dragonpilot/selfdrive/modeld/runners/keras_runner.py

#!/usr/bin/env python3
# TODO: why are the keras models saved with python 2?
from __future__ import print_function

import tensorflow as tf
import os
import sys
import tensorflow.keras as keras
import numpy as np
from tensorflow.keras.models import Model
from tensorflow.keras.models import load_model

def read(sz):
  dd = []
  gt = 0
  while gt < sz*4:
    st = os.read(0, sz*4 - gt)
    assert(len(st) > 0)
    dd.append(st)
    gt += len(st)
  return np.fromstring(b''.join(dd), dtype=np.float32)

def write(d):
  os.write(1, d.tobytes())

def run_loop(m):
  isize = m.inputs[0].shape[1]
  osize = m.outputs[0].shape[1]
  print("ready to run keras model %d -> %d" % (isize, osize), file=sys.stderr)
  while 1:
    idata = read(isize).reshape((1, isize))
    ret = m.predict_on_batch(idata)
    write(ret)

if __name__ == "__main__":
  print(tf.__version__, file=sys.stderr)
  # limit gram alloc
  gpus = tf.config.experimental.list_physical_devices('GPU')
  if len(gpus) > 0:
    tf.config.experimental.set_virtual_device_configuration(gpus[0], [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=2048)])

  m = load_model(sys.argv[1])
  print(m, file=sys.stderr)
  bs = [int(np.product(ii.shape[1:])) for ii in m.inputs]
  ri = keras.layers.Input((sum(bs),))

  tii = []
  acc = 0
  for i, ii in enumerate(m.inputs):
    print(ii, file=sys.stderr)
    ti = keras.layers.Lambda(lambda x: x[:,acc:acc+bs[i]], output_shape=(1, bs[i]))(ri)
    acc += bs[i]
    tr = keras.layers.Reshape(ii.shape[1:])(ti)
    tii.append(tr)
  no = keras.layers.Concatenate()(m(tii))
  m = Model(inputs=ri, outputs=[no])
  run_loop(m)