def v1like_filter_cufft(image,filter_source,model_config):
conv_mode = model_config['conv_mode']
filter_shape = model_config['filter']['kshape']
image_shape = image.shape
full_shape = tuple( np.array(image_shape) + np.array(filter_shape) - 1 )
image_fft = v1_pyfft.cufft(image,full_shape)
filter_key = (full_shape,repr(model_config['filter']))
if filter_key not in FILTER_FFT_CACHE:
filterbank = filter_source()
original_dtype = filterbank.dtype
filter_fft = np.empty(full_shape + (filterbank.shape[2],),dtype=np.complex64)
for i in range(filterbank.shape[2]):
filter_fft[:,:,i] = v1_pyfft.fft(filterbank[:,:,i],full_shape)
FILTER_FFT_CACHE[filter_key] = (filter_fft,original_dtype)
else:
(filter_fft,original_dtype) = FILTER_FFT_CACHE[filter_key]
res_fft = np.empty(fft_shape + (filter_fft.shape[2],), dtype=filter_fft.dtype)
for i in range(res_fft.shape[2]):
res_fft[:,:,i] = v1_pyfft.cufft(image_fft * filter_fft[:,:,i],inverse=True)
myslice = get_bounds(image_shape,filter_shape,conv_mode)
res_fft = res_fft[myslice]
return np.cast[original_dtype](res_fft)
def v1like_filter_pyfft(image,filter_source,model_config,device_id=0):
conv_mode = model_config['conv_mode']
filter_shape = model_config['filter']['kshape']
image_shape = image.shape
full_shape = tuple( np.array(image_shape) + np.array(filter_shape) - 1 )
fft_shape = power2(full_shape)
image_fft = v1_pyfft.fft(image,fft_shape,device_id=device_id)
filter_key = (fft_shape,repr(model_config['filter']))
if filter_key not in FILTER_FFT_CACHE:
filterbank = filter_source()
original_dtype = filterbank.dtype
filter_fft = np.empty((filterbank.shape[0],) + fft_shape ,dtype=np.complex64)
for i in range(filterbank.shape[2]):
filter_fft[i] = v1_pyfft.fft(filterbank[i],fft_shape,device_id=device_id)
FILTER_FFT_CACHE[filter_key] = (filter_fft,original_dtype)
else:
(filter_fft,original_dtype) = FILTER_FFT_CACHE[filter_key]
res_fft = np.empty((filter_fft.shape[0],) + fft_shape , dtype=filter_fft.dtype)
for i in range(res_fft.shape[2]):
res_fft[:,:,i] = v1_pyfft.fft(image_fft * filter_fft[i],inverse=True,device_id=device_id)
delta = np.array(fft_shape) - np.array(full_shape)
myslice = tuple([slice(0,f) for (d,f) in zip(delta,full_shape)])
res_fft = res_fft[myslice]
myslice = get_bounds(image_shape,filter_shape,conv_mode)
res_fft = res_fft[myslice]
return np.cast[original_dtype](res_fft)
