testadmin
/
ParakeetEricRoss
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Upgrade to paddle 1.7.1

This commit is contained in:
liuyibing01 2020-03-12 14:20:46 +00:00
parent 2664398127
commit bfbc5e2046
3 changed files with 15 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
README.md
View File

@ -40,7 +40,7 @@ sudo apt-get install libsndfile1
### Install PaddlePaddle ### 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 ### Install Parakeet
@ -82,7 +82,7 @@ Parakeet also releases some well-trained parameters for the example models, whic
<table> <table>
<thead> <thead>
<tr> <tr>
<th style="width: 250px"> <th style="width: 250px">
WaveFlow WaveFlow
</th> </th>
<th style="width: 250px"> <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"> <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> <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"> <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> </th>
</tr> </tr>
</tbody> </tbody>

23
parakeet/models/waveflow/waveflow_modules.py
View File

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

13
parakeet/modules/weight_norm.py
View File

@ -65,15 +65,6 @@ def compute_weight(v, g, dim, power):
return weight 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): class WeightNormWrapper(dg.Layer):
def __init__(self, layer, param_name="weight", dim=0, power=2): def __init__(self, layer, param_name="weight", dim=0, power=2):
super(WeightNormWrapper, self).__init__() super(WeightNormWrapper, self).__init__()
@ -93,13 +84,13 @@ class WeightNormWrapper(dg.Layer):
w_v, w_v,
self.create_parameter( self.create_parameter(
shape=original_weight.shape, dtype=original_weight.dtype)) 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) delattr(layer, param_name)
temp = norm_except(getattr(self, w_v), self.dim, self.power) temp = norm_except(getattr(self, w_v), self.dim, self.power)
self.add_parameter( self.add_parameter(
w_g, self.create_parameter( w_g, self.create_parameter(
shape=temp.shape, dtype=temp.dtype)) 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 # also set this when setting up
setattr(self.layer, self.param_name, setattr(self.layer, self.param_name,