def compute_features_core(image_fh,filter_fh,model_config,convolve_func):
m_config = model_config['config']['model']
conv_mode = m_config['conv_mode']
#preprocessing
array = v1e.image2array(m_config ,image_fh)
preprocessed,orig_imga = v1e.preprocess(array,m_config)
#input normalization
norm_in = v1e.norm(preprocessed,conv_mode,m_config.get('normin'))
#filtering
filtered = v1e.convolve(norm_in, filter_fh, m_config , convolve_func)
#nonlinear activation
activ = v1e.activate(filtered,m_config.get('activ'))
#output normalization
norm_out = v1e.norm(activ,conv_mode,m_config.get('normout'))
#pooling
pooled = v1e.pool(norm_out,conv_mode,m_config.get('pool'))
return pooled
def extract_features(image_config, image_fs, filterbank, model_config, convolve_func):
cached_val = get_from_cache((image_config,model_config),FEATURE_CACHE)
if cached_val is not None:
output = cached_val
else:
print('extracting', image_config, model_config)
image_fh = image_fs.get_version(image_config['filename'])
m_config = model_config
conv_mode = m_config['conv_mode']
#preprocessing
array = v1e.image2array(m_config ,image_fh)
preprocessed,orig_imga = v1e.preprocess(array,m_config )
#input normalization
norm_in = v1e.norm(preprocessed,conv_mode,m_config.get('normin'))
#filtering
filtered = v1e.convolve(norm_in, filterbank, m_config , convolve_func)
#nonlinear activation
activ = v1e.activate(filtered,m_config.get('activ'))
#output normalization
norm_out = v1e.norm(activ,conv_mode,m_config.get('normout'))
#pooling
pooled = v1e.pool(norm_out,conv_mode,m_config.get('pool'))
#postprocessing
fvector_l = v1e.postprocess(norm_in,filtered,activ,norm_out,pooled,orig_imga,m_config.get('featsel'))
output = sp.concatenate(fvector_l).ravel()
put_in_cache((image_config,m_config),output,FEATURE_CACHE)
return output
