def _init_geometry(self, batch_win_size):
"""
Initializes:
self.enc_in_len
self.trim_ups_out
self.trim_dec_out
self.trim_dec_in
"""
# Calculate max length of mfcc encoder input and wav decoder input
w = batch_win_size
mfcc_vc = self.encoder.vc['beg'].parent
beg_grcc_vc = self.decoder.vc['beg_grcc']
end_grcc_vc = self.decoder.vc['end_grcc']
end_ups_vc = self.decoder.vc['last_upsample']
end_enc_vc = self.encoder.vc['end']
do = vconv.GridRange((0, 100000), (0, w), 1)
di = vconv.input_range(beg_grcc_vc, end_grcc_vc, do)
ei = vconv.input_range(mfcc_vc, end_grcc_vc, do)
mi = vconv.input_range(mfcc_vc.child, end_grcc_vc, do)
eo = vconv.output_range(mfcc_vc, end_enc_vc, ei)
uo = vconv.output_range(mfcc_vc, end_ups_vc, ei)
# Needed for trimming various tensors
self.enc_in_len = ei.sub_length()
self.enc_in_mel_len = mi.sub_length()
self.embed_len = eo.sub_length()
self.dec_in_len = di.sub_length()
self.trim_dec_in = torch.tensor([di.sub[0] - ei.sub[0], di.sub[1] -
ei.sub[0]], dtype=torch.long)
self.decoder.trim_ups_out = torch.tensor([di.sub[0] - uo.sub[0],
di.sub[1] - uo.sub[0]], dtype=torch.long)
self.trim_dec_out = torch.tensor([do.sub[0] - di.sub[0], do.sub[1] -
di.sub[0]], dtype=torch.long)
def _init_geometry(self, batch_win_size):
"""
Initializes lengths and trimming needed to produce batch_win_size
output
self.enc_in_len - encoder input length (timesteps)
self.dec_in_len - decoder input length (timesteps)
self.trim_ups_out - trims decoder lc_dense before use
self.trim_dec_out - trims wav_dec_input to wav_dec_output
self.trim_dec_in - trims wav_enc_input to wav_dec_input
The trimming vectors are needed because, due to striding geometry,
output tensors cannot be produced in single-increment sizes, therefore
must be over-produced in some cases.
"""
# Calculate max length of mfcc encoder input and wav decoder input
w = batch_win_size
mfcc_vc = self.encoder.vc['beg'].parent
end_enc_vc = self.encoder.vc['end']
end_ups_vc = self.decoder.vc['last_upsample']
beg_grcc_vc = self.decoder.vc['beg_grcc']
end_grcc_vc = self.decoder.vc['end_grcc']
# naming: (d: decoder, e: encoder, u: upsample), (o: output, i:input)
do = vconv.GridRange((0, 100000), (0, w), 1)
di = vconv.input_range(beg_grcc_vc, end_grcc_vc, do)
ei = vconv.input_range(mfcc_vc, end_grcc_vc, do)
mi = vconv.input_range(mfcc_vc.child, end_grcc_vc, do)
eo = vconv.output_range(mfcc_vc, end_enc_vc, ei)
uo = vconv.output_range(mfcc_vc, end_ups_vc, ei)
# Needed for trimming various tensors
self.enc_in_len = ei.sub_length()
self.enc_in_mel_len = mi.sub_length()
# used by jitter_index
self.embed_len = eo.sub_length()
# sets size for wav_dec_in
self.dec_in_len = di.sub_length()
# trims wav_enc_input to wav_dec_input
self.trim_dec_in = torch.tensor(
[di.sub[0] - ei.sub[0], di.sub[1] - ei.sub[0]], dtype=torch.long)
# needed by wavenet to trim upsampled local conditioning tensor
self.decoder.trim_ups_out = torch.tensor(
[di.sub[0] - uo.sub[0], di.sub[1] - uo.sub[0]], dtype=torch.long)
#
self.trim_dec_out = torch.tensor(
[do.sub[0] - di.sub[0], do.sub[1] - di.sub[0]], dtype=torch.long)
def autoenc_test(vcs, in_len, slice_beg):
enc = vcs['MFCC'], vcs['Upsampling_3']
dec = vcs['GRCC_0,0'], vcs['GRCC_1,9']
mfcc = vcs['MFCC'], vcs['MFCC']
autoenc = vcs['MFCC'], vcs['GRCC_1,9']
full_in = vconv.GridRange((0, in_len), (0, in_len), 1)
full_mfcc = vconv.output_range(*mfcc, full_in)
full_out = vconv.output_range(*autoenc, full_in)
out_req = vconv.GridRange(full_out.full, (slice_beg, slice_beg + 100), 1)
mid_req = vconv.input_range(*dec, out_req)
in_req = vconv.input_range(*enc, mid_req)
in_act = in_req
mfcc_act = vconv.output_range(*mfcc, in_act)
mid_act = vconv.output_range(*enc, in_act)
# wav -> wav_mid
wav_mid_sl = vconv.tensor_slice(in_act, mid_req.sub)
# wav_mid_ten = wav_ten[wav_mid_sl]
# lcond -> lcond_sl
lcond_sl = vconv.tensor_slice(mid_act, mid_req.sub)
# lcond_sl_ten = lcond_ten[lcond_sl]
# wav -> wav_out
# +1 since it is predicting the next step
wav_out_sl = vconv.tensor_slice(in_act, out_req.sub)
# wav_out_ten = wav_ten[sl_b+1:sl_e+1]
mfcc_in_sl = vconv.tensor_slice(full_mfcc, mfcc_act.sub)
print('{:10}: {}'.format('full_in', full_in))
print('{:10}: {}'.format('full_mfcc', full_mfcc))
print('{:10}: {}'.format('in_req', in_req))
print('{:10}: {}'.format('mfcc_req', mfcc_act))
print('{:10}: {}'.format('mid_req', mid_req))
print('{:10}: {}'.format('mid_act', mid_act))
print('{:10}: {}'.format('out_req', out_req))
print('{:10}: {}'.format('full_out', full_out))
print('wav_mid_sl: {} len: {}'.format(wav_mid_sl, wav_mid_sl[1] -
wav_mid_sl[0]))
print('mfcc_in_sl: {} len: {}'.format(mfcc_in_sl, mfcc_in_sl[1] -
mfcc_in_sl[0]))
print('lcond_sl: {} len: {}'.format(lcond_sl, lcond_sl[1] - lcond_sl[0]))
print('wav_out_sl: {} len: {}'.format(wav_out_sl, wav_out_sl[1] - wav_out_sl[0]))
def _init_geometry(self, batch_win_size):
"""
Initializes:
self.enc_in_len - timesteps of encoder input needed to
produce batch_win_size decoder output timesteps
self.trim_ups_out - offsets for trimming the upsampler output tensor
self.trim_dec_out - offsets for trimming the decoder output
self.trim_dec_in - offsets for trimming the decoder input
The trimming vectors are needed because, due to striding geometry,
output tensors cannot be produced in single-increment sizes, therefore
must be over-produced in some cases.
"""
# Calculate max length of mfcc encoder input and wav decoder input
w = batch_win_size
mfcc_vc = self.encoder.vc['beg'].parent
beg_grcc_vc = self.decoder.vc['beg_grcc']
end_grcc_vc = self.decoder.vc['end_grcc']
end_ups_vc = self.decoder.vc['last_upsample']
end_enc_vc = self.encoder.vc['end']
# naming: (d: decoder, e: encoder, u: upsample), (o: output, i:input)
do = vconv.GridRange((0, 100000), (0, w), 1)
di = vconv.input_range(beg_grcc_vc, end_grcc_vc, do)
ei = vconv.input_range(mfcc_vc, end_grcc_vc, do)
mi = vconv.input_range(mfcc_vc.child, end_grcc_vc, do)
eo = vconv.output_range(mfcc_vc, end_enc_vc, ei)
uo = vconv.output_range(mfcc_vc, end_ups_vc, ei)
# Needed for trimming various tensors
self.enc_in_len = ei.sub_length()
self.enc_in_mel_len = mi.sub_length()
self.embed_len = eo.sub_length()
self.dec_in_len = di.sub_length()
self.trim_dec_in = torch.tensor(
[di.sub[0] - ei.sub[0], di.sub[1] - ei.sub[0]], dtype=torch.long)
self.decoder.trim_ups_out = torch.tensor(
[di.sub[0] - uo.sub[0], di.sub[1] - uo.sub[0]], dtype=torch.long)
self.trim_dec_out = torch.tensor(
[do.sub[0] - di.sub[0], do.sub[1] - di.sub[0]], dtype=torch.long)
def downsample_test(vc, x):
try:
y = vconv.output_range(vc, vc, x)
except RuntimeError:
return Result.NO_OUTPUT
try:
xn = vconv.input_range(vc, vc, y)
except RuntimeError:
return Result.NO_INPUT
try:
yt = vconv.output_range(vc, vc, xn)
except RuntimeError:
return Result.NO_OUTPUT
try:
xt = vconv.input_range(vc, vc, yt)
except RuntimeError:
return Result.NO_INPUT
if xn != xt:
return Result.UNEQUAL
else:
return Result.SUCCESS
def same_or_upsample_test(vc, x):
try:
y = vconv.output_range(vc, vc, x)
except RuntimeError:
return Result.NO_OUTPUT
try:
xn = vconv.input_range(vc, vc, y)
except RuntimeError:
return Result.NO_INPUT
if xn != x:
return Result.UNEQUAL
else:
return Result.SUCCESS
def sample(self, wav_onehot, lc_sparse, speaker_inds, jitter_index, n_rep):
"""
Generate n_rep samples, using lc_sparse and speaker_inds for local and global
conditioning.
wav_onehot: full length wav vector
lc_sparse: full length local conditioning vector derived from full
wav_onehot
"""
# initialize model geometry
mfcc_vc = self.vc['beg'].parent
up_vc = self.vc['pre_upsample'].child
beg_grcc_vc = self.vc['beg_grcc']
end_vc = self.vc['end_grcc']
# calculate full output range
wav_gr = vconv.GridRange((0, 1e12), (0, wav_onehot.size()[2]), 1)
full_out_gr = vconv.output_range(mfcc_vc, end_vc, wav_gr)
n_ts = full_out_gr.sub_length()
# calculate starting input range for single timestep
one_gr = vconv.GridRange((0, 1e12), (0, 1), 1)
vconv.compute_inputs(end_vc, one_gr)
# calculate starting position of wav
wav_beg = int(beg_grcc_vc.input_gr.sub[0] - mfcc_vc.input_gr.sub[0])
# wav_end = int(beg_grcc_vc.input_gr.sub[1] - mfcc_vc.input_gr.sub[0])
wav_onehot = wav_onehot[:,:,wav_beg:]
# !!! hack - I'm not sure why the int() cast is necessary
n_init_ts = int(beg_grcc_vc.in_len())
lc_sparse = lc_sparse.repeat(n_rep, 1, 1)
jitter_index = jitter_index.repeat(n_rep, 1)
speaker_inds = speaker_inds.repeat(n_rep)
# precalculate conditioning vector for all timesteps
D1 = lc_sparse.size()[1]
lc_jitter = torch.take(lc_sparse,
jitter_index.unsqueeze(1).expand(-1, D1, -1))
lc_conv = self.lc_conv(lc_jitter)
lc_dense = self.lc_upsample(lc_conv)
cond = self.cond(lc_dense, speaker_inds)
n_ts = cond.size()[2]
# cond_loff, cond_roff = vconv.output_offsets(mfcc_vc, up_end_vc)
# zero out
start_pos = 26000
n_samples = 20000
end_pos = start_pos + n_samples
# wav_onehot[...,n_init_ts:] = 0
wav_onehot = wav_onehot.repeat(n_rep, 1, 1)
# wav_onehot[...,start_pos:end_pos] = 0
# assert cond.size()[2] == wav_onehot.size()[2]
# loop through timesteps
# inrange = torch.tensor((0, n_init_ts), dtype=torch.int32)
inrange = torch.tensor((start_pos - n_init_ts, start_pos), dtype=torch.int32)
# end_ind = torch.tensor([n_ts], dtype=torch.int32)
end_ind = torch.tensor([end_pos], dtype=torch.int32)
# inefficient - this recalculates intermediate activations for the
# entire receptive fields, rather than just the advancing front
while not torch.equal(inrange[1], end_ind[0]):
# while inrange[1] != end_ind[0]:
sig = self.base_layer(wav_onehot[:,:,inrange[0]:inrange[1]])
sig, skp_sum = self.conv_layers[0](sig, cond[:,:,inrange[0]:inrange[1]])
for layer in self.conv_layers[1:]:
sig, skp = layer(sig, cond[:,:,inrange[0]:inrange[1]])
skp_sum += skp
post1 = self.post1(self.relu(skp_sum))
quant = self.post2(self.relu(post1))
cat = dcat.OneHotCategorical(logits=quant.squeeze(2))
wav_onehot[1:,:,inrange[1]] = cat.sample()[1:,...]
inrange += 1
if inrange[0] % 100 == 0:
print(inrange, end_ind[0])
# convert to value format
quant_range = wav_onehot.new(list(range(self.n_quant)))
wav = torch.matmul(wav_onehot.permute(0,2,1), quant_range)
torch.set_printoptions(threshold=100000)
pad = 5
print('padding = {}'.format(pad))
print('original')
print(wav[0,start_pos-pad:end_pos+pad])
print('synth')
print(wav[1,start_pos-pad:end_pos+pad])
# print(wav[:,end_pos:end_pos + 10000])
print('synth range std: {}, baseline std: {}'.format(
wav[:,start_pos:end_pos].std(), wav[:,end_pos:].std()
))
return wav
for vc, x in input_gen(t):
if vc.stride_ratio.numerator > 1:
res = downsample_test(vc, x)
else:
res = same_or_upsample_test(vc, x)
results[res] += 1
if c > 0 and c % t.report_freq == 0:
print(results)
c += 1
print('Finished')
print('Results: {}'.format(results))
x = vconv.GridRange((0, 250000), (0, 250000), 1)
y = vconv.output_range(vcs['MFCC'], vcs['GRCC_1,9'], x)
xi = vconv.input_range(vcs['MFCC'], vcs['GRCC_1,9'], y)
#print('x0: {}'.format(x))
#print('y0: {}'.format(y))
#print('xi: {}'.format(xi))
def autoenc_test(vcs, in_len, slice_beg):
enc = vcs['MFCC'], vcs['Upsampling_3']
dec = vcs['GRCC_0,0'], vcs['GRCC_1,9']
mfcc = vcs['MFCC'], vcs['MFCC']
autoenc = vcs['MFCC'], vcs['GRCC_1,9']
full_in = vconv.GridRange((0, in_len), (0, in_len), 1)
full_mfcc = vconv.output_range(*mfcc, full_in)