def local_gpua_pool_dnn_grad_stride(op, ctx_name, inputs, outputs):
if not dnn_available(ctx_name):
return
if not op.ignore_border:
return
inp, out, out_grad, ws, stride, pad = inputs
nd = op.ndim
if nd not in (2, 3):
return
inp = gpu_contiguous(as_gpuarray_variable(inp, ctx_name))
out = gpu_contiguous(as_gpuarray_variable(out, ctx_name))
out_grad = gpu_contiguous(as_gpuarray_variable(out_grad, ctx_name))
mode = op.mode
# the GPU ops expect exactly 2 non-pooling dimensions
if inp.ndim == nd + 2:
return GpuDnnPoolGrad(mode=mode)(inp, out, out_grad, ws, stride, pad)
else:
# reshape to 4D or 5D with 2 non-pooling dimensions
inp_padded = pad_dims(inp, 2, nd)
out_padded = pad_dims(out, 2, nd)
out_grad_padded = pad_dims(out_grad, 2, nd)
ret_padded = GpuDnnPoolGrad(mode=mode)(inp_padded, out_padded,
out_grad_padded, ws, stride,
pad)
return unpad_dims(ret_padded, inp, 2, nd)
def local_gpua_avg_pool_dnn_grad_stride(op, ctx_name, inputs, outputs):
if not dnn_available(ctx_name):
return
if not op.ignore_border:
return
inp, out_grad, ws, stride, pad = inputs
nd = op.ndim
if nd not in (2, 3):
return
inp = gpu_contiguous(as_gpuarray_variable(inp, ctx_name))
out_grad = gpu_contiguous(as_gpuarray_variable(out_grad, ctx_name))
mode = op.mode
# the GPU ops expect exactly 2 non-pooling dimensions
if inp.ndim == nd + 2:
# We reuse out_grad because cuDNN does not use the value of the `out`
# argument but still checks its shape for average pooling. This
# has been observed in v2 and v3 as far as I know.
return GpuDnnPoolGrad(mode=mode)(inp, out_grad, out_grad, ws, stride,
pad)
else:
# reshape to 4D or 5D with 2 non-pooling dimensions
inp_padded = pad_dims(inp, 2, nd)
out_grad_padded = pad_dims(out_grad, 2, nd)
ret_padded = GpuDnnPoolGrad(mode=mode)(inp_padded, out_grad_padded,
out_grad_padded, ws, stride,
pad)
return unpad_dims(ret_padded, inp, 2, nd)
def local_gpua_pool_dnn_alternative(fgraph, op, ctx_name, inputs, outputs):
if not dnn_available(ctx_name):
return
if not op.ignore_border:
return
img, ws, stride, pad = inputs
nd = op.ndim
if nd not in (2, 3):
return
img = gpu_contiguous(as_gpuarray_variable(img, ctx_name))
mode = op.mode
# dnn_pool expects exactly 2 non-pooling dimensions
if img.ndim == nd + 2:
return dnn_pool(img, ws, stride=stride, pad=pad, mode=mode)
else:
# reshape to 4D or 5D with 2 non-pooling dimensions
img_padded = pad_dims(img, 2, nd)
ret_padded = dnn_pool(img_padded, ws, stride=stride, pad=pad, mode=mode)
return unpad_dims(ret_padded, img, 2, nd)
