def three():
image_size = 32
batch_size = 32
input_filters = 512
output_filters = 512
np.random.seed(123)
with make_backend(batch_size=batch_size,
datatype=np.float32, device_id=0) as be:
W = np.random.randn(input_filters,3,3,output_filters).astype(np.float32)
W_cuda = MyTensor.from_np(W)
print('type(W_cuda)', type(W_cuda))
inputs = np.zeros((input_filters,image_size, image_size,batch_size), dtype=np.float32)
inputs[:] = np.random.randn(*inputs.shape)
inputs_cuda = MyTensor.from_np(inputs)
print('type(inputs_cuda)', type(inputs_cuda))
conv = Convolution((3, 3, output_filters), strides=1, padding=1, be=be) #, init=init)
print('created conv')
conv.W = W_cuda
conv.configure((input_filters,image_size, image_size))
conv.W = W_cuda
print('configure done')
outputs = np.zeros((image_size * image_size * output_filters, batch_size), dtype=np.float32)
outputs_cuda = MyTensor.from_np(outputs)
conv.outputs = outputs_cuda
conv.fprop(inputs_cuda)
cuda.Context.synchronize()
for it in range(3):
start = time.time()
conv.fprop(inputs_cuda)
cuda.Context.synchronize()
print('time=', time.time() - start)
# outputs = outputs_cuda.get()
outputs_cuda.to_host()
print(outputs[1:3,1:3])
print('outputs.shape', outputs.shape)
printDims(W=W, I=inputs)
check(W=W, I=inputs, O=outputs, c=0, h=0, w=0, n=0, eps=1e-3)
check(W=W, I=inputs, O=outputs, c=0, h=0, w=0, n=1, eps=1e-3)
check(W=W, I=inputs, O=outputs, c=0, h=0, w=1, n=0, eps=1e-3)
check(W=W, I=inputs, O=outputs, c=0, h=1, w=0, n=0, eps=1e-3)
check(W=W, I=inputs, O=outputs, c=1, h=0, w=0, n=0, eps=1e-3)
check(W=W, I=inputs, O=outputs, c=3, h=2, w=1, n=27, eps=1e-3)
check(W=W, I=inputs, O=outputs, c=17, h=25, w=7, n=27, eps=1e-3)
W_cuda = gpuarray.to_gpu(W)
conv.W = W_cuda
print('type(W_cuda)', type(W_cuda))
inputs = np.zeros((input_filters,image_size, image_size,batch_size), dtype=np.float32)
inputs[:] = np.random.randn(*inputs.shape)
inputs_cuda = gpuarray.to_gpu(inputs)
print('type(inputs_cuda)', type(inputs_cuda))
conv.configure((input_filters,image_size, image_size))
print('configure done')
outputs = np.zeros((image_size * image_size * output_filters, batch_size), dtype=np.float32)
outputs_cuda = gpuarray.to_gpu(outputs)
conv.outputs = outputs_cuda
conv.fprop(inputs_cuda)
for it in range(3):
start = time.time()
for i in range(10):
conv.fprop(inputs_cuda)
cuda.Context.synchronize()
print('time=', time.time() - start)
outputs = outputs_cuda.get()
print(outputs[1:3,1:3])
assert abs(outputs[1,1] - 1.33960593) < 1e-4
assert abs(outputs[1,2] + 6.06682396) < 1e-4
assert abs(outputs[2,2] - 8.76905346) < 1e-4
