I’m trying to convert my model from pytorch to caffe2 using onnx. I’m performing the conversion as shown below:
from torch.autograd import Variable
import torch.onnx
import _init_paths
from nets.resnet_v1 import resnetv1
import numpy as np
num_classes=30
model = resnetv1(batch_size=1, num_layers=101)
model.create_architecture(num_classes, tag='default',anchor_scales=[8,16,32],anchor_ratios=[0.5,1.0,2.0])
model.eval()
model.cuda()
model_path = "res101_faster_rcnn_iter_300000.pth"
model.load_state_dict(torch.load(model_path))
x = np.random.randn(1,224,224,3)
im_info = np.array([224.0,224.0,1.0], dtype=np.float32)
torch_out = torch.onnx.export(model, # model being run
[x,im_info], # model input (or a tuple for multiple inputs)
"resnet101_fasterrcnn.onnx", # where to save the model (can be a file or file-like object)
export_params=True) # store the trained parameter weights inside the model file
This throws the error: ValueError: NestedIOFunction doesn't know how to process an input object of type ndarray.
I know that I could easily convert the above numpy arrays to torch Variables, but my forward() method expects numpy arrays (and a string) as input (see below):
def forward(self, image, im_info, gt_boxes=None, mode='TRAIN'):
self._image = Variable(torch.from_numpy(image.transpose([0, 3, 1, 2])).cuda(), volatile=mode == 'TEST')
self._im_info = im_info # No need to change; actually it can be an list
self._gt_boxes = Variable(torch.from_numpy(gt_boxes).cuda()) if gt_boxes is not None else None
self._mode = mode
rois, cls_prob, bbox_pred = self._predict(mode)
Is there a way to perform the conversion without modifying the original source code (for example, by allowing onnx.export() to take numpy arrays as input)?