-
-
Save jmduarte/5c9cf2607a25633246e6fc85cbca89de to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import numpy as np | |
import random as rnd | |
import hls4ml | |
from tensorflow.keras.layers import Conv2D, MaxPooling2D, BatchNormalization | |
from qkeras.qconvolutional import QConv2D | |
from qkeras.quantizers import quantized_bits, quantized_relu | |
from tensorflow.keras.models import Sequential | |
from qkeras import QActivation | |
height = 8 | |
width = 8 | |
chan = 3 | |
input_shape = (height,width,chan) | |
num_classes = 32 | |
keras_model = Sequential() | |
keras_model.add(QConv2D(4, kernel_size=(3, 3), strides=(2, 2), activation='linear', input_shape=input_shape, | |
kernel_quantizer=quantized_bits(15, 7, alpha=1), | |
bias_quantizer=quantized_bits(15, 7, alpha=1), padding='valid')) | |
keras_model.add(BatchNormalization()) | |
keras_model.add(QActivation("quantized_bits(15, 7, alpha=1)")) | |
print(keras_model.summary()) | |
# Set some random weights | |
rnd.seed(42) | |
for layer in keras_model.layers: | |
old_weights = layer.get_weights() | |
if len(old_weights) > 0: | |
new_weights = [] | |
for w in old_weights: | |
new_w = [] | |
for i in range(np.prod(w.shape)): | |
new_w.append(rnd.uniform(0, 2)) | |
new_w = np.asarray(new_w).reshape(w.shape) | |
new_weights.append(new_w) | |
layer.set_weights(new_weights) | |
# Create conversion config | |
yaml_config = {} | |
yaml_config['KerasModel'] = keras_model | |
yaml_config['OutputDir'] = 'pr299_test' | |
yaml_config['ProjectName'] = 'myproject' | |
yaml_config['XilinxPart'] = 'xcvu9p-flgb2104-2-e' | |
yaml_config['ClockPeriod'] = 5 | |
yaml_config['IOType'] = 'io_stream' | |
yaml_config['HLSConfig'] = { | |
'Model': { | |
'Precision': 'ap_fixed<16,8>', | |
'ReuseFactor': 1 | |
} | |
} | |
# Prepare some data | |
np.random.seed(42) | |
x = np.random.rand(np.prod(keras_model.input.shape[1:])).reshape(keras_model.input.shape[1:]) | |
# Predict with Keras | |
y_gold = keras_model.predict(np.expand_dims(x, axis=0)) | |
print('Keras predictions:') | |
print(y_gold) | |
# Predict with hls4ml (strategy: latency) | |
hls_model = hls4ml.converters.keras_to_hls(yaml_config) | |
hls_model.compile() | |
y_latency = hls_model.predict(x) | |
print('HLS4ML predictions (strategy: latency):') | |
print(y_latency) | |
# Predict with hls4ml (strategy: resource) | |
yaml_config['HLSConfig']['Model']['Strategy'] = 'Resource' | |
hls_model = hls4ml.converters.keras_to_hls(yaml_config) | |
hls_model.compile() | |
y_resource = hls_model.predict(x) | |
print('HLS4ML predictions (strategy: resource):') | |
print(y_resource) | |
print('HLS4ML strategies are equal:', np.array_equal(y_latency, y_resource)) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment