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

[codegen/shift] substantial cleaning of triton-c shift-conv code

Browse Source
This commit is contained in:
Philippe Tillet
2019-07-11 16:38:58 -07:00
parent 75cf2df110
commit 207e021973
4 changed files with 191 additions and 154 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
examples/python/pytorch/run.py
View File

@@ -68,11 +68,11 @@ def ShiftConv2d(in_planes, out_planes, kernel_size=3, stride=1, groups=1, dilati
class NetReference(nn.Module):
def __init__(self):
super(NetReference, self).__init__()
self.conv1 = ShiftConv2d(1, 32, 3, 2)
#self.conv1 = triton.ShiftConv2d(1, 32, 3, 2)
#self.conv1 = ShiftConv2d(1, 32, 3, 2)
self.conv1 = triton.ShiftConv2d(1, 32, 3, 2)
self.bn1 = nn.BatchNorm2d(32)
#self.conv2 = triton.ShiftConv2d(32, 32, 3, 2)
self.conv2 = ShiftConv2d(32, 32, 3, 2)
self.conv2 = triton.ShiftConv2d(32, 32, 3, 2)
#self.conv2 = ShiftConv2d(32, 32, 3, 2)
self.bn2 = nn.BatchNorm2d(32)
self.fc1 = nn.Linear(32*7*7, 500)
self.fc2 = nn.Linear(500, 10)

12
examples/python/tensorflow/run.py
View File

@@ -59,15 +59,15 @@ def blocksparse_matmul_grad(op, dy):
def run_shift():
B, C, H, W = 16, 16, 4, 4
R, S, F = 3, 3, 16
R, S, F = 3, 3, 32
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.zeros(C, dtype=np.int32)
hshift_w = np.zeros(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_h = np.zeros(C, dtype=np.int32)
#hshift_w = np.zeros(C, dtype=np.int32)
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)
@@ -122,8 +122,6 @@ def run_batchnorm():
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
result = sess.run([y, m, v], feed_dict = {x: hx, g: hg, b: hb})
print(result[1])
print(np.mean(hx, (1, 2, 3)))
grads = tf.test.compute_gradient([x, g, b], [(C, H, W, B), (C, ), (C, )], y, (C, H, W, B),
extra_feed_dict = {x: hx, g: hg, b: hb})
dx_t, dx_n = grads[0]