def test_callback():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
input_image = tf.constant(0.69, tf.float64, [2, 2, 5, 5, 2])
conv_filter = tf.constant(0.01, tf.float64, [1, 1, 1, 2, 2])
tests = []
tests.append(
tf.nn.conv3d(input_image,
conv_filter,
strides=[1, 1, 1, 1, 1],
padding="VALID"))
myinput = tf.constant(0.69, tf.float64, [2, 2])
tests.append(tf.sigmoid(myinput))
myinput = np.random.rand(2, 3, 4).astype(np.float64)
tests.append(tf.reduce_max(myinput))
myinput = np.random.rand(10).astype(np.float64)
tests.append(tf.nn.top_k(myinput, 4)[0])
tests.append(tf.nn.top_k(myinput, 4)[1])
sess_tf = tf.Session()
sess_dace = TFSession()
for test in tests:
output_tf = sess_tf.run(test)
output_dace = sess_dace.run(test)
print(output_dace)
print(output_tf)
assert np.linalg.norm(output_dace - output_tf) < 1e-8
def test_mean():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
shape = [10, 11, 12, 13]
inp = tf.placeholder(tf.float64, shape)
outp_1 = tf.reduce_mean(inp, keepdims=True)
outp_3 = tf.reduce_mean(inp, axis=[0, 2], keepdims=True)
outp_0 = tf.reduce_mean(inp, axis=[0, 2])
outp_2 = tf.reduce_mean(inp, axis=[-2, -1])
outp_4 = tf.reduce_mean(inp, axis=[0, -1], keepdims=True)
sess_tf = tf.Session()
sess_dace = TFSession()
real_inp = np.random.rand(*shape)
for index, op in enumerate([outp_0, outp_1, outp_2, outp_3, outp_4]):
output_tf = sess_tf.run(op, feed_dict={inp: real_inp})
output_dace = sess_dace.run(op, feed_dict={inp: real_inp})
try:
assert tf.norm(output_dace -
output_tf).eval(session=sess_tf) < 1e-10
except:
print(output_dace)
print(output_tf)
print(tf.norm(output_dace - output_tf).eval(session=sess_tf))
raise AssertionError("mean test {i} failed".format(i=index))
print("mean tests passed!")
def test_shapen():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
myshape = [69, 96, 666]
num_inputs = 5
inpList = [tf.ones(myshape) for _ in range(num_inputs)]
sess_tf = tf.Session()
sess_dace = TFSession()
shapes_tf = sess_tf.run(tf.shape_n(inpList))
shapes_dace = sess_dace.run(tf.shape_n(inpList))
for dc, tf in zip(shapes_dace, shapes_tf):
try:
assert (dc == tf).all()
except (AssertionError):
print(dc)
print(tf)
def test_addn():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
shape = [10, 11, 12, 13]
inputs = [np.random.rand(*shape) for _ in range(10)]
addn_test_0 = tf.add_n(inputs)
sess_tf = tf.Session()
sess_dace = TFSession()
output_tf = sess_tf.run(addn_test_0)
output_dace = sess_dace.run(addn_test_0)
try:
assert tf.norm(output_dace - output_tf).eval(session=sess_tf) < 1e-10
except:
print(output_dace)
print(output_tf)
print(tf.norm(output_dace - output_tf).eval(session=sess_tf))
raise AssertionError("AddN test failed")
print("AddN test passed!")
def test_slice():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
t = tf.placeholder(tf.int32, [3, 2, 3])
b = tf.placeholder(tf.int32, [3])
s = tf.placeholder(tf.int32, [3])
output = tf.placeholder(tf.int32, [1, 1, 3])
output = tf.slice(t, b, s)
input_tensor = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]],
[[5, 5, 5], [6, 6, 6]]])
sess_tf = tf.Session()
sess_dace = TFSession()
begin_tensor = tf.constant([1, 0, 0])
size_tensor_1 = tf.constant([1, 2, 2])
size_tensor_2 = tf.constant([1, 2, 3])
size_tensor_3 = tf.constant([2, 1, 3])
tf_out = sess_tf.run(tf.slice(input_tensor, begin_tensor, size_tensor_3))
dace_out = sess_dace.run(
tf.slice(input_tensor, begin_tensor, size_tensor_3))
print(tf_out)
print(dace_out)
assert (tf_out == dace_out).all()
def test_simple():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
print('DaCe Tensorflow frontend test')
A = np.random.rand(16, 16).astype(np.float32)
B = np.random.rand(16, 16).astype(np.float32)
A_tf = tf.placeholder(tf.float32, shape=[16, 16])
B_tf = tf.placeholder(tf.float32, shape=[16, 16])
with TFSession() as sess:
# Simple matrix multiplication
C = sess.run(A_tf @ B_tf, feed_dict={A_tf: A, B_tf: B})
diff = np.linalg.norm(C - (A @ B)) / (16 * 16)
print("Difference:", diff)
print("==== Program end ====")
assert diff <= 1e-5
try:
import tensorflow as tf
except ImportError:
print("WARNING: Tensorflow not found, skipping test")
exit(0)
from dace.frontend.tensorflow import TFSession
if __name__ == '__main__':
t = tf.placeholder(tf.int32, [3, 2, 3])
b = tf.placeholder(tf.int32, [3])
s = tf.placeholder(tf.int32, [3])
output = tf.placeholder(tf.int32, [1, 1, 3])
output = tf.slice(t, b, s)
input_tensor = tf.constant([[[1, 1, 1], [2, 2, 2]], [[3, 3, 3], [4, 4, 4]],
[[5, 5, 5], [6, 6, 6]]])
sess_tf = tf.Session()
sess_dace = TFSession()
begin_tensor = tf.constant([1, 0, 0])
size_tensor_1 = tf.constant([1, 2, 2])
size_tensor_2 = tf.constant([1, 2, 3])
size_tensor_3 = tf.constant([2, 1, 3])
tf_out = sess_tf.run(tf.slice(input_tensor, begin_tensor, size_tensor_3))
dace_out = sess_dace.run(
tf.slice(input_tensor, begin_tensor, size_tensor_3))
print(tf_out)
print(dace_out)
assert (tf_out == dace_out).all()
input_placeholder = tf.placeholder(tf.float32, size_in)
ksize = [1, 3, 3, 1]
stride = [1, 2, 2, 1]
# need to fix bug in padding SAME
max_pool_outp = tf.nn.max_pool(input_placeholder,
ksize,
stride,
"VALID",
data_format="NHWC")
avg_pool_outp = tf.nn.avg_pool(input_placeholder,
ksize,
stride,
"VALID",
data_format="NHWC")
sess_tf = tf.Session()
sess_dace = TFSession()
# MAX pool test
tf_output = sess_tf.run(max_pool_outp,
feed_dict={input_placeholder: input_tensor})
dace_output = sess_dace.run(max_pool_outp,
feed_dict={input_placeholder: input_tensor})
try:
assert tf.norm(dace_output - tf_output).eval(session=sess_tf) < 1e-10
except:
print(dace_output.shape)
print(tf_output.shape)
print(tf.norm(dace_output - tf_output).eval(session=sess_tf))
raise AssertionError("max pool test failed")
print("Max pool test passed")
# AVG pool test
if __name__ == '__main__':
input_image = tf.constant(0.69, tf.float64, [2, 2, 5, 5, 2])
conv_filter = tf.constant(0.01, tf.float64, [1, 1, 1, 2, 2])
tests = []
tests.append(
tf.nn.conv3d(input_image,
conv_filter,
strides=[1, 1, 1, 1, 1],
padding="VALID"))
myinput = tf.constant(0.69, tf.float64, [2, 2])
tests.append(tf.sigmoid(myinput))
myinput = np.random.rand(2, 3, 4).astype(np.float64)
tests.append(tf.reduce_max(myinput))
myinput = np.random.rand(10).astype(np.float64)
tests.append(tf.nn.top_k(myinput, 4)[0])
tests.append(tf.nn.top_k(myinput, 4)[1])
sess_tf = tf.Session()
sess_dace = TFSession()
for test in tests:
output_tf = sess_tf.run(test)
output_dace = sess_dace.run(test)
print(output_dace)
print(output_tf)
assert np.linalg.norm(output_dace - output_tf) < 1e-8
def test_pooling():
import tensorflow as tf
from dace.frontend.tensorflow import TFSession
size_in = [1, 112, 112, 3]
# size_in = [4, 4, 4, 4]
np.random.seed(0)
input_tensor = np.random.uniform(size=size_in).astype(np.float32)
input_placeholder = tf.placeholder(tf.float32, size_in)
ksize = [1, 3, 3, 1]
stride = [1, 2, 2, 1]
# need to fix bug in padding SAME
max_pool_outp = tf.nn.max_pool(input_placeholder,
ksize,
stride,
"VALID",
data_format="NHWC")
avg_pool_outp = tf.nn.avg_pool(input_placeholder,
ksize,
stride,
"VALID",
data_format="NHWC")
sess_tf = tf.Session()
sess_dace = TFSession()
# MAX pool test
tf_output = sess_tf.run(max_pool_outp,
feed_dict={input_placeholder: input_tensor})
dace_output = sess_dace.run(max_pool_outp,
feed_dict={input_placeholder: input_tensor})
try:
assert tf.norm(dace_output - tf_output).eval(session=sess_tf) < 1e-10
except:
print(dace_output.shape)
print(tf_output.shape)
print(tf.norm(dace_output - tf_output).eval(session=sess_tf))
raise AssertionError("max pool test failed")
print("Max pool test passed")
# AVG pool test
tf_output = sess_tf.run(avg_pool_outp,
feed_dict={input_placeholder: input_tensor})
dace_output = sess_dace.run(avg_pool_outp,
feed_dict={input_placeholder: input_tensor})
try:
assert tf.norm(dace_output - tf_output).eval(session=sess_tf) < 1e-5
except:
print(dace_output.shape)
print(tf_output.shape)
print(tf.norm(dace_output - tf_output).eval(session=sess_tf))
raise AssertionError("avg pool test failed")
print("Average pool test passed")
# AVG pool gradient test
np.random.seed(0)
loss_placeholder = tf.placeholder(tf.float32, avg_pool_outp.shape)
loss_tensor = np.random.uniform(size=avg_pool_outp.shape).astype(
np.float32)
grads_avg = tf.gradients(avg_pool_outp,
input_placeholder,
grad_ys=loss_placeholder)
dace_output = sess_dace.run(grads_avg,
feed_dict={loss_placeholder: loss_tensor})
tf_output = sess_tf.run(grads_avg,
feed_dict={loss_placeholder: loss_tensor})
try:
assert tf.norm(dace_output[0] -
tf_output[0]).eval(session=sess_tf) < 1e-5
except:
print(dace_output)
print(tf_output)
print(tf.norm(dace_output[0] - tf_output[0]).eval(session=sess_tf))
raise AssertionError("avg pool gradient test failed")
# Max pool gradient test
loss_placeholder = tf.placeholder(tf.float32, max_pool_outp.shape)
np.random.seed(0)
loss_tensor = np.random.uniform(size=max_pool_outp.shape).astype(
np.float32)
grads_max = tf.gradients(max_pool_outp,
input_placeholder,
grad_ys=loss_placeholder)
dace_output = sess_dace.run(
grads_max,
feed_dict={
input_placeholder: input_tensor,
loss_placeholder: loss_tensor
},
)
tf_output = sess_tf.run(
grads_max,
feed_dict={
input_placeholder: input_tensor,
loss_placeholder: loss_tensor
},
)
try:
assert tf.norm(dace_output[0] -
tf_output[0]).eval(session=sess_tf) < 1e-5
except:
print(dace_output)
print(tf_output)
print(tf.norm(dace_output[0] - tf_output[0]).eval(session=sess_tf))
raise AssertionError("max pool gradient test failed")
import numpy as np
import dace
from dace.frontend.tensorflow import TFSession
if __name__ == '__main__':
shape = [10, 11, 12, 13]
inp = tf.placeholder(tf.float64, shape)
outp_1 = tf.reduce_mean(inp, keepdims=True)
outp_3 = tf.reduce_mean(inp, axis=[0, 2], keepdims=True)
outp_0 = tf.reduce_mean(inp, axis=[0, 2])
outp_2 = tf.reduce_mean(inp, axis=[-2, -1])
outp_4 = tf.reduce_mean(inp, axis=[0, -1], keepdims=True)
sess_tf = tf.Session()
sess_dace = TFSession()
real_inp = np.random.rand(*shape)
for index, op in enumerate([outp_0, outp_1, outp_2, outp_3, outp_4]):
output_tf = sess_tf.run(op, feed_dict={inp: real_inp})
output_dace = sess_dace.run(op, feed_dict={inp: real_inp})
try:
assert tf.norm(output_dace -
output_tf).eval(session=sess_tf) < 1e-10
except:
print(output_dace)
print(output_tf)
print(tf.norm(output_dace - output_tf).eval(session=sess_tf))
raise AssertionError("mean test {i} failed".format(i=index))
print("mean tests passed!")
inputs = [np.random.rand(*shape) for _ in range(10)]
import numpy as np
import dace
from dace.frontend.tensorflow import TFSession
inp_shape = [10, 10, 10, 10]
filter_shape = [3, 3, 10, 3]
strides = [1, 3, 3, 1]
inp = tf.placeholder(tf.float64, inp_shape)
filter = tf.placeholder(tf.float64, filter_shape)
outp = tf.nn.conv2d(inp, filter, strides, padding="SAME", data_format="NHWC")
test_in = np.random.uniform(size=tuple(inp_shape)).astype(np.float64)
test_filter = np.random.uniform(size=tuple(filter_shape)).astype(np.float64)
sess_dace = TFSession()
sess_tf = tf.Session()
output_dace = sess_dace.run(outp,
feed_dict={
inp: test_in,
filter: test_filter
})
output_tf = sess_tf.run(outp, feed_dict={inp: test_in, filter: test_filter})
try:
assert tf.norm(output_dace - output_tf).eval(session=sess_tf) < 1e-10
except:
print(output_tf)
print(output_dace)
print(tf.linalg.norm(output_tf - output_dace).eval(session=sess_tf))
raise AssertionError("Convolution test failed")
try:
import tensorflow as tf
except ImportError:
print("WARNING: Tensorflow not found, skipping test")
exit(0)
from dace.frontend.tensorflow import TFSession
myshape = [69, 96, 666]
num_inputs = 5
inpList = [tf.ones(myshape) for _ in range(num_inputs)]
sess_tf = tf.Session()
sess_dace = TFSession()
shapes_tf = sess_tf.run(tf.shape_n(inpList))
shapes_dace = sess_dace.run(tf.shape_n(inpList))
for dc, tf in zip(shapes_dace, shapes_tf):
try:
assert (dc == tf).all()
except (AssertionError):
print(dc)
print(tf)
# Copyright 2019-2020 ETH Zurich and the DaCe authors. All rights reserved.
try:
import tensorflow as tf
except ImportError:
print("WARNING: Tensorflow not found, skipping test")
exit(0)
import numpy as np
from dace.frontend.tensorflow import TFSession
if __name__ == '__main__':
print('DaCe Tensorflow frontend test')
A = np.random.rand(16, 16).astype(np.float32)
B = np.random.rand(16, 16).astype(np.float32)
A_tf = tf.placeholder(tf.float32, shape=[16, 16])
B_tf = tf.placeholder(tf.float32, shape=[16, 16])
with TFSession() as sess:
# Simple matrix multiplication
C = sess.run(A_tf @ B_tf, feed_dict={A_tf: A, B_tf: B})
diff = np.linalg.norm(C - (A @ B)) / (16 * 16)
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
scale = tf.placeholder(tf.float32, [num_channels])
offset = tf.placeholder(tf.float32, [num_channels])
populationMean = tf.placeholder(tf.float32, [num_channels])
populationVariance = tf.placeholder(tf.float32, [num_channels])
y, mean, var, _, var_sqrt = gen_nn_ops._fused_batch_norm(inp,
scale,
offset, [], [],
epsilon=0.1,
is_training=True)
outputs = [y, mean, var]
test_in = np.random.uniform(size=size).astype(np.float32)
test_scale = np.random.uniform(size=[num_channels]).astype(np.float32)
test_offset = np.random.uniform(size=[num_channels]).astype(np.float32)
sess_tf = tf.Session(config=config)
sess_dace = TFSession()
outputs_dace = sess_dace.run(
outputs,
feed_dict={
inp: test_in,
scale: test_scale,
offset: test_offset,
},
)
outputs_tf = sess_tf.run(
outputs,
feed_dict={
inp: test_in,
scale: test_scale,
offset: test_offset,
def test_fused_batch_norm():
import tensorflow as tf
from tensorflow.python.ops import gen_nn_ops
from dace.frontend.tensorflow import TFSession
num_channels = 3
size = [8, 224, 224, num_channels]
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
inp = tf.placeholder(tf.float32, size)
scale = tf.placeholder(tf.float32, [num_channels])
offset = tf.placeholder(tf.float32, [num_channels])
populationMean = tf.placeholder(tf.float32, [num_channels])
populationVariance = tf.placeholder(tf.float32, [num_channels])
y, mean, var, _, var_sqrt = gen_nn_ops._fused_batch_norm(inp,
scale,
offset, [], [],
epsilon=0.1,
is_training=True)
outputs = [y, mean, var]
test_in = np.random.uniform(size=size).astype(np.float32)
test_scale = np.random.uniform(size=[num_channels]).astype(np.float32)
test_offset = np.random.uniform(size=[num_channels]).astype(np.float32)
sess_tf = tf.Session(config=config)
sess_dace = TFSession()
outputs_dace = sess_dace.run(
outputs,
feed_dict={
inp: test_in,
scale: test_scale,
offset: test_offset,
},
)
outputs_tf = sess_tf.run(
outputs,
feed_dict={
inp: test_in,
scale: test_scale,
offset: test_offset,
},
)
try:
assert (tf.linalg.norm(outputs_tf[0] -
outputs_dace[0]).eval(session=sess_tf) < 1e-1
and tf.linalg.norm(outputs_dace[2] -
outputs_tf[2]).eval(session=sess_tf) < 1e-4
and tf.linalg.norm(outputs_dace[1] -
outputs_tf[1]).eval(session=sess_tf) < 1e-4)
except:
print("FBN test failed")
print(
tf.linalg.norm(outputs_tf[0] -
outputs_dace[0]).eval(session=sess_tf))
print(
tf.linalg.norm(outputs_tf[1] -
outputs_dace[1]).eval(session=sess_tf))
print(
tf.linalg.norm(outputs_tf[2] -
outputs_dace[2]).eval(session=sess_tf))
################# FBN GRADIENT TEST ###############################
outputGrad = tf.placeholder(tf.float32, size)
x_grad, gamma_grad, beta_grad, _, _ = gen_nn_ops.fused_batch_norm_grad(
outputGrad,
inp,
scale,
outputs[1],
var_sqrt,
epsilon=0.1,
is_training=True)
gradients = [x_grad, gamma_grad, beta_grad]
test_outputgrad = np.random.uniform(size=size).astype(np.float32)
outputs_dace = sess_dace.run(
gradients,
feed_dict={
inp: test_in,
outputGrad: test_outputgrad,
scale: test_scale,
offset: test_offset,
},
)
# TF
x_grad, gamma_grad, beta_grad, _, _ = gen_nn_ops.fused_batch_norm_grad(
outputGrad,
inp,
scale,
outputs[1],
tf.math.rsqrt(outputs[2] + float(0.1))
if tf.test.is_built_with_cuda() else outputs[2],
epsilon=0.1,
is_training=True,
)
gradients = [x_grad, gamma_grad, beta_grad]
# writer = tf.summary.FileWriter("./", sess_tf.graph)
outputs_tf = sess_tf.run(
gradients,
feed_dict={
inp: test_in,
outputGrad: test_outputgrad,
scale: test_scale,
offset: test_offset,
},
)
try:
assert (tf.linalg.norm(outputs_tf[0] -
outputs_dace[0]).eval(session=sess_tf) < 1e-1
and tf.linalg.norm(outputs_dace[2] -
outputs_tf[2]).eval(session=sess_tf) < 10
and tf.linalg.norm(outputs_dace[1] -
outputs_tf[1]).eval(session=sess_tf) < 10)
except:
print("FBN Gradient test failed")
print(
tf.linalg.norm(outputs_tf[0] -
outputs_dace[0]).eval(session=sess_tf))
print(
tf.linalg.norm(outputs_tf[1] -
outputs_dace[1]).eval(session=sess_tf))
print(
tf.linalg.norm(outputs_tf[2] -
outputs_dace[2]).eval(session=sess_tf))
print(
tf.linalg.norm(outputs_tf[2] -
np.sum(test_outputgrad, axis=(0, 1, 2))).eval(
session=sess_tf))
def test_conv():
import tensorflow as tf
from tensorflow.python.ops import gen_nn_ops
from dace.frontend.tensorflow import TFSession
inp_shape = [10, 10, 10, 10]
filter_shape = [3, 3, 10, 3]
strides = [1, 3, 3, 1]
inp = tf.placeholder(tf.float64, inp_shape)
filter = tf.placeholder(tf.float64, filter_shape)
outp = tf.nn.conv2d(inp,
filter,
strides,
padding="SAME",
data_format="NHWC")
test_in = np.random.uniform(size=tuple(inp_shape)).astype(np.float64)
test_filter = np.random.uniform(size=tuple(filter_shape)).astype(
np.float64)
sess_dace = TFSession()
sess_tf = tf.Session()
output_dace = sess_dace.run(outp,
feed_dict={
inp: test_in,
filter: test_filter
})
output_tf = sess_tf.run(outp,
feed_dict={
inp: test_in,
filter: test_filter
})
try:
assert tf.norm(output_dace - output_tf).eval(session=sess_tf) < 1e-10
except:
print(output_tf)
print(output_dace)
print(tf.linalg.norm(output_tf - output_dace).eval(session=sess_tf))
raise AssertionError("Convolution test failed")
##### Conv backprop grad ######
inp_shape = [10, 10, 10, 10]
filters = [[2, 2, 10, 3]]
strides = [[1, 3, 3, 1]]
paddings = ["VALID"]
for p in paddings:
for f in filters:
for s in strides:
print(p, f, s)
filter = tf.placeholder(tf.float64, f)
outp = tf.nn.conv2d(inp,
filter,
s,
padding=p,
data_format="NHWC")
out_backprop = tf.placeholder(tf.float64, outp.shape)
inp_gradients = gen_nn_ops.conv2d_backprop_input(inp_shape,
filter,
out_backprop,
s,
padding=p)
test_grads = np.random.uniform(size=outp.shape).astype(
np.float64)
test_filter = np.random.uniform(size=tuple(f)).astype(
np.float64)
output_tf = sess_tf.run(inp_gradients,
feed_dict={
filter: test_filter,
out_backprop: test_grads
})
output_dace = sess_dace.run(inp_gradients,
feed_dict={
filter: test_filter,
out_backprop: test_grads
})
try:
assert tf.norm(output_dace -
output_tf).eval(session=sess_tf) < 1e-10
except:
print(p)
print(f)
print(s)
print(output_tf)
print(output_dace)
print(
tf.linalg.norm(output_tf -
output_dace).eval(session=sess_tf))
raise AssertionError("Convolution grad test failed")
##### Conv filter backprop ##################
inp_shape = [10, 10, 10, 10]
filters = [[4, 4, 10, 3]]
strides = [[1, 1, 1, 1]]
paddings = ["SAME"]
for p in paddings:
for f in filters:
for s in strides:
input_placeholder = tf.placeholder(tf.float64, inp_shape)
filter = tf.placeholder(tf.float64, f)
outp = tf.nn.conv2d(inp,
filter,
s,
padding=p,
data_format="NHWC")
out_backprop = tf.placeholder(tf.float64, outp.shape)
filter_gradients = gen_nn_ops.conv2d_backprop_filter(
input_placeholder, f, out_backprop, s, padding=p)
test_grads = np.random.uniform(size=outp.shape).astype(
np.float64)
test_input = np.random.uniform(size=tuple(inp_shape)).astype(
np.float64)
output_tf = sess_tf.run(filter_gradients,
feed_dict={
input_placeholder: test_input,
out_backprop: test_grads
})
output_dace = sess_dace.run(filter_gradients,
feed_dict={
input_placeholder: test_input,
out_backprop: test_grads
})
try:
assert tf.norm(output_dace -
output_tf).eval(session=sess_tf) < 1e-10
except:
print(p)
print(f)
print(s)
print(output_tf)
print(output_dace)
print(
tf.linalg.norm(output_tf -
output_dace).eval(session=sess_tf))
raise AssertionError("Convolution filter grad test failed")
softmax = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_node, logits=logits)
loss = tf.reduce_mean(softmax, name="loss")
# Get gradient tensors and initializer operation
gradients = tf.gradients(loss, tf.trainable_variables())
init = tf.global_variables_initializer()
# Compute gradients and compare
# Tensorflow
with tf.Session() as sess:
sess.run(init)
tf_gradients = sess.run(gradients)
# DaCe
with TFSession(seed=SEED) as sess:
sess.run(init)
dace_gradients = sess.run(gradients)
# Compare
for tfgrad, dacegrad in zip(tf_gradients, dace_gradients):
inf_norm = np.linalg.norm((tfgrad - dacegrad).flatten())
print("Max. Diff:", inf_norm)
if (inf_norm <= 1e-4):
continue
else:
print("==== Program end====")
print("Error: norm too large")
exit(1)
print("==== Program end ====")
exit(0)
softmax = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
loss = tf.reduce_mean(softmax, name="loss")
gradients = tf.train.GradientDescentOptimizer(
learning_rate).compute_gradients(loss)
gradient_tensors = []
for tup in gradients:
gradient_tensors.append(tup[0])
update_op = tf.train.GradientDescentOptimizer(
learning_rate).apply_gradients(gradients)
return logits, update_op
# DaCe
sess = TFSession(seed=SEED)
y = build_resnet(input_placeholder, label_placeholder)
# TensorFlow + XLA
#sess = tf.Session()
#[y] = xla.compile(build_resnet, inputs=[input_placeholder, label_placeholder])
init = tf.global_variables_initializer()
sess.run(init)
images, labels = random_batch(batch_size)
# Warmup run
sess_run = sess.compile(y, gpu=True) # Change to gpu=True to run on the GPU
start = time.time()
