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

dx works but that makes no sense?

Browse Source
This commit is contained in:
Philippe Tillet
2019-07-03 20:21:32 -07:00
parent 1b2ceadf0d
commit bd1040510f
2 changed files with 19 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
examples/python/tensorflow/run.py
View File

@@ -58,12 +58,15 @@ def blocksparse_matmul_grad(op, dy):
def run_shift(): def run_shift():
B, C, H, W = 1, 16, 4, 4 B, C, H, W = 1, 16, 4, 4
R, S, F = 3, 3, 16 R, S, F = 3, 3, 16
np.random.seed(2)
a = tf.placeholder(tf.float32, shape=[C, H, W, B]) a = tf.placeholder(tf.float32, shape=[C, H, W, B])
b = tf.placeholder(tf.float32, shape=[C, F]) b = tf.placeholder(tf.float32, shape=[C, F])
#hshift_h = np.random.randint(-R//2, R//2 + 1, size=C, dtype=np.int32) 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_w = np.random.randint(- (S//2), R//2 + 1, size=C, dtype=np.int32)
hshift_h = np.ones(C, dtype=np.int32) print(hshift_h)
hshift_w = np.ones(C, dtype=np.int32) print(hshift_w)
#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)) c = module.shift_conv(a, b, shift_h=tf.make_tensor_proto(hshift_h), shift_w=tf.make_tensor_proto(hshift_w))
# Reference # Reference
ha = np.random.rand(C, H, W, B) ha = np.random.rand(C, H, W, B)
@@ -75,13 +78,16 @@ def run_shift():
extra_feed_dict={a: ha, b: hb}) extra_feed_dict={a: ha, b: hb})
dx_t, dx_n = grads[0] dx_t, dx_n = grads[0]
dw_t, dw_n = grads[1] dw_t, dw_n = grads[1]
print(dx_t) print(dw_t)
print(dx_n) print(dw_n)
#print(np.max(dw_t - dw_n)) print(np.max(dw_t - dw_n))
#np.savetxt('diff.dat', dw_t - dw_n, fmt='%2.4f')
#np.savetxt('theoretical.dat', dw_t, fmt='%2.4f')
#np.savetxt('numerical.dat', dw_n, fmt='%2.4f')
print(np.max(dx_t - dx_n)) print(np.max(dx_t - dx_n))
np.savetxt('diff.dat', dx_t - dx_n, fmt='%2.4f') #np.savetxt('diff.dat', dx_t - dx_n, fmt='%2.4f')
np.savetxt('theoretical.dat', dx_t, fmt='%2.4f') #np.savetxt('theoretical.dat', dx_t, fmt='%2.4f')
np.savetxt('numerical.dat', dx_n, fmt='%2.4f') #np.savetxt('numerical.dat', dx_n, fmt='%2.4f')
# Run # Run
sess.run(tf.global_variables_initializer()) sess.run(tf.global_variables_initializer())
result = sess.run([c], feed_dict = {a: ha, result = sess.run([c], feed_dict = {a: ha,

6
lib/dnn/shift.cpp
View File

@@ -287,15 +287,15 @@ else{
if(ty_ == BPROP){ if(ty_ == BPROP){
os << R"( os << R"(
int32 rcwhc[TM] = rxc / ABS; int32 rcwhc[TM] = rxc / ABS;
int32 rcw[TM] = (rcwhc % AW); int32 rcw[TM] = rcwhc % AW;
int32 rchc[TM] = rcwhc / AW; int32 rchc[TM] = rcwhc / AW;
int32 rch[TM] = (rchc % AH); int32 rch[TM] = rchc % AH;
int1 maskh[TM] = (rch >= pad_h) && (rch < (AH - pad_h)); int1 maskh[TM] = (rch >= pad_h) && (rch < (AH - pad_h));
int1 maskw[TM] = (rcw >= pad_w) && (rcw < (AW - pad_w)); int1 maskw[TM] = (rcw >= pad_w) && (rcw < (AW - pad_w));
int1 interior[TM, TN] = maskh[:, newaxis] && maskw[:, newaxis]; int1 interior[TM, TN] = maskh[:, newaxis] && maskw[:, newaxis];
__constant__ int32* pd[TN] = delta + ryc; __constant__ int32* pd[TN] = delta + ryc;
fp32* shift_pc[TM, TN] = pc + (*pd)[newaxis, :]; fp32* shift_pc[TM, TN] = pc + (*pd)[newaxis, :];
pc = interior ? shift_pc : pc; pc = interior ? pc : shift_pc;
@checkc __atomic_add(pc, C); @checkc __atomic_add(pc, C);
)"; )";
} }