def make_tn(
c,
transport_id,
):
""" Test """
from cargonet.visualization.gmtplot import GMTTransportPlot
from cargonet.preprocessing.datalake.retrieval import Retriever
from cargonet.preprocessing.graphs.tgraph import TransportGraph
r = Retriever()
stations = r.retrieve_stations(keep_ids=True)
# Load the transport
tg = TransportGraph(r.retrieve_transport(transport_id=transport_id)[0],
stations=stations)
GMTTransportPlot(
tg.nx_actual_route,
check=False,
filename="predictions/compare/%s.pdf" % transport_id,
node_size=17,
thickness=4,
fontsize=25,
node_border_color="black",
node_border_width=3,
node_color="white",
).plot(fit_factor=1.9)
def _receive_tg(transport_id):
from cargonet.preprocessing.graphs.tgraph import TransportGraph
from cargonet.preprocessing.datalake.retrieval import Retriever
r = Retriever()
t = r.retrieve_transport(transport_id=transport_id)[0]
s = r.retrieve_stations(keep_ids=True)
return TransportGraph(t, stations=s)
def plot_prediction(self):
import cargonet.preprocessing.tasks.debug_transport as dt
from cargonet.preprocessing.datalake.retrieval import Retriever
from cargonet.preprocessing.graphs.tgraph import TransportGraph
from cargonet.visualization.delays import DelayProgressPlot
r = Retriever()
s = r.retrieve_stations(keep_ids=True)
t = r.retrieve_transport(transport_id=transport_id)[0]
tg = TransportGraph(t, stations=s)
DelayProgressPlot(stations=s, smooth=smooth).plot_route(
tg, save=True, show_stations=True
)
def debug_transport_sections(c, transport_id):
"""Debug transport sections
"""
import cargonet.preprocessing.tasks.debug_transport as dt
from cargonet.preprocessing.datalake.retrieval import Retriever
r = Retriever()
s = r.retrieve_stations(keep_ids=True)
t_raw = r.retrieve_transport(transport_id=transport_id, raw_sections=True)
all_live = []
for sec in t_raw.get("sections"):
for l in sec.get("live"):
all_live.append(l)
dt.debug_live_sections(t_raw.get("sections"), all_live)
def plot_predicted_delay(c, transport_id):
"""Debug transport delays
"""
import cargonet.visualization.delays as d
from cargonet.preprocessing.graphs.tgraph import TransportGraph
from cargonet.preprocessing.datalake.retrieval import Retriever
r = Retriever()
s = r.retrieve_stations(keep_ids=True)
t = r.retrieve_transport(transport_id=transport_id)[0]
tg = TransportGraph(t, stations=s)
d.DelayProgressPlot(stations=s).plot_predictions(tg,
save=True,
show_stations=True)
def test_models(
c,
plot=True,
limit=32, # 8 months
device=None,
pred_seq_len=10,
horizons=None,
linear=False,
):
"""Train activeroutesv1 model
"""
import torch
import matplotlib.pyplot as plt
from datetime import timedelta
from cargonet.models.model import MLModel
from cargonet.models.activeroutesv1 import ActiveRoutesModelV1
from cargonet.models.baselines.fc2 import FCModelV2
from cargonet.models.baselines.lstm import BaselineLSTMModelV1
from cargonet.models.baselines.timeshift import BaselineTimeshiftModelV1
from cargonet.dataset.activeroutesv1 import ActiveRoutesV1
from cargonet.dataset.simulator import Simulation
from cargonet.models.normalization import Scaler
from cargonet.visualization.delays import DelayProgressPlot
from cargonet.visualization.gmtplot import GMTTransportPlot
from cargonet.preprocessing.datalake.retrieval import Retriever
from cargonet.preprocessing.graphs.tgraph import TransportGraph
from cargonet.models.utils import rec_dd
import networkx as nx
base_path = os.path.dirname(os.path.realpath(__file__))
dataset_base_path = os.path.join(base_path, "datasets")
models_base_path = os.path.join(base_path, "trained")
assert os.path.exists(dataset_base_path)
assert os.path.exists(models_base_path)
dataset_name = "active-routes-v1"
dataset_path = os.path.join(dataset_base_path, dataset_name)
simulation_dataset_name = "simulation-v1"
simulation_dataset_path = os.path.join(dataset_base_path,
simulation_dataset_name)
pred_seq_len = 10
ds_options = dict(seq_len=10, pred_seq_len=pred_seq_len)
batch_hours = 7 * 24
horizons = [0, 3, 6, 9]
use_simulation = False
if use_simulation:
dataset = Simulation(root=simulation_dataset_path,
name=simulation_dataset_name,
limit=32 * 10 * 2,
**ds_options)
else:
dataset = ActiveRoutesV1(root=dataset_path,
name=dataset_name,
limit=limit,
batch=timedelta(hours=batch_hours),
**ds_options)
model_options = dict(
node_input_dim=len(dataset.encoder.seq_route_node_fts),
edge_input_dim=len(dataset.encoder.route_edge_fts),
shuffle=False,
shuffle_after_split=False,
)
delay_stddev = None
def normalize_func(data, means, stds, **kwargs):
data.x = Scaler.zscore(data.x, mean=means["x"], std=stds["x"])
nonlocal delay_stddev
delay_stddev = stds["x"][-1]
data.temporal_edge_attr = Scaler.zscore(
data.temporal_edge_attr,
mean=means["temporal_edge_attr"],
std=stds["temporal_edge_attr"],
)
assert not torch.isnan(data.temporal_edge_attr).any()
assert not torch.isnan(data.x).any()
return data
class MockModel:
name = "Missing"
print("Creating models")
ar_model = sf_ar_model = fc2_model = lstm_model = ts_model = MockModel()
if True: # YES
ar_model = ActiveRoutesModelV1(dataset,
device=device,
use_rnn=False,
**ds_options,
**model_options)
ar_model.load()
if False:
sf_ar_model = ActiveRoutesModelV1(dataset,
device=device,
use_rnn=True,
**ds_options,
**model_options)
sf_ar_model.load()
if True: # YES
fc2_model = FCModelV2(dataset,
device=device,
**ds_options,
**model_options)
fc2_model.load()
if False:
lstm_model = BaselineLSTMModelV1(dataset,
device=device,
**ds_options,
**model_options)
lstm_model.load()
if True: # YES
ts_model = BaselineTimeshiftModelV1(dataset,
device=device,
**ds_options,
**model_options)
ts_model.load()
models = [
ar_model,
sf_ar_model,
ts_model,
fc2_model,
lstm_model,
]
models = [m for m in models if not isinstance(m, MockModel)]
print("Evaluating %d models for horizons %s" % (len(models), horizons))
trained_limit = 32
cache = "%s_norm_%d_%d" % (dataset.name, batch_hours, trained_limit)
print("fitting normalization", cache)
z_score_norm = Scaler.fit(models[0].train_data,
normalize=normalize_func,
attrs=dict(
temporal_edge_attr=1,
x=1,
y=1,
),
cache=cache)
for model in models:
if isinstance(model, BaselineTimeshiftModelV1):
continue
model.dataset.transform = z_score_norm
model.init_loaders()
print("done fitting normalization")
# DEBUG
if False:
for data in models[0].val_data:
print(data.x)
break
return
if False:
distr = []
for j, data in enumerate(ts_model.data):
if data.x is None or torch.isnan(data.x).any():
distr.append(0)
continue
distr.append(data.x.size(0))
fig, ax = plt.subplots(tight_layout=True)
ax.fill_between(range(len(distr)), 0, distr)
# ax.plot(range(len(distr)), distr) # , bins=int(len(distr) * 0.5))
plt.show()
return
long_val = [34877359, 34904458]
wrong = [34813294, 34834374]
plot_limit = 100_000
COLORS = {
# a2de96 light green
# 01a9b4 blue
ts_model.name: "#a2de96", # light green
getattr(fc2_model, "name", "FC2"): "#fc7e2f", # orange
getattr(lstm_model, "name", "LSTM"): "#fbd46d", # yellow
ar_model.name: "#f40552", # red
sf_ar_model.name: "#c3edea", # light blue
}
RENAME = {
ts_model.name: "Timeshift",
getattr(fc2_model, "name", "FC2"): "FCNN",
getattr(lstm_model, "name", "LSTM"): "LSTM",
ar_model.name: "RailSTGCNN",
sf_ar_model.name: "Stateful RailSTGCNN",
}
r = Retriever()
stations = r.retrieve_stations(keep_ids=True)
summary = MLModel.test_models(models,
pred_seq_len=pred_seq_len,
debug=long_val)
lengths = []
# print(summary)
for transport, results in summary.items():
# if not transport in long_val:
# continue
# if not transport in []:
# continue
if plot_limit < 1:
continue
# Load the transport
tg = TransportGraph(r.retrieve_transport(transport_id=transport)[0],
stations=stations)
# Sort the results by time first first
ts = sorted(results.keys())
plot_trajs = True
if plot_trajs and not use_simulation:
GMTTransportPlot(
tg.nx_actual_route,
check=False,
filename="predictions/compare/%s.pdf" % transport,
node_size=17,
thickness=4,
fontsize=25,
node_border_color="black",
node_border_width=3,
node_color="white",
).plot(fit_factor=1.9)
plot_limit -= 1
for hor in (horizons if horizons is not None else range(pred_seq_len)):
try:
route = list(nx.topological_sort(tg.nx_actual_route))
transport_preds = defaultdict(lambda: [None] * len(route))
raw_transport_preds = rec_dd()
for i, n in enumerate(route):
for t in ts:
for mdl, predictions in results[t].items():
# predictions: p_s_i+1, p_s_i+2, ..., p_s_i+n
for j, pred in enumerate(reversed(predictions)):
# pred: p_s_i+n
s, p = pred
if n == s and len(predictions) - j > hor:
# print("Found %d/%d" % (-i-1, len(route)))
# print("Found %d/%d at %d from %s" % (i+1, len(route), len(predictions) - j - 1, t))
transport_preds[mdl][i] = (s, p)
for t in ts:
for mdl, predictions in results[t].items():
# Here the prediction values are added
if len(predictions) <= hor:
continue
s, p = predictions[hor]
raw_transport_preds[mdl][s] = p
assert all([
len(route) == len(preds)
for preds in transport_preds.values()
])
timeseries = []
for mdl, preds in raw_transport_preds.items():
if mdl == "labeled":
continue
if len(preds) < 20:
continue
lengths.append(len(preds))
# times = [pt for pt, pp in preds]
# values = [pp for pt, pp in preds]
# print(times)
# print(values)
times, values = [], []
for s, p in preds.items():
for n in nx.topological_sort(tg.nx_actual_route):
if n == s:
position = tg.nx_actual_route.nodes[n].get(
"arrivalTime")
times.append(position)
values.append(p)
assert len(times) == len(values)
# Sort items and times
pls = zip(times, values)
pls = sorted(pls, key=lambda x: x[0])
times = [x[0] for x in pls]
values = [x[1] for x in pls]
timeseries.append(
dict(
label=RENAME.get(mdl, mdl),
# times = np.linspace(0, len(preds), len(preds))
times=
times, # if not linear else np.linspace(0, len(preds), len(preds)),
values=values,
index=0,
style="solid", # "dashed",
color=COLORS.get(mdl, "black"),
width=2,
), )
if len(timeseries) < 1:
continue
size, aspect = 10, 1.5
fig, ax = plt.subplots(nrows=1,
ncols=1,
figsize=(size * aspect, size))
ds_name = "ELETA" if not use_simulation else "SIM"
DelayProgressPlot(
smooth=False, stations=stations,
fontsize=25).plot_predictions(
tg=tg,
fig=fig,
ax=ax,
predictions=timeseries,
save=True,
markers=False,
show_stations=True,
has_time_axis=not linear,
filename="predictions/compare/%s_%s_prediction_%s.pdf"
% (ds_name, transport, str(hor)),
)
plt.close()
except Exception as e:
print(e)
if len(lengths) > 0:
lens = torch.FloatTensor(lengths)
print("mean: %f min: %f max: %f" %
(lens.mean(), lens.min(), lens.max()))
print("Delay stddev is", delay_stddev)