def keras_graph_to_onnx_graph(cls, model, name):
# some import attribute:keras_version, backend
# this list record outputs of this graph
graph_outputs = []
for o in model.outputs:
graph_outputs.append(
make_tensor_value_info(name=o.name,
elem_type=STR_TO_ONNX_TYPE[K.dtype(o)],
shape=convert_shape(K.int_shape(o))))
# this list record input of this graph
graph_inputs = []
for i in model.inputs:
shape = convert_shape(K.int_shape(i))
if shape[0] == -1:
shape[0] = 1
# channels_last to channels_first
if K.image_data_format() == "channels_last" and len(shape) == 4:
shape = shape[:1] + shape[3:] + shape[1:3]
print(shape)
graph_inputs.append(
make_tensor_value_info(name=i.name,
elem_type=STR_TO_ONNX_TYPE[K.dtype(i)],
shape=shape))
# save structure of the graph (all layers) into nodes
nodes = []
# save all weights into initializer
initializer = []
for i, layer in enumerate(model.layers):
# the InputLayer only contain input data of model
if isinstance(layer, InputLayer):
continue
else:
handler = cls.switch_onnx_node_creater(layer)
graph_input_list, weight_list, node_list = handler(layer)
nodes.extend(node_list)
graph_inputs.extend(graph_input_list)
initializer.extend(weight_list)
return make_graph(nodes=nodes,
name=name,
inputs=graph_inputs,
outputs=graph_outputs,
initializer=initializer)
im_l = im_l[..., 0:3]
im_input = im_l / 255.0
im_input = np.transpose(im_input, (2, 0, 1))
im_input = im_input[np.newaxis, ...]
im_input = torch.from_numpy(im_input).float()
if opt.cuda:
model = model.cuda()
im_input = im_input.cuda()
with torch.no_grad():
out_c, out_s, out_p = model(im_input)
out_c, out_s, out_p = out_c.cpu(), out_s.cpu(), out_p.cpu()
out_img_c = out_c.detach().numpy().squeeze()
out_img_c = utils.convert_shape(out_img_c)
out_img_s = out_s.detach().numpy().squeeze()
out_img_s = utils.convert_shape(out_img_s)
out_img_p = out_p.detach().numpy().squeeze()
out_img_p = utils.convert_shape(out_img_p)
if opt.isHR:
if opt.only_y is True:
im_label = utils.quantize(sc.rgb2ycbcr(im_gt)[:, :, 0])
im_pre = utils.quantize(sc.rgb2ycbcr(out_img_c)[:, :, 0])
else:
im_label = im_gt
im_pre = out_img_c
def make_symbolic_weights(cls, name, weights):
return make_tensor_value_info(
name=name,
elem_type=NP_TYPE_TO_TENSOR_TYPE[weights.dtype],
shape=convert_shape(weights.shape))
def make_weights(cls, name, weight):
return make_tensor(name=name,
data_type=NP_TYPE_TO_TENSOR_TYPE[weight.dtype],
dims=convert_shape(weight.shape),
vals=weight.flatten().tolist())