[fix] Fix TypeError: numpy.float64 cannot be interpreted as integer (#716)

Cast encoder.n_feats_out and similar values to int across all model constructors to prevent TypeError when numpy returns float64 scalars from arithmetic operations. Newer numpy versions enforce stricter integer type checking, causing nn.Conv1d, nn.Parameter, and other PyTorch layers to reject numpy.float64 where a Python int is expected.

Fixes #713

Author: deekshaNVIDIA

asteroid/masknn/recurrent.py

@@ -676,7 +676,7 @@ def __init__(self, encoders, decoders, n_freqs, **kwargs):
         super().__init__(
             encoders=[
                 *(DCUNetComplexEncoderBlock(*args, activation="prelu") for args in encoders),
-                DCCRMaskNetRNN(np.prod(last_encoder_out_shape)),
+                DCCRMaskNetRNN(int(np.prod(last_encoder_out_shape))),
             ],
             decoders=[
                 torch.nn.Identity(),

asteroid/models/conv_tasnet.py

@@ -76,7 +76,7 @@ def __init__(
             sample_rate=sample_rate,
             **fb_kwargs,
         )
-        n_feats = encoder.n_feats_out
+        n_feats = int(encoder.n_feats_out)
         if in_chan is not None:
             assert in_chan == n_feats, (
                 "Number of filterbank output channels"

asteroid/models/dccrnet.py

@@ -24,7 +24,7 @@ class DCCRNet(BaseDCUNet):
     def __init__(
         self, *args, stft_n_filters=512, stft_kernel_size=400, stft_stride=100, **masknet_kwargs
     ):
-        masknet_kwargs.setdefault("n_freqs", stft_n_filters // 2)
+        masknet_kwargs.setdefault("n_freqs", int(stft_n_filters) // 2)
         super().__init__(
             *args,
             stft_n_filters=stft_n_filters,

asteroid/models/demask.py

@@ -61,8 +61,9 @@ def __init__(
             **fb_kwargs,
         )
 
-        n_masker_in = self._get_n_feats_input(input_type, encoder.n_feats_out)
-        n_masker_out = self._get_n_feats_output(output_type, encoder.n_feats_out)
+        n_feats_out = int(encoder.n_feats_out)
+        n_masker_in = self._get_n_feats_input(input_type, n_feats_out)
+        n_masker_out = self._get_n_feats_output(output_type, n_feats_out)
         masker = build_demask_masker(
             n_masker_in,
             n_masker_out,

asteroid/models/dprnn_tasnet.py

@@ -82,7 +82,7 @@ def __init__(
             sample_rate=sample_rate,
             **fb_kwargs,
         )
-        n_feats = encoder.n_feats_out
+        n_feats = int(encoder.n_feats_out)
         if in_chan is not None:
             assert in_chan == n_feats, (
                 "Number of filterbank output channels"

asteroid/models/dptnet.py

@@ -79,7 +79,7 @@ def __init__(
             sample_rate=sample_rate,
             **fb_kwargs,
         )
-        n_feats = encoder.n_feats_out
+        n_feats = int(encoder.n_feats_out)
         if in_chan is not None:
             assert in_chan == n_feats, (
                 "Number of filterbank output channels"

asteroid/models/lstm_tasnet.py

@@ -68,7 +68,7 @@ def __init__(
             sample_rate=sample_rate,
             **fb_kwargs,
         )
-        n_feats = encoder.n_feats_out
+        n_feats = int(encoder.n_feats_out)
         if in_chan is not None:
             assert in_chan == n_feats, (
                 "Number of filterbank output channels"

asteroid/models/sudormrf.py

@@ -62,7 +62,7 @@ def __init__(
             output_padding=(kernel_size // 2) - 1,
             **fb_kwargs,
         )
-        n_feats = enc.n_feats_out
+        n_feats = int(enc.n_feats_out)
         enc = _Padder(enc, upsampling_depth=upsampling_depth, kernel_size=kernel_size)
 
         if in_chan is not None:
@@ -136,7 +136,7 @@ def __init__(
             output_padding=(kernel_size // 2) - 1,
             **fb_kwargs,
         )
-        n_feats = enc.n_feats_out
+        n_feats = int(enc.n_feats_out)
         enc = _Padder(enc, upsampling_depth=upsampling_depth, kernel_size=kernel_size)
 
         if in_chan is not None:

asteroid/models/x_umx.py

@@ -73,7 +73,7 @@ def __init__(
         self.nb_channels = nb_channels
         self.nb_layers = nb_layers
         self.bidirectional = bidirectional
-        self.nb_output_bins = in_chan // 2 + 1
+        self.nb_output_bins = int(in_chan) // 2 + 1
         if max_bin:
             self.max_bin = max_bin
         else:

egs/dns_challenge_INTERSPEECH2020/baseline/model.py

@@ -28,7 +28,7 @@ def make_model_and_optimizer(conf):
     # Because we concatenate (re, im, mag) as input and compute a complex mask.
     if conf["main_args"]["is_complex"]:
         inp_size = int(stft.n_feats_out * 3 / 2)
-        output_size = stft.n_feats_out
+        output_size = int(stft.n_feats_out)
     else:
         inp_size = output_size = int(stft.n_feats_out / 2)
     # Add these fields to the mask model dict