params.quantization_steps = 10
params.causal_conv_no_bias = True
params.causal_conv_filter_width = 3
params.causal_conv_channels = [32]
params.residual_conv_dilation_no_bias = True
params.residual_conv_projection_no_bias = True
params.residual_conv_filter_width = 3
params.residual_conv_channels = [16, 16]
params.residual_num_blocks = 1
params.softmax_conv_no_bias = False
params.softmax_conv_channels = [24, 10]
params.optimizer = "eve"
params.learning_rate = 0.001
params.momentum = 0.9
params.weight_decay = 0.00001
params.gradient_clipping = 10.0
wavenet = WaveNet(params)
f = open(filename, "w")
json.dump(params.to_dict(), open(filename, "w"), indent=4)
params.dump()
wavenet.load(args.model_dir)
if args.gpu_device != -1:
cuda.get_device(args.gpu_device).use()
wavenet.to_gpu()
params.causal_conv_no_bias = True
params.causal_conv_filter_width = 2
params.causal_conv_channels = [256]
params.residual_conv_dilation_no_bias = True
params.residual_conv_projection_no_bias = True
params.residual_conv_filter_width = 2
params.residual_conv_channels = [
128, 128, 128, 128, 128, 128, 128, 128, 128, 128
]
params.residual_num_blocks = 1
params.softmax_conv_no_bias = False
params.softmax_conv_channels = [256, 256]
params.optimizer = "adam"
params.momentum = 0.9
params.weight_decay = 0
params.gradient_clipping = 1.0
with open(filename, "w") as f:
json.dump(params.to_dict(), f, indent=4)
if args.fast:
wavenet = FasterWaveNet(params)
else:
wavenet = WaveNet(params)
params.dump()
wavenet.load(args.model_dir)