playground.py

#%%
from __future__ import print_function
from __future__ import absolute_import
from __future__ import division
from __future__ import unicode_literals

import argparse
import sys
import scipy.misc
import numpy as np
import importlib

import matplotlib.pyplot as plt

import chainer
import chainer.training
import chainer.training.extensions as extensions
import chainer.links as L

import wavenet.utils as utils
import wavenet.models as models


#%%
def simple(image):
    return (image > 0.5).astype('f')

#%%
test, train = chainer.datasets.get_mnist(False, ndim=2)

def plot(sample):
    plt.subplot(141)
    plt.imshow(sample, cmap='Greys', interpolation='none')
    plt.subplot(142)
    plt.imshow(utils.binarize(sample), cmap='Greys', interpolation='none')
    plt.subplot(143)
    plt.imshow(utils.quantisize(sample, 2), cmap='Greys', interpolation='none')
    plt.subplot(144)
    plt.imshow(simple(sample), cmap='Greys', interpolation='none')


plot(test[0])
plot(test[1])

#%%
B, CHANNELS, DIM, H, W = 16, 256, 3, 27, 26
input = np.zeros([B, CHANNELS * DIM, H, W])
indices = np.arange(CHANNELS * DIM)
indices % 3 == 2

input[:, indices % 3 == 0, :, :] = 1.0
input[:, indices % 3 == 1, :, :] = 2.0
input[:, indices % 3 == 2, :, :] = 3.0

r = np.reshape(input, [B, CHANNELS, DIM, H, W])
rt = np.transpose(r, [0, 2, 1, 3, 4])

r.shape
r[0, :, 2, 0, 0]

rt.shape


#%%

mask = np.ones([Cout, Cin, kh, kw])

yc, xc = kh // 2, kw // 2

mask[:, :, yc+1:, :] = 0.0
mask[:, :, yc:, xc+1:] = 0.0

mtype = 'B'
value = 0.0 if mtype == 'A' else 1.0




def bmask(i_out, i_in):
    cout_idx = np.expand_dims(np.arange(Cout) % 3 == i_out, 1)
    cin_idx = np.expand_dims(np.arange(Cin) % 3 == i_in, 0)
    a1, a2 = np.broadcast_arrays(cout_idx, cin_idx)
    return a1 * a2


for j in range(3):
    mask[bmask(j, j), yc, xc] = value

mask[bmask(1, 0), yc, xc] = 0.0
mask[bmask(2, 0), yc, xc] = 0.0
mask[bmask(2, 1), yc, xc] = 0.0


mask[bmask(0, 1), yc, xc] = .0
mask[bmask(0, 2), yc, xc] = .0

mask[bmask(1, 2), yc, xc] = .0

mask[:, :, yc, xc]


#%%
import importlib
importlib.reload(models)

Cin, Cout, kh, kw = 5, 12, 3, 3
link = models.MaskedConvolution2D(Cin, Cout, 3, mask='B', pad=1)
link.W.data = np.ones_like(link.W.data)

# model = models.PixelCNN(3, 16, 5, 8, 100)
# zeros = np.zeros([1, 3, 11, 13], dtype='f')

img = np.ones([1, 5, 6, 7]).astype('f')
img[:, 1, :, :] *= 10
img[:, 2, :, :] *= 100
img[:, 4, :, :] *= 10
img[:, 5, :, :] *= 100
img[:, 7, :, :] *= 10
img[:, 8, :, :] *= 100

h = link(img)


out = link(img).data.astype('i')
batch_size, channels, height, width = out.shape
print(batch_size, channels, height, width)
print(out[0, 0], out[0, 1], out[0, 2], out[0, 3], out[0, 11],  sep='\n')
out = np.reshape(out, [batch_size, 4, 3, height, width])
out = np.transpose(out, [0, 2, 1, 3, 4])
print(out[0, :, 0], out[0, :, 1], out[0, :, 2], sep='\n')

out.shape

122*4

#%%
import chainer.links as L

N = 3
conv1 = L.Convolution2D(1, 1, )
conv1.W.data

conv2 = L.Convolution2D(1, 1, [1, N//2+1], pad=[0, N//2+1], initialW=1.0)
conv2.W.data

# this filter rank is 1, so this convolution is separable
conv_combined = L.Convolution2D(1, 1, N, pad=1, initialW=1.0)
conv_combined.W.data


dims = 1, 1, 5, 6  # B, Cin, H, W
input = np.arange(np.prod(dims)).astype('f').reshape(dims)

input
conv2(conv1(input)).data
conv_combined(input).data

(conv2(conv1(input)) - conv_combined(input)).data
(conv1(conv2(input)) - conv_combined(input)).data


import wavenet.models as models
import importlib
importlib.reload(models)

conv1 = models.CroppedConvolution(1, 1, [1, 2], pad=[0, 2], initialW=1.0)
conv1.pad, conv1.ksize

input, conv1(input).data

#%%

import matplotlib.pyplot as plt
import chainer

plt.rcParams["figure.figsize"] = (16, 12)
train, test = chainer.datasets.get_cifar10()

def show(samples):
    for col, sample in enumerate(samples):
        image = sample[0].transpose(1, 2, 0)
        plt.subplot(1, len(samples), col+1)
        plt.imshow(image)
    plt.show()

show([sample for sample in train[:100] if sample[1] == 2])


#%%
import numpy as np
import matplotlib.pyplot as plt
import scipy.io.wavfile as wavfile
import scipy.signal as signal

rate, audio = wavfile.read('p225_001.wav')
audio_downsampled = audio[np.arange(len(audio)) % 4 == 0]
wavfile.write('p225_001_d.wav', 12000, audio_downsampled)

audio.shape, audio_downsampled.shape, audio.shape[0] / 4
_ = plt.hist(audio_downsampled, bins=100)

def split(audio, chunk):


#%%
import numpy as np

with open('test', 'wb') as ofile:
    for _ in range(2):
        array_to_save = np.random.randn(1, 1024)
        np.save(ofile, array_to_save)


total = []
with open('test', 'rb') as ifile:
    for _ in range(4):
        data = np.load(ifile)
        total.append(data)


#%%
importlib.reload(models)
layer = models.CausalDilatedConvolution1D(1, 1, pad=4, dilate=4, kernel_width=2, initialW=1.0)
input = np.arange(1, 9).reshape([1, 1, 1, 8]).astype(np.float32)

layer(input).data, layer(input).shape

layer.zerograds()

grads = np.zeros_like(output.data)
grads[0, 0, 0, 6] = 1
output.grad = grads

output.backward()
layer.W.grad


#%%
import importlib
import matplotlib.pyplot as plt
import scipy.io.wavfile as wavfile
importlib.reload(utils)

rate, audio.shape[0] // 6 / 1024  = wavfile.read('p225_001.wav')
_ = plt.hist(audio, bins=100)

transformed = utils.mulaw(utils.wav_to_float(audio))
_ = plt.hist(transformed, bins=100)

restored = utils.inverse_mulaw(transformed)
_ = plt.hist(restored, bins=100)

value = np.iinfo(np.int16).max
audio2 = (restored * value).astype(np.int16)
_ = plt.hist(audio2, bins=100)

#%%
overlap = 256
data = utils.VCTK('.', 1024)
audio, labels, _ = data.get_example(0)
audio2, labels2, _ = data.get_example(1)

for i in range(1024, 1050, 2):
    print(audio[0, 0, i], labels[0, 0, i])

a1, _, _ = data.get_example(0)
a2, _, _ = data.get_example(1)
a3, _, _ = data.get_example(2)
a4, _, _ = data.get_example(3)

_ = plt.hist(np.squeeze(audio), bins=100)
_ = plt.hist(np.squeeze(labels), bins=100)

batch = list(utils._preprocess('p225_001.wav', 8000, 1024))
_ = plt.hist(np.squeeze(batch[0]), bins=100)
_ = plt.hist(np.squeeze(utils.quantisize(batch[0], 256)), bins=100)

#%%
import os
import numpy as np

_levels = 256

with open('vctk_0', 'rb') as ifile:
    d = np.load(ifile)

count, width = d.shape
data = d
labels = utils.quantisize(data, _levels)
data = np.eye(_levels)[labels]
data = np.expand_dims(np.transpose(data, [0, 2, 1]), 2)
labels = np.reshape(labels, [count, 1, 1, width])
labels[:, :, :, :_receptive_field_size] = -1


data.shape, labels.shape