Replies: 2 comments 4 replies
-
|
Can you post the full code snippet here ? I'll take a closer look |
Beta Was this translation helpful? Give feedback.
-
|
Sure, here you go import keras
import tf2onnx
import tensorflow as tf
import os
import ezkl
import cv2
model_keras = keras.saving.load_model('../../models/one-class.keras')
model_keras.summary(expand_nested=True)
# ONNX Conversion
model_path = './one-class.onnx'
onnx_model, _ = tf2onnx.convert.from_keras(
model_keras,
input_signature=(
tf.TensorSpec(
[
1,
model_keras.inputs[0].shape[1],
model_keras.inputs[0].shape[2],
model_keras.inputs[0].shape[3]
],
dtype=model_keras.inputs[0].dtype,
name=model_keras.inputs[0].name,
),
),
inputs_as_nchw=['input_layer_1'],
output_path=model_path,
opset=12, # the ONNX version to export the model to
)
# Create circuit
output_root = './zk'
settings_path = os.path.join(output_root, 'settings.json')
data_path = os.path.join(output_root, 'input.json')
cal_path = os.path.join(output_root, 'cal_data.json')
compiled_model_path = os.path.join(output_root, 'network.compiled')
pk_path = os.path.join(output_root, 'test.pk')
vk_path = os.path.join(output_root, 'test.vk')
py_run_args = ezkl.PyRunArgs()
py_run_args.input_visibility = "public"
py_run_args.output_visibility = "public"
py_run_args.param_visibility = "fixed" # private by default
res = ezkl.gen_settings(model_path, settings_path, py_run_args=py_run_args, )
assert res == True
# Create calibration data
# Capture set of data points
num_data_points = 4
dataset_path = '../../models/one_class'
data_points = []
for i, file in enumerate(os.listdir(os.path.join(dataset_path, 'target'))):
if i >= num_data_points:
break
if '.png' in file:
image = cv2.imread(os.path.join(dataset_path, 'cells', file), cv2.IMREAD_GRAYSCALE)
image = cv2.resize(image, (128, 128))
data_points.append(image)
for i, file in enumerate(os.listdir(os.path.join(dataset_path, 'not_target'))):
if i >= num_data_points:
break
if '.png' in file:
image = cv2.imread(os.path.join(dataset_path, 'not', file), cv2.IMREAD_GRAYSCALE)
image = cv2.resize(image, (128, 128))
data_points.append(image)
# Stack the data points to create a batch
train_data_batch = tf.cast(tf.stack(data_points), tf.float32)
x = train_data_batch.numpy().reshape([-1]).tolist()
data = dict(input_data = [x])
# Serialize data into file:
json.dump( data, open(cal_path, 'w' ))
# Calibrate
ezkl calibrate-settings --data ./zk/cal_data.json --model ./one-class.onnx --settings-path ./zk/settings.json
res = await ezkl.calibrate_settings(cal_path, model_path, settings_path, "resources")
# Compile
res = ezkl.compile_circuit(model_path, compiled_model_path, settings_path)
assert res == True
# Setup
res = ezkl.get_srs(settings_path)
res = ezkl.setup(
compiled_model_path,
vk_path,
pk_path,
)
assert res == True
assert os.path.isfile(vk_path)
assert os.path.isfile(pk_path)
assert os.path.isfile(settings_path)The result of this (i I skip the calibration step) is the following error : (it was executed on Jupyter notebook) |
Beta Was this translation helpful? Give feedback.
-
|
ty ty :) will get back to u |
Beta Was this translation helpful? Give feedback.
-
|
can I get the code where you train and save the keras model so that i can run the thing end to end ? |
Beta Was this translation helpful? Give feedback.
-
|
Yes, give me some time and I'll create a gist for you. |
Beta Was this translation helpful? Give feedback.
-
|
You will find all the model definition here I cannot freely share the data I used (not quickly at least). |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hi, first of all, awesome project! really easy to use!
Now to my problem, I have a small Keras conv net:
Model Summary
Wich I converted to ONNX and created a setting files following the basic procedure:
Conversion procedure
ONNX file creation:Generated Settings
Calibrated (cannot finish, too long)
Compile and get SRS
Now, I have some problems and questions.
Question 1
First, and most important, the setup function is not working:
it returns:
thread '<unnamed>' panicked at src/graph/vars.rs:444:21: dynamic lookup or shuffle should only have one blockI don't know why it happened or how to proceed. The network itself is really simple, no exotic layers.
Question 2
The
ezkl.calibrate_settingsis very slow, taking ~4 hs per step on a i9 CPU (I never finished the calibration process but I read it is optional).Question 3
Can calibration be avoided if the model is created using quantization-aware training?
If so, where can I find more data on the quantization types recommended?
Beta Was this translation helpful? Give feedback.
All reactions