneer.parentdir_prefix = 'nilmtools-'
@@ -61,14 +61,13 @@ setup(name='nilmtools',
       long_description = "NILM Database Tools",
       license = "Proprietary",
       author_email = 'jim@jtan.com',
-      install_requires = [ 'nilmdb >= 1.5.0',
+      install_requires = [ 'nilmdb >= 1.8.1',
                            'numpy',
                            'scipy',
-                           'matplotlib',
+                           #'matplotlib',
                            ],
       packages = [ 'nilmtools',
                    ],
-      package_dir = { 'nilmtools': 'src' },
       entry_points = {
           'console_scripts': [
               'nilm-decimate = nilmtools.decimate:main',
@@ -79,6 +78,8 @@ setup(name='nilmtools',
               'nilm-copy-wildcard = nilmtools.copy_wildcard:main',
               'nilm-sinefit = nilmtools.sinefit:main',
               'nilm-cleanup = nilmtools.cleanup:main',
+              'nilm-median = nilmtools.median:main',
+              'nilm-trainola = nilmtools.trainola:main',
               ],
           },
       zip_safe = False,

src/sinefit.py

@@ -1,187 +0,0 @@
-#!/usr/bin/python
-
-# Sine wave fitting.  This runs about 5x faster than realtime on raw data.
-
-import nilmtools.filter
-import nilmdb.client
-from numpy import *
-from scipy import *
-#import pylab as p
-import operator
-
-def main(argv = None):
-    f = nilmtools.filter.Filter()
-    parser = f.setup_parser("Sine wave fitting")
-    group = parser.add_argument_group("Sine fit options")
-    group.add_argument('-c', '--column', action='store', type=int,
-                       help='Column number (first data column is 1)')
-    group.add_argument('-f', '--frequency', action='store', type=float,
-                       default=60.0,
-                       help='Approximate frequency (default: %(default)s)')
-
-    # Parse arguments
-    try:
-        args = f.parse_args(argv)
-    except nilmtools.filter.MissingDestination as e:
-        rec = "float32_3"
-        print "Source is %s (%s)" % (e.src.path, e.src.layout)
-        print "Destination %s doesn't exist" % (e.dest.path)
-        print "You could make it with a command like:"
-        print "  nilmtool -u %s create %s %s" % (e.dest.url, e.dest.path, rec)
-        raise SystemExit(1)
-
-    if args.column is None or args.column < 1:
-        parser.error("need a column number >= 1")
-    if args.frequency < 0.1:
-        parser.error("frequency must be >= 0.1")
-
-    f.check_dest_metadata({ "sinefit_source": f.src.path,
-                            "sinefit_column": args.column })
-    f.process_numpy(process, args = (args.column, args.frequency))
-
-def process(data, interval, args, insert_function, final):
-    (column, f_expected) = args
-    rows = data.shape[0]
-
-    # Estimate sampling frequency from timestamps
-    fs = 1e6 * (rows-1) / (data[-1][0] - data[0][0])
-
-    # Pull out about 3.5 periods of data at once;
-    # we'll expect to match 3 zero crossings in each window
-    N = max(int(3.5 * fs / f_expected), 10)
-
-    # If we don't have enough data, don't bother processing it
-    if rows < N:
-        return 0
-
-    # Process overlapping windows
-    start = 0
-    num_zc = 0
-    while start < (rows - N):
-        this = data[start:start+N, column]
-        t_min = data[start, 0]/1e6
-        t_max = data[start+N-1, 0]/1e6
-
-        # Do 4-parameter sine wave fit
-        (A, f0, phi, C) = sfit4(this, fs)
-
-        # Check bounds.  If frequency is too crazy, ignore this window
-        if f0 < (f_expected/2) or f0 > (f_expected*2):
-            print "frequency", f0, "too far from expected value", f_expected
-            start += N
-            continue
-
-        #p.plot(arange(N), this)
-        #p.plot(arange(N), A * cos(f0/fs * 2 * pi * arange(N) + phi) + C, 'g')
-
-        # Period starts when the argument of cosine is 3*pi/2 degrees,
-        # so we're looking for sample number:
-        #     n = (3 * pi / 2 - phi) / (f0/fs * 2 * pi)
-        zc_n = (3 * pi / 2 - phi) / (f0 / fs * 2 * pi)
-        period_n = fs/f0
-
-        # Add periods to make N positive
-        while zc_n < 0:
-            zc_n += period_n
-
-        last_zc = None
-        # Mark the zero crossings until we're a half period away
-        # from the end of the window
-        while zc_n < (N - period_n/2):
-            #p.plot(zc_n, C, 'ro')
-            t = t_min + zc_n / fs
-            insert_function([[t * 1e6, f0, A, C]])
-            num_zc += 1
-            last_zc = zc_n
-            zc_n += period_n
-
-        # Advance the window one quarter period past the last marked
-        # zero crossing, or advance the window by half its size if we
-        # didn't mark any.
-        if last_zc is not None:
-            advance = min(last_zc + period_n/4, N)
-        else:
-            advance = N/2
-        #p.plot(advance, C, 'go')
-        #p.show()
-
-        start = int(round(start + advance))
-
-    # Return the number of rows we've processed
-    print "Marked", num_zc, "zero-crossings in", start, "rows"
-    return start
-
-def sfit4(data, fs):
-    """(A, f0, phi, C) = sfit4(data, fs)
-
-    Compute 4-parameter (unknown-frequency) least-squares fit to
-    sine-wave data, according to IEEE Std 1241-2010 Annex B
-
-    Input:
-      data  vector of input samples
-      fs    sampling rate (Hz)
-
-    Output:
-      Parameters [A, f0,  phi, C] to fit the equation
-        x[n] = A * cos(f0/fs * 2 * pi * n + phi) + C
-      where n is sample number.  Or, as a function of time:
-        x(t) = A * cos(f0 * 2 * pi * t + phi) + C
-
-    by Jim Paris
-    (Verified to match sfit4.m)
-    """
-    N = len(data)
-    t = linspace(0, (N-1) / fs, N)
-
-    ## Estimate frequency using FFT (step b)
-    Fc = fft(data)
-    F = abs(Fc)
-    F[0] = 0   # eliminate DC
-
-    # Find pair of spectral lines with largest amplitude:
-    # resulting values are in F(i) and F(i+1)
-    i = argmax(F[0:int(N/2)] + F[1:int(N/2+1)])
-
-    # Interpolate FFT to get a better result (from Markus [B37])
-    U1 = real(Fc[i])
-    U2 = real(Fc[i+1])
-    V1 = imag(Fc[i])
-    V2 = imag(Fc[i+1])
-    n = 2 * pi / N
-    ni1 = n * i
-    ni2 = n * (i+1)
-    K = ((V2-V1)*sin(ni1) + (U2-U1)*cos(ni1)) / (U2-U1)
-    Z1 = V1 * (K - cos(ni1)) / sin(ni1) + U1
-    Z2 = V2 * (K - cos(ni2)) / sin(ni2) + U2
-    i = arccos((Z2*cos(ni2) - Z1*cos(ni1)) / (Z2-Z1)) / n
-
-    # Convert to Hz
-    f0 = i * fs / N
-
-    ## Fit it
-    # first guess for A0, B0 using 3-parameter fit (step c)
-    w = 2*pi*f0
-    D = c_[cos(w*t), sin(w*t), ones(N)]
-    s = linalg.lstsq(D, data)[0]
-
-    # Now iterate 6 times (step i)
-    for idx in range(6):
-        D = c_[cos(w*t), sin(w*t), ones(N),
-              -s[0] * t * sin(w*t) + s[1] * t * cos(w*t) ] # eqn B.16
-        s = linalg.lstsq(D, data)[0] # eqn B.18
-        w = w + s[3]	# update frequency estimate
-
-    ## Extract results
-    A = sqrt(s[0]*s[0] + s[1]*s[1]) # eqn B.21
-    f0 = w / (2*pi)
-    try:
-        phi = -arctan2(s[1], s[0]) # eqn B.22
-    except TypeError:
-        # something broke down, just return zeros
-        return (0, 0, 0, 0)
-    C = s[2]
-
-    return (A, f0, phi, C)
-
-if __name__ == "__main__":
-    main()