Merge branch 'kl/stereo_and_scaling_fix' into 'master'

Kl/stereo and scaling fix

See merge request machine-learning/bonito!180

bonito/crf/model.py

@@ -147,11 +147,11 @@ def conv(c_in, c_out, ks, stride=1, bias=False, activation=None, norm=None):
     return Convolution(c_in, c_out, ks, stride=stride, padding=ks//2, bias=bias, activation=activation, norm=norm)
 
 
-def rnn_encoder(n_base, state_len, insize=1, stride=5, winlen=19, activation='swish', rnn_type='lstm', features=768, scale=5.0, blank_score=None, expand_blanks=True, num_layers=5, norm=None):
+def rnn_encoder(n_base, state_len, insize=1, first_conv_size=4, stride=5, winlen=19, activation='swish', rnn_type='lstm', features=768, scale=5.0, blank_score=None, expand_blanks=True, num_layers=5, norm=None):
     rnn = layers[rnn_type]
     return Serial([
-        conv(insize, 4, ks=5, bias=True, activation=activation, norm=norm),
-        conv(4, 16, ks=5, bias=True, activation=activation, norm=norm),
+        conv(insize, first_conv_size, ks=5, bias=True, activation=activation, norm=norm),
+        conv(first_conv_size, 16, ks=5, bias=True, activation=activation, norm=norm),
         conv(16, features, ks=winlen, stride=stride, bias=True, activation=activation, norm=norm),
         Permute([2, 0, 1]),
         *(rnn(features, features, reverse=(num_layers - i) % 2) for i in range(num_layers)),

bonito/fast5.py

@@ -77,6 +77,8 @@ def __init__(self, read, filename, meta=False, do_trim=True, scaling_strategy=No
         self.scaled = np.array(self.scaling * (raw + self.offset), dtype=np.float32)
         self.num_samples = len(self.scaled)
 
+        self.scaling_strategy = ("quantile" if scaling_strategy is None else
+                                 scaling_strategy.get("strategy","quantile")) 
         self.shift, self.scale = bonito.reader.normalisation(self.scaled, scaling_strategy, norm_params)
         self.trimmed_samples = bonito.reader.trim(self.scaled, threshold=self.scale * 2.4 + self.shift) if do_trim else 0
         self.template_start = self.start + (self.trimmed_samples / self.sample_rate)

bonito/pod5.py

@@ -55,7 +55,9 @@ def __init__(self, read, filename, meta=False, do_trim=True, scaling_strategy=No
         self.scaling = self.calibration.scale
         self.offset = self.calibration.offset
         self.scaled = self.scaling * (self.raw.astype(np.float32) + self.offset)
-
+        self.scaling_strategy = ("quantile" if scaling_strategy is None else
+                                 scaling_strategy.get("strategy","quantile")) 
+
         self.shift, self.scale = bonito.reader.normalisation(self.scaled, scaling_strategy, norm_params)
         self.trimmed_samples = bonito.reader.trim(self.scaled, threshold=self.scale * 2.4 + self.shift) if do_trim else 0
 

bonito/reader.py

@@ -82,7 +82,7 @@ def tagdata(self):
             f"f5:Z:{self.filename}",
             f"sm:f:{self.shift}",
             f"sd:f:{self.scale}",
-            f"sv:Z:quantile",
+            f"sv:Z:{self.scaling_strategy}",
         ]