def train_supervised():
initialize_experiment()
setup = get_current_parameters()["Setup"]
supervised_params = get_current_parameters()["Supervised"]
num_epochs = supervised_params["num_epochs"]
model, model_loaded = load_model()
print("Loading data")
train_envs, dev_envs, test_envs = get_all_env_id_lists(max_envs=setup["max_envs"])
if "split_train_data" in supervised_params and supervised_params["split_train_data"]:
split_name = supervised_params["train_data_split"]
split = load_env_split()[split_name]
train_envs = [env_id for env_id in train_envs if env_id in split]
print("Using " + str(len(train_envs)) + " envs from dataset split: " + split_name)
filename = "supervised_" + setup["model"] + "_" + setup["run_name"]
start_filename = "tmp/" + filename + "_epoch_" + str(supervised_params["start_epoch"])
if supervised_params["start_epoch"] > 0:
if file_exists(start_filename):
load_pytorch_model(model, start_filename)
else:
print("Couldn't continue training. Model file doesn't exist at:")
print(start_filename)
exit(-1)
if setup["restore_weights_name"]:
restore_pretrained_weights(model, setup["restore_weights_name"], setup["fix_restored_weights"])
trainer = Trainer(model, epoch=supervised_params["start_epoch"], name=setup["model"], run_name=setup["run_name"])
print("Beginning training...")
best_test_loss = 1000
for epoch in range(num_epochs):
train_loss = trainer.train_epoch(train_data=None, train_envs=train_envs, eval=False)
trainer.model.correct_goals = 0
trainer.model.total_goals = 0
test_loss = trainer.train_epoch(train_data=None, train_envs=dev_envs, eval=True)
print("GOALS: ", trainer.model.correct_goals, trainer.model.total_goals)
if test_loss < best_test_loss:
best_test_loss = test_loss
save_pytorch_model(trainer.model, filename)
print("Saved model in:", filename)
print ("Epoch", epoch, "train_loss:", train_loss, "test_loss:", test_loss)
save_pytorch_model(trainer.model, "tmp/" + filename + "_epoch_" + str(epoch))
if hasattr(trainer.model, "save"):
trainer.model.save(epoch)
save_pretrained_weights(trainer.model, setup["run_name"])
def evaluate():
P.initialize_experiment()
params = P.get_current_parameters()
setup = params["Setup"]
# import pdb;pdb.set_trace()
models = []
for i in range(setup["num_workers"]):
model, model_loaded = load_model()
if setup["restore_weights_name"]:
restore_pretrained_weights(model, setup["restore_weights_name"],
setup["fix_restored_weights"])
models.append(model)
eval_envs = get_correct_eval_env_id_list()
roll_out_params = RollOutParams() \
.setModelName(setup["model"]) \
.setModelFile(setup["model_file"]) \
.setRunName(setup["run_name"]) \
.setSetupName(P.get_setup_name()) \
.setEnvList(eval_envs) \
.setMaxDeviation(400) \
.setHorizon(100) \
.setStepsToForceStop(10) \
.setPlot(False) \
.setShowAction(False) \
.setIgnorePolicyStop(False) \
.setPlotDir("evaluate/" + setup["run_name"]) \
.setSavePlots(True) \
.setRealtimeFirstPerson(False) \
.setSaveSamples(False) \
.setBuildTrainData(False) \
.setSegmentReset("always") \
.setSegmentLevel(True) \
.setFirstSegmentOnly(False) \
.setDebug(setup["debug"]) \
.setCuda(setup["cuda"])
custom_eval = "Eval" in params and params["Eval"]["custom_eval"]
instructions = None
if custom_eval:
examples = params["Eval"]["examples"]
eval_envs, eval_sets, eval_segs, instructions = tuple(
map(lambda m: list(m), list(zip(*examples))))
print("!! Running custom evaluation with the following setup:")
print(examples)
roll_out_params.setEnvList(eval_envs)
roll_out_params.setSegList(eval_segs)
roll_out_params.setCustomInstructions(instructions)
if setup["num_workers"] > 1:
roller = ParallelPolicyRoller(num_workers=setup["num_workers"])
else:
roller = PolicyRoller()
dataset = roller.roll_out_policy(roll_out_params)
results = {}
if setup["eval_landmark_side"]:
evaler = DataEvalLandmarkSide(setup["run_name"])
evaler.evaluate_dataset(dataset)
results = evaler.get_results()
if setup["eval_nl"]:
evaler = DataEvalNL(setup["run_name"],
save_images=True,
entire_trajectory=False,
custom_instr=instructions)
evaler.evaluate_dataset(dataset)
results = evaler.get_results()
print("Results:", results)
def train_top_down_pred(args, max_epoch=SUPERVISED_EPOCHS):
initialize_experiment(args.run_name, args.setup_name)
model, model_loaded = load_model()
# TODO: Get batch size from global parameter server when it exists
batch_size = 1 if \
args.model == "top_down" or \
args.model == "top_down_prior" or \
args.model == "top_down_sm" or \
args.model == "top_down_pretrain" or \
args.model == "top_down_goal_pretrain" or \
args.model == "top_down_nav" or \
args.model == "top_down_cond" \
else BATCH_SIZE
lr = 0.001 # * batch_size
trainer = Trainer(model,
epoch=args.start_epoch,
name=args.model,
run_name=args.run_name)
train_envs, dev_envs, test_envs = get_all_env_id_lists(
max_envs=args.max_envs)
filename = "top_down_" + args.model + "_" + args.run_name
if args.restore_weights_name is not None:
restore_pretrained_weights(model, args.restore_weights_name,
args.fix_restored_weights)
print("Beginning training...")
best_test_loss = 1000
validation_loss = []
for epoch in range(SUPERVISED_EPOCHS):
train_loss = -1
if not args.eval_pretrain:
train_loss = trainer.train_epoch(train_envs=train_envs, eval=False)
test_loss = trainer.train_epoch(train_envs=dev_envs, eval=True)
validation_loss.append([epoch, test_loss])
if not args.eval_pretrain:
if test_loss < best_test_loss:
best_test_loss = test_loss
save_pytorch_model(trainer.model, filename)
print("Saved model in:", filename)
print("Epoch", epoch, "train_loss:", train_loss, "test_loss:",
test_loss)
save_pytorch_model(trainer.model,
"tmp/" + filename + "_epoch_" + str(epoch))
save_pretrained_weights(trainer.model, args.run_name)
else:
break
if max_epoch is not None and epoch > max_epoch:
print("Reached epoch limit!")
break
test_loss_dir = get_model_dir(
) + "/test_loss/" + filename + "_test_loss.csv"
validation_loss = pd.DataFrame(validation_loss,
columns=['epoch', "test_loss"])
validation_loss.to_csv(test_loss_dir, index=False)
def train_supervised():
initialize_experiment()
setup = get_current_parameters()["Setup"]
supervised_params = get_current_parameters()["Supervised"]
num_epochs = supervised_params["num_epochs"]
model, model_loaded = load_model()
# import pdb; pdb.set_trace()
# import pickle
# with open('/storage/dxsun/model_input.pickle', 'rb') as f: data = pickle.load(f)
# g = model(data['images'], data['states'], data['instructions'], data['instr_lengths'], data['has_obs'], data['plan'], data['save_maps_only'], data['pos_enc'], data['noisy_poses'], data['start_poses'], data['firstseg'])
print("model:", model)
print("model type:", type(model))
print("Loading data")
# import pdb;pdb.set_trace()
train_envs, dev_envs, test_envs = get_all_env_id_lists(
max_envs=setup["max_envs"])
if "split_train_data" in supervised_params and supervised_params[
"split_train_data"]:
split_name = supervised_params["train_data_split"]
split = load_env_split()[split_name]
train_envs = [env_id for env_id in train_envs if env_id in split]
print("Using " + str(len(train_envs)) + " envs from dataset split: " +
split_name)
filename = "supervised_" + setup["model"] + "_" + setup["run_name"]
# Code looks weird here because load_pytorch_model adds ".pytorch" to end of path, but
# file_exists doesn't
model_path = "tmp/" + filename + "_epoch_" + str(
supervised_params["start_epoch"])
model_path_with_extension = model_path + ".pytorch"
print("model path:", model_path_with_extension)
if supervised_params["start_epoch"] > 0:
if file_exists(model_path_with_extension):
print("THE FILE EXISTS code1")
load_pytorch_model(model, model_path)
else:
print("Couldn't continue training. Model file doesn't exist at:")
print(model_path_with_extension)
exit(-1)
# import pdb;pdb.set_trace()
## If you just want to use the pretrained model
# load_pytorch_model(model, "supervised_pvn_stage1_train_corl_pvn_stage1")
# all_train_data, all_test_data = data_io.train_data.load_supervised_data(max_envs=100)
if setup["restore_weights_name"]:
restore_pretrained_weights(model, setup["restore_weights_name"],
setup["fix_restored_weights"])
# Add a tensorboard logger to the model and trainer
tensorboard_dir = get_current_parameters(
)['Environment']['tensorboard_dir']
logger = Logger(tensorboard_dir)
model.logger = logger
if hasattr(model, "goal_good_criterion"):
print("gave logger to goal evaluator")
model.goal_good_criterion.logger = logger
trainer = Trainer(model,
epoch=supervised_params["start_epoch"],
name=setup["model"],
run_name=setup["run_name"])
trainer.logger = logger
# import pdb;pdb.set_trace()
print("Beginning training...")
best_test_loss = 1000
continue_epoch = supervised_params["start_epoch"] + 1 if supervised_params[
"start_epoch"] > 0 else 0
rng = range(0, num_epochs)
print("filename:", filename)
import pdb
pdb.set_trace()
for epoch in rng:
# import pdb;pdb.set_trace()
train_loss = trainer.train_epoch(train_data=None,
train_envs=train_envs,
eval=False)
# train_loss = trainer.train_epoch(train_data=all_train_data, train_envs=train_envs, eval=False)
trainer.model.correct_goals = 0
trainer.model.total_goals = 0
test_loss = trainer.train_epoch(train_data=None,
train_envs=dev_envs,
eval=True)
print("GOALS: ", trainer.model.correct_goals,
trainer.model.total_goals)
if test_loss < best_test_loss:
best_test_loss = test_loss
save_pytorch_model(trainer.model, filename)
print("Saved model in:", filename)
print("Epoch", epoch, "train_loss:", train_loss, "test_loss:",
test_loss)
save_pytorch_model(trainer.model,
"tmp/" + filename + "_epoch_" + str(epoch))
if hasattr(trainer.model, "save"):
trainer.model.save(epoch)
save_pretrained_weights(trainer.model, setup["run_name"])