mnist.py

#
# DirectML MNIST sample
# Based on the following model: https://github.com/onnx/models/blob/master/vision/classification/mnist/model/mnist-8.onnx
#

import pydirectml as dml
import numpy as np
from PIL import Image, ImageOps
import sys
import os

argument_count = len(sys.argv)

image_file_path = "mnist_image1.png"
tensor_data_path = "mnist-8_tensor_data"

if (argument_count >= 2):
    image_file_path = sys.argv[1]

if (argument_count >= 3):
    tensor_data_path = sys.argv[2]

if (os.path.exists(image_file_path) == False):
    print("File not found at: " + str(image_file_path))
    sys.exit(1)

# Opens image, converts to grayscale, resizes, and crops to the input size.
image = ImageOps.fit(ImageOps.grayscale(Image.open(image_file_path)), (28, 28), method = 0, bleed = 0, centering = (0.5, 0.5))

# Check top left pixel's color, and inverts image color if pixel is bright.
x, y = 0, 0
coordinate = x, y
if (image.getpixel(coordinate) >= 128):
    image = ImageOps.invert(image)

# Expand dimensions to 4d tensor format, and rescale values in range of 0.0 and 1.0.
img_array = np.array(image, np.float32)
ndarray_image = np.expand_dims(img_array, axis = (0, 1))
rescaled_image = ndarray_image / ndarray_image.max()

input_bindings = []

def append_input_tensor(builder: dml.GraphBuilder, input_bindings: list, input_tensor: dml.TensorDesc, file_name: str):
    tensor = dml.input_tensor(builder, len(input_bindings), input_tensor)
    if file_name == "":
        input_bindings.append(dml.Binding(tensor, np.zeros(tensor.get_output_desc().sizes)))
    else:
        input_bindings.append(dml.Binding(tensor, np.load(tensor_data_path + "/" + file_name)))
    return tensor

# Create a GPU device, and build a model graph.
device = dml.Device(True, True)
builder = dml.GraphBuilder(device)

data_type = dml.TensorDataType.FLOAT32
input = dml.input_tensor(builder, 0, dml.TensorDesc(data_type, [1, 1, 28, 28]))
flags = dml.TensorFlags.OWNED_BY_DML

input_bindings.append(dml.Binding(input, rescaled_image))

# convolution28
convolution28_weight = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [8, 1, 5, 5]), "Parameter5.npy")
convolution28_bias = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [1, 8, 1, 1]), "")
convolution28 = dml.convolution(input, convolution28_weight, convolution28_bias, strides = [1, 1], start_padding = [2, 2], end_padding = [2, 2])

# plus30
plus30_param6 = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, [1, 8, 28, 28], [8, 1, 0, 0]), "Parameter6.npy")
plus30 = dml.add(convolution28, plus30_param6)

# relu32
relu32 = dml.activation_relu(plus30)

# pooling66
pooling66 = dml.max_pooling(relu32, strides = [2, 2], window_sizes = [2, 2])

# convolution110
convolution110_weight = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [16, 8, 5, 5]), "Parameter87.npy")
convolution110_bias = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [1, 16, 1, 1]), "")
convolution110 = dml.convolution(pooling66.values, convolution110_weight, convolution110_bias, strides = [1, 1], start_padding = [2, 2], end_padding = [2, 2])

# plus112
plus112_param88 = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, [1, 16, 14, 14], [16, 1, 0, 0]), "Parameter88.npy")
plus112 = dml.add(convolution110, plus112_param88)

# relu114
relu114 = dml.activation_relu(plus112)

# pooling160
pooling160 = dml.max_pooling(relu114, strides = [3, 3], window_sizes = [3, 3])

# times212_reshape0
times212_reshape0 = dml.reinterpret(pooling160.values, dml.TensorDataType.FLOAT32, [1, 1, 1, 256], [256, 256, 256, 1])

# times212_reshape1
times212_reshape1_param193 = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [16, 4, 4, 10]), "Parameter193.npy")
identity = dml.activation_identity(times212_reshape1_param193)
times212_reshape1 = dml.reinterpret(identity, dml.TensorDataType.FLOAT32, [1, 1, 256, 10], [2560, 2560, 10, 1])

# times212
times212 = dml.gemm(times212_reshape0, times212_reshape1)

# plus214
plus214_param194 = append_input_tensor(builder, input_bindings, dml.TensorDesc(data_type, flags, [1, 1, 1, 10]), "Parameter194.npy")
plus214 = dml.add(times212, plus214_param194)

softmax = dml.activation_soft_max(plus214)
# Compile the expression graph into a compiled operator
op = builder.build(dml.ExecutionFlags.NONE, [softmax])

# Compute the result
output_data = device.compute(op, input_bindings, [softmax])
output_tensor = np.array(output_data[0], np.float32)

number = np.argmax(output_tensor)
print("\nNumber is: {}".format(number, end=''))
print("Confidence: {:2.2f}%".format(np.amax(output_tensor) * 100))