Upgrade to paddle 1.7.1

README.md

@@ -40,7 +40,7 @@ sudo apt-get install libsndfile1
 
 ### Install PaddlePaddle
 
-See [install](https://www.paddlepaddle.org.cn/install/quick) for more details. This repo requires PaddlePaddle **1.7.0** or above.
+See [install](https://www.paddlepaddle.org.cn/install/quick) for more details. This repo requires PaddlePaddle **1.7.1** or above.
 
 ### Install Parakeet
 
@@ -82,7 +82,7 @@ Parakeet also releases some well-trained parameters for the example models, whic
 <table>
     <thead>
         <tr>
-            <th  style="width: 250px"> 
+            <th  style="width: 250px">
             WaveFlow
             </th>
             <th  style="width: 250px">
@@ -109,7 +109,7 @@ Parakeet also releases some well-trained parameters for the example models, whic
             <a href="https://paddlespeech.bj.bcebos.com/Parakeet/clarinet_ljspeech_samples_1.0/step_500000_sentence_3.wav">
             <img src="images/audio_icon.png" width=250 /></a><br>
             <a href="https://paddlespeech.bj.bcebos.com/Parakeet/clarinet_ljspeech_samples_1.0/step_500000_sentence_4.wav">
-            <img src="images/audio_icon.png" width=250 /></a>            
+            <img src="images/audio_icon.png" width=250 /></a>  
             </th>
         </tr>
     </tbody>

parakeet/models/waveflow/waveflow_modules.py

@@ -96,7 +96,7 @@ class Conditioner(dg.Layer):
                 x = fluid.layers.cast(x, "float16")
             x = fluid.layers.leaky_relu(x, alpha=0.4)
 
-        return fluid.layers.reshape(x, [x.shape[0], x.shape[2], x.shape[3]])
+        return fluid.layers.squeeze(x, [1])
 
     def infer(self, x):
         x = fluid.layers.unsqueeze(x, 1)
@@ -111,7 +111,7 @@ class Conditioner(dg.Layer):
             time_cutoff = layer._filter_size[1] - layer._stride[1]
             x = fluid.layers.leaky_relu(x[:, :, :, :-time_cutoff], alpha=0.4)
 
-        return fluid.layers.reshape(x, [x.shape[0], x.shape[2], x.shape[3]])
+        return fluid.layers.squeeze(x, [1])
 
 
 class Flow(dg.Layer):
@@ -220,10 +220,9 @@ class Flow(dg.Layer):
             # Pad width dim (time): dialated non-causal convolution
             pad_top, pad_bottom = (self.kernel_h - 1) * dilation_h, 0
             pad_left = pad_right = int((self.kernel_w - 1) * dilation_w / 2)
-            self.in_layers[i].layer._padding = [
-                pad_top, pad_bottom, pad_left, pad_right
-            ]
-            hidden = self.in_layers[i](audio)
+            audio_pad = fluid.layers.pad2d(
+                audio, paddings=[pad_top, pad_bottom, pad_left, pad_right])
+            hidden = self.in_layers[i](audio_pad)
             cond_hidden = self.cond_layers[i](mel)
             in_acts = hidden + cond_hidden
             out_acts = fluid.layers.tanh(in_acts[:, :self.n_channels, :]) * \
@@ -268,9 +267,8 @@ class Flow(dg.Layer):
             pad_top, pad_bottom = 0, 0
             pad_left = int((self.kernel_w - 1) * dilation_w / 2)
             pad_right = int((self.kernel_w - 1) * dilation_w / 2)
-            self.in_layers[i].layer._padding = [
-                pad_top, pad_bottom, pad_left, pad_right
-            ]
+            state = fluid.layers.pad2d(
+                state, paddings=[pad_top, pad_bottom, pad_left, pad_right])
             hidden = self.in_layers[i](state)
             cond_hidden = self.cond_layers[i](mel)
             in_acts = hidden + cond_hidden
@@ -301,6 +299,7 @@ class WaveFlowModule(dg.Layer):
     Returns:
         WaveFlowModule
     """
+
     def __init__(self, config):
         super(WaveFlowModule, self).__init__()
         self.n_flows = config.n_flows
@@ -380,8 +379,7 @@ class WaveFlowModule(dg.Layer):
             mel_slices = [mel[:, :, j, :] for j in self.perms[i]]
             mel = fluid.layers.stack(mel_slices, axis=2)
 
-        z = fluid.layers.reshape(
-            audio, [audio.shape[0], audio.shape[2], audio.shape[3]])
+        z = fluid.layers.squeeze(audio, [1])
         return z, log_s_list
 
     def synthesize(self, mel, sigma=1.0):
@@ -442,8 +440,7 @@ class WaveFlowModule(dg.Layer):
             audio = fluid.layers.concat(audio_list, axis=2)
 
         # audio: [bs, n_group, time/n_group]
-        audio = fluid.layers.reshape(
-            audio, [audio.shape[0], audio.shape[2], audio.shape[3]])
+        audio = fluid.layers.squeeze(audio, [1])
         # audio: [bs, time]
         audio = fluid.layers.reshape(
             fluid.layers.transpose(audio, [0, 2, 1]), [audio.shape[0], -1])

parakeet/modules/weight_norm.py

@@ -65,15 +65,6 @@ def compute_weight(v, g, dim, power):
     return weight
 
 
-def assign_by_cast(i, o):
-    fluid.default_main_program().current_block().append_op(
-        type="cast",
-        inputs={"X": i},
-        outputs={"Out": o},
-        attrs={"in_dtype": i.dtype,
-               "out_dtype": o.dtype})
-
-
 class WeightNormWrapper(dg.Layer):
     def __init__(self, layer, param_name="weight", dim=0, power=2):
         super(WeightNormWrapper, self).__init__()
@@ -93,13 +84,13 @@ class WeightNormWrapper(dg.Layer):
             w_v,
             self.create_parameter(
                 shape=original_weight.shape, dtype=original_weight.dtype))
-        assign_by_cast(original_weight, getattr(self, w_v))
+        F.assign(original_weight, getattr(self, w_v))
         delattr(layer, param_name)
         temp = norm_except(getattr(self, w_v), self.dim, self.power)
         self.add_parameter(
             w_g, self.create_parameter(
                 shape=temp.shape, dtype=temp.dtype))
-        assign_by_cast(temp, getattr(self, w_g))
+        F.assign(temp, getattr(self, w_g))
 
         # also set this when setting up
         setattr(self.layer, self.param_name,