from deep500.frameworks import reference as d5ref
from deep500 import test_op_forward, test_op_gradient
if __name__ == '__main__':
with tf.Session() as sess:
# Create two matrices A and B
A = np.random.rand(5, 40).astype(np.float64)
B = np.random.rand(40, 5).astype(np.float64)
# Native (TensorFlow) versions of A and B
native_A = tf.Variable(A, expected_shape=A.shape, dtype=A.dtype)
native_B = tf.Variable(B, expected_shape=B.shape, dtype=B.dtype)
# Create a Deep500 Operator from the tf.matmul function.
# The native TF operator accepts two inputs and one output, whose
# tensor descriptors (i.e., shape and data type) are created below.
op = d5tf.custom_op_from_native(
tf.matmul, [d5ref.desc_from_tensor(A),
d5ref.desc_from_tensor(B)],
[d5ref.desc_from_tensor(A @ B)])
# Test forward (first using Deep500, which copies to/from numpy,
# then natively through the framework)
test_op_forward(op, [A, B], [A @ B])
d5tf.test_nativeop_forward(tf.matmul, [native_A, native_B], [A @ B])
# Test backward (gradient) using Deep500 and native tests.
test_op_gradient(op, [A, B], eps=1e-7)
d5tf.test_nativeop_gradient(tf.matmul, [native_A, native_B], eps=1e-7)
kernel_size=[test.r, test.s],
strides=[test.hstride, test.wstride],
data_format='channels_first',
kernel_initializer=tf.constant_initializer(W))
# Create a conv2d function that only receives one variable
cnv2d = lambda x: tf.layers.conv2d(
x,
filters=test.k,
kernel_size=[test.r, test.s],
strides=[test.hstride, test.wstride],
data_format='channels_first',
kernel_initializer=tf.constant_initializer(W))
times, = \
d5tf.test_nativeop_forward(cnv2d, [var_X], [None],
metrics=[d5.WallclockTime(RUNS*AVG_OVER, AVG_OVER)])
except Exception as ex:
print('Exception:', ex)
times = [-1.0]
# Clear memory
tf.reset_default_graph()
with open('0_d5tf_conv_deepbench.log', 'a') as fp:
fp.writelines([
'{test.n},{test.c},{test.h},{test.w},'
'{test.k},{test.r},{test.s},'
'{test.hstride},{test.wstride},'
'{time:.15f}\n'.format(test=test, time=time) for time in times
])
if test.b_trans:
B = np.random.rand(test.n, test.k).astype(np.float32)
var_B = tf.Variable(B, name="var_B", dtype=tf.float32)
Bmult = B.transpose()
else:
B = np.random.rand(test.k, test.n).astype(np.float32)
var_B = tf.Variable(B, name="var_B", dtype=tf.float32)
Bmult = B
# Create a matmul function that only receives two variables
mm = lambda a,b: tf.matmul(a,b, test.a_trans, test.b_trans)
# Run Deep500 test
l2err, maxerr, times = \
d5tf.test_nativeop_forward(mm, [var_A, var_B], [Amult @ Bmult],
metrics=[d5.L2Error(), d5.MaxError(),
d5.WallclockTime(RUNS*AVG_OVER, AVG_OVER)])
except:
l2err = -1.0
maxerr = -1.0
times = [-1.0]
# Clear memory
tf.reset_default_graph()
with open('0_d5tf_gemm_deepbench.log', 'a') as fp:
fp.writelines(['{test.m},{test.n},{test.k},'
'{test.a_trans},{test.b_trans},'
'{time:.15f},{l2err},{maxerr}\n'.format(test=test,
time=time, l2err=l2err, maxerr=maxerr)
for time in times])