Getting started: using the new features of MIGraphX 0.5

New Features in MIGraphX 0.5

MIGraphX 0.5 is primarily a performance and bugfix release. The release includes:

• Additional operators: split, ceil, floor
• Support for additional models including NASNet-a_Large for Tensorflow
• Simplified python interface
• Performance improvements
• A driver for exercising migraphx

driver

There is a new driver as /opt/rocm/bin/migraphx-driver. The --help command shows options available

prompt$ /opt/rocm/bin/migraphx-driver --help

    -h, --help
        Show help

Commands:
    perf
    params
    read
    run
    verify
    compile

For example, the read command will read a file and print the internal graph from MIGraphx

/opt/rocm/bin/migraphx-driver read --onnx /home/mev/source/migraphx_onnx/torchvision/resnet50i64.onnx 

Another example, the following command measures performance running an ONNX file using the driver

/opt/rocm/bin/migraphx-driver perf --onnx /home/mev/source/migraphx_onnx/torchvision/resnet50i64.onnx 

The verify command checks internal consistency once read into MIGraphX

/opt/rocm/bin/migraphx-driver perf --onnx /home/mev/source/migraphx_onnx/torchvision/resnet50i64.onnx 

python

The Python interface has been simplified so that it is no longer necessary to copy in parameters. An updated version of the webcam example from MIGraphX examples v0.2 is shown below.

A careful comparison with previous example shows it is no longer necessary to allocate the parameters on the GPU or to explicitly copy parameters to the GPU or results from the GPU. The Python interface handles this by default.

import numpy as np
import cv2
import json
import migraphx

# video settings
cap = cv2.VideoCapture(0)
cap.set(cv2.cv.CV_CAP_PROP_FRAME_WIDTH, 320)
cap.set(cv2.cv.CV_CAP_PROP_FRAME_HEIGHT, 240)
ret, frame = cap.read()

# neural network settings
model = migraphx.parse_onnx("resnet50.onnx")
model.compile(migraphx.get_target("gpu"))

# get labels
with open('imagenet_class_index.json') as json_data:
    class_idx = json.load(json_data)
    idx2label = [class_idx[str(k)][1] for k in range(len(class_idx))]

# primary loop to read webcam images
count = 0
while (True):
    # capture frame by frame
    ret, frame = cap.read()

    if ret: # check - some webcams need warmup operations on the frame
        cropped = frame[16:304,8:232]    # 224x224

        trans = cropped.transpose(2,0,1) # convert HWC to CHW

        # convert to float, normalize and make batch size = 1
        image = np.ascontiguousarray(
            np.expand_dims(trans.astype('float32')/256.0,0))

        # display the frame
        cv2.imshow('frame',cropped)

        migraphx_result = model.run({'0':migraphx.argument(image)})

        result = np.array(migraphx_result,copy=False)

        idx = np.argmax(result[0])

        print(idx2label[idx], " ", result[0][idx])

    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

# when everything is done, release the capture
cap.release()
cv2.destroyAllWindows()