Cleanup simulator and add usage instructions (#1050)

* cleanup simulator files

* minor updates

* update readme

* keras runner builds

* hmm, still doesn't work

* keras runner works

* should work with python3 keras mod

* touchups

external/tensorflow/.gitignore

@@ -1,4 +0,0 @@
-*.tar.gz
-include
-lib
-*LICENSE*

external/tensorflow/download.sh

@@ -1,11 +0,0 @@
-#!/bin/bash
-TF=libtensorflow-gpu-linux-x86_64-1.13.1.tar.gz
-#TF=libtensorflow-gpu-linux-x86_64-1.14.0.tar.gz
-#TF=libtensorflow-gpu-linux-x86_64-1.15.0.tar.gz
-
-if [ ! -f $TF ]; then
-  wget https://storage.googleapis.com/tensorflow/libtensorflow/$TF
-fi
-rm -rf include lib
-tar xvf $TF
-

selfdrive/modeld/SConscript

@@ -14,13 +14,11 @@ if arch == "aarch64":
 else:
   libs += ['symphony-cpu', 'pthread']
 
-  if FindFile('libtensorflow.so', env['LIBPATH']):
-    # for tensorflow support
-    common_src += ['runners/tfmodel.cc']
-    libs += ['tensorflow']
-    # tell runners to use it
-    lenv['CFLAGS'].append("-DUSE_TF_MODEL")
-    lenv['CXXFLAGS'].append("-DUSE_TF_MODEL")
+  # for tensorflow support
+  common_src += ['runners/tfmodel.cc']
+  # tell runners to use it
+  lenv['CFLAGS'].append("-DUSE_TF_MODEL")
+  lenv['CXXFLAGS'].append("-DUSE_TF_MODEL")
 
 common = lenv.Object(common_src)
 

selfdrive/modeld/models/dmonitoring.cc

@@ -108,7 +108,7 @@ DMonitoringResult dmonitoring_eval_frame(DMonitoringModelState* s, void* stream_
 
   delete[] cropped_buf;
   delete[] resized_buf;
-  s->m->execute(net_input_buf);
+  s->m->execute(net_input_buf, yuv_buf_len);
   delete[] net_input_buf;
 
   DMonitoringResult ret = {0};

selfdrive/modeld/models/driving.cc

@@ -70,7 +70,7 @@ ModelDataRaw model_eval_frame(ModelState* s, cl_command_queue q,
   float *new_frame_buf = frame_prepare(&s->frame, q, yuv_cl, width, height, transform);
   memmove(&s->input_frames[0], &s->input_frames[MODEL_FRAME_SIZE], sizeof(float)*MODEL_FRAME_SIZE);
   memmove(&s->input_frames[MODEL_FRAME_SIZE], new_frame_buf, sizeof(float)*MODEL_FRAME_SIZE);
-  s->m->execute(s->input_frames);
+  s->m->execute(s->input_frames, MODEL_FRAME_SIZE*2);
 
   #ifdef DUMP_YUV
     FILE *dump_yuv_file = fopen("/sdcard/dump.yuv", "wb");

selfdrive/modeld/runners/keras_runner.py

@@ -0,0 +1,53 @@
+#!/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)
+  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)
+

selfdrive/modeld/runners/runmodel.h

@@ -5,7 +5,7 @@ class RunModel {
 public:
   virtual void addRecurrent(float *state, int state_size) {}
   virtual void addDesire(float *state, int state_size) {}
-  virtual void execute(float *net_input_buf) {}
+  virtual void execute(float *net_input_buf, int buf_size) {}
 };
 
 #endif

selfdrive/modeld/runners/snpemodel.cc

@@ -117,7 +117,7 @@ std::unique_ptr<zdl::DlSystem::IUserBuffer> SNPEModel::addExtra(float *state, in
   return ret;
 }
 
-void SNPEModel::execute(float *net_input_buf) {
+void SNPEModel::execute(float *net_input_buf, int buf_size) {
   assert(inputBuffer->setBufferAddress(net_input_buf));
   if (!snpe->execute(inputMap, outputMap)) {
     PrintErrorStringAndExit();

selfdrive/modeld/runners/snpemodel.h

@@ -25,7 +25,7 @@ public:
   }
   void addRecurrent(float *state, int state_size);
   void addDesire(float *state, int state_size);
-  void execute(float *net_input_buf);
+  void execute(float *net_input_buf, int buf_size);
 private:
   uint8_t *model_data = NULL;
 

selfdrive/modeld/runners/tfmodel.cc

@@ -1,160 +1,98 @@
 #include "tfmodel.h"
+#include <stdio.h>
 #include <string>
 #include <string.h>
+#include <signal.h>
+#include <unistd.h>
 #include <stdlib.h>
 #include <stdexcept>
 #include "common/util.h"
+#include "common/utilpp.h"
 #include "common/swaglog.h"
 #include <cassert>
 
-void TFModel::status_check() const {
-  if (TF_GetCode(this->status) != TF_OK) {
-    throw std::runtime_error(TF_Message(status));
-  }
-}
-
-TF_Tensor *TFModel::allocate_tensor_for_output(TF_Output out, float *dat) {
-  int num_dims = TF_GraphGetTensorNumDims(graph, out, status);
-  status_check();
-  int64_t *dims = new int64_t[num_dims];
-  TF_GraphGetTensorShape(graph, out, dims, num_dims, status);
-  status_check();
-  dims[0] = 1;
-
-  int total = 1;
-  for (int i = 0; i < num_dims; i++) total *= dims[i];
-  //printf("dims %d total %d wdat %p\n", num_dims, total, dat);
-
-  // don't deallocate the buffers
-  auto d = [](void* ddata, size_t, void* arg) {};
-  TF_Tensor *ret = TF_NewTensor(TF_FLOAT, dims, num_dims, (void*)dat, sizeof(float)*total, d, NULL);
-
-  //TF_Tensor *ret = TF_AllocateTensor(TF_FLOAT, dims, num_dims, sizeof(float)*total);
-  //memcpy(TF_TensorData(ret), dat, sizeof(float)*total);
-
-  assert(ret);
-  delete[] dims;
-
-  return ret;
-}
 
 TFModel::TFModel(const char *path, float *_output, size_t _output_size, int runtime) {
-  // load model
-  {
-    TF_Buffer* buf;
-    size_t model_size;
-    char tmp[1024];
-    snprintf(tmp, sizeof(tmp), "%s.pb", path);
-    LOGD("loading model %s", tmp);
-    uint8_t *model_data = (uint8_t *)read_file(tmp, &model_size);
-    assert(model_data);
-    buf = TF_NewBuffer();
-    buf->data = model_data;
-    buf->length = model_size;
-    buf->data_deallocator = [](void *data, size_t) { free(data); };
-    LOGD("loaded model of size %d", model_size);
-
-    // import graph
-    status = TF_NewStatus();
-    graph = TF_NewGraph();
-    TF_ImportGraphDefOptions *opts = TF_NewImportGraphDefOptions();
-    // TODO: fix the GPU, currently it hangs if you set this to /gpu:0
-    //TF_ImportGraphDefOptionsSetDefaultDevice(opts, "/cpu:0");
-    TF_GraphImportGraphDef(graph, buf, opts, status);
-    TF_DeleteImportGraphDefOptions(opts);
-    TF_DeleteBuffer(buf);
-    status_check();
-    LOGD("imported graph");
-  }
-
-  // set up session
-  TF_SessionOptions* sess_opts = TF_NewSessionOptions();
-
-  // don't use all GPU memory
-  /*uint8_t config[15] = {0x32, 0xb, 0x9, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x20, 0x1, 0x38, 0x1};
-  double gpu_memory_fraction = 0.2;
-  auto bytes = reinterpret_cast<std::uint8_t*>(&gpu_memory_fraction);
-  for (std::size_t i = 0; i < sizeof(gpu_memory_fraction); ++i) {
-    config[i + 3] = bytes[i];
-  }
-  TF_SetConfig(sess_opts, config, sizeof(config), status);
-  status_check();*/
-
-  // make session
-  session = TF_NewSession(graph, sess_opts, status);
-  TF_DeleteSessionOptions(sess_opts);
-  status_check();
-
-  // find tensors
-  // TODO: make this generic
-  input_operation = {TF_GraphOperationByName(graph, "lambda/div"), 0};
-  if (input_operation.oper == NULL) {
-    input_operation = {TF_GraphOperationByName(graph, "vision_lambda/div"), 0};
-  }
-  assert(input_operation.oper != NULL);
-
-  output_operation = {TF_GraphOperationByName(graph, "outputs/outputs/Identity"), 0};
-  if (output_operation.oper == NULL) {
-    output_operation = {TF_GraphOperationByName(graph, "outputs/concat"), 0};
-  }
-  assert(output_operation.oper != NULL);
-
-  // output tensor is good to bind now
   output = _output;
   output_size = _output_size;
+
+  char tmp[1024];
+  strncpy(tmp, path, sizeof(tmp));
+  strstr(tmp, ".dlc")[0] = '\0';
+  strcat(tmp, ".keras");
+  LOGD("loading model %s", tmp);
+
+  assert(pipe(pipein) == 0);
+  assert(pipe(pipeout) == 0);
+
+  std::string exe_dir = util::dir_name(util::readlink("/proc/self/exe"));
+  std::string keras_runner = exe_dir + "/runners/keras_runner.py";
+
+  proc_pid = fork();
+  if (proc_pid == 0) {
+    LOGD("spawning keras process %s", keras_runner.c_str());
+    char *argv[] = {(char*)keras_runner.c_str(), tmp, NULL};
+    dup2(pipein[0], 0);
+    dup2(pipeout[1], 1);
+    close(pipein[0]);
+    close(pipein[1]);
+    close(pipeout[0]);
+    close(pipeout[1]);
+    execvp(keras_runner.c_str(), argv);
+  }
+
+  // parent
+  close(pipein[0]);
+  close(pipeout[1]);
 }
 
 TFModel::~TFModel() {
-  TF_DeleteSession(session, status);
-  status_check();
-  TF_DeleteGraph(graph);
-  TF_DeleteStatus(status);
+  close(pipein[1]);
+  close(pipeout[0]);
+  kill(proc_pid, SIGTERM);
+}
+
+void TFModel::pwrite(float *buf, int size) {
+  char *cbuf = (char *)buf;
+  int tw = size*sizeof(float);
+  while (tw > 0) {
+    int err = write(pipein[1], cbuf, tw);
+    //printf("host write %d\n", err);
+    assert(err >= 0);
+    cbuf += err;
+    tw -= err;
+  }
+  //printf("host write done\n");
+}
+
+void TFModel::pread(float *buf, int size) {
+  char *cbuf = (char *)buf;
+  int tr = size*sizeof(float);
+  while (tr > 0) {
+    int err = read(pipeout[0], cbuf, tr);
+    //printf("host read %d/%d\n", err, tr);
+    assert(err >= 0);
+    cbuf += err;
+    tr -= err;
+  }
+  //printf("host read done\n");
 }
 
 void TFModel::addRecurrent(float *state, int state_size) {
-  rnn_operation.oper = TF_GraphOperationByName(graph, "rnn_state");
-  rnn_operation.index = 0;
-  assert(rnn_operation.oper != NULL);
-
   rnn_input_buf = state;
+  rnn_state_size = state_size;
 }
 
 void TFModel::addDesire(float *state, int state_size) {
-  desire_operation.oper = TF_GraphOperationByName(graph, "desire");
-  desire_operation.index = 0;
-  assert(desire_operation.oper != NULL);
-
   desire_input_buf = state;
+  desire_state_size = state_size;
 }
 
-void TFModel::execute(float *net_input_buf) {
-  TF_Tensor *input_tensor = allocate_tensor_for_output(input_operation, net_input_buf);
-  assert(input_tensor);
-  TF_Tensor *output_tensor = NULL;
-
-  if (rnn_input_buf == NULL) {
-    TF_SessionRun(session, NULL,
-      &input_operation, &input_tensor, 1,
-      &output_operation, &output_tensor, 1,
-      NULL, 0, NULL, status);
-  } else {
-    //printf("%f %f %f\n", net_input_buf[0], rnn_input_buf[0], desire_input_buf[0]);
-    TF_Tensor *rnn_tensor = allocate_tensor_for_output(rnn_operation, rnn_input_buf);
-    TF_Tensor *desire_tensor = allocate_tensor_for_output(desire_operation, desire_input_buf);
-    TF_Output io[] = {input_operation, rnn_operation, desire_operation};
-    TF_Tensor* id[] = {input_tensor, rnn_tensor, desire_tensor};
-    TF_SessionRun(session, NULL,
-      io, id, 3,
-      &output_operation, &output_tensor, 1,
-      NULL, 0, NULL, status);
-    TF_DeleteTensor(rnn_tensor);
-    TF_DeleteTensor(desire_tensor);
-  }
-  TF_DeleteTensor(input_tensor);
-  status_check();
-  assert(output_tensor);
-  memcpy((void*)output, TF_TensorData(output_tensor), output_size*sizeof(float));
-  TF_DeleteTensor(output_tensor);
+void TFModel::execute(float *net_input_buf, int buf_size) {
+  // order must be this
+  pwrite(net_input_buf, buf_size);
+  pwrite(desire_input_buf, desire_state_size);
+  pwrite(rnn_input_buf, rnn_state_size);
+  pread(output, output_size);
 }
 
-

selfdrive/modeld/runners/tfmodel.h

@@ -4,8 +4,6 @@
 #include <stdlib.h>
 #include "runmodel.h"
 
-#include "tensorflow/c/c_api.h"
-
 struct TFState;
 
 class TFModel : public RunModel {
@@ -14,25 +12,23 @@ public:
 	~TFModel();
   void addRecurrent(float *state, int state_size);
   void addDesire(float *state, int state_size);
-  void execute(float *net_input_buf);
+  void execute(float *net_input_buf, int buf_size);
 private:
-  void status_check() const;
-  TF_Tensor *allocate_tensor_for_output(TF_Output out, float *dat);
+  int proc_pid;
 
   float *output;
   size_t output_size;
 
-  TF_Session* session;
-  TF_Graph* graph;
-  TF_Status* status;
-
-  TF_Output input_operation;
-  TF_Output rnn_operation;
-  TF_Output desire_operation;
-  TF_Output output_operation;
-
   float *rnn_input_buf = NULL;
+  int rnn_state_size;
   float *desire_input_buf = NULL;
+  int desire_state_size;
+
+  // pipe to communicate to keras subprocess
+  void pread(float *buf, int size);
+  void pwrite(float *buf, int size);
+  int pipein[2];
+  int pipeout[2];
 };
 
 #endif

tools/sim/.gitignore

@@ -1,3 +1,4 @@
 CARLA_*.tar.gz
 carla
+carla_tmp
 

tools/sim/README

@@ -0,0 +1,24 @@
+Needs Ubuntu 16.04
+
+== Checkout openpilot ==
+cd ~/
+git clone https://github.com/commaai/openpilot.git
+# Add export PYTHONPATH=$HOME/openpilot to your bashrc
+# Have a working tensorflow+keras in python2
+
+== Install (in tab 1) ==
+
+cd ~/openpilot/tools/sim
+./start_carla.sh  # install CARLA 0.9.7 and start the server
+
+== openpilot (in tab 2) ==
+
+cd ~/openpilot/selfdrive/
+PASSIVE=0 NOBOARD=1 ./manager.py
+
+== bridge (in tab 3) ==
+
+# links carla to openpilot, will "start the car" according to manager
+cd ~/openpilot/tools/sim
+./bridge.py
+

tools/sim/bridge.py

@@ -7,11 +7,16 @@ import numpy as np
 import threading
 import random
 import cereal.messaging as messaging
+import argparse
 from common.params import Params
 from common.realtime import Ratekeeper
-from can import can_function, sendcan_function
+from lib.can import can_function, sendcan_function
 import queue
 
+parser = argparse.ArgumentParser(description='Bridge between CARLA and openpilot.')
+parser.add_argument('--autopilot', action='store_true')
+args = parser.parse_args()
+
 pm = messaging.PubMaster(['frame', 'sensorEvents', 'can'])
 
 W,H = 1164, 874
@@ -99,7 +104,9 @@ def go():
 
   vehicle_bp = random.choice(blueprint_library.filter('vehicle.bmw.*'))
   vehicle = world.spawn_actor(vehicle_bp, random.choice(world_map.get_spawn_points()))
-  #vehicle.set_autopilot(True)
+
+  if args.autopilot:
+    vehicle.set_autopilot(True)
 
   blueprint = blueprint_library.find('sensor.camera.rgb')
   blueprint.set_attribute('image_size_x', str(W))
@@ -123,7 +130,6 @@ def go():
     print("done")
   atexit.register(destroy)
 
-
   # can loop
   sendcan = messaging.sub_sock('sendcan')
   rk = Ratekeeper(100)
@@ -148,6 +154,7 @@ if __name__ == "__main__":
   from selfdrive.version import terms_version, training_version
   params.put("HasAcceptedTerms", terms_version)
   params.put("CompletedTrainingVersion", training_version)
+  params.put("CommunityFeaturesToggle", "1")
 
   threading.Thread(target=health_function).start()
   threading.Thread(target=fake_driver_monitoring).start()

tools/sim/get.sh

@@ -1,10 +0,0 @@
-#!/bin/bash -e
-FILE=CARLA_0.9.7.tar.gz
-if [ ! -f $FILE ]; then
-  curl -O http://carla-assets-internal.s3.amazonaws.com/Releases/Linux/$FILE
-fi
-mkdir -p carla
-cd carla
-tar xvf ../$FILE
-easy_install PythonAPI/carla/dist/carla-0.9.7-py3.5-linux-x86_64.egg
-

tools/sim/start.sh

@@ -1,4 +0,0 @@
-#!/bin/bash
-cd carla
-./CarlaUE4.sh
-

tools/sim/start_carla.sh

@@ -0,0 +1,19 @@
+#!/bin/bash -e
+
+FILE=CARLA_0.9.7.tar.gz
+if [ ! -f $FILE ]; then
+  curl -O http://carla-assets-internal.s3.amazonaws.com/Releases/Linux/$FILE
+fi
+if [ ! -d carla ]; then
+  rm -rf carla_tmp
+  mkdir -p carla_tmp
+  cd carla_tmp
+  tar xvf ../$FILE
+  easy_install PythonAPI/carla/dist/carla-0.9.7-py3.5-linux-x86_64.egg
+  cd ../
+  mv carla_tmp carla
+fi
+
+cd carla
+./CarlaUE4.sh
+