Hi, I will be appreciate it if I get any hints on this issue. 
I’d like to tune pre-trained tflite models such as mobilenet-v1 from google (https://www.tensorflow.org/lite/guide/hosted_models) on an android mobile device.
But there’s a format issue. (NHWC)
I followed your guide.
- I can register and connect the android device.
- rpc test is done. (including opencl)
- “Auto-tuning a convolutional network for Mobile GPU” in tutorials, the default model is resnet.
There’s a compile issue while running tune_relay_mobile_gpu.py. - When I change the model as “mobilenet”, no errors but the device is disconnected then try to connect during auto-tuning. It seems working.
So, I modified “tune_relay_mobile_gpu.py” a little bit to load tflite model.
def get_network( batch_size):
output_shape = (batch_size, 1001)
tflite_model_file = os.path.join("./", “mobilenet_v1_1.0_224.tflite”)
tflite_model_buf = open(tflite_model_file, “rb”).read()
import tflite.Model
tflite_model = tflite.Model.Model.GetRootAsModel(tflite_model_buf, 0)
input_tensor = “input”
input_shape = (batch_size, 224, 224, 3)
from tvm import relay
mod, params = relay.frontend.from_tflite(tflite_model,
shape_dict={input_tensor: input_shape}
dtype_dict={input_tensor: dtype})
return mod, params, input_shape, output_shape
Loading is okay but I got an assertion error after that.
File “/home/dongjin/Desktop/fall/tvm/python/tvm/autotvm/task/topi_integration.py”, line 179, in _topi_nn_conv2d
assert layout == ‘NCHW’, “only support NCHW currently”
Because all data formats are NHWC in tflite file… so I’m wondering if there is a way to transpose this NHWC format to the NCHW format before calling autotvm.task.extract_from_program method.
Regards,
nadongguri.
