def test_dconv_zeros(backend, zeros_convargs):
fshape, nofm, batch_size = zeros_convargs
NervanaObject.be.bsz = NervanaObject.be.bs = batch_size
dtypeu = np.float32
init_unif = Uniform(low=0.0, high=0.0)
inshape = (64, 28, 28)
insize = np.prod(inshape)
neon_layer = Deconv(fshape=(fshape, fshape, nofm), strides=1, padding=0, init=init_unif)
inp_arr_shape = (insize, batch_size)
inp = np.random.random(inp_arr_shape).astype(dtypeu)
inp = neon_layer.be.array(inp)
inp.lshape = inshape
outa = neon_layer.fprop(inp)
out = outa.asnumpyarray()
assert np.min(out) == 0.0 and np.max(out) == 0.0
err = dtypeu(np.zeros(outa.shape))
deltas = neon_layer.bprop(NervanaObject.be.array(err)).asnumpyarray()
assert np.min(deltas) == 0.0 and np.max(deltas) == 0.0
dw = neon_layer.dW.asnumpyarray()
assert np.min(dw) == 0.0 and np.max(dw) == 0.0
return
def test_dconv_zeros(backend, zeros_convargs):
fshape, nofm, batch_size = zeros_convargs
NervanaObject.be.bsz = NervanaObject.be.bs = batch_size
dtypeu = np.float32
init_unif = Uniform(low=0.0, high=0.0)
inshape = (64, 28, 28)
insize = np.prod(inshape)
neon_layer = Deconv(fshape=(fshape, fshape, nofm),
strides=1,
padding=0,
init=init_unif)
inp_arr_shape = (insize, batch_size)
inp = np.random.random(inp_arr_shape).astype(dtypeu)
inp = neon_layer.be.array(inp)
inp.lshape = inshape
outa = neon_layer.fprop(inp)
out = outa.asnumpyarray()
assert np.min(out) == 0.0 and np.max(out) == 0.0
err = dtypeu(np.zeros(outa.shape))
deltas = neon_layer.bprop(NervanaObject.be.array(err)).asnumpyarray()
assert np.min(deltas) == 0.0 and np.max(deltas) == 0.0
dw = neon_layer.dW.asnumpyarray()
assert np.min(dw) == 0.0 and np.max(dw) == 0.0
return
def test_dconv_ones(backend, ones_convargs):
indim, nifm, fshape, nofm, batch_size = ones_convargs
NervanaObject.be.bsz = NervanaObject.be.bs = batch_size
dtypeu = np.float32
# weights set to one
init_unif = Uniform(low=1.0, high=1.0)
inshape = (nifm, indim, indim)
insize = np.prod(inshape)
neon_layer = Deconv(fshape=(fshape, fshape, nofm),
strides=1,
padding=0,
init=init_unif)
inp = neon_layer.be.array(np.ones((insize, batch_size)).astype(dtypeu))
inp.lshape = inshape
# run fprop
out = neon_layer.fprop(inp).asnumpyarray()
out_exp_min = nifm
out_exp_max = fshape * fshape * nifm
assert np.min(out) == out_exp_min and np.max(out) == out_exp_max
# generate err array
err = np.ones(out.shape).astype(dtypeu)
# run bprop
neon_layer.bprop(NervanaObject.be.array(err)).asnumpyarray()
dw = neon_layer.dW.asnumpyarray()
# generate the reference layer
ref_layer = DeconvRefLayer(1, batch_size, identity, inshape[0],
inshape[1:3], (fshape, fshape), nofm, 1, dtypeu)
ref_layer.weights = np.ones(neon_layer.W.shape).T.astype(dtypeu)
# run bprop
ref_layer.bprop(err)
# expected output for updates is uniform matrix with
# all elements == ofmsize*batch_size
updates_exp = ref_layer.ofmsize * batch_size
# check dw from neon layer
assert np.max(dw) == updates_exp and np.min(dw) == updates_exp
# no tolerence here should be exact
assert np.max(np.abs(ref_layer.y.T - neon_layer.deltas.get())) == 0.0
return
def test_dconv_rand(backend, rand_convargs):
indim, nifm, fshape, nofm, batch_size, rngmax, w_rng = rand_convargs
NervanaObject.be.bsz = NervanaObject.be.bs = batch_size
dtypeu = np.float32
inp_rng = [0.0, rngmax]
init_unif = Uniform(low=w_rng[0], high=w_rng[1])
inshape = (indim, indim, nifm)
insize = np.prod(inshape)
# generate neon deconv layer
# need to switch to nofm here...
neon_layer = Deconv(fshape=(fshape, fshape, nofm),
strides=1,
padding=0,
init=init_unif)
insize = np.prod(inshape)
# generate reference deconv layer
ref_layer = DeconvRefLayer(1, batch_size, identity, inshape[0],
inshape[1:3], (fshape, fshape), nofm, 1, dtypeu)
# setup input in range inp_rng
inpa = np.random.random((insize, batch_size))
inpa *= (inp_rng[1] - inp_rng[0])
inpa += inp_rng[0]
inpa = inpa.astype(dtypeu)
inp = neon_layer.be.array(inpa)
inp.lshape = inshape
# run fprop on neon
neon_out = neon_layer.fprop(inp).asnumpyarray()
# pull neon weights into ref layer weights
ref_layer.weights = neon_layer.W.asnumpyarray().T
ref_out = np.copy(ref_layer.berror)
# estimate the numerical precision
ref_layer.fprop(inpa.T, permute=True)
ref_out2 = ref_layer.berror
atol = 10 * np.max(np.abs(ref_out - ref_out2))
assert (np.allclose(ref_out.T, neon_out, atol=atol, rtol=0.0),
'%e %e' % (np.max(np.abs(ref_out.T - neon_out)), atol))
# generate err array
erra = np.random.random(neon_out.shape)
erra *= (inp_rng[1] - inp_rng[0])
erra += inp_rng[0]
erra = erra.astype(dtypeu)
def test_dconv_rand(backend, rand_convargs):
indim, nifm, fshape, nofm, batch_size, rngmax, w_rng = rand_convargs
NervanaObject.be.bsz = NervanaObject.be.bs = batch_size
dtypeu = np.float32
inp_rng = [0.0, rngmax]
init_unif = Uniform(low=w_rng[0], high=w_rng[1])
inshape = (indim, indim, nifm)
insize = np.prod(inshape)
# generate neon deconv layer
# need to switch to nofm here...
neon_layer = Deconv(fshape=(fshape, fshape, nofm), strides=1, padding=0, init=init_unif)
insize = np.prod(inshape)
# generate reference deconv layer
ref_layer = DeconvRefLayer(1, batch_size, identity, inshape[0], inshape[1:3], (fshape, fshape), nofm, 1, dtypeu)
# setup input in range inp_rng
inpa = np.random.random((insize, batch_size))
inpa *= inp_rng[1] - inp_rng[0]
inpa += inp_rng[0]
inpa = inpa.astype(dtypeu)
inp = neon_layer.be.array(inpa)
inp.lshape = inshape
# run fprop on neon
neon_out = neon_layer.fprop(inp).asnumpyarray()
# pull neon weights into ref layer weights
ref_layer.weights = neon_layer.W.asnumpyarray().T
ref_out = np.copy(ref_layer.berror)
# estimate the numerical precision
ref_layer.fprop(inpa.T, permute=True)
ref_out2 = ref_layer.berror
atol = 10 * np.max(np.abs(ref_out - ref_out2))
assert (
np.allclose(ref_out.T, neon_out, atol=atol, rtol=0.0),
"%e %e" % (np.max(np.abs(ref_out.T - neon_out)), atol),
)
# generate err array
erra = np.random.random(neon_out.shape)
erra *= inp_rng[1] - inp_rng[0]
erra += inp_rng[0]
erra = erra.astype(dtypeu)
def test_dconv_ones(backend, ones_convargs):
indim, nifm, fshape, nofm, batch_size = ones_convargs
NervanaObject.be.bsz = NervanaObject.be.bs = batch_size
dtypeu = np.float32
# weights set to one
init_unif = Uniform(low=1.0, high=1.0)
inshape = (nifm, indim, indim)
insize = np.prod(inshape)
neon_layer = Deconv(fshape=(fshape, fshape, nofm), strides=1, padding=0, init=init_unif)
inp = neon_layer.be.array(np.ones((insize, batch_size)).astype(dtypeu))
inp.lshape = inshape
# run fprop
out = neon_layer.fprop(inp).asnumpyarray()
out_exp_min = nifm
out_exp_max = fshape * fshape * nifm
assert np.min(out) == out_exp_min and np.max(out) == out_exp_max
# generate err array
err = np.ones(out.shape).astype(dtypeu)
# run bprop
neon_layer.bprop(NervanaObject.be.array(err)).asnumpyarray()
dw = neon_layer.dW.asnumpyarray()
# generate the reference layer
ref_layer = DeconvRefLayer(1, batch_size, identity, inshape[0], inshape[1:3], (fshape, fshape), nofm, 1, dtypeu)
ref_layer.weights = np.ones(neon_layer.W.shape).T.astype(dtypeu)
# run bprop
ref_layer.bprop(err)
# expected output for updates is uniform matrix with
# all elements == ofmsize*batch_size
updates_exp = ref_layer.ofmsize * batch_size
# check dw from neon layer
assert np.max(dw) == updates_exp and np.min(dw) == updates_exp
# no tolerence here should be exact
assert np.max(np.abs(ref_layer.y.T - neon_layer.deltas.get())) == 0.0
return
# Set input and target to X_train
train = ArrayIterator(X_train, lshape=(1, 28, 28))
# Initialize the weights and the learning rule
init_uni = Uniform(low=-0.1, high=0.1)
opt_gdm = GradientDescentMomentum(learning_rate=0.001, momentum_coef=0.9)
# Strided conv autoencoder
bn = False
layers = [
Conv((4, 4, 8), init=init_uni, activation=Rectlin(), batch_norm=bn),
Pooling(2),
Conv((4, 4, 32), init=init_uni, activation=Rectlin(), batch_norm=bn),
Pooling(2),
Deconv(fshape=(4, 4, 8),
init=init_uni,
activation=Rectlin(),
batch_norm=bn),
Deconv(fshape=(3, 3, 8),
init=init_uni,
activation=Rectlin(),
strides=2,
batch_norm=bn),
Deconv(fshape=(2, 2, 1), init=init_uni, strides=2, padding=1)
]
# Define the cost
cost = GeneralizedCost(costfunc=SumSquared())
model = Model(layers=layers)
# configure callbacks
# setup weight initialization function
init = Gaussian(scale=0.05)
# generator using "decovolution" layers
relu = Rectlin(slope=0) # relu for generator
conv = dict(init=init, batch_norm=True, activation=relu)
convp1 = dict(init=init, batch_norm=True, activation=relu, padding=1)
convp2 = dict(init=init, batch_norm=True, activation=relu, padding=2)
convp1s2 = dict(init=init,
batch_norm=True,
activation=relu,
padding=1,
strides=2)
G_layers = [
Deconv((1, 1, 16), name="G11", **conv),
Deconv((3, 3, 192), name="G12", **convp1),
Deconv((3, 3, 192), name="G21", **convp1s2),
Deconv((3, 3, 192), name="G22", **convp1),
Deconv((3, 3, 96), name="G31", **convp1s2),
Deconv((3, 3, 96), name="G32", **conv),
Deconv((3, 3, 1),
name="G_out",
init=init,
batch_norm=False,
padding=1,
activation=Logistic(shortcut=False))
]
# discriminiator using convolution layers
lrelu = Rectlin(slope=0.1) # leaky relu for discriminator
pad1 = dict(pad_h=2, pad_w=2, pad_d=2)
str1 = dict(str_h=2, str_w=2, str_d=2)
conv1 = dict(init=init_gen, batch_norm=False, activation=lrelu, padding=pad1, strides=str1, bias=init_gen)
pad2 = dict(pad_h=2, pad_w=2, pad_d=2)
str2 = dict(str_h=2, str_w=2, str_d=2)
conv2 = dict(init=init_gen, batch_norm=False, activation=lrelu, padding=pad2, strides=str2, bias=init_gen)
pad3 = dict(pad_h=0, pad_w=0, pad_d=0)
str3 = dict(str_h=1, str_w=1, str_d=1)
conv3 = dict(init=init_gen, batch_norm=False, activation=Tanh(), padding=pad3, strides=str3, bias=init_gen)
bg = BranchNode("bg")
branchg = [bg,
Affine(1024, init=init_gen, bias=init_gen, activation=relu),
BatchNorm(),
Affine(8 * 7 * 7 * 7, init=init_gen, bias=init_gen),
Reshape((8, 7, 7, 7)),
Deconv((6, 6, 6, 6), **conv1), #14x14x14
BatchNorm(),
# Linear(5 * 14 * 14 * 14, init=init),
# Reshape((5, 14, 14, 14)),
Deconv((5, 5, 5, 64), **conv2), #27x27x27
BatchNorm(),
Conv((3, 3, 3, 1), **conv3)
]
G_layers = Tree([branchg], name="Generator")
print D_layers
print G_layers
layers = myGenerativeAdversarial(generator=G_layers, discriminator=D_layers)
#discriminator=Sequential(D_layers, name="Discriminator"))
print 'layers defined'
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=args.data_dir)
# Set input and target to X_train
train = DataIterator(X_train, lshape=(1, 28, 28))
# Initialize the weights and the learning rule
init_uni = Uniform(low=-0.1, high=0.1)
opt_gdm = GradientDescentMomentum(learning_rate=0.001, momentum_coef=0.9)
# Define the layers
layers = [
Conv((4, 4, 8), init=init_uni, activation=Rectlin()),
Pooling(2),
Conv((4, 4, 32), init=init_uni, activation=Rectlin()),
Pooling(2),
Deconv(fshape=(3, 3, 8), init=init_uni, strides=2, padding=1),
Deconv(fshape=(3, 3, 8), init=init_uni, strides=2, padding=1),
Deconv(fshape=(4, 4, 1), init=init_uni, strides=2, padding=0)
]
# Define the cost
cost = GeneralizedCost(costfunc=SumSquared())
mlp = Model(layers=layers)
# Fit the model
# configure callbacks
callbacks = Callbacks(mlp, train, **args.callback_args)
mlp.fit(train,
optimizer=opt_gdm,
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=args.data_dir)
# Set input and target to X_train
train = DataIterator(X_train, lshape=(1, 28, 28))
# Initialize the weights and the learning rule
init_uni = Uniform(low=-0.1, high=0.1)
opt_gdm = GradientDescentMomentum(learning_rate=0.001, momentum_coef=0.9)
# Define the layers
layers = []
layers.append(Conv((4, 4, 8), init=init_uni, activation=Rectlin()))
layers.append(Pooling(2))
layers.append(Conv((4, 4, 32), init=init_uni, activation=Rectlin()))
layers.append(Pooling(2))
layers.append(Deconv(fshape=(4, 4, 8), init=init_uni))
layers.append(Deconv(fshape=(2, 2, 8), init=init_uni, strides=2))
layers.append(Deconv(fshape=(2, 2, 1), init=init_uni, strides=2))
# Define the cost
cost = GeneralizedCost(costfunc=SumSquared())
mlp = Model(layers=layers)
# Fit the model
# configure callbacks
callbacks = Callbacks(mlp,
train,
output_file=args.output_file,
progress_bar=args.progress_bar)
train.init_batch_provider()
test.init_batch_provider()
init = Gaussian(scale=0.1)
opt = Adadelta(decay=0.9)
common = dict(init=init, batch_norm=True, activation=Rectlin())
# Set up the model layers
layers = []
nchan = 128
layers.append(Conv((2, 2, nchan), strides=2, **common))
for idx in range(16):
layers.append(Conv((3, 3, nchan), **common))
if nchan > 16:
nchan /= 2
for idx in range(15):
layers.append(Deconv((3, 3, nchan), **common))
layers.append(Deconv((4, 4, nchan), strides=2, **common))
layers.append(Deconv((3, 3, 1), init=init, activation=Logistic(shortcut=True)))
cost = GeneralizedCost(costfunc=SumSquared())
mlp = Model(layers=layers)
callbacks = Callbacks(mlp, train, **args.callback_args)
evaluator = Evaluator(callbacks.callback_data, mlp, test, imwidth, args.epochs,
args.data_dir, point_num)
callbacks.add_callback(evaluator)
mlp.fit(train,
optimizer=opt,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
train.exit_batch_provider()
