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

weight gradient seem to work

Browse Source
This commit is contained in:
Philippe Tillet
2019-07-03 20:04:38 -07:00
parent 39aa22babb
commit 1b2ceadf0d
2 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
examples/python/tensorflow/run.py
View File

@@ -75,13 +75,13 @@ def run_shift():
extra_feed_dict={a: ha, b: hb})
dx_t, dx_n = grads[0]
dw_t, dw_n = grads[1]
print(dw_t)
print(dw_n)
print(np.max(dw_t - dw_n))
#print(np.max(dx_t - dx_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(dx_t)
print(dx_n)
#print(np.max(dw_t - dw_n))
print(np.max(dx_t - dx_n))
np.savetxt('diff.dat', dx_t - dx_n, fmt='%2.4f')
np.savetxt('theoretical.dat', dx_t, fmt='%2.4f')
np.savetxt('numerical.dat', dx_n, fmt='%2.4f')
# Run
sess.run(tf.global_variables_initializer())
result = sess.run([c], feed_dict = {a: ha,

2
lib/dnn/shift.cpp
View File

@@ -228,7 +228,7 @@ if(ty_ == WGRAD){
int1 maskh[TK] = (rbh >= pad_h) && (rbh < (AH - pad_h));
int1 maskw[TK] = (rbw >= pad_w) && (rbw < (AW - pad_w));
int1 mask[TK, TN] = maskh[:, newaxis] && maskw[:, newaxis];
int32 inc[TK, TN] = mask ? 0 : shift;
int32 inc[TK, TN] = mask ? shift : 0;
)" << b_ty_ << R"(* shifted_pb[TK, TN] = pb + inc;
)" << b_ty_ << R"( b[TK, TN] = *shifted_pb;)";
}