Merge pull request mila-iqia#28 from jbornschein/plot

Various minor fixes for blocks-plot

bin/blocks-plot

@@ -34,14 +34,14 @@ def plot_dataframe(dataframe):
     print("Plotting {} channels:".format(len(dataframe.columns)))
     for cname, series in iteritems(dataframe):
         print("    {}".format(cname))
-        pylab.plot(t, series.interpolate(), label=cname)
+        pylab.plot(t, series.interpolate(method='nearest'), label=cname)
     pylab.legend()
     pylab.show(block=True)
 
 
 def main(args):
     import readline
-    import blocks.scripts.plot as plot
+    import blocks.extras.scripts.plot as plot
 
     from six import iteritems
     from six.moves import input
@@ -121,6 +121,7 @@ def main(args):
                 break
 
             column_specs = column_spec.split(',')
+            column_specs = [s.strip() for s in column_specs]
             matched = plot.match_column_specs(experiments, column_specs)
 
             if len(matched.columns) == 0:

blocks/extras/scripts/plot.py

@@ -13,7 +13,7 @@
 from blocks.serialization import load
 
 try:
-    from pandas import DataFrame
+    import pandas
     PANDAS_AVAILABLE = True
 except ImportError:
     PANDAS_AVAILABLE = False
@@ -95,7 +95,7 @@ def match_column_specs(experiments, column_specs):
                           " install it with pip.")
     # We iterate over all column and match each spec to the
     # channels of all experiments.
-    df = DataFrame()
+    df = pandas.DataFrame()
     for spec in column_specs:
         if ":" in spec:
             exp_spec, column_spec = spec.split(":")
@@ -111,6 +111,8 @@ def match_column_specs(experiments, column_specs):
                     continue
 
                 column_name = "{}:{}".format(i, column)
-                df[column_name] = exp[column]
+
+                exp = exp.rename(columns={column: column_name})
+                df = pandas.concat((df, exp[column_name]), axis=1)
 
     return df

tests/scripts/test_plot.py

@@ -3,6 +3,7 @@
 
 from collections import OrderedDict
 from tests import silence_printing, skip_if_not_available
+from numpy import nan, isfinite
 
 from blocks.log import TrainingLog
 from blocks.main_loop import MainLoop
@@ -22,8 +23,8 @@ def some_experiments():
     experiments['exp0']['col0'] = (0, 1, 2)
     experiments['exp0']['col1'] = (3, 4, 5)
     experiments['exp1'] = DataFrame()
-    experiments['exp1']['col0'] = (6, 7, 8)
-    experiments['exp1']['col1'] = (9, 9, 9)
+    experiments['exp1']['col0'] = (6, 7, 8, 9)
+    experiments['exp1']['col1'] = (9, 9, 9, 9)
     return experiments
 
 
@@ -66,3 +67,23 @@ def test_match_column_specs():
 
     assert isinstance(df, DataFrame)
     assert list(df.columns) == ['0:col0', '0:col1', '1:col1']
+    assert list(df.index) == [0, 1, 2, 3]
+
+
+def test_interpolate():
+    skip_if_not_available(modules=['pandas'])
+    """ Ensure tha DataFrame.interpolate(method='nearest') has the
+    desired properties.
+
+    It is used by blocks-plot and should:
+
+    * interpolate missing/NaN datapoints between valid ones
+    * not replace any NaN before/after the first/last finite datapoint
+    """
+    y = [nan, nan, 2., 3., nan, 5, nan, nan]
+    df = DataFrame(y)
+    df_ = df.interpolate(method='nearest')[0]
+
+    assert all(isfinite(df_[2:6]))
+    assert all(~isfinite(df_[0:2]))
+    assert all(~isfinite(df_[6:8]))