x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
num_classes = 10
model = create_keras_model(num_classes)
model.load_weights(args.model)
if args.train:
train(model, x_train, y_train, x_test, y_test)
cpu_out = mlp_common.predict_cpu(model, x_test)
if args.engine == 'fcn':
fpga_out = mlp_common.predict_fpga(model, x_test, xclbin_prop,
g_in_scale, g_wgt_scale,
g_bias_scale, g_post_scale)
else:
fpga_out = mlp_common.predict_uspmv_fpga(model, x_test, xclbin_prop)
print("compare real data with cpu:")
mlp_common.compare_real_results(y_test, np.argmax(cpu_out, axis=1))
print("compare real data with fpga:")
mlp_common.compare_real_results(y_test, np.argmax(fpga_out, axis=1))
print("compare cpu with fpga:")
mlp_common.compare_real_results(np.argmax(cpu_out, axis=1),
np.argmax(fpga_out, axis=1))
gemx.createUSPMVHandle( args, xclbin_prop )
else:
gemx.createSPMVHandle( args, xclbin_prop )
train_fd = pd.read_csv(args.data) # Load training data.
IDcol = 'Run' # A column used to identified the run for data collection; not an independent variable.
target = 'Class' # The column name for our dependent variable.
predictors = [x for x in train_fd.columns if x not in [target, IDcol]] # Define column names to use as independent variables.
# Encode class values as integers
encoder = LabelEncoder()
encoder.fit(train_fd[target])
encoded_Y = encoder.transform(train_fd[target])
# Convert integers to dummy variables (i.e. one hot encoded)
train_y = np_utils.to_categorical(encoded_Y)
num_classes = len(train_fd[target].unique())
model = create_keras_model(train_fd[predictors].values.shape[1],num_classes)
model.load_weights(args.model)
if args.train:
train(train_fd, predictors, train_y, len(train_fd[target].unique()))
if args.engine == 'fcn':
fpga_out = mlp_common.predict_fpga(model, train_fd[predictors].values, xclbin_prop, g_in_scale, g_wgt_scale, g_wgt_scale, g_post_scale)
elif args.engine == 'uspmv':
fpga_out = mlp_common.predict_uspmv_fpga(model, train_fd[predictors].values, xclbin_prop)
else:
fpga_out = mlp_common.predict_spmv_fpga(model, train_fd[predictors].values, xclbin_prop)
cpu_out = mlp_common.predict_cpu( model, train_fd[predictors].values)
compare_results ( cpu_out, fpga_out)
if args.train:
train(train_fd, predictors, train_y, len(train_fd[target].unique()))
fpga_rt = []
fpga_out = []
for i in range(numKernels):
fpga_rt.append(
mlp_common.init_fpga(model, xclbin_opts, g_wgt_scale, g_bias_scale,
g_post_scale, None, i, 0))
inp = train_fd[predictors].values
if not args.run_async: # for larger batch size, run multi-kernels in parallel will bring up to 4x better performance
for i in range(numKernels):
fpga_out.append(
mlp_common.predict_fpga(fpga_rt[i], inp, g_in_scale))
else:
for i in range(numKernels):
fpga_rt[i].send_matrices(inp, None)
xfblas.executeAsync(numKernels, 0)
for i in range(numKernels):
fpga_out.append(fpga_rt[i].get_result())
fpga_out[i] = fpga_out[i].astype(np.float32)
for j in range(fpga_out[i].shape[0]):
fpga_out[i][j, :] = mlp_common.softmax(fpga_out[i][j, :])
cpu_out = mlp_common.predict_cpu(model, train_fd[predictors].values)