gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[dnn/shift]: added stride to shift

Browse Source
This commit is contained in:
Philippe Tillet
2019-07-09 14:08:51 -07:00
parent cc41604784
commit 066ae338f1
4 changed files with 74 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
examples/python/tensorflow/run.py
View File

@@ -49,35 +49,35 @@ def run_conv():
def blocksparse_matmul_grad(op, dy):
shift_h = op.get_attr('shift_h')
shift_w = op.get_attr('shift_w')
stride_h = op.get_attr('stride_h')
stride_w = op.get_attr('stride_w')
x = op.inputs[0]
w = op.inputs[1]
dx = module.shift_conv_dx(dy, w, shift_h=shift_h, shift_w=shift_w)
dw = module.shift_conv_dw(dy, x, shift_h=shift_h, shift_w=shift_w)
dx = module.shift_conv_dx(dy, w, stride_h=stride_h, stride_w=stride_w, shift_h=shift_h, shift_w=shift_w)
dw = module.shift_conv_dw(dy, x, stride_h=stride_h, stride_w=stride_w, shift_h=shift_h, shift_w=shift_w)
return (dx, dw)
def run_shift():
B, C, H, W = 16, 1024, 8, 8
R, S, F = 3, 3, 1024
B, C, H, W = 16, 16, 4, 4
R, S, F = 3, 3, 4
stride_h, stride_w = 2, 2
np.random.seed(2)
a = tf.placeholder(tf.float32, shape=[C, H, W, B])
b = tf.placeholder(tf.float32, shape=[C, F])
hshift_h = np.random.randint(- (R//2), R//2 + 1, size=C, dtype=np.int32)
hshift_w = np.random.randint(- (S//2), R//2 + 1, size=C, dtype=np.int32)
#hshift_h = np.ones(C, dtype=np.int32)
#hshift_w = np.ones(C, dtype=np.int32)
c = module.shift_conv(a, b, shift_h=tf.make_tensor_proto(hshift_h), shift_w=tf.make_tensor_proto(hshift_w))
# Reference
c = module.shift_conv(a, b, stride_h=stride_h, stride_w=stride_w, shift_h=tf.make_tensor_proto(hshift_h), shift_w=tf.make_tensor_proto(hshift_w))
# feed values
ha = np.random.rand(C, H, W, B)
hb = np.random.rand(C, F)
#ha = np.ones((C, H, W, B), dtype=np.int32)
#hb = np.ones((C, F), dtype=np.int32)
sess = tf.InteractiveSession()
#grads = tf.test.compute_gradient([a, b], [(C, H, W, B), (C, F)], c, (F, H, W, B),
# extra_feed_dict = {a: ha, b: hb})
#dw_t, dw_n = grads[1]
#dx_t, dx_n = grads[0]
#print(np.max(np.abs(dw_t - dw_n)))
#print(np.max(np.abs(dx_t - dx_n)))
# test
grads = tf.test.compute_gradient([a, b], [(C, H, W, B), (C, F)], c, (F, H//stride_h, W//stride_w, B),
extra_feed_dict = {a: ha, b: hb})
dw_t, dw_n = grads[1]
dx_t, dx_n = grads[0]
print(np.max(np.abs(dw_t - dw_n)))
print(np.max(np.abs(dx_t - dx_n)))
# Run
sess.run(tf.global_variables_initializer())
result = sess.run([c], feed_dict = {a: ha,
@@ -127,4 +127,4 @@ def run_batchnorm():
print(np.max(np.abs(dg_t - dg_n)))
print(np.max(np.abs(db_t - db_n)))
run_batchnorm()
run_shift()