def parse_pooling(self, layer, attributes):
attributes.update({
'input_shape': layer.input_shape,
'pool_size': layer.pool_size,
'strides': layer.stride,
'padding': padding_string(layer.pad, layer.pool_size)
})
def parse_convolution(self, layer, attributes):
idx_offset = 0 # Normally: params = [w, b]
# With the low-precision version of caffe (ADAPTIoN), there is a second
# copy of quantized parameters available.
if hasattr(layer, 'lpfp_param'):
attributes['Qm.f'] = (layer.lpfp_param.bd, layer.lpfp_param.ad)
attributes['quantize_bias'] = layer.lpfp_param.round_bias
if attributes['Qm.f'] != (0, 0):
idx_offset = 1 # params = [w, w_lp, b, b_lp]
weights = self.input_model[0].params[layer.name][0].data
bias = self.input_model[0].params[layer.name][1 + idx_offset].data
weights = weights[:, :, ::-1, ::-1]
weights = np.transpose(weights, (2, 3, 1, 0))
print("Flipped kernels.")
attributes['parameters'] = [weights, bias]
p = layer.convolution_param
# Take maximum here because sometimes not not all fields are set
# (e.g. kernel_h == 0 even though kernel_size == [3])
filter_size = [
max(p.kernel_w, p.kernel_size[0]),
max(p.kernel_h, p.kernel_size[-1])
]
pad = p.pad[0] if len(p.pad) > 0 else 0
pad = (max(p.pad_w, pad), max(p.pad_h, pad))
padding = padding_string(pad, filter_size)
strides = [max(p.stride_w, p.stride[0]), max(p.stride_h, p.stride[0])]
attributes.update({
'filters': p.num_output,
'kernel_size': filter_size,
'padding': padding,
'strides': strides
})
def parse_convolution(self, layer, attributes):
weights = k.get_value(layer.W)
weights = np.transpose(weights, (2, 3, 1, 0))
bias = k.get_value(layer.b)
if bias is None:
bias = np.zeros(layer.num_filters)
attributes['parameters'] = [weights, bias]
padding = padding_string(layer.pad, layer.filter_size)
attributes.update({
'input_shape': layer.input_shape,
'filters': layer.num_filters,
'kernel_size': layer.filter_size,
'padding': padding,
'strides': layer.stride
})
def parse_pooling(self, layer, attributes):
p = layer.pooling_param
# Take maximum here because sometimes not not all fields are set
# (e.g. kernel_h == 0 even though kernel_size == 2)
pool_size = [
max(p.kernel_w, p.kernel_size),
max(p.kernel_h, p.kernel_size)
]
pad = (max(p.pad_w, p.pad), max(p.pad_h, p.pad))
padding = padding_string(pad, pool_size)
strides = [max(p.stride_w, p.stride), max(p.stride_h, p.stride)]
attributes.update({
'pool_size': pool_size,
'padding': padding,
'strides': strides
})
def parse_convolution(self, layer, attributes):
weights = layer.W.get_value()
weights = np.transpose(weights, (2, 3, 1, 0))
if keras.backend.backend() != 'theano': # Assumes lasagne uses theano.
weights = keras.utils.conv_utils.convert_kernel(weights)
bias = layer.b.get_value()
if bias is None:
bias = np.zeros(layer.num_filters)
attributes['parameters'] = [weights, bias]
padding = padding_string(layer.pad, layer.filter_size)
attributes.update({
'input_shape': layer.input_shape,
'filters': layer.num_filters,
'kernel_size': layer.filter_size,
'padding': padding,
'strides': layer.stride
})