def verifyValues(filter_in_sizes,
rho_filter=1,
dim=5,
scale_val=0.1,
bias_val=0.1,
num_trials=3):
out_channel_count = filter_in_sizes[-1]
if isinstance(bias_val, collections.Iterable):
bias = np.array(bias_val, dtype=np.float32)
else:
bias = np.array([bias_val] * out_channel_count, dtype=np.float32)
scale = np.array(scale_val, dtype=np.float32)
[t2ind, t2val, t2sh] = sp.createRandomSparseTensor(rho_filter,
filter_in_sizes)
s2 = tf.SparseTensor(indices=t2ind, values=t2val, dense_shape=t2sh)
d2 = sp.sparse_to_dense(t2ind, t2val, t2sh)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
with tf.device("/gpu:0"):
convf = sc_module.direct_sparse_filter_conversion(
t2ind, t2val, t2sh, t2sh)
with tf.Session(config=config) as sess:
pf = sess.run(convf)
tf.reset_default_graph()
ts = 0
with tf.device("/gpu:0"):
creg = sc_module.direct_sparse_channelwise_biased_l2_regularization(
pf.out_indices, pf.out_values, pf.out_shape,
pf.out_channel_mapping, scale, bias, dim)
with tf.Session(config=config) as sess:
t6 = time.time()
sv3 = sess.run(creg)
t5 = time.time()
for i in range(0, num_trials):
sess.run(creg)
t6 = time.time()
ts = abs(t6 - t5) / max(num_trials, 1)
print("time approx sparse: ", ts)
tf.reset_default_graph()
time.sleep(1)
reg_loss = 0
for out_channel in range(out_channel_count):
reg_loss += np.sum(
np.power(d2[:, :, :, :, out_channel] + bias[out_channel],
2)) * scale / 2.
print(sv3, reg_loss)
if abs(sv3 - reg_loss) > 0.001:
print("error")
return 1
#bp_sfg = sp.sparse1d_to_dense(pf.out_indices, res_bp3.filter_grads, pf.out_shape, pf.out_channel_mapping[-1])
return 0
def verifyValues(tensor_in_sizes,
filter_in_sizes,
stride,
rho_data=0.1,
rho_filter=1,
padding='SAME',
dim=5,
max_density=0.1,
num_trials=3,
filter_type="K-RELU",
test_type=""):
if isinstance(stride, collections.Iterable):
strides = [1] + list(stride) + [1]
else:
strides = [1, stride, stride, stride, 1]
out_sizes = np.copy(tensor_in_sizes)
out_sizes[-1] = filter_in_sizes[-1]
out_entry_count = np.prod(out_sizes) * max_density
bias = np.zeros([filter_in_sizes[-1]], dtype=np.float32)
no_strides = [1, 1, 1, 1, 1]
[t1ind, t1val, t1sh] = sp.createRandomSparseTensor(rho_data,
tensor_in_sizes, -3, 3)
s1 = tf.SparseTensor(indices=t1ind, values=t1val, dense_shape=t1sh)
d1 = sp.sparse_to_dense(t1ind, t1val, t1sh)
[t2ind, t2val, t2sh] = sp.createRandomSparseTensor(rho_filter,
filter_in_sizes)
s2 = tf.SparseTensor(indices=t2ind, values=t2val, dense_shape=t2sh)
d2 = sp.sparse_to_dense(t2ind, t2val, t2sh)
print("strides: \n", strides)
print("input shape", tensor_in_sizes)
print("filter shape", filter_in_sizes)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
with tf.device("/gpu:0"):
convd = sc_module.direct_sparse_data_conversion(t1ind, t1val, t1sh)
convf = sc_module.direct_sparse_filter_conversion(
t2ind, t2val, t2sh, t1sh)
with tf.Session(config=config) as sess:
pd = sess.run(convd)
pf = sess.run(convf)
tf.reset_default_graph()
ts = 0
with tf.device("/gpu:0"):
approx_scskconv = sc_module.direct_sparse_conv_kd(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, pf.out_indices, pf.out_values,
pf.out_shape, pf.out_channel_mapping, bias, strides, padding,
out_entry_count, dim, max_density, filter_type)
with tf.Session(config=config) as sess:
t6 = time.time()
sv3 = sess.run(approx_scskconv)
t5 = time.time()
for i in range(0, num_trials):
sess.run(approx_scskconv)
t6 = time.time()
ts = abs(t6 - t5) / max(num_trials, 1)
print("time approx sparse: ", ts)
tf.reset_default_graph()
time.sleep(1)
td = 0
with tf.device("/gpu:0"):
conv = nn_ops.conv3d(d1, d2, strides, padding)
with tf.Session(config=config) as sess:
t22 = time.time()
expected = sess.run(conv)
t11 = time.time()
for i in range(0, num_trials):
sess.run(conv)
t22 = time.time()
td = abs(t22 - t11) / max(num_trials, 1)
print("time dense gpu: ", td)
tf.reset_default_graph()
print("time ratio: ", ts / td)
return
[bp_ind, sv3_bp_val,
bp_sh] = sp.createRandomSparseTensor(1, [len(sv3.out_values)], 1, 9)
d3_ = sp.sparse1d_to_dense(sv3.out_indices, sv3_bp_val, sv3.out_shape,
sv3.out_block_channel_mapping[-1])
out_backprop_val = constant_op.constant(d3_)
t_bp1 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = nn_ops.conv3d_backprop_filter_v2(d1, filter_in_sizes,
out_backprop_val, strides,
padding)
res_bp1 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp1 = t_bp1 + t2 - t1
t_bp1 = t_bp1 / float(num_trials)
print("time bp1: ", t_bp1)
t_bp2 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = nn_ops.conv3d_backprop_input_v2(tensor_in_sizes, d2,
out_backprop_val, strides,
padding)
res_bp2 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp2 = t_bp2 + t2 - t1
t_bp2 = t_bp2 / float(num_trials)
print("time bp2: ", t_bp2)
t_bp3 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = sc_module.direct_sparse_conv_kd_backprop(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, pf.out_indices, pf.out_values,
pf.out_shape, pf.out_channel_mapping, sv3.out_indices,
sv3.out_values, sv3.out_shape, sv3.out_block_channel_mapping,
sv3_bp_val, strides, padding, dim, max_density)
res_bp3 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp3 = t_bp3 + t2 - t1
t_bp3 = t_bp3 / float(num_trials)
print("time bp3: ", t_bp3)
print("sparse ratio: ", t_bp3 / (t_bp2 + t_bp1))
bp_sfg = sp.sparse1d_to_dense(pf.out_indices, res_bp3.filter_grads,
pf.out_shape, pf.out_channel_mapping[-1])
bp_sig = sp.sparse1d_to_dense(pd.out_indices, res_bp3.input_grads,
pd.out_shape,
pd.out_block_channel_mapping[-1])
value3 = sp.sparse1d_to_dense(sv3.out_indices, sv3.out_values,
sv3.out_shape,
sv3.out_block_channel_mapping[-1])
print("expected", expected)
print("sv3", value3)
print("out densities", sv3.out_channel_densities)
has_error = False
approx_cmp = expected.flatten()
approx = value3.flatten()
non_zero_count = 0
for i in range(len(approx_cmp)):
non_zero_count = non_zero_count + 1
print("entry count: ", non_zero_count)
error_cnt = 0
first_error = 0
correct_cnt = 0
for i in range(len(approx_cmp)):
if approx_cmp[i] > 0 and abs(approx_cmp[i] - approx[i]) > 1e-3:
if has_error == False:
first_error = i
has_error = True
error_cnt = error_cnt + 1
elif approx[i] != 0:
correct_cnt = correct_cnt + 1
bp_sig_flat = bp_sig.flatten()
res_bp2_flat = res_bp2.flatten()
bp_i_error_cnt = 0
bp_i_correct_cnt = 0
for i in range(len(bp_sig_flat)):
if bp_sig_flat[i] != 0:
if bp_sig_flat[i] == res_bp2_flat[i]:
bp_i_correct_cnt = bp_i_correct_cnt + 1
else:
bp_i_error_cnt = bp_i_error_cnt + 1
filter_flat = d2.flatten()
bp_sfg_flat = bp_sfg.flatten()
res_bp1_flat = res_bp1.flatten()
bp_f_error_cnt = 0
bp_f_correct_cnt = 0
for i in range(len(filter_flat)):
if filter_flat[i] != 0:
if bp_sfg_flat[i] == res_bp1_flat[i]:
bp_f_correct_cnt = bp_f_correct_cnt + 1
else:
bp_f_error_cnt = bp_f_error_cnt + 1
print("total number of non-zero corrects: ", correct_cnt)
print("sparse input size: ", len(t1ind))
print("total number of bpi corrects: ", bp_i_correct_cnt)
print("sparse filter size: ", len(t2ind))
print("total number of bpf corrects: ", bp_f_correct_cnt)
if has_error:
print("total number of errors: ", error_cnt)
print("first error: ", first_error)
return 1
if bp_i_error_cnt > 0:
print("total number of bpi errors: ", bp_i_error_cnt)
if bp_f_error_cnt > 0:
print("total number of bpf errors: ", bp_f_error_cnt)
print("OK")
return 0
def verifyValues(tensor_in_sizes,
filter_in_sizes,
stride,
rho_data=0.1,
rho_filter=1,
padding='SAME',
dim=5,
max_density=0.1,
num_trials=3,
filter_type='K-RELU',
test_type='',
dense=True):
if isinstance(stride, collections.Iterable):
strides = [1] + list(stride) + [1]
else:
strides = [1, stride, stride, stride, 1]
out_sizes = np.copy(tensor_in_sizes)
out_sizes[-1] = filter_in_sizes[-1]
out_entry_count = np.prod(out_sizes) * max_density
bias = np.zeros([filter_in_sizes[-1]], dtype=np.float32)
no_strides = [1, 1, 1, 1, 1]
[t1ind, t1val, t1sh] = sp.createRandomSparseTensor(rho_data,
tensor_in_sizes, -3, 3)
s1 = tf.SparseTensor(indices=t1ind, values=t1val, dense_shape=t1sh)
d1 = sp.sparse_to_dense(t1ind, t1val, t1sh)
[t2ind, t2val, t2sh] = sp.createRandomSparseTensor(rho_filter,
filter_in_sizes)
s2 = tf.SparseTensor(indices=t2ind, values=t2val, dense_shape=t2sh)
d2 = sp.sparse_to_dense(t2ind, t2val, t2sh)
print("strides: \n", strides)
print("input shape", tensor_in_sizes)
print("filter shape", filter_in_sizes)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
with tf.device("/gpu:0"):
convd = sc_module.direct_sparse_data_conversion(t1ind, t1val, t1sh)
convf = sc_module.direct_sparse_filter_conversion(
t2ind, t2val, t2sh, t1sh)
with tf.Session(config=config) as sess:
pd = sess.run(convd)
pf = sess.run(convf)
tf.reset_default_graph()
ts = 0
with tf.device("/gpu:0"):
approx_scskconv = sc_module.direct_sparse_conv_kd(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, pf.out_indices, pf.out_values,
pf.out_shape, pf.out_channel_mapping, bias, strides, padding,
out_entry_count, dim, max_density, filter_type)
with tf.Session(config=config) as sess:
t6 = time.time()
sv3 = sess.run(approx_scskconv)
t5 = time.time()
for i in range(0, num_trials):
sess.run(approx_scskconv)
t6 = time.time()
ts = abs(t6 - t5) / max(num_trials, 1)
print("time approx sparse: ", ts)
tf.reset_default_graph()
time.sleep(1)
if dense:
td = 0
with tf.device("/gpu:0"):
conv = nn_ops.conv3d(d1, d2, strides, padding)
with tf.Session(config=config) as sess:
t22 = time.time()
expected = sess.run(conv)
t11 = time.time()
for i in range(0, num_trials):
sess.run(conv)
t22 = time.time()
td = abs(t22 - t11) / max(num_trials, 1)
print("time dense gpu: ", td)
tf.reset_default_graph()
print("time ratio: ", ts / td)
return [expected, sv3, ts, td]
def verifyValues(tensor_in_sizes,
rho_data=0.1,
dim=5,
num_trials=3,
test_type=""):
[t1ind, t1val, t1sh] = sp.createRandomSparseTensor(rho_data,
tensor_in_sizes)
s1 = tf.SparseTensor(indices=t1ind, values=t1val, dense_shape=t1sh)
d1 = sp.sparse_to_dense(t1ind, t1val, t1sh)
#print("ind in: \n", t1ind)
#print("input: \n", d1)
# Initializes the input tensor with array containing incrementing
# numbers from 1.
print("input shape", tensor_in_sizes)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
#reorder data and generate block index lookup table
td = 0
with tf.device("/gpu:0"):
dts = sc_module.direct_dense_to_sparse(d1, tensor_in_sizes, dim)
with tf.Session(config=config) as sess:
t22 = time.time()
pd = sess.run(dts)
t11 = time.time()
for i in range(0, num_trials):
sess.run(dts)
t22 = time.time()
td = abs(t22 - t11) / max(num_trials, 1)
print("time dense to sparse gpu: ", td)
tf.reset_default_graph()
expected = d1
td = 0
with tf.device("/gpu:0"):
s2d = sc_module.direct_sparse_to_dense(pd.out_indices, pd.out_values,
pd.out_shape,
pd.out_block_channel_mapping)
with tf.Session(config=config) as sess:
t22 = time.time()
sv3 = sess.run(s2d)
t11 = time.time()
for i in range(0, num_trials):
sess.run(s2d)
t22 = time.time()
td = abs(t22 - t11) / max(num_trials, 1)
print("time sparse to dense gpu: ", td)
tf.reset_default_graph()
[bp_ind, sv3_bp_val,
bp_sh] = sp.createRandomSparseTensor(1, tensor_in_sizes, 1, 9)
d3_ = sp.sparse1d_to_dense(pd.out_indices, sv3_bp_val, pd.out_shape,
pd.out_block_channel_mapping[-1])
out_backprop_val = constant_op.constant(d3_)
t_bp3 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = sc_module.direct_sparse_to_dense_backprop(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, sv3, out_backprop_val)
res_bp3 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp3 = t_bp3 + t2 - t1
t_bp3 = t_bp3 / float(num_trials)
print("time bp sparse to dense: ", t_bp3)
t_bp4 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = sc_module.direct_dense_to_sparse_backprop(
sv3, pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, res_bp3)
res_bp4 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp4 = t_bp3 + t2 - t1
t_bp4 = t_bp4 / float(num_trials)
print("time bp dense to sparse: ", t_bp4)
bp_sig = sp.sparse1d_to_dense(pd.out_indices, res_bp3, pd.out_shape,
pd.out_block_channel_mapping[-1])
#print("dense bp ", res_bp1)
#print("sparse bp: ", bp_sig)
has_error = False
approx_cmp = expected.flatten()
approx = sv3.flatten()
non_zero_count = 0
for i in range(len(approx_cmp)):
non_zero_count = non_zero_count + 1
print("entry count: ", non_zero_count)
error_cnt = 0
first_error = 0
correct_cnt = 0
for i in range(len(approx_cmp)):
if abs(approx_cmp[i] - approx[i]) > 1e-3:
if has_error == False:
first_error = i
has_error = True
error_cnt = error_cnt + 1
elif approx[i] != 0:
correct_cnt = correct_cnt + 1
ebp = d3_.flatten()
#rbp = bp_sig.flatten()
rbp = res_bp4.flatten()
bperror_cnt = 0
bpcorrect_cnt = 0
for i in range(len(ebp)):
if abs(ebp[i] - rbp[i]) > 1e-3:
bperror_cnt = bperror_cnt + 1
elif rbp[i] != 0:
bpcorrect_cnt = bpcorrect_cnt + 1
print("total number of non-zero corrects: ", correct_cnt)
print("total number of backprop corrects: ", bpcorrect_cnt)
if has_error:
print("total number of errors: ", error_cnt)
print("first error: ", first_error)
return 1
if bperror_cnt > 0:
print("total number of backprop errors: ", bperror_cnt)
print("OK")
return 0
def verifyValues(tensor_in_sizes,
filter_in_sizes,
stride,
rho_data=0.1,
rho_filter=1,
padding='SAME',
dim=5,
max_density=0.1,
num_trials=3,
filter_type="K-RELU",
test_type=""):
if isinstance(stride, collections.Iterable):
strides = [1] + list(stride) + [1]
else:
strides = [1, stride, stride, stride, 1]
bias = np.zeros([filter_in_sizes[-1]], dtype=np.float32)
no_strides = [1, 1, 1, 1, 1]
[t1ind, t1val, t1sh] = sp.createRandomSparseTensor(rho_data,
tensor_in_sizes, -3, 3)
s1 = tf.SparseTensor(indices=t1ind, values=t1val, dense_shape=t1sh)
d1 = sp.sparse_to_dense(t1ind, t1val, t1sh)
[t2ind, t2val, t2sh] = sp.createRandomSparseTensor(rho_filter,
filter_in_sizes, -3, 3)
s2 = tf.SparseTensor(indices=t2ind, values=t2val, dense_shape=t2sh)
d2 = sp.sparse_to_dense(t2ind, t2val, t2sh)
filter_in_sizes2 = filter_in_sizes[:]
filter_in_sizes2[-2] = filter_in_sizes2[-1]
[t3ind, t3val, t3sh] = sp.createRandomSparseTensor(rho_filter,
filter_in_sizes2, -3, 3)
s3 = tf.SparseTensor(indices=t3ind, values=t3val, dense_shape=t3sh)
d3 = sp.sparse_to_dense(t3ind, t3val, t3sh)
[t4ind, t4val, t4sh] = sp.createRandomSparseTensor(rho_filter,
filter_in_sizes2, -3, 3)
s4 = tf.SparseTensor(indices=t4ind, values=t4val, dense_shape=t4sh)
d4 = sp.sparse_to_dense(t4ind, t4val, t4sh)
print("strides: \n", strides)
print("input shape", tensor_in_sizes)
print("filter shape", filter_in_sizes)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.device("/gpu:0"):
convd = sc_module.direct_sparse_data_conversion(t1ind, t1val, t1sh)
convf = sc_module.direct_sparse_filter_conversion(
t2ind, t2val, t2sh, t1sh)
convf2 = sc_module.direct_sparse_filter_conversion(
t3ind, t3val, t3sh, t3sh)
convf3 = sc_module.direct_sparse_filter_conversion(
t4ind, t4val, t4sh, t4sh)
with tf.Session(config=config) as sess:
pd = sess.run(convd)
pf = sess.run(convf)
pf2 = sess.run(convf2)
pf3 = sess.run(convf3)
tf.reset_default_graph()
ts = 0
with tf.device("/gpu:0"):
net = sc_module.direct_sparse_conv_kd(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, pf.out_indices, pf.out_values,
pf.out_shape, pf.out_channel_mapping, bias, strides, padding, dim,
max_density, filter_type)
net = sc_module.direct_sparse_conv_kd(
net.out_indices, net.out_values, net.out_shape,
net.out_block_channel_mapping, pf2.out_indices, pf2.out_values,
pf2.out_shape, pf2.out_channel_mapping, bias, strides, padding,
dim, max_density, filter_type)
net = sc_module.direct_sparse_conv_kd(
net.out_indices, net.out_values, net.out_shape,
net.out_block_channel_mapping, pf3.out_indices, pf3.out_values,
pf3.out_shape, pf3.out_channel_mapping, bias, strides, padding,
dim, max_density, filter_type)
with tf.Session(config=config) as sess:
t6 = time.time()
sv3 = sess.run(net)
t5 = time.time()
for i in range(0, num_trials):
sess.run(net)
t6 = time.time()
ts = abs(t6 - t5) / max(num_trials, 1)
print("time approx sparse: ", ts)
tf.reset_default_graph()
td = 0
with tf.device("/gpu:0"):
net = nn_ops.conv3d(d1, d2, strides, padding)
if filter_type == "K-RELU":
net = nn_ops.relu(net)
net = nn_ops.conv3d(net, d3, strides, padding)
if filter_type == "K-RELU":
net = nn_ops.relu(net)
net = nn_ops.conv3d(net, d4, strides, padding)
if filter_type == "K-RELU":
net = nn_ops.relu(net)
with tf.Session(config=config) as sess:
t22 = time.time()
expected = sess.run(net)
t11 = time.time()
for i in range(0, num_trials):
sess.run(net)
t22 = time.time()
td = abs(t22 - t11) / max(num_trials, 1)
print("time dense gpu: ", td)
tf.reset_default_graph()
value3 = sp.sparse1d_to_dense(sv3.out_indices, sv3.out_values,
sv3.out_shape,
sv3.out_block_channel_mapping[-1])
#print("expected: ", expected)
#print("sparse: ", value3, sv3)
has_error = False
approx_cmp = expected.flatten()
approx = value3.flatten()
non_zero_count = 0
for i in range(len(approx_cmp)):
non_zero_count = non_zero_count + 1
print("entry count: ", non_zero_count)
error_cnt = 0
first_error = 0
correct_cnt = 0
for i in range(len(approx_cmp)):
if abs(approx_cmp[i] - approx[i]) > 1e-3:
if has_error == False:
first_error = i
has_error = True
error_cnt = error_cnt + 1
elif approx[i] != 0:
correct_cnt = correct_cnt + 1
print("total number of non-zero corrects: ", correct_cnt)
print("sparse input size: ", len(t1ind))
if has_error:
print("total number of errors: ", error_cnt)
print("first error: ", first_error)
return 1
print("OK")
return 0
def verifyValues(tensor_in_sizes,
stride,
rho_data=0.1,
padding='SAME',
dim=5,
max_density=1,
num_trials=3,
test_type=""):
if isinstance(stride, collections.Iterable):
strides = [1] + list(stride) + [1]
else:
strides = [1, stride, stride, stride, 1]
no_strides = [1, 1, 1, 1, 1]
[t1ind, t1val, t1sh] = sp.createRandomSparseTensor(rho_data,
tensor_in_sizes)
d1 = sp.sparse_to_dense(t1ind, t1val, t1sh)
print("strides: \n", strides)
print("input shape", tensor_in_sizes)
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.7
#reorder data and generate block index lookup table
with tf.device("/gpu:0"):
convd = sc_module.direct_sparse_data_conversion(t1ind, t1val, t1sh)
with tf.Session(config=config) as sess:
pd = sess.run(convd)
tf.reset_default_graph()
ts = 0
with tf.device("/gpu:0"):
approx_scskconv = sc_module.direct_sparse_max_pooling_kd(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, strides, dim)
with tf.Session(config=config) as sess:
t6 = time.time()
sv3 = sess.run(approx_scskconv)
t5 = time.time()
for i in range(0, num_trials):
sess.run(approx_scskconv)
t6 = time.time()
ts = abs(t6 - t5) / max(num_trials, 1)
print("time approx sparse: ", ts)
tf.reset_default_graph()
td = 0
with tf.device("/gpu:0"):
pooling = tf.nn.max_pool3d(d1, strides, strides, "SAME")
with tf.Session(config=config) as sess:
t22 = time.time()
expected = sess.run(pooling)
t11 = time.time()
for i in range(0, num_trials):
sess.run(pooling)
t22 = time.time()
td = abs(t22 - t11) / max(num_trials, 1)
print("time dense gpu: ", td)
tf.reset_default_graph()
print("time ratio: ", ts / td)
[bp_ind, sv3_bp_val,
bp_sh] = sp.createRandomSparseTensor(1, [len(sv3.out_values)], 1, 9)
d3_ = sp.sparse1d_to_dense(sv3.out_indices, sv3_bp_val, sv3.out_shape,
sv3.out_block_channel_mapping[-1])
out_backprop_val = constant_op.constant(d3_)
t_bp1 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = gen_nn_ops._max_pool3d_grad(d1, expected, out_backprop_val,
strides, strides, "SAME")
#fbp = nn_ops.conv3d_backprop_filter_v2(d1, filter_in_sizes, out_backprop_val, strides, padding)
res_bp1 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp1 = t_bp1 + t2 - t1
t_bp1 = t_bp1 / float(num_trials)
print("time bp1: ", t_bp1)
t_bp3 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = sc_module.direct_sparse_max_pooling_kd_backprop(
pd.out_indices, pd.out_values, pd.out_shape,
pd.out_block_channel_mapping, sv3.out_indices, sv3.out_values,
sv3.out_shape, sv3.out_block_channel_mapping, sv3_bp_val,
strides, dim)
res_bp3 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp3 = t_bp3 + t2 - t1
t_bp3 = t_bp3 / float(num_trials)
print("time bp3: ", t_bp3)
print("bp ratio: ", t_bp1 / t_bp3)
return 0
bp_sig = sp.sparse1d_to_dense(pd.out_indices, res_bp3, pd.out_shape,
pd.out_block_channel_mapping[-1])
#print("dense bp ", res_bp1)
#print("sparse bp: ", bp_sig)
tsu = 0
with tf.device("/gpu:0"):
unpooling = sc_module.direct_sparse_unpooling_kd(
sv3.out_indices, sv3.out_values, sv3.out_shape,
sv3.out_block_channel_mapping, pd.out_indices, pd.out_shape,
pd.out_block_channel_mapping, strides, dim)
with tf.Session(config=config) as sess:
t6 = time.time()
sv4 = sess.run(unpooling)
t5 = time.time()
for i in range(0, num_trials):
sess.run(unpooling)
t6 = time.time()
tsu = abs(t6 - t5) / max(num_trials, 1)
print("time sparse unpooling: ", tsu)
tf.reset_default_graph()
'''print("sparse bp", bp_sig)
print("sv3 obcm", sv3.out_block_channel_mapping)
print("len", len(sv3.out_indices))
print("pd obcm", pd.out_block_channel_mapping)
print("len", len(pd.out_indices))
'''
t_bp4 = 0
with tf.Session(config=config) as sess:
with tf.device("/gpu:0"):
fbp = sc_module.direct_sparse_unpooling_kd_backprop(
sv3.out_indices, sv3.out_values, sv3.out_shape,
sv3.out_block_channel_mapping, pd.out_indices, pd.out_values,
pd.out_shape, pd.out_block_channel_mapping, res_bp3, strides,
dim)
res_bp4 = sess.run(fbp)
for i in range(num_trials):
t1 = time.time()
sess.run(fbp)
t2 = time.time()
t_bp4 = t_bp4 + t2 - t1
t_bp4 = t_bp4 / float(num_trials)
print("time bp3: ", t_bp4)
bp_unpool = sp.sparse1d_to_dense(sv3.out_indices, res_bp4, sv3.out_shape,
sv3.out_block_channel_mapping[-1])
#print("bp unpool", bp_unpool)
value3 = sp.sparse1d_to_dense(sv3.out_indices, sv3.out_values,
sv3.out_shape,
sv3.out_block_channel_mapping[-1])
#print("result sparse ", value3)
has_error = False
approx_cmp = expected.flatten()
approx = value3.flatten()
non_zero_count = 0
for i in range(len(approx_cmp)):
#if approx[i] == 0:
#approx_cmp[i] = 0
#else:
non_zero_count = non_zero_count + 1
print("entry count: ", non_zero_count)
error_cnt = 0
first_error = 0
correct_cnt = 0
for i in range(len(approx_cmp)):
if abs(approx_cmp[i] - approx[i]) > 1e-3:
#print("error: %d != %d " % (approx_cmp[i], approx[i]))
#print("at id ", i)
if has_error == False:
first_error = i
has_error = True
error_cnt = error_cnt + 1
elif approx[i] != 0:
correct_cnt = correct_cnt + 1
bp_sig_flat = bp_sig.flatten()
res_bp2_flat = res_bp1.flatten()
bp_i_error_cnt = 0
bp_i_correct_cnt = 0
for i in range(len(approx_cmp)):
if approx[i] != 0:
if bp_sig_flat[i] == res_bp2_flat[i]:
bp_i_correct_cnt = bp_i_correct_cnt + 1
else:
bp_i_error_cnt = bp_i_error_cnt + 1
p_flat = pd.out_values.flatten()
up_flat = sv4.flatten()
up_i_error_cnt = 0
up_i_correct_cnt = 0
for i in range(len(p_flat)):
if p_flat[i] <= up_flat[i]:
up_i_correct_cnt = up_i_correct_cnt + 1
else:
up_i_error_cnt = up_i_error_cnt + 1
if dim == 5:
up_bp_cor = 0
up_bp_err = 0
for batch in range(0, tensor_in_sizes[0]):
for channel in range(0, tensor_in_sizes[4]):
for x in range(
0,
int(ceil(
float(tensor_in_sizes[1]) / float(strides[1])))):
for y in range(
0,
int(
ceil(
float(tensor_in_sizes[2]) /
float(strides[2])))):
for z in range(
0,
int(
ceil(
float(tensor_in_sizes[3]) /
float(strides[3])))):
id_in = (batch, x, y, z, channel)
inval = value3.item(id_in)
max_out_val = -100000000000
for dx in range(0, strides[1]):
xout = x * strides[1] + dx
if xout >= d1.shape[1]:
continue
for dy in range(0, strides[2]):
yout = y * strides[2] + dy
if yout >= d1.shape[2]:
continue
for dz in range(0, strides[3]):
zout = z * strides[3] + dz
if zout >= d1.shape[3]:
continue
id_out = (batch, xout, yout, zout,
channel)
out_val = d1.item(id_out)
max_out_val = max(max_out_val, out_val)
if max_out_val == -100000000000 or max_out_val == inval:
up_bp_cor = up_bp_cor + 1
else:
up_bp_err = up_bp_err + 1
print("total number of pooling corrects: ", up_bp_cor)
print("total number of pooling errors: ", up_bp_err)
if dim == 5:
up_bp_cor = 0
up_bp_err = 0
tmp = np.copy(bp_unpool)
for batch in range(0, tensor_in_sizes[0]):
for channel in range(0, tensor_in_sizes[4]):
for x in range(
0,
int(ceil(
float(tensor_in_sizes[1]) / float(strides[1])))):
for y in range(
0,
int(
ceil(
float(tensor_in_sizes[2]) /
float(strides[2])))):
for z in range(
0,
int(
ceil(
float(tensor_in_sizes[3]) /
float(strides[3])))):
id_in = (batch, x, y, z, channel)
inval = bp_unpool.item(id_in)
sum_out_val = 0
for dx in range(0, strides[1]):
xout = x * strides[1] + dx
if xout >= bp_sig.shape[1]:
continue
for dy in range(0, strides[2]):
yout = y * strides[2] + dy
if yout >= bp_sig.shape[2]:
continue
for dz in range(0, strides[3]):
zout = z * strides[3] + dz
if zout >= bp_sig.shape[3]:
continue
id_out = (batch, xout, yout, zout,
channel)
out_val = bp_sig.item(id_out)
sum_out_val = sum_out_val + out_val
if sum_out_val == inval:
up_bp_cor = up_bp_cor + 1
else:
up_bp_err = up_bp_err + 1
tmp[id_in] = sum_out_val
#print("pbup: ", bp_unpool)
#print("epbup: ", tmp)
print("total number of unpooling bp corrects: ", up_bp_cor)
print("total number of unpooling bp errors: ", up_bp_err)
print("total number of non-zero corrects: ", correct_cnt)
print("total number of bpi corrects: ", bp_i_correct_cnt)
print("total number of unpooling corrects: ", up_i_correct_cnt)
if has_error:
print("total number of errors: ", error_cnt)
print("first error: ", first_error)
if bp_i_error_cnt > 0:
print("total number of bpi errors: ", bp_i_error_cnt)
if up_i_error_cnt > 0:
print("total number of up errors: ", up_i_error_cnt)
return 1
print("OK")
return 0