def main(
out_dir=None, preds_dir=None, data_dir=None, metric_names=None,
plot_output=None, results_file=None, sweep_param_name=None):
if metric_names is None:
metric_names = ('accuracy', 'edit_score', 'overlap_score')
preds_dir = os.path.expanduser(preds_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
logger.info(f"Writing to: {out_dir}")
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
if os.path.exists(results_file):
os.remove(results_file)
else:
results_file = os.path.expanduser(results_file)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name, data_dir):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'trial={seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
assembly_vocab = joblib.load(os.path.join(data_dir, 'assembly-vocab.pkl'))
trial_ids = utils.getUniqueIds(preds_dir, prefix='trial=')
pred_seqs = loadAll(trial_ids, 'pred-label-seq.pkl', preds_dir)
true_seqs = loadAll(trial_ids, 'true-label-seq.pkl', preds_dir)
input_seqs = loadAll(trial_ids, 'score-seq.pkl', preds_dir)
action_vocab = []
action_batch = []
assembly_batch = []
for i, trial_id in enumerate(trial_ids):
logger.info(f"VIDEO {trial_id}:")
pred_assembly_index_seq = pred_seqs[i]
true_assembly_index_seq = true_seqs[i]
pred_action_segs, pred_action_index_seq = actionsFromAssemblies(
pred_assembly_index_seq, assembly_vocab, action_vocab
)
lib_assembly.writeAssemblies(
os.path.join(out_data_dir, f'trial={trial_id}_pred-actions.txt'),
pred_action_segs
)
true_action_segs, true_action_index_seq = actionsFromAssemblies(
true_assembly_index_seq, assembly_vocab, action_vocab
)
lib_assembly.writeAssemblies(
os.path.join(out_data_dir, f'trial={trial_id}_true-actions.txt'),
true_action_segs
)
metric_dict = {}
for name in metric_names:
key = f"{name}_action"
value = getattr(LCTM.metrics, name)(pred_action_index_seq, true_action_index_seq) / 100
metric_dict[key] = value
logger.info(f" {key}: {value * 100:.1f}%")
key = f"{name}_assembly"
value = getattr(LCTM.metrics, name)(
pred_assembly_index_seq, true_assembly_index_seq
) / 100
metric_dict[key] = value
logger.info(f" {key}: {value * 100:.1f}%")
utils.writeResults(results_file, metric_dict, sweep_param_name, {})
input_seq = input_seqs[i]
action_batch.append((pred_action_index_seq, input_seq, true_action_index_seq, trial_id))
assembly_batch.append(
(pred_assembly_index_seq, input_seq, true_assembly_index_seq, trial_id)
)
lib_assembly.writeAssemblies(os.path.join(out_data_dir, 'action-vocab.txt'), action_vocab)
label_names = ('true', 'pred')
for preds, inputs, gt_labels, seq_id in action_batch:
fn = os.path.join(fig_dir, f"trial={seq_id}_model-io_action.png")
utils.plot_array(None, (gt_labels, preds), label_names, fn=fn, labels_together=True)
label_names = ('true', 'pred')
for preds, inputs, gt_labels, seq_id in assembly_batch:
fn = os.path.join(fig_dir, f"trial={seq_id}_model-io_assembly.png")
utils.plot_array(None, (gt_labels, preds), label_names, fn=fn, labels_together=True)
def main(
out_dir=None, data_dir=None, model_name=None,
model_params={}, cv_params={}, train_params={}, viz_params={},
plot_predictions=None, results_file=None, sweep_param_name=None):
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, f'results.csv')
write_mode = 'w'
else:
results_file = os.path.expanduser(results_file)
write_mode = 'a'
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name, data_dir):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'trial={seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
# Load data
trial_ids = utils.getUniqueIds(data_dir, prefix='trial=')
feature_seqs = loadAll(trial_ids, 'feature-seq.pkl', data_dir)
label_seqs = loadAll(trial_ids, 'label-seq.pkl', data_dir)
# Define cross-validation folds
dataset_size = len(trial_ids)
cv_folds = utils.makeDataSplits(dataset_size, **cv_params)
def getSplit(split_idxs):
split_data = tuple(
tuple(s[i] for i in split_idxs)
for s in (feature_seqs, label_seqs, trial_ids)
)
return split_data
for cv_index, cv_splits in enumerate(cv_folds):
train_data, val_data, test_data = tuple(map(getSplit, cv_splits))
train_feats, train_labels, train_ids = train_data
test_feats, test_labels, test_ids = test_data
val_feats, val_labels, val_ids = val_data
logger.info(
f'CV fold {cv_index + 1} / {len(cv_folds)}: {len(trial_ids)} total '
f'({len(train_ids)} train, {len(val_ids)} val, {len(test_ids)} test)'
)
input_dim = train_set.num_obsv_dims
output_dim = train_set.num_label_types
elif model_name == 'dummy':
# FIXME
model = None
else:
raise AssertionError()
metric_str = ' '.join(str(m) for m in metric_dict.values())
logger.info('[TST] ' + metric_str)
d = {k: v.value for k, v in metric_dict.items()}
utils.writeResults(
results_file, d, sweep_param_name, model_params,
write_mode=write_mode
)
if plot_predictions:
io_fig_dir = os.path.join(fig_dir, 'model-io')
if not os.path.exists(io_fig_dir):
os.makedirs(io_fig_dir)
label_names = ('gt', 'pred')
preds, scores, inputs, gt_labels, ids = zip(*test_io_history)
for batch in test_io_history:
batch = map(lambda x: x.cpu().numpy(), batch)
for preds, _, inputs, gt_labels, seq_id in zip(*batch):
fn = os.path.join(io_fig_dir, f"trial={seq_id}_model-io.png")
utils.plot_array(inputs, (gt_labels, preds), label_names, fn=fn)
def saveTrialData(pred_seq, score_seq, feat_seq, label_seq, trial_id):
saveVariable(pred_seq, f'trial={trial_id}_pred-label-seq')
saveVariable(score_seq, f'trial={trial_id}_score-seq')
saveVariable(label_seq, f'trial={trial_id}_true-label-seq')
for batch in test_io_history:
batch = map(lambda x: x.cpu().numpy(), batch)
for io in zip(*batch):
saveTrialData(*io)
saveVariable(train_ids, f'cvfold={cv_index}_train-ids')
saveVariable(test_ids, f'cvfold={cv_index}_test-ids')
saveVariable(val_ids, f'cvfold={cv_index}_val-ids')
saveVariable(train_epoch_log, f'cvfold={cv_index}_{model_name}-train-epoch-log')
saveVariable(val_epoch_log, f'cvfold={cv_index}_{model_name}-val-epoch-log')
train_fig_dir = os.path.join(fig_dir, 'train-plots')
if not os.path.exists(train_fig_dir):
os.makedirs(train_fig_dir)
def main(out_dir=None,
data_dir=None,
event_scores_dir=None,
action_scores_dir=None,
part_scores_dir=None,
as_atomic_events=False,
only_fold=None,
plot_io=None,
prefix='seq=',
results_file=None,
sweep_param_name=None,
model_params={},
cv_params={}):
data_dir = os.path.expanduser(data_dir)
event_scores_dir = os.path.expanduser(event_scores_dir)
action_scores_dir = os.path.expanduser(action_scores_dir)
part_scores_dir = os.path.expanduser(part_scores_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
loader = DataLoader(
{
'event': os.path.join(data_dir, 'event-dataset'),
'action': os.path.join(data_dir, 'action-dataset'),
'part': os.path.join(data_dir, 'part-dataset')
}, {
'event': event_scores_dir,
'action': action_scores_dir,
'part': part_scores_dir
},
prefix=prefix)
logger.info(
f"Loaded ids for {len(loader.seq_ids)} sequences from {data_dir}")
all_metrics = collections.defaultdict(list)
if as_atomic_events:
map_df = pd.read_csv(os.path.join(data_dir, 'labels',
'event-vocab.csv'),
index_col=False,
keep_default_na=False)
else:
map_df = pd.DataFrame(
[[f"{action}({part})", action] +
[part == n for n in loader.vocabs['part'] if n]
for action in loader.vocabs['action']
for part in loader.vocabs['part']],
columns=['event', 'action'] +
[f"{n}_active" for n in loader.vocabs['part'] if n])
event_vocab = map_df['event'].to_list()
utils.saveVariable(event_vocab, 'vocab', out_data_dir)
# Define cross-validation folds
cv_folds = utils.makeDataSplits(len(loader.seq_ids), **cv_params)
utils.saveVariable(cv_folds, 'cv-folds', out_data_dir)
for cv_index, cv_fold in enumerate(cv_folds):
if only_fold is not None and cv_index != only_fold:
continue
train_indices, val_indices, test_indices = cv_fold
logger.info(
f"CV FOLD {cv_index + 1} / {len(cv_folds)}: "
f"{len(train_indices)} train, {len(val_indices)} val, {len(test_indices)} test"
)
action_part_counts = sum(
count_action_part_bigrams(loader.vocabs['action'],
loader.vocabs['part'], la, lp)
for la, lp in loader._getLabels(train_indices,
label_name=['action', 'part'],
load_from_data=True))
model = AttributeClassifier(action_part_counts)
plot_action_part_bigrams(
loader.vocabs['action'], loader.vocabs['part'],
(model.action_part_counts > 0).astype(int),
os.path.join(fig_dir,
f'cvfold={cv_index}_action-part-coocurrence.png'))
for i in test_indices:
seq_id = loader.seq_ids[i]
logger.info(f" Processing sequence {seq_id}...")
action_score_seq, part_score_seq, true_action_seq, true_part_seq = loader[
seq_id]
event_score_seq = model.forward(action_score_seq, part_score_seq)
pred_action_seq, pred_part_seq = model.predict(event_score_seq)
true_event_seq = ap_to_event(true_action_seq, true_part_seq,
map_df, loader.vocabs)
pred_event_seq = ap_to_event(pred_action_seq, pred_part_seq,
map_df, loader.vocabs)
metric_dict = {}
metric_dict = eval_metrics(pred_event_seq,
true_event_seq,
name_prefix='Event ',
append_to=metric_dict)
metric_dict = eval_metrics(pred_action_seq,
true_action_seq,
name_prefix='Action ',
append_to=metric_dict)
metric_dict = eval_metrics(pred_part_seq,
true_part_seq,
name_prefix='Part ',
append_to=metric_dict)
for name, value in metric_dict.items():
logger.info(f" {name}: {value * 100:.2f}%")
all_metrics[name].append(value)
seq_id_str = f"seq={seq_id}"
utils.saveVariable(event_score_seq, f'{seq_id_str}_score-seq',
out_data_dir)
utils.saveVariable(true_event_seq, f'{seq_id_str}_true-label-seq',
out_data_dir)
utils.saveVariable(pred_event_seq, f'{seq_id_str}_pred-label-seq',
out_data_dir)
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
if plot_io:
utils.plot_array(
event_score_seq.reshape(event_score_seq.shape[0], -1).T,
(true_event_seq, pred_event_seq), ('true', 'pred'),
fn=os.path.join(fig_dir, f"seq={seq_id:03d}_event.png"))
utils.plot_array(action_score_seq.T,
(true_action_seq, pred_action_seq),
('true', 'pred'),
fn=os.path.join(
fig_dir, f"seq={seq_id:03d}_action.png"))
utils.plot_array(part_score_seq.T,
(true_part_seq, pred_part_seq),
('true', 'pred'),
fn=os.path.join(fig_dir,
f"seq={seq_id:03d}_part.png"))
def __test(fig_dir, num_classes=2, num_samples=10, min_dur=2, max_dur=4):
eps_str = 'ε'
bos_str = '<BOS>'
eos_str = '<EOS>'
dur_internal_str = 'I'
dur_final_str = 'F'
aux_symbols = (eps_str, bos_str, eos_str)
dur_vocab = (dur_internal_str, dur_final_str)
sample_vocab = tuple(i for i in range(num_samples))
dur_vocab = tuple(i for i in range(1, max_dur + 1))
class_vocab = tuple(i for i in range(num_classes))
class_dur_vocab = tuple((c, s) for c in class_vocab for s in dur_vocab)
def to_strings(vocab):
return tuple(map(str, vocab))
def to_integerizer(vocab):
return {item: i for i, item in enumerate(vocab)}
sample_vocab_str = to_strings(sample_vocab)
dur_vocab_str = to_strings(dur_vocab)
class_vocab_str = to_strings(class_vocab)
class_dur_vocab_str = to_strings(class_dur_vocab)
# sample_integerizer = to_integerizer(sample_vocab)
dur_integerizer = to_integerizer(dur_vocab)
class_integerizer = to_integerizer(class_vocab)
# class_dur_integerizer = to_integerizer(class_dur_vocab)
sample_str_integerizer = to_integerizer(sample_vocab_str)
class_str_integerizer = to_integerizer(class_vocab_str)
class_dur_str_integerizer = to_integerizer(class_dur_vocab_str)
def get_parts(class_dur_key):
c, s = class_dur_vocab[class_dur_str_integerizer[class_dur_key]]
c_str = class_vocab_str[class_integerizer[c]]
s_str = dur_vocab_str[dur_integerizer[s]]
return c_str, s_str
class_dur_to_str = {get_parts(name): name for name in class_dur_vocab_str}
sample_symbols = libfst.makeSymbolTable(aux_symbols + sample_vocab,
prepend_epsilon=False)
# dur_symbols = libfst.makeSymbolTable(aux_symbols + dur_vocab, prepend_epsilon=False)
class_symbols = libfst.makeSymbolTable(aux_symbols + class_vocab,
prepend_epsilon=False)
# dur_symbols = libfst.makeSymbolTable(aux_symbols + dur_vocab, prepend_epsilon=False)
class_dur_symbols = libfst.makeSymbolTable(aux_symbols + class_dur_vocab,
prepend_epsilon=False)
obs_scores = np.zeros((num_samples, num_classes))
dur_scores = np.array([[0 if d >= min_dur else np.inf for d in dur_vocab]
for c in class_vocab],
dtype=float)
def score_transition(c_prev, s_prev, c_cur, s_cur):
if c_prev != c_cur:
if s_prev == dur_final_str and s_cur == dur_internal_str:
score = 0
else:
score = np.inf
else:
if s_prev == dur_internal_str and s_cur == dur_final_str:
score = 0
elif s_prev == dur_internal_str and s_cur == dur_internal_str:
score = 0
else:
score = np.inf
return score
transition_scores = np.array([[
score_transition(c_prev, s_prev, c_cur, s_cur)
for (c_cur, s_cur) in class_dur_vocab
] for (c_prev, s_prev) in class_dur_vocab],
dtype=float)
init_scores = np.array([
0 if c == 0 and s == dur_internal_str else np.inf
for (c, s) in class_dur_vocab
],
dtype=float)
final_scores = np.array([
0 if c == 1 and s == dur_final_str else np.inf
for (c, s) in class_dur_vocab
],
dtype=float)
def score_arc_state(class_dur_key, class_key):
c, s = class_dur_vocab[class_dur_str_integerizer[class_dur_key]]
c_prime = class_str_integerizer[class_key]
if c == c_prime:
score = 0
else:
score = np.inf
return score
class_dur_to_class_scores = np.array([[
score_arc_state(class_dur_key, class_key)
for class_key in class_vocab_str
] for class_dur_key in class_dur_vocab_str],
dtype=float)
def log_normalize(arr, axis=1):
denom = -scipy.special.logsumexp(-arr, axis=axis, keepdims=True)
return arr - denom
obs_scores = log_normalize(obs_scores)
dur_scores = log_normalize(dur_scores)
transition_scores = log_normalize(transition_scores)
init_scores = log_normalize(init_scores, axis=None)
final_scores = log_normalize(final_scores, axis=None)
obs_fst = add_endpoints(fromArray(obs_scores,
sample_vocab_str,
class_vocab_str,
input_symbols=sample_symbols,
output_symbols=class_symbols,
arc_type='standard'),
bos_str=bos_str,
eos_str=eos_str).arcsort(sort_type='ilabel')
dur_fst = add_endpoints(make_duration_fst(
dur_scores,
class_vocab_str,
class_dur_to_str,
dur_internal_str=dur_internal_str,
dur_final_str=dur_final_str,
input_symbols=class_symbols,
output_symbols=class_dur_symbols,
allow_self_transitions=False,
arc_type='standard',
),
bos_str=bos_str,
eos_str=eos_str).arcsort(sort_type='ilabel')
transition_fst = fromTransitions(
transition_scores,
class_dur_vocab_str,
class_dur_vocab_str,
init_weights=init_scores,
final_weights=final_scores,
input_symbols=class_dur_symbols,
output_symbols=class_dur_symbols).arcsort(sort_type='ilabel')
class_dur_to_class_fst = single_state_transducer(
class_dur_to_class_scores,
class_dur_vocab_str,
class_vocab_str,
input_symbols=class_dur_symbols,
output_symbols=class_symbols,
arc_type='standard').arcsort(sort_type='ilabel')
seq_model = openfst.compose(dur_fst, transition_fst)
decode_lattice = openfst.compose(obs_fst, seq_model).rmepsilon()
# Result is in the log semiring (ie weights are negative log probs)
arc_scores = libfst.fstArcGradient(decode_lattice).arcsort(
sort_type='ilabel')
best_arcs = openfst.shortestpath(decode_lattice).arcsort(
sort_type='ilabel')
state_scores = openfst.compose(
arc_scores, openfst.arcmap(class_dur_to_class_fst, map_type='to_log'))
best_states = openfst.compose(best_arcs, class_dur_to_class_fst)
state_scores_arr, weight_type = toArray(state_scores,
sample_str_integerizer,
class_str_integerizer)
draw_fst(os.path.join(fig_dir, 'obs_fst'),
obs_fst,
vertical=True,
width=50,
height=50,
portrait=True)
draw_fst(os.path.join(fig_dir, 'dur_fst'),
dur_fst,
vertical=True,
width=50,
height=50,
portrait=True)
draw_fst(os.path.join(fig_dir, 'transition_fst'),
transition_fst,
vertical=True,
width=50,
height=50,
portrait=True)
draw_fst(os.path.join(fig_dir, 'seq_model'),
seq_model,
vertical=True,
width=50,
height=50,
portrait=True)
draw_fst(os.path.join(fig_dir, 'decode_lattice'),
decode_lattice,
vertical=True,
width=50,
height=50,
portrait=True)
draw_fst(os.path.join(fig_dir, 'state_scores'),
state_scores,
vertical=True,
width=50,
height=50,
portrait=True)
draw_fst(os.path.join(fig_dir, 'best_states'),
best_states,
vertical=True,
width=50,
height=50,
portrait=True)
utils.plot_array(obs_scores.T, (state_scores_arr.T, ), ('-logprobs', ),
fn=os.path.join(fig_dir, "test_io.png"))
def main(out_dir=None,
data_dir=None,
segs_dir=None,
scores_dir=None,
vocab_dir=None,
label_type='edges',
gpu_dev_id=None,
start_from=None,
stop_at=None,
num_disp_imgs=None,
results_file=None,
sweep_param_name=None,
model_params={},
cv_params={}):
data_dir = os.path.expanduser(data_dir)
segs_dir = os.path.expanduser(segs_dir)
scores_dir = os.path.expanduser(scores_dir)
vocab_dir = os.path.expanduser(vocab_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
io_dir_images = os.path.join(fig_dir, 'model-io_images')
if not os.path.exists(io_dir_images):
os.makedirs(io_dir_images)
io_dir_plots = os.path.join(fig_dir, 'model-io_plots')
if not os.path.exists(io_dir_plots):
os.makedirs(io_dir_plots)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
seq_ids = utils.getUniqueIds(scores_dir,
prefix='trial=',
suffix='score-seq.*',
to_array=True)
logger.info(
f"Loaded scores for {len(seq_ids)} sequences from {scores_dir}")
link_vocab = {}
joint_vocab = {}
joint_type_vocab = {}
vocab, parts_vocab, part_labels = load_vocab(link_vocab, joint_vocab,
joint_type_vocab, vocab_dir)
pred_vocab = [] # FIXME
if label_type == 'assembly':
logger.info("Converting assemblies -> edges")
state_pred_seqs = tuple(
utils.loadVariable(f"trial={seq_id}_pred-label-seq", scores_dir)
for seq_id in seq_ids)
state_true_seqs = tuple(
utils.loadVariable(f"trial={seq_id}_true-label-seq", scores_dir)
for seq_id in seq_ids)
edge_pred_seqs = tuple(part_labels[seq] for seq in state_pred_seqs)
edge_true_seqs = tuple(part_labels[seq] for seq in state_true_seqs)
elif label_type == 'edge':
logger.info("Converting edges -> assemblies (will take a few minutes)")
edge_pred_seqs = tuple(
utils.loadVariable(f"trial={seq_id}_pred-label-seq", scores_dir)
for seq_id in seq_ids)
edge_true_seqs = tuple(
utils.loadVariable(f"trial={seq_id}_true-label-seq", scores_dir)
for seq_id in seq_ids)
state_pred_seqs = tuple(
edges_to_assemblies(seq, pred_vocab, parts_vocab, part_labels)
for seq in edge_pred_seqs)
state_true_seqs = tuple(
edges_to_assemblies(seq, vocab, parts_vocab, part_labels)
for seq in edge_true_seqs)
device = torchutils.selectDevice(gpu_dev_id)
dataset = sim2real.LabeledConnectionDataset(
utils.loadVariable('parts-vocab', vocab_dir),
utils.loadVariable('part-labels', vocab_dir),
utils.loadVariable('vocab', vocab_dir),
device=device)
all_metrics = collections.defaultdict(list)
# Define cross-validation folds
cv_folds = utils.makeDataSplits(len(seq_ids), **cv_params)
utils.saveVariable(cv_folds, 'cv-folds', out_data_dir)
for cv_index, cv_fold in enumerate(cv_folds):
train_indices, val_indices, test_indices = cv_fold
logger.info(
f"CV FOLD {cv_index + 1} / {len(cv_folds)}: "
f"{len(train_indices)} train, {len(val_indices)} val, {len(test_indices)} test"
)
train_states = np.hstack(
tuple(state_true_seqs[i] for i in (train_indices)))
train_edges = part_labels[train_states]
# state_train_vocab = np.unique(train_states)
# edge_train_vocab = part_labels[state_train_vocab]
train_freq_bigram, train_freq_unigram = edge_joint_freqs(train_edges)
# state_probs = utils.makeHistogram(len(vocab), train_states, normalize=True)
test_states = np.hstack(
tuple(state_true_seqs[i] for i in (test_indices)))
test_edges = part_labels[test_states]
# state_test_vocab = np.unique(test_states)
# edge_test_vocab = part_labels[state_test_vocab]
test_freq_bigram, test_freq_unigram = edge_joint_freqs(test_edges)
f, axes = plt.subplots(1, 2)
axes[0].matshow(train_freq_bigram)
axes[0].set_title('Train')
axes[1].matshow(test_freq_bigram)
axes[1].set_title('Test')
plt.tight_layout()
plt.savefig(
os.path.join(fig_dir, f"edge-freqs-bigram_cvfold={cv_index}.png"))
f, axis = plt.subplots(1)
axis.stem(train_freq_unigram,
label='Train',
linefmt='C0-',
markerfmt='C0o')
axis.stem(test_freq_unigram,
label='Test',
linefmt='C1--',
markerfmt='C1o')
plt.legend()
plt.tight_layout()
plt.savefig(
os.path.join(fig_dir, f"edge-freqs-unigram_cvfold={cv_index}.png"))
for i in test_indices:
seq_id = seq_ids[i]
logger.info(f" Processing sequence {seq_id}...")
trial_prefix = f"trial={seq_id}"
# I include the '.' to differentiate between 'rgb-frame-seq' and
# 'rgb-frame-seq-before-first-touch'
# rgb_seq = utils.loadVariable(f"{trial_prefix}_rgb-frame-seq.", data_dir)
# seg_seq = utils.loadVariable(f"{trial_prefix}_seg-labels-seq", segs_dir)
score_seq = utils.loadVariable(f"{trial_prefix}_score-seq",
scores_dir)
# if score_seq.shape[0] != rgb_seq.shape[0]:
# err_str = f"scores shape {score_seq.shape} != data shape {rgb_seq.shape}"
# raise AssertionError(err_str)
edge_pred_seq = edge_pred_seqs[i]
edge_true_seq = edge_true_seqs[i]
state_pred_seq = state_pred_seqs[i]
state_true_seq = state_true_seqs[i]
num_types = np.unique(state_pred_seq).shape[0]
num_samples = state_pred_seq.shape[0]
num_total = len(pred_vocab)
logger.info(
f" {num_types} assemblies predicted ({num_total} total); "
f"{num_samples} samples")
# edge_freq_bigram, edge_freq_unigram = edge_joint_freqs(edge_true_seq)
# dist_shift = np.linalg.norm(train_freq_unigram - edge_freq_unigram)
metric_dict = {
# 'State OOV rate': oov_rate_state(state_true_seq, state_train_vocab),
# 'Edge OOV rate': oov_rate_edges(edge_true_seq, edge_train_vocab),
# 'State avg prob, true': state_probs[state_true_seq].mean(),
# 'State avg prob, pred': state_probs[state_pred_seq].mean(),
# 'Edge distribution shift': dist_shift
}
metric_dict = eval_edge_metrics(edge_pred_seq,
edge_true_seq,
append_to=metric_dict)
metric_dict = eval_state_metrics(state_pred_seq,
state_true_seq,
append_to=metric_dict)
for name, value in metric_dict.items():
logger.info(f" {name}: {value * 100:.2f}%")
all_metrics[name].append(value)
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
if num_disp_imgs is not None:
pred_images = tuple(
render(dataset, vocab[seg_label])
for seg_label in utils.computeSegments(state_pred_seq)[0])
imageprocessing.displayImages(
*pred_images,
file_path=os.path.join(
io_dir_images,
f"seq={seq_id:03d}_pred-assemblies.png"),
num_rows=None,
num_cols=5)
true_images = tuple(
render(dataset, vocab[seg_label])
for seg_label in utils.computeSegments(state_true_seq)[0])
imageprocessing.displayImages(
*true_images,
file_path=os.path.join(
io_dir_images,
f"seq={seq_id:03d}_true-assemblies.png"),
num_rows=None,
num_cols=5)
utils.plot_array(score_seq.T,
(edge_true_seq.T, edge_pred_seq.T),
('true', 'pred'),
fn=os.path.join(io_dir_plots,
f"seq={seq_id:03d}.png"))
def main(out_dir=None,
video_data_dir=None,
features_dir=None,
activity_labels_dir=None,
gt_keyframes_dir=None,
use_gt_activity_labels=False):
out_dir = os.path.expanduser(out_dir)
video_data_dir = os.path.expanduser(video_data_dir)
features_dir = os.path.expanduser(features_dir)
activity_labels_dir = os.path.expanduser(activity_labels_dir)
if gt_keyframes_dir is not None:
gt_keyframes_dir = os.path.expanduser(gt_keyframes_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
def loadFromDir(var_name, dir_name):
return joblib.load(os.path.join(dir_name, f"{var_name}.pkl"))
def saveToWorkingDir(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f"{var_name}.pkl"))
trial_ids = utils.getUniqueIds(features_dir,
prefix='trial=',
suffix='.pkl')
seq_lens = []
for seq_idx, trial_id in enumerate(trial_ids):
logger.info(
f"Processing video {seq_idx + 1} / {len(trial_ids)} (trial {trial_id})"
)
logger.info(f" Loading data...")
feature_seq = loadFromDir(f"trial={trial_id}_feature-seq",
features_dir)
raw_labels = loadFromDir(f"trial-{trial_id}_action-seq",
video_data_dir)
timestamp_seq = loadFromDir(
f"trial-{trial_id}_rgb-frame-timestamp-seq", video_data_dir)
if use_gt_activity_labels:
activity_label_fn = f"trial={trial_id}_true-label-seq"
else:
activity_label_fn = f"trial={trial_id}_pred-label-seq"
activity_labels = loadFromDir(activity_label_fn, activity_labels_dir)
action_labels = makeActionLabels(raw_labels,
seq_len=feature_seq.shape[0])
if timestamp_seq.shape[0] != feature_seq.shape[0]:
logger.warning(
f"Video dimensions don't match: "
f"{feature_seq.shape} scores, {timestamp_seq.shape} timestamps"
)
continue
# trim sequences
is_activity = activity_labels == 1
if not is_activity.any():
logger.warning(f"No activity detected: skipping trial")
continue
action_labels = action_labels[is_activity, ...]
timestamp_seq = timestamp_seq[is_activity, ...]
feature_seq = feature_seq[is_activity, ...]
if feature_seq.shape[0] != action_labels.shape[0]:
err_str = (
"Data dimensions don't match: "
f"{feature_seq.shape} features, {action_labels.shape} labels")
raise AssertionError(err_str)
seq_lens.append(feature_seq.shape[0])
logger.info(f" Saving output...")
trial_str = f"trial={trial_id}"
saveToWorkingDir(timestamp_seq, f'{trial_str}_timestamp-seq')
saveToWorkingDir(feature_seq, f'{trial_str}_feature-seq')
saveToWorkingDir(action_labels, f'{trial_str}_label-seq')
fn = os.path.join(fig_dir, f"trial={trial_id}.png")
utils.plot_array(feature_seq.T, (action_labels, ), ('action', ), fn=fn)
logger.info(f"min seq len: {min(seq_lens)} max seq len: {max(seq_lens)}")
def main(
out_dir=None, modalities=['rgb', 'imu'], gpu_dev_id=None, plot_io=None,
rgb_data_dir=None, rgb_attributes_dir=None, imu_data_dir=None, imu_attributes_dir=None):
out_dir = os.path.expanduser(out_dir)
rgb_data_dir = os.path.expanduser(rgb_data_dir)
rgb_attributes_dir = os.path.expanduser(rgb_attributes_dir)
imu_data_dir = os.path.expanduser(imu_data_dir)
imu_attributes_dir = os.path.expanduser(imu_attributes_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
# Load data
if modalities == ['rgb']:
trial_ids = utils.getUniqueIds(rgb_data_dir, prefix='trial=', to_array=True)
logger.info(f"Processing {len(trial_ids)} videos")
else:
rgb_trial_ids = utils.getUniqueIds(rgb_data_dir, prefix='trial=', to_array=True)
imu_trial_ids = utils.getUniqueIds(imu_data_dir, prefix='trial=', to_array=True)
trial_ids = np.array(sorted(set(rgb_trial_ids.tolist()) & set(imu_trial_ids.tolist())))
logger.info(
f"Processing {len(trial_ids)} videos common to "
f"RGB ({len(rgb_trial_ids)} total) and IMU ({len(imu_trial_ids)} total)"
)
device = torchutils.selectDevice(gpu_dev_id)
dataset = FusionDataset(
trial_ids, rgb_attributes_dir, rgb_data_dir, imu_attributes_dir, imu_data_dir,
device=device, modalities=modalities,
)
utils.saveMetadata(dataset.metadata, out_data_dir)
utils.saveVariable(dataset.vocab, 'vocab', out_data_dir)
for i, trial_id in enumerate(trial_ids):
logger.info(f"Processing sequence {trial_id}...")
true_label_seq = dataset.loadTargets(trial_id)
attribute_feats = dataset.loadInputs(trial_id)
# (Process the samples here if we need to)
attribute_feats = attribute_feats.cpu().numpy()
true_label_seq = true_label_seq.cpu().numpy()
trial_prefix = f"trial={trial_id}"
utils.saveVariable(attribute_feats, f'{trial_prefix}_feature-seq', out_data_dir)
utils.saveVariable(true_label_seq, f'{trial_prefix}_label-seq', out_data_dir)
if plot_io:
fn = os.path.join(fig_dir, f'{trial_prefix}.png')
utils.plot_array(
attribute_feats.T,
(true_label_seq,),
('gt',),
fn=fn
)
def main(out_dir=None,
data_dir=None,
model_name=None,
predict_mode='classify',
gpu_dev_id=None,
batch_size=None,
learning_rate=None,
independent_signals=None,
active_only=None,
output_dim_from_vocab=False,
prefix='trial=',
feature_fn_format='feature-seq.pkl',
label_fn_format='label_seq.pkl',
dataset_params={},
model_params={},
cv_params={},
train_params={},
viz_params={},
metric_names=['Loss', 'Accuracy', 'Precision', 'Recall', 'F1'],
plot_predictions=None,
results_file=None,
sweep_param_name=None):
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def saveVariable(var, var_name, to_dir=out_data_dir):
return utils.saveVariable(var, var_name, to_dir)
# Load data
device = torchutils.selectDevice(gpu_dev_id)
trial_ids = utils.getUniqueIds(data_dir,
prefix=prefix,
suffix=feature_fn_format,
to_array=True)
dataset = utils.CvDataset(
trial_ids,
data_dir,
prefix=prefix,
feature_fn_format=feature_fn_format,
label_fn_format=label_fn_format,
)
utils.saveMetadata(dataset.metadata, out_data_dir)
utils.saveVariable(dataset.vocab, 'vocab', out_data_dir)
# Define cross-validation folds
cv_folds = utils.makeDataSplits(len(trial_ids), **cv_params)
utils.saveVariable(cv_folds, 'cv-folds', out_data_dir)
if predict_mode == 'binary multiclass':
# criterion = torchutils.BootstrappedCriterion(
# 0.25, base_criterion=torch.nn.functional.binary_cross_entropy_with_logits,
# )
criterion = torch.nn.BCEWithLogitsLoss()
labels_dtype = torch.float
elif predict_mode == 'multiclass':
criterion = torch.nn.CrossEntropyLoss()
labels_dtype = torch.long
elif predict_mode == 'classify':
criterion = torch.nn.CrossEntropyLoss()
labels_dtype = torch.long
else:
raise AssertionError()
def make_dataset(feats, labels, ids, shuffle=True):
dataset = torchutils.SequenceDataset(feats,
labels,
device=device,
labels_dtype=labels_dtype,
seq_ids=ids,
**dataset_params)
loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
shuffle=True)
return dataset, loader
for cv_index, cv_fold in enumerate(cv_folds):
train_data, val_data, test_data = dataset.getFold(cv_fold)
if independent_signals:
train_data = splitSeqs(*train_data, active_only=active_only)
val_data = splitSeqs(*val_data, active_only=active_only)
test_data = splitSeqs(*test_data, active_only=False)
train_set, train_loader = make_dataset(*train_data, shuffle=True)
test_set, test_loader = make_dataset(*test_data, shuffle=False)
val_set, val_loader = make_dataset(*val_data, shuffle=True)
logger.info(
f'CV fold {cv_index + 1} / {len(cv_folds)}: {len(dataset.trial_ids)} total '
f'({len(train_set)} train, {len(val_set)} val, {len(test_set)} test)'
)
logger.info(f'{train_set.num_label_types} unique labels in train set; '
f'vocab size is {len(dataset.vocab)}')
input_dim = train_set.num_obsv_dims
output_dim = train_set.num_label_types
if output_dim_from_vocab:
output_dim = len(dataset.vocab)
if model_name == 'linear':
model = torchutils.LinearClassifier(
input_dim, output_dim, **model_params).to(device=device)
elif model_name == 'conv':
model = ConvClassifier(input_dim, output_dim,
**model_params).to(device=device)
elif model_name == 'TCN':
if predict_mode == 'multiclass':
num_multiclass = train_set[0][1].shape[-1]
output_dim = max([
train_set.num_label_types, test_set.num_label_types,
val_set.num_label_types
])
else:
num_multiclass = None
model = TcnClassifier(input_dim,
output_dim,
num_multiclass=num_multiclass,
**model_params).to(device=device)
elif model_name == 'LSTM':
if predict_mode == 'multiclass':
num_multiclass = train_set[0][1].shape[-1]
output_dim = max([
train_set.num_label_types, test_set.num_label_types,
val_set.num_label_types
])
else:
num_multiclass = None
model = LstmClassifier(input_dim,
output_dim,
num_multiclass=num_multiclass,
**model_params).to(device=device)
else:
raise AssertionError()
optimizer_ft = torch.optim.Adam(model.parameters(),
lr=learning_rate,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=0,
amsgrad=False)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer_ft,
step_size=1,
gamma=1.00)
train_epoch_log = collections.defaultdict(list)
val_epoch_log = collections.defaultdict(list)
metric_dict = {name: metrics.makeMetric(name) for name in metric_names}
model, last_model_wts = torchutils.trainModel(
model,
criterion,
optimizer_ft,
lr_scheduler,
train_loader,
val_loader,
device=device,
metrics=metric_dict,
train_epoch_log=train_epoch_log,
val_epoch_log=val_epoch_log,
**train_params)
# Test model
metric_dict = {name: metrics.makeMetric(name) for name in metric_names}
test_io_history = torchutils.predictSamples(
model.to(device=device),
test_loader,
criterion=criterion,
device=device,
metrics=metric_dict,
data_labeled=True,
update_model=False,
seq_as_batch=train_params['seq_as_batch'],
return_io_history=True)
if independent_signals:
test_io_history = tuple(joinSeqs(test_io_history))
logger.info('[TST] ' +
' '.join(str(m) for m in metric_dict.values()))
utils.writeResults(results_file,
{k: v.value
for k, v in metric_dict.items()}, sweep_param_name,
model_params)
if plot_predictions:
io_fig_dir = os.path.join(fig_dir, 'model-io')
if not os.path.exists(io_fig_dir):
os.makedirs(io_fig_dir)
label_names = ('gt', 'pred')
preds, scores, inputs, gt_labels, ids = zip(*test_io_history)
for batch in test_io_history:
batch = tuple(
x.cpu().numpy() if isinstance(x, torch.Tensor) else x
for x in batch)
for preds, _, inputs, gt_labels, seq_id in zip(*batch):
fn = os.path.join(io_fig_dir,
f"{prefix}{seq_id}_model-io.png")
utils.plot_array(inputs, (gt_labels.T, preds.T),
label_names,
fn=fn)
for batch in test_io_history:
batch = tuple(x.cpu().numpy() if isinstance(x, torch.Tensor) else x
for x in batch)
for pred_seq, score_seq, feat_seq, label_seq, trial_id in zip(
*batch):
saveVariable(pred_seq, f'{prefix}{trial_id}_pred-label-seq')
saveVariable(score_seq, f'{prefix}{trial_id}_score-seq')
saveVariable(label_seq, f'{prefix}{trial_id}_true-label-seq')
saveVariable(model, f'cvfold={cv_index}_{model_name}-best')
train_fig_dir = os.path.join(fig_dir, 'train-plots')
if not os.path.exists(train_fig_dir):
os.makedirs(train_fig_dir)
if train_epoch_log:
torchutils.plotEpochLog(train_epoch_log,
subfig_size=(10, 2.5),
title='Training performance',
fn=os.path.join(
train_fig_dir,
f'cvfold={cv_index}_train-plot.png'))
if val_epoch_log:
torchutils.plotEpochLog(val_epoch_log,
subfig_size=(10, 2.5),
title='Heldout performance',
fn=os.path.join(
train_fig_dir,
f'cvfold={cv_index}_val-plot.png'))
def main(out_dir=None,
data_dir=None,
assembly_data_dir=None,
scores_dir=None,
event_attr_fn=None,
connection_attr_fn=None,
assembly_attr_fn=None,
only_fold=None,
plot_io=None,
prefix='seq=',
stop_after=None,
background_action='',
model_params={},
cv_params={},
stride=None,
results_file=None,
sweep_param_name=None):
data_dir = os.path.expanduser(data_dir)
assembly_data_dir = os.path.expanduser(assembly_data_dir)
scores_dir = os.path.expanduser(scores_dir)
event_attr_fn = os.path.expanduser(event_attr_fn)
connection_attr_fn = os.path.expanduser(connection_attr_fn)
assembly_attr_fn = os.path.expanduser(assembly_attr_fn)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
misc_dir = os.path.join(out_dir, 'misc')
if not os.path.exists(misc_dir):
os.makedirs(misc_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
seq_ids = utils.getUniqueIds(data_dir,
prefix=prefix,
suffix='labels.*',
to_array=True)
dataset = utils.FeaturelessCvDataset(seq_ids,
data_dir,
prefix=prefix,
label_fn_format='labels')
logger.info(
f"Loaded scores for {len(seq_ids)} sequences from {scores_dir}")
# Define cross-validation folds
cv_folds = utils.makeDataSplits(len(seq_ids), **cv_params)
utils.saveVariable(cv_folds, 'cv-folds', out_data_dir)
# Load event, connection attributes
assembly_vocab = tuple(
tuple(sorted(tuple(sorted(joint)) for joint in a))
for a in utils.loadVariable('vocab', assembly_data_dir))
(*probs, assembly_transition_probs), vocabs = loadPartInfo(
event_attr_fn,
connection_attr_fn,
assembly_attr_fn,
background_action=background_action,
assembly_vocab=assembly_vocab)
event_assembly_scores = event_to_assembly_scores(*probs, vocabs)
assembly_transition_scores = np.log(assembly_transition_probs)
viz_transition_probs(os.path.join(fig_dir, 'action-transitions'),
np.exp(event_assembly_scores), vocabs['event_vocab'])
write_transition_probs(os.path.join(misc_dir, 'action-transitions'),
np.exp(event_assembly_scores),
vocabs['event_vocab'], vocabs['assembly_vocab'])
for cv_index, cv_fold in enumerate(cv_folds):
if only_fold is not None and cv_index != only_fold:
continue
train_indices, val_indices, test_indices = cv_fold
logger.info(
f"CV FOLD {cv_index + 1} / {len(cv_folds)}: "
f"{len(train_indices)} train, {len(val_indices)} val, {len(test_indices)} test"
)
train_data, val_data, test_data = dataset.getFold(cv_fold)
cv_str = f'cvfold={cv_index}'
class_priors, event_dur_probs = count_priors(train_data[0],
len(dataset.vocab),
stride=stride,
approx_upto=0.95,
support_only=True)
event_dur_scores = np.log(event_dur_probs)
event_dur_scores = np.zeros_like(event_dur_scores)
scores = (event_dur_scores, event_assembly_scores,
assembly_transition_scores)
model = decode.AssemblyActionRecognizer(scores, vocabs, model_params)
viz_priors(os.path.join(fig_dir, f'{cv_str}_priors'), class_priors,
event_dur_probs)
model.write_fsts(os.path.join(misc_dir, f'{cv_str}_fsts'))
model.save_vocabs(os.path.join(out_data_dir, f'{cv_str}_model-vocabs'))
for i, (_, seq_id) in enumerate(zip(*test_data)):
if stop_after is not None and i >= stop_after:
break
trial_prefix = f"{prefix}{seq_id}"
if model_params['return_label'] == 'input':
true_seq = utils.loadVariable(f"{trial_prefix}_true-label-seq",
scores_dir)
elif model_params['return_label'] == 'output':
try:
true_seq = utils.loadVariable(f"{trial_prefix}_label-seq",
assembly_data_dir)
true_seq = true_seq[::stride]
except AssertionError:
# logger.info(f' Skipping sequence {seq_id}: {e}')
continue
logger.info(f" Processing sequence {seq_id}...")
event_score_seq = utils.loadVariable(f"{trial_prefix}_score-seq",
scores_dir)
if event_score_seq.shape[0] != true_seq.shape[0]:
err_str = (f'Event scores shape {event_score_seq.shape} '
f'!= labels shape {true_seq.shape}')
raise AssertionError(err_str)
# FIXME: the serialized variables are probs, not log-probs
# event_score_seq = suppress_nonmax(event_score_seq)
# event_score_seq = np.log(event_score_seq)
decode_score_seq = model.forward(event_score_seq)
pred_seq = model.predict(decode_score_seq)
metric_dict = eval_metrics(pred_seq, true_seq)
for name, value in metric_dict.items():
logger.info(f" {name}: {value * 100:.2f}%")
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
utils.saveVariable(decode_score_seq, f'{trial_prefix}_score-seq',
out_data_dir)
utils.saveVariable(pred_seq, f'{trial_prefix}_pred-label-seq',
out_data_dir)
utils.saveVariable(true_seq, f'{trial_prefix}_true-label-seq',
out_data_dir)
if plot_io:
utils.plot_array(event_score_seq.T, (pred_seq.T, true_seq.T),
('pred', 'true'),
fn=os.path.join(fig_dir,
f"seq={seq_id:03d}.png"))
write_labels(
os.path.join(misc_dir, f"seq={seq_id:03d}_pred-seq.txt"),
pred_seq, model.output_vocab.as_raw())
write_labels(
os.path.join(misc_dir, f"seq={seq_id:03d}_true-seq.txt"),
true_seq, model.output_vocab.as_raw())
def main(out_dir=None,
data_dir=None,
actions_dir=None,
parts_dir=None,
events_dir=None,
edges_dir=None,
prefix='seq=',
feature_fn_format='score-seq',
label_fn_format='true-label-seq',
stop_after=None,
only_fold=None,
plot_io=None,
dataset_params={},
model_params={},
cv_params={}):
data_dir = os.path.expanduser(data_dir)
actions_dir = os.path.expanduser(actions_dir)
parts_dir = os.path.expanduser(parts_dir)
events_dir = os.path.expanduser(events_dir)
edges_dir = os.path.expanduser(edges_dir)
out_dir = os.path.expanduser(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
# vocab = utils.loadVariable(
# 'assembly-action-vocab',
# os.path.join(data_dir, 'event-dataset')
# )
# vocab = [BlockAssembly()] + list(abs(x) for x in vocab)
dataset = FusionDataset(
actions_dir,
parts_dir,
events_dir,
edges_dir,
prefix=prefix,
**dataset_params,
# vocab=vocab,
)
utils.saveMetadata(dataset.metadata, out_data_dir)
utils.saveVariable(dataset.vocab, 'vocab', out_data_dir)
seq_ids = dataset.trial_ids
logger.info(f"Loaded scores for {len(seq_ids)} sequences from {data_dir}")
# Define cross-validation folds
# cv_folds = utils.makeDataSplits(len(seq_ids), **cv_params)
# utils.saveVariable(cv_folds, 'cv-folds', out_data_dir)
for i, seq_id in enumerate(seq_ids):
try:
labels = dataset.loadTargets(seq_id)
features = dataset.loadInputs(seq_id)
except AssertionError as e:
logger.warning(f'Skipping sequence {seq_id}: {e}')
continue
logger.info(f"Processing sequence {seq_id}")
if labels.shape[0] != features.shape[0]:
message = f'Label shape {labels.shape} != feature shape {features.shape}'
raise AssertionError(message)
seq_prefix = f"seq={seq_id}"
utils.saveVariable(features, f'{seq_prefix}_feature-seq', out_data_dir)
utils.saveVariable(labels, f'{seq_prefix}_label-seq', out_data_dir)
if plot_io:
fn = os.path.join(fig_dir, f'{seq_prefix}.png')
utils.plot_array(features.T, (labels.T, ), ('gt', ), fn=fn)
def main(out_dir=None,
data_dir=None,
scores_dir=None,
frames_dir=None,
vocab_from_scores_dir=None,
only_fold=None,
plot_io=None,
prefix='seq=',
results_file=None,
sweep_param_name=None,
model_params={},
cv_params={}):
data_dir = os.path.expanduser(data_dir)
scores_dir = os.path.expanduser(scores_dir)
frames_dir = os.path.expanduser(frames_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
io_dir_images = os.path.join(fig_dir, 'model-io_images')
if not os.path.exists(io_dir_images):
os.makedirs(io_dir_images)
io_dir_plots = os.path.join(fig_dir, 'model-io_plots')
if not os.path.exists(io_dir_plots):
os.makedirs(io_dir_plots)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
seq_ids = utils.getUniqueIds(scores_dir,
prefix=prefix,
suffix='pred-label-seq.*',
to_array=True)
logger.info(
f"Loaded scores for {len(seq_ids)} sequences from {scores_dir}")
if vocab_from_scores_dir:
vocab = utils.loadVariable('vocab', scores_dir)
else:
vocab = utils.loadVariable('vocab', data_dir)
all_metrics = collections.defaultdict(list)
# Define cross-validation folds
cv_folds = utils.makeDataSplits(len(seq_ids), **cv_params)
utils.saveVariable(cv_folds, 'cv-folds', out_data_dir)
all_pred_seqs = []
all_true_seqs = []
for cv_index, cv_fold in enumerate(cv_folds):
if only_fold is not None and cv_index != only_fold:
continue
train_indices, val_indices, test_indices = cv_fold
logger.info(
f"CV FOLD {cv_index + 1} / {len(cv_folds)}: "
f"{len(train_indices)} train, {len(val_indices)} val, {len(test_indices)} test"
)
for i in test_indices:
seq_id = seq_ids[i]
logger.info(f" Processing sequence {seq_id}...")
trial_prefix = f"{prefix}{seq_id}"
score_seq = utils.loadVariable(f"{trial_prefix}_score-seq",
scores_dir)
pred_seq = utils.loadVariable(f"{trial_prefix}_pred-label-seq",
scores_dir)
true_seq = utils.loadVariable(f"{trial_prefix}_true-label-seq",
scores_dir)
metric_dict = eval_metrics(pred_seq, true_seq)
for name, value in metric_dict.items():
logger.info(f" {name}: {value * 100:.2f}%")
all_metrics[name].append(value)
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
all_pred_seqs.append(pred_seq)
all_true_seqs.append(true_seq)
if plot_io:
utils.plot_array(score_seq.T, (true_seq, pred_seq),
('true', 'pred'),
fn=os.path.join(io_dir_plots,
f"seq={seq_id:03d}.png"))
if False:
confusions = metrics.confusionMatrix(all_pred_seqs, all_true_seqs,
len(vocab))
utils.saveVariable(confusions, "confusions", out_data_dir)
per_class_acc, class_counts = metrics.perClassAcc(confusions,
return_counts=True)
class_preds = confusions.sum(axis=1)
logger.info(f"MACRO ACC: {np.nanmean(per_class_acc) * 100:.2f}%")
metrics.plotConfusions(os.path.join(fig_dir, 'confusions.png'),
confusions, vocab)
metrics.plotPerClassAcc(os.path.join(fig_dir,
'per-class-results.png'), vocab,
per_class_acc, class_preds, class_counts)
def main(
out_dir=None, hand_detections_dir=None, labels_dir=None,
plot_output=None, results_file=None, sweep_param_name=None):
hand_detections_dir = os.path.expanduser(hand_detections_dir)
labels_dir = os.path.expanduser(labels_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
logger.info(f"Reading from: {hand_detections_dir}")
logger.info(f"Writing to: {out_dir}")
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
# write_mode = 'w'
else:
results_file = os.path.expanduser(results_file)
# write_mode = 'a'
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
part_names, part_names_to_idxs, part_idxs_to_bins = airplanecorpus.loadParts()
for i, fn in enumerate(glob.glob(os.path.join(hand_detections_dir, '*.txt'))):
video_id = utils.stripExtension(fn).split('.handsdetections')[0]
logger.info(f"Processing video {video_id}")
hand_detections = airplanecorpus.loadHandDetections(
video_id, dir_name=hand_detections_dir, unflatten=True
)
hand_detections = hand_detections.reshape(hand_detections.shape[0], -1)
mean_detection = np.nanmean(hand_detections)
hand_detections[np.isnan(hand_detections)] = mean_detection
action_labels = airplanecorpus.loadLabels(
video_id, dir_name=labels_dir,
part_names_to_idxs=part_names_to_idxs
)
bin_labels = makeBinLabels(action_labels, part_idxs_to_bins, hand_detections.shape[0])
fig_fn = os.path.join(fig_dir, f"{video_id}.png")
utils.plot_array(hand_detections.T, (bin_labels,), ('bin',), fn=fig_fn)
video_id = video_id.replace('_', '-')
saveVariable(hand_detections, f'trial={video_id}_feature-seq')
saveVariable(bin_labels, f'trial={video_id}_label-seq')
def main(
out_dir=None, data_dir=None, results_file=None, cv_file=None,
take_log=False, col_format=None, win_params={}, slowfast_csv_params={}):
out_dir = os.path.expanduser(out_dir)
data_dir = os.path.expanduser(data_dir)
results_file = os.path.expanduser(results_file)
cv_file = os.path.expanduser(cv_file)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
vocab = utils.loadVariable('vocab', data_dir)
metadata = utils.loadMetadata(data_dir)
slowfast_labels = pd.read_csv(
cv_file, keep_default_na=False, index_col=0,
**slowfast_csv_params
)
seg_ids = slowfast_labels.index.to_numpy()
vid_names = slowfast_labels['video_name'].unique().tolist()
metadata['seq_id'] = metadata.index
vid_ids = metadata.set_index('dir_name').loc[vid_names].set_index('seq_id').index
metadata = metadata.drop('seq_id', axis=1)
with open(results_file, 'rb') as file_:
model_probs, gt_labels = pickle.load(file_)
model_probs = model_probs.numpy()
gt_labels = gt_labels.numpy()
if len(model_probs) != len(seg_ids):
err_str = f"{len(model_probs)} segment scores != {slowfast_labels.shape[0]} CSV rows"
raise AssertionError(err_str)
logger.info(f"Loaded {len(seg_ids)} segments, {len(vid_ids)} videos")
for vid_id, vid_name in zip(vid_ids, vid_names):
matches_video = (slowfast_labels['video_name'] == vid_name).to_numpy()
win_labels = gt_labels[matches_video]
win_probs = model_probs[matches_video, :]
if win_labels.shape == win_probs.shape:
win_preds = (win_probs > 0.5).astype(int)
else:
win_preds = win_probs.argmax(axis=1)
if take_log:
win_probs = np.log(win_probs)
seq_id_str = f"seq={vid_id}"
utils.saveVariable(win_probs, f'{seq_id_str}_score-seq', out_data_dir)
utils.saveVariable(win_labels, f'{seq_id_str}_true-label-seq', out_data_dir)
utils.saveVariable(win_preds, f'{seq_id_str}_pred-label-seq', out_data_dir)
utils.plot_array(
win_probs.T, (win_labels.T, win_preds.T), ('true', 'pred'),
tick_names=vocab,
fn=os.path.join(fig_dir, f"{seq_id_str}.png"),
subplot_width=12, subplot_height=5
)
utils.saveVariable(vocab, 'vocab', out_data_dir)
utils.saveMetadata(metadata, out_data_dir)
def main(out_dir=None,
bin_scores_dir=None,
action_labels_dir=None,
subsample_period=None):
bin_scores_dir = os.path.expanduser(bin_scores_dir)
action_labels_dir = os.path.expanduser(action_labels_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
logger.info(f"Writing to: {out_dir}")
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name, data_dir):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'trial={seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
part_names, part_names_to_idxs, part_idxs_to_bins = airplanecorpus.loadParts(
)
trial_ids = utils.getUniqueIds(bin_scores_dir, prefix='trial=')
bin_score_seqs = loadAll(trial_ids, 'score-seq.pkl', bin_scores_dir)
transition_to_index = {}
assembly_label_seqs = []
for i, trial_id in enumerate(trial_ids):
video_id = utils.stripExtension(trial_id).replace('-', '_')
action_labels = airplanecorpus.loadLabels(
video_id,
dir_name=action_labels_dir,
part_names_to_idxs=part_names_to_idxs)
# Make assembly labels from action labels
assembly_labels = makeAssemblyLabels(action_labels,
part_names,
bin_score_seqs[i].shape[0],
vocabulary=transition_to_index)
assembly_label_seqs.append(assembly_labels)
transition_scores, initial_scores, final_scores = su.smoothCounts(
*su.countSeqs(assembly_label_seqs), as_scores=True)
fn = os.path.join(fig_dir, "transitions.png")
su.plot_transitions(fn, transition_scores, initial_scores, final_scores)
num_items = len(transition_to_index)
transition_vocabulary = {v: k for k, v in transition_to_index.items()}
transition_vocabulary = tuple(transition_vocabulary[i]
for i in range(num_items))
saveVariable(transition_vocabulary, 'transition-vocabulary')
for i, trial_id in enumerate(trial_ids):
video_id = utils.stripExtension(trial_id).replace('-', '_')
bin_score_seq = bin_score_seqs[i]
assembly_labels = assembly_label_seqs[i]
assembly_scores = makeAssemblyScores(
bin_score_seq,
part_names_to_idxs,
part_idxs_to_bins,
transition_vocabulary,
)
assembly_scores = assembly_scores[::subsample_period, :]
assembly_labels = assembly_labels[::subsample_period]
fig_fn = os.path.join(fig_dir, f"{video_id}.png")
utils.plot_array(assembly_scores.T, (assembly_labels, ),
('assembly', ),
fn=fig_fn)
video_id = video_id.replace('_', '-')
saveVariable(assembly_scores, f'trial={video_id}_score-seq')
saveVariable(assembly_scores, f'trial={video_id}_feature-seq')
saveVariable(assembly_labels, f'trial={video_id}_label-seq')
def main(out_dir=None,
data_dir=None,
cv_data_dir=None,
score_dirs=[],
fusion_method='sum',
prune_imu=None,
standardize=None,
decode=None,
plot_predictions=None,
results_file=None,
sweep_param_name=None,
gpu_dev_id=None,
model_params={},
cv_params={},
train_params={},
viz_params={}):
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
score_dirs = tuple(map(os.path.expanduser, score_dirs))
if cv_data_dir is not None:
cv_data_dir = os.path.expanduser(cv_data_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name, data_dir):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'trial={seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
device = torchutils.selectDevice(gpu_dev_id)
# Load data
dir_trial_ids = tuple(
set(utils.getUniqueIds(d, prefix='trial=', to_array=True))
for d in score_dirs)
dir_trial_ids += (set(
utils.getUniqueIds(data_dir, prefix='trial=', to_array=True)), )
trial_ids = np.array(list(sorted(set.intersection(*dir_trial_ids))))
for dir_name, t_ids in zip(score_dirs + (data_dir, ), dir_trial_ids):
logger.info(f"{len(t_ids)} trial ids from {dir_name}:")
logger.info(f" {t_ids}")
logger.info(f"{len(trial_ids)} trials in intersection: {trial_ids}")
assembly_seqs = loadAll(trial_ids, 'assembly-seq.pkl', data_dir)
feature_seqs = tuple(
loadAll(trial_ids, 'data-scores.pkl', d) for d in score_dirs)
feature_seqs = tuple(zip(*feature_seqs))
# Combine feature seqs
include_indices = []
for i, seq_feats in enumerate(feature_seqs):
feat_shapes = tuple(f.shape for f in seq_feats)
include_seq = all(f == feat_shapes[0] for f in feat_shapes)
if include_seq:
include_indices.append(i)
else:
warn_str = (
f'Excluding trial {trial_ids[i]} with mismatched feature shapes: '
f'{feat_shapes}')
logger.warning(warn_str)
trial_ids = trial_ids[include_indices]
assembly_seqs = tuple(assembly_seqs[i] for i in include_indices)
feature_seqs = tuple(feature_seqs[i] for i in include_indices)
feature_seqs = tuple(np.stack(f) for f in feature_seqs)
# Define cross-validation folds
if cv_data_dir is None:
dataset_size = len(trial_ids)
cv_folds = utils.makeDataSplits(dataset_size, **cv_params)
cv_fold_trial_ids = tuple(
tuple(map(lambda x: trial_ids[x], splits)) for splits in cv_folds)
else:
fn = os.path.join(cv_data_dir, 'cv-fold-trial-ids.pkl')
cv_fold_trial_ids = joblib.load(fn)
def getSplit(split_idxs):
split_data = tuple(
tuple(s[i] for i in split_idxs)
for s in (feature_seqs, assembly_seqs, trial_ids))
return split_data
gt_scores = []
all_scores = []
num_keyframes_total = 0
num_rgb_errors_total = 0
num_correctable_errors_total = 0
num_oov_total = 0
num_changed_total = 0
for cv_index, (train_ids, test_ids) in enumerate(cv_fold_trial_ids):
try:
test_idxs = np.array(
[trial_ids.tolist().index(i) for i in test_ids])
include_indices.append(cv_index)
except ValueError:
logger.info(
f" Skipping fold {cv_index}: missing test data {test_ids}")
logger.info(f'CV fold {cv_index + 1}: {len(trial_ids)} total '
f'({len(train_ids)} train, {len(test_ids)} test)')
# TRAIN PHASE
if cv_data_dir is None:
train_idxs = np.array([trial_ids.index(i) for i in train_ids])
train_assembly_seqs = tuple(assembly_seqs[i] for i in train_idxs)
train_assemblies = []
for seq in train_assembly_seqs:
list(
labels.gen_eq_classes(seq,
train_assemblies,
equivalent=None))
model = None
else:
fn = f'cvfold={cv_index}_train-assemblies.pkl'
train_assemblies = joblib.load(os.path.join(cv_data_dir, fn))
train_idxs = [
i for i in range(len(trial_ids)) if i not in test_idxs
]
fn = f'cvfold={cv_index}_model.pkl'
model = joblib.load(os.path.join(cv_data_dir, fn))
train_features, train_assembly_seqs, train_ids = getSplit(train_idxs)
if False:
train_labels = tuple(
np.array(
list(
labels.gen_eq_classes(assembly_seq,
train_assemblies,
equivalent=None)), )
for assembly_seq in train_assembly_seqs)
train_set = torchutils.SequenceDataset(train_features,
train_labels,
seq_ids=train_ids,
device=device)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=1,
shuffle=True)
train_epoch_log = collections.defaultdict(list)
# val_epoch_log = collections.defaultdict(list)
metric_dict = {
'Avg Loss': metrics.AverageLoss(),
'Accuracy': metrics.Accuracy()
}
criterion = torch.nn.CrossEntropyLoss()
optimizer_ft = torch.optim.Adam(model.parameters(),
lr=1e-3,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=0,
amsgrad=False)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer_ft,
step_size=1,
gamma=1.00)
model = FusionClassifier(num_sources=train_features[0].shape[0])
model, last_model_wts = torchutils.trainModel(
model,
criterion,
optimizer_ft,
lr_scheduler,
train_loader, # val_loader,
device=device,
metrics=metric_dict,
train_epoch_log=train_epoch_log,
# val_epoch_log=val_epoch_log,
**train_params)
torchutils.plotEpochLog(train_epoch_log,
subfig_size=(10, 2.5),
title='Training performance',
fn=os.path.join(
fig_dir,
f'cvfold={cv_index}_train-plot.png'))
test_assemblies = train_assemblies.copy()
for feature_seq, gt_assembly_seq, trial_id in zip(
*getSplit(test_idxs)):
gt_seq = np.array(
list(
labels.gen_eq_classes(gt_assembly_seq,
test_assemblies,
equivalent=None)))
if plot_predictions:
assembly_fig_dir = os.path.join(fig_dir, 'assembly-imgs')
if not os.path.exists(assembly_fig_dir):
os.makedirs(assembly_fig_dir)
for i, assembly in enumerate(test_assemblies):
assembly.draw(assembly_fig_dir, i)
# TEST PHASE
accuracies = []
for feature_seq, gt_assembly_seq, trial_id in zip(
*getSplit(test_idxs)):
gt_seq = np.array(
list(
labels.gen_eq_classes(gt_assembly_seq,
test_assemblies,
equivalent=None)))
num_labels = gt_seq.shape[0]
num_features = feature_seq.shape[-1]
if num_labels != num_features:
err_str = (f"Skipping trial {trial_id}: "
f"{num_labels} labels != {num_features} features")
logger.info(err_str)
continue
# Ignore OOV states in ground-truth
sample_idxs = np.arange(feature_seq.shape[-1])
score_idxs = gt_seq[gt_seq < feature_seq.shape[1]]
sample_idxs = sample_idxs[gt_seq < feature_seq.shape[1]]
gt_scores.append(feature_seq[:, score_idxs, sample_idxs])
all_scores.append(feature_seq.reshape(feature_seq.shape[0], -1))
if fusion_method == 'sum':
score_seq = feature_seq.sum(axis=0)
elif fusion_method == 'rgb_only':
score_seq = feature_seq[1]
elif fusion_method == 'imu_only':
score_seq = feature_seq[0]
else:
raise NotImplementedError()
if not decode:
model = None
if model is None:
pred_seq = score_seq.argmax(axis=0)
elif isinstance(model, torch.nn.Module):
inputs = torch.tensor(feature_seq[None, ...],
dtype=torch.float,
device=device)
outputs = model.forward(inputs)
pred_seq = model.predict(outputs)[0].cpu().numpy()
else:
dummy_samples = np.arange(score_seq.shape[1])
pred_seq, _, _, _ = model.viterbi(dummy_samples,
log_likelihoods=score_seq,
ml_decode=(not decode))
pred_assemblies = [train_assemblies[i] for i in pred_seq]
gt_assemblies = [test_assemblies[i] for i in gt_seq]
acc = metrics.accuracy_upto(pred_assemblies,
gt_assemblies,
equivalence=None)
accuracies.append(acc)
rgb_pred_seq = feature_seq[1].argmax(axis=0)
num_changed = np.sum(rgb_pred_seq != pred_seq)
rgb_is_wrong = rgb_pred_seq != gt_seq
num_rgb_errors = np.sum(rgb_is_wrong)
imu_scores = feature_seq[0]
imu_scores_gt = np.array([
imu_scores[s_idx,
t] if s_idx < imu_scores.shape[0] else -np.inf
for t, s_idx in enumerate(gt_seq)
])
imu_scores_rgb = np.array([
imu_scores[s_idx,
t] if s_idx < imu_scores.shape[0] else -np.inf
for t, s_idx in enumerate(rgb_pred_seq)
])
# imu_scores_gt = imu_scores[gt_seq, range(len(rgb_pred_seq))]
best_imu_scores = imu_scores.max(axis=0)
imu_is_right = imu_scores_gt >= best_imu_scores
rgb_pred_score_is_lower = imu_scores_gt > imu_scores_rgb
is_correctable_error = rgb_is_wrong & imu_is_right & rgb_pred_score_is_lower
num_correctable_errors = np.sum(is_correctable_error)
prop_correctable = num_correctable_errors / num_rgb_errors
num_oov = np.sum(gt_seq >= len(train_assemblies))
num_states = len(gt_seq)
num_keyframes_total += num_states
num_rgb_errors_total += num_rgb_errors
num_correctable_errors_total += num_correctable_errors
num_oov_total += num_oov
num_changed_total += num_changed
logger.info(f" trial {trial_id}: {num_states} keyframes")
logger.info(f" accuracy (fused): {acc * 100:.1f}%")
logger.info(
f" {num_oov} OOV states ({num_oov / num_states * 100:.1f}%)"
)
logger.info(
f" {num_rgb_errors} RGB errors; "
f"{num_correctable_errors} correctable from IMU ({prop_correctable * 100:.1f}%)"
)
saveVariable(score_seq, f'trial={trial_id}_data-scores')
saveVariable(pred_assemblies,
f'trial={trial_id}_pred-assembly-seq')
saveVariable(gt_assemblies, f'trial={trial_id}_gt-assembly-seq')
if plot_predictions:
io_figs_dir = os.path.join(fig_dir, 'system-io')
if not os.path.exists(io_figs_dir):
os.makedirs(io_figs_dir)
fn = os.path.join(io_figs_dir, f'trial={trial_id:03}.png')
utils.plot_array(feature_seq, (gt_seq, pred_seq, score_seq),
('gt', 'pred', 'scores'),
fn=fn)
score_figs_dir = os.path.join(fig_dir, 'modality-scores')
if not os.path.exists(score_figs_dir):
os.makedirs(score_figs_dir)
plot_scores(feature_seq,
k=25,
fn=os.path.join(score_figs_dir,
f"trial={trial_id:03}.png"))
paths_dir = os.path.join(fig_dir, 'path-imgs')
if not os.path.exists(paths_dir):
os.makedirs(paths_dir)
assemblystats.drawPath(pred_seq, trial_id,
f"trial={trial_id}_pred-seq", paths_dir,
assembly_fig_dir)
assemblystats.drawPath(gt_seq, trial_id,
f"trial={trial_id}_gt-seq", paths_dir,
assembly_fig_dir)
label_seqs = (gt_seq, ) + tuple(
scores.argmax(axis=0) for scores in feature_seq)
label_seqs = np.row_stack(label_seqs)
k = 10
for i, scores in enumerate(feature_seq):
label_score_seqs = tuple(
np.array([
scores[s_idx,
t] if s_idx < scores.shape[0] else -np.inf
for t, s_idx in enumerate(label_seq)
]) for label_seq in label_seqs)
label_score_seqs = np.row_stack(label_score_seqs)
drawPaths(label_seqs,
f"trial={trial_id}_pred-scores_modality={i}",
paths_dir,
assembly_fig_dir,
path_scores=label_score_seqs)
topk_seq = (-scores).argsort(axis=0)[:k, :]
path_scores = np.column_stack(
tuple(scores[idxs, i]
for i, idxs in enumerate(topk_seq.T)))
drawPaths(topk_seq,
f"trial={trial_id}_topk_modality={i}",
paths_dir,
assembly_fig_dir,
path_scores=path_scores)
label_score_seqs = tuple(
np.array([
score_seq[s_idx,
t] if s_idx < score_seq.shape[0] else -np.inf
for t, s_idx in enumerate(label_seq)
]) for label_seq in label_seqs)
label_score_seqs = np.row_stack(label_score_seqs)
drawPaths(label_seqs,
f"trial={trial_id}_pred-scores_fused",
paths_dir,
assembly_fig_dir,
path_scores=label_score_seqs)
topk_seq = (-score_seq).argsort(axis=0)[:k, :]
path_scores = np.column_stack(
tuple(score_seq[idxs, i]
for i, idxs in enumerate(topk_seq.T)))
drawPaths(topk_seq,
f"trial={trial_id}_topk_fused",
paths_dir,
assembly_fig_dir,
path_scores=path_scores)
if accuracies:
fold_accuracy = float(np.array(accuracies).mean())
# logger.info(f' acc: {fold_accuracy * 100:.1f}%')
metric_dict = {'Accuracy': fold_accuracy}
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
num_unexplained_errors = num_rgb_errors_total - (
num_oov_total + num_correctable_errors_total)
prop_correctable = num_correctable_errors_total / num_rgb_errors_total
prop_oov = num_oov_total / num_rgb_errors_total
prop_unexplained = num_unexplained_errors / num_rgb_errors_total
prop_changed = num_changed_total / num_keyframes_total
logger.info("PERFORMANCE ANALYSIS")
logger.info(
f" {num_rgb_errors_total} / {num_keyframes_total} "
f"RGB errors ({num_rgb_errors_total / num_keyframes_total * 100:.1f}%)"
)
logger.info(f" {num_oov_total} / {num_rgb_errors_total} "
f"RGB errors are OOV ({prop_oov * 100:.1f}%)")
logger.info(f" {num_correctable_errors_total} / {num_rgb_errors_total} "
f"RGB errors are correctable ({prop_correctable * 100:.1f}%)")
logger.info(f" {num_unexplained_errors} / {num_rgb_errors_total} "
f"RGB errors are unexplained ({prop_unexplained * 100:.1f}%)")
logger.info(
f" {num_changed_total} / {num_keyframes_total} "
f"Predictions changed after fusion ({prop_changed * 100:.1f}%)")
gt_scores = np.hstack(tuple(gt_scores))
plot_hists(np.exp(gt_scores),
fn=os.path.join(fig_dir, "score-hists_gt.png"))
all_scores = np.hstack(tuple(all_scores))
plot_hists(np.exp(all_scores),
fn=os.path.join(fig_dir, "score-hists_all.png"))
def main(out_dir=None,
rgb_data_dir=None,
rgb_attributes_dir=None,
rgb_vocab_dir=None,
imu_data_dir=None,
imu_attributes_dir=None,
modalities=['rgb', 'imu'],
gpu_dev_id=None,
plot_predictions=None,
results_file=None,
sweep_param_name=None,
model_params={},
cv_params={},
train_params={},
viz_params={}):
out_dir = os.path.expanduser(out_dir)
rgb_data_dir = os.path.expanduser(rgb_data_dir)
rgb_attributes_dir = os.path.expanduser(rgb_attributes_dir)
rgb_vocab_dir = os.path.expanduser(rgb_vocab_dir)
imu_data_dir = os.path.expanduser(imu_data_dir)
imu_attributes_dir = os.path.expanduser(imu_attributes_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def saveVariable(var, var_name, to_dir=out_data_dir):
utils.saveVariable(var, var_name, to_dir)
# Load data
if modalities == ['rgb']:
trial_ids = utils.getUniqueIds(rgb_data_dir,
prefix='trial=',
to_array=True)
logger.info(f"Processing {len(trial_ids)} videos")
else:
rgb_trial_ids = utils.getUniqueIds(rgb_data_dir,
prefix='trial=',
to_array=True)
imu_trial_ids = utils.getUniqueIds(imu_data_dir,
prefix='trial=',
to_array=True)
trial_ids = np.array(
sorted(set(rgb_trial_ids.tolist()) & set(imu_trial_ids.tolist())))
logger.info(
f"Processing {len(trial_ids)} videos common to "
f"RGB ({len(rgb_trial_ids)} total) and IMU ({len(imu_trial_ids)} total)"
)
device = torchutils.selectDevice(gpu_dev_id)
dataset = FusionDataset(trial_ids,
rgb_attributes_dir,
rgb_data_dir,
imu_attributes_dir,
imu_data_dir,
device=device,
modalities=modalities)
utils.saveMetadata(dataset.metadata, out_data_dir)
saveVariable(dataset.vocab, 'vocab')
# parts_vocab = loadVariable('parts-vocab')
edge_labels = {
'rgb':
utils.loadVariable('part-labels', rgb_vocab_dir),
'imu':
np.stack([
labels.inSameComponent(a, lower_tri_only=True)
for a in dataset.vocab
])
}
# edge_labels = revise_edge_labels(edge_labels, input_seqs)
attribute_labels = tuple(edge_labels[name] for name in modalities)
logger.info('Making transition probs...')
transition_probs = make_transition_scores(dataset.vocab)
saveVariable(transition_probs, 'transition-probs')
model = AttributeModel(*attribute_labels, device=device)
if plot_predictions:
figsize = (12, 3)
fig, axis = plt.subplots(1, figsize=figsize)
axis.imshow(edge_labels['rgb'].T, interpolation='none', aspect='auto')
plt.savefig(os.path.join(fig_dir, "edge-labels.png"))
plt.close()
for i, trial_id in enumerate(trial_ids):
logger.info(f"Processing sequence {trial_id}...")
trial_prefix = f"trial={trial_id}"
true_label_seq = dataset.loadTargets(trial_id)
attribute_feats = dataset.loadInputs(trial_id)
score_seq = model(attribute_feats)
pred_label_seq = model.predict(score_seq)
attribute_feats = attribute_feats.cpu().numpy()
score_seq = score_seq.cpu().numpy()
true_label_seq = true_label_seq.cpu().numpy()
pred_label_seq = pred_label_seq.cpu().numpy()
saveVariable(score_seq.T, f'{trial_prefix}_score-seq')
saveVariable(true_label_seq.T, f'{trial_prefix}_label-seq')
if plot_predictions:
fn = os.path.join(fig_dir, f'{trial_prefix}.png')
utils.plot_array(attribute_feats.T,
(true_label_seq, pred_label_seq, score_seq),
('gt', 'pred', 'scores'),
fn=fn)
metric_dict = eval_metrics(pred_label_seq, true_label_seq)
for name, value in metric_dict.items():
logger.info(f" {name}: {value * 100:.2f}%")
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
def main(out_dir=None,
data_dir=None,
attributes_dir=None,
use_gt_segments=None,
segments_dir=None,
cv_data_dir=None,
ignore_trial_ids=None,
gpu_dev_id=None,
plot_predictions=None,
results_file=None,
sweep_param_name=None,
model_params={},
cv_params={},
train_params={},
viz_params={}):
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
attributes_dir = os.path.expanduser(attributes_dir)
if segments_dir is not None:
segments_dir = os.path.expanduser(segments_dir)
if cv_data_dir is not None:
cv_data_dir = os.path.expanduser(cv_data_dir)
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name, data_dir, prefix='trial='):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'{prefix}{seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
# Load data
trial_ids = utils.getUniqueIds(data_dir, prefix='trial=', to_array=True)
assembly_seqs = loadAll(trial_ids,
'assembly-seq.pkl',
data_dir,
prefix='trial=')
# Define cross-validation folds
if cv_data_dir is None:
dataset_size = len(trial_ids)
cv_folds = utils.makeDataSplits(dataset_size, **cv_params)
cv_fold_trial_ids = tuple(
tuple(map(lambda x: trial_ids[x], splits)) for splits in cv_folds)
else:
fn = os.path.join(cv_data_dir, 'cv-fold-trial-ids.pkl')
cv_fold_trial_ids = joblib.load(fn)
test_ids = set().union(*tuple(set(fold[-1]) for fold in cv_fold_trial_ids))
logger.info(f"{len(test_ids)} final test ids: {test_ids}")
def getSplit(split_idxs):
split_data = tuple(
tuple(s[i] for i in split_idxs)
for s in (assembly_seqs, trial_ids))
return split_data
for cv_index, (train_ids, test_ids) in enumerate(cv_fold_trial_ids):
logger.info(f'CV fold {cv_index + 1}: {len(trial_ids)} total '
f'({len(train_ids)} train, {len(test_ids)} test)')
try:
test_idxs = torch.tensor(
[trial_ids.tolist().index(i) for i in test_ids])
except ValueError:
logger.info(" Skipping fold: missing test data")
continue
# TRAIN PHASE
if cv_data_dir is None:
train_idxs = torch.tensor([trial_ids.index(i) for i in train_ids])
train_assembly_seqs = tuple(assembly_seqs[i] for i in train_idxs)
train_assemblies = []
for seq in train_assembly_seqs:
list(
labels.gen_eq_classes(seq,
train_assemblies,
equivalent=None))
else:
fn = os.path.join(cv_data_dir,
f'cvfold={cv_index}_train-assemblies.pkl')
train_assemblies = joblib.load(fn)
signatures = make_signatures(train_assemblies)
if plot_predictions:
figsize = (12, 3)
fig, axis = plt.subplots(1, figsize=figsize)
axis.imshow(signatures.numpy().T,
interpolation='none',
aspect='auto')
plt.savefig(
os.path.join(fig_dir, f"cvfold={cv_index}_signatures.png"))
plt.close()
# TEST PHASE
accuracies = {'assembly': [], 'assembly_upto_eq': []}
for gt_assembly_seq, trial_id in zip(*getSplit(test_idxs)):
# if ignore_trial_ids is not None and trial_id in ignore_trial_ids:
if False:
logger.info(f" Ignoring trial {trial_id} in test fold")
continue
# FIXME: implement consistent data dimensions during serialization
# (ie samples along rows)
# feature_seq shape is (pairs, samples, classes)
# should be (samples, pairs, classes)
try:
fn = os.path.join(attributes_dir,
f'trial={trial_id}_score-seq.pkl')
feature_seq = joblib.load(fn)
except FileNotFoundError:
logger.info(f" File not found: {fn}")
continue
test_assemblies = train_assemblies.copy()
gt_assembly_id_seq = list(
labels.gen_eq_classes(gt_assembly_seq,
test_assemblies,
equivalent=None))
gt_seq = makeAssemblyLabels(gt_assembly_id_seq, gt_assembly_seq)
if use_gt_segments:
segments = utils.makeSegmentLabels(gt_seq)
elif segments_dir is not None:
var_name = 'segment-seq-imu.pkl'
fn = os.path.join(segments_dir, f'trial={trial_id}_{var_name}')
try:
segments = joblib.load(fn)
except FileNotFoundError:
logger.info(f" File not found: {fn}")
continue
else:
segments = None
feature_seq = torch.tensor(feature_seq, dtype=torch.float)
if segments is not None:
feature_seq = feature_seq.transpose(0, 1)
feature_seq, _ = utils.reduce_over_segments(
feature_seq.numpy(),
segments,
reduction=lambda x: x.mean(axis=0))
feature_seq = torch.tensor(feature_seq.swapaxes(0, 1),
dtype=torch.float)
gt_seq, _ = utils.reduce_over_segments(gt_seq.numpy(),
segments)
gt_seq = torch.tensor(gt_seq, dtype=torch.long)
feature_seq = make_features(feature_seq)
score_seq = score(feature_seq, signatures)
pred_seq = predict(score_seq)
pred_assemblies = [train_assemblies[i] for i in pred_seq]
gt_assemblies = [test_assemblies[i] for i in gt_seq]
acc = metrics.accuracy_upto(pred_assemblies,
gt_assemblies,
equivalence=components_equivalent)
accuracies['assembly_upto_eq'].append(acc)
acc = metrics.accuracy_upto(pred_assemblies, gt_assemblies)
accuracies['assembly'].append(acc)
# FIXME: Improve data naming convention in decode_keyframes.py
saveVariable(score_seq, f'trial={trial_id}_data-scores')
if plot_predictions:
fn = os.path.join(fig_dir, f'trial={trial_id:03}.png')
utils.plot_array(
feature_seq,
(gt_seq.numpy(), pred_seq.numpy(), score_seq.numpy()),
('gt', 'pred', 'scores'),
fn=fn)
if not any(v for v in accuracies.values()):
continue
metric_dict = {}
for name, fold_accuracies in accuracies.items():
fold_accuracy = float(torch.tensor(fold_accuracies).mean())
logger.info(f' {name}: {fold_accuracy * 100:.1f}%')
metric_dict[name] = fold_accuracy
utils.writeResults(results_file, metric_dict, sweep_param_name,
model_params)
def main(
out_dir=None, data_dir=None, part_symmetries=None,
plot_output=None, results_file=None, sweep_param_name=None, start_from=None):
if part_symmetries is None:
part_symmetries = [
['frontbeam_hole_1', 'frontbeam_hole_2', 'backbeam_hole_1', 'backbeam_hole_2'],
['cushion_hole_1', 'cushion_hole_2']
]
part_symmetries = {
symms[i]: symms
for symms in part_symmetries
for i in range(len(symms))
}
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
debug_dir = os.path.join(out_dir, 'debug')
if not os.path.exists(debug_dir):
os.makedirs(debug_dir)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
logger.info(f"Reading from: {data_dir}")
logger.info(f"Writing to: {out_dir}")
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
# write_mode = 'w'
else:
results_file = os.path.expanduser(results_file)
# write_mode = 'a'
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
labels_dir = os.path.join(data_dir, 'labels')
with open(os.path.join(labels_dir, 'action_and_part_names.yaml'), 'rt') as f:
names = yaml.safe_load(f)
action_names = names['action_names']
action_name_to_index = {name: i for i, name in enumerate(action_names)}
part_names = names['part_names']
part_name_to_index = {name: i for i, name in enumerate(part_names)}
video_ids = []
all_label_arrs = []
for label_fn in glob.glob(os.path.join(labels_dir, "*.csv")):
video_id = utils.stripExtension(label_fn)
labels_arr = pd.read_csv(label_fn)
all_label_arrs.append(labels_arr)
video_ids.append(video_id)
pose_dir = os.path.join(data_dir, 'poses')
pose_ids = tuple(
video_id
for video_id in video_ids
if os.path.exists(os.path.join(pose_dir, video_id))
)
keep_ids = tuple(v_id in pose_ids for v_id in video_ids)
logger.info(
f"Ignoring {len(keep_ids) - sum(keep_ids)} video(s) with missing data: "
f"{', '.join([v_id for v_id, keep in zip(video_ids, keep_ids) if not keep])}"
)
def filterSeq(seq):
return tuple(x for x, keep_id in zip(seq, keep_ids) if keep_id)
all_label_arrs = filterSeq(all_label_arrs)
video_ids = filterSeq(video_ids)
assembly_vocab = []
label_seqs = []
for i, video_id in enumerate(video_ids):
if start_from is not None and i < start_from:
continue
logger.info("PROCESSING VIDEO {0}: {1}".format(i, video_id))
labels_arr = all_label_arrs[i]
video_dir = os.path.join(pose_dir, video_id)
def loadFile(part_name):
path = os.path.join(video_dir, f'{part_name}.csv')
arr = pd.read_csv(path)
return arr
part_data = tuple(loadFile(part_name) for part_name in part_names)
poses_seq = np.stack(tuple(arr.values for arr in part_data), axis=-1)
feature_seq = relativePose(poses_seq, lower_tri_only=True, magnitude_only=True)
label_seq = actionLabels(
labels_arr, feature_seq.shape[0],
action_name_to_index, part_name_to_index
)
part_pair_names = makePairs(part_names, lower_tri_only=True)
is_possible = possibleConnections(part_pair_names)
feature_seq = feature_seq[:, is_possible, :]
label_seq = label_seq[:, is_possible]
part_pair_names = tuple(n for (b, n) in zip(is_possible, part_pair_names) if b)
utils.plot_multi(
np.moveaxis(feature_seq, (0, 1, 2), (-1, 0, 1)), label_seq.T,
axis_names=part_pair_names, label_name='action',
feature_names=('translation_dist', 'rotation_dist'),
tick_names=[''] + action_names,
fn=os.path.join(fig_dir, f"{video_id}_actions.png")
)
label_seq = assemblyLabels(
labels_arr, feature_seq.shape[0],
assembly_vocab=assembly_vocab, symmetries=part_symmetries
)
# expanded_label_seq = _assemblyLabels(
# labels_arr, feature_seq.shape[0],
# assembly_vocab=assembly_vocab, symmetries=part_symmetries
# )
utils.plot_array(
feature_seq.sum(axis=-1).T, (label_seq,), ('assembly',),
fn=os.path.join(fig_dir, f"{video_id}_assemblies.png")
)
label_seqs.append(label_seq)
label_segments, __ = utils.computeSegments(label_seq)
assembly_segments = [assembly_vocab[i] for i in label_segments]
lib_assembly.writeAssemblies(
os.path.join(debug_dir, f'trial={video_id}_assembly-seq.txt'),
assembly_segments
)
video_id = video_id.replace('_', '-')
saveVariable(feature_seq, f'trial={video_id}_feature-seq')
saveVariable(label_seq, f'trial={video_id}_label-seq')
# saveVariable(expanded_label_seq, f'trial={video_id}_expanded-label-seq')
if False:
from seqtools import utils as su
transition_probs, start_probs, end_probs = su.smoothCounts(
*su.countSeqs(label_seqs)
)
# import pdb; pdb.set_trace()
lib_assembly.writeAssemblies(
os.path.join(debug_dir, 'assembly-vocab.txt'),
assembly_vocab
)
saveVariable(assembly_vocab, 'assembly-vocab')
with open(os.path.join(out_data_dir, 'action-vocab.yaml'), 'wt') as f:
yaml.dump(action_names, f)
with open(os.path.join(out_data_dir, 'part-vocab.yaml'), 'wt') as f:
yaml.dump(part_names, f)
def main(out_dir=None,
data_dir=None,
metadata_file=None,
metric_names=None,
ignore_initial_state=None,
draw_paths=None,
plot_predictions=None,
results_file=None,
sweep_param_name=None):
if metric_names is None:
metric_names = ('accuracy', 'edit_score', 'overlap_score')
logger.info(f"Reading from: {data_dir}")
logger.info(f"Writing to: {out_dir}")
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def loadVariable(var_name):
return joblib.load(os.path.join(data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'trial={seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
trial_ids = utils.getUniqueIds(data_dir, prefix='trial=', to_array=True)
pred_assembly_seqs = loadAll(trial_ids, "pred-assembly-seq.pkl")
gt_assembly_seqs = loadAll(trial_ids, "gt-assembly-seq.pkl")
all_assemblies = []
gt_assembly_label_seqs = tuple(
np.array(list(labels.gen_eq_classes(gt_assembly_seq, all_assemblies)))
for gt_assembly_seq in gt_assembly_seqs)
pred_assembly_label_seqs = tuple(
np.array(list(labels.gen_eq_classes(pred_assembly_seq,
all_assemblies)))
for pred_assembly_seq in pred_assembly_seqs)
def checkPredictions(pred_assembly_label_seqs, gt_assembly_label_seqs,
trial_ids):
for i, trial_id in enumerate(trial_ids):
pred_seq = pred_assembly_label_seqs[i]
true_seq = gt_assembly_label_seqs[i]
train_seqs = gt_assembly_label_seqs[:i] + gt_assembly_label_seqs[
i + 1:]
train_vocab = np.unique(np.hstack(train_seqs))
pred_is_correct = pred_seq == true_seq
pred_in_test_vocab = np.array(
[np.any(train_vocab == i) for i in pred_seq])
pred_is_oov_and_correct = ~pred_in_test_vocab * pred_is_correct
if not pred_in_test_vocab.all():
warn_str = (
f"trial {trial_id}: {np.sum(~pred_in_test_vocab)} / "
f"{len(pred_in_test_vocab)} preds not in test; "
f"{np.sum(pred_is_oov_and_correct)} oov preds are correct")
logger.warning(warn_str)
checkPredictions(pred_assembly_label_seqs, gt_assembly_label_seqs,
trial_ids)
def estimate_oov_rate(label_seqs):
num_oov = 0
num_items = 0
for i in range(len(label_seqs)):
test_seq = label_seqs[i]
train_seqs = label_seqs[:i] + label_seqs[i + 1:]
train_vocab = np.unique(np.hstack(train_seqs))
num_oov += sum(
float(not np.any(train_vocab == i)) for i in test_seq)
num_items += len(test_seq)
return num_oov / num_items
oov_rate = estimate_oov_rate(gt_assembly_label_seqs)
logger.info(f"OOV rate: {oov_rate * 100:.2f}%")
pred_action_seqs = tuple(map(actionsFromAssemblies, pred_assembly_seqs))
gt_action_seqs = tuple(map(actionsFromAssemblies, gt_assembly_seqs))
all_actions = []
gt_action_label_seqs = tuple(
np.array(list(labels.gen_eq_classes(gt_action_seq, all_actions)))
for gt_action_seq in gt_action_seqs)
pred_action_label_seqs = tuple(
np.array(list(labels.gen_eq_classes(pred_action_seq, all_actions)))
for pred_action_seq in pred_action_seqs)
if draw_paths:
assembly_fig_dir = os.path.join(fig_dir, 'assembly-imgs')
if not os.path.exists(assembly_fig_dir):
os.makedirs(assembly_fig_dir)
for i, assembly in enumerate(all_assemblies):
assembly.draw(assembly_fig_dir, i)
action_fig_dir = os.path.join(fig_dir, 'action-imgs')
if not os.path.exists(action_fig_dir):
os.makedirs(action_fig_dir)
for i, action in enumerate(all_actions):
action.draw(action_fig_dir, i)
logger.info(f"Evaluating {len(trial_ids)} sequence predictions")
batch = []
for i, trial_id in enumerate(trial_ids):
logger.info(f"VIDEO {trial_id}:")
pred_assembly_index_seq = pred_assembly_label_seqs[i]
true_assembly_index_seq = gt_assembly_label_seqs[i]
if ignore_initial_state:
pred_assembly_index_seq = pred_assembly_index_seq[1:]
true_assembly_index_seq = true_assembly_index_seq[1:]
pred_action_index_seq = pred_action_label_seqs[i]
true_action_index_seq = gt_action_label_seqs[i]
if draw_paths:
drawPaths([pred_assembly_index_seq, true_assembly_index_seq],
f"trial={trial_id}_assembly-paths",
fig_dir,
assembly_fig_dir,
path_labels=None,
img_ext='png')
drawPaths([pred_action_index_seq, true_action_index_seq],
f"trial={trial_id}_action-paths",
fig_dir,
action_fig_dir,
path_labels=None,
img_ext='png')
metric_dict = {}
for name in metric_names:
key = f"{name}_action"
value = getattr(LCTM.metrics, name)(pred_action_index_seq,
true_action_index_seq) / 100
metric_dict[key] = value
logger.info(f" {key}: {value * 100:.1f}%")
key = f"{name}_assembly"
value = getattr(LCTM.metrics, name)(pred_assembly_index_seq,
true_assembly_index_seq) / 100
metric_dict[key] = value
logger.info(f" {key}: {value * 100:.1f}%")
utils.writeResults(results_file, metric_dict, sweep_param_name, {})
batch.append(
(pred_action_index_seq, None, true_action_index_seq, trial_id))
if plot_predictions:
label_names = ('gt', 'pred')
for preds, inputs, gt_labels, seq_id in batch:
fn = os.path.join(fig_dir, f"trial={seq_id}_model-io.png")
utils.plot_array(inputs, (gt_labels, preds),
label_names,
fn=fn,
labels_together=True)
def main(out_dir=None,
data_dir=None,
model_name=None,
part_symmetries=None,
gpu_dev_id=None,
batch_size=None,
learning_rate=None,
model_params={},
cv_params={},
train_params={},
viz_params={},
plot_predictions=None,
results_file=None,
sweep_param_name=None):
if part_symmetries is None:
part_symmetries = {
'beam_side': ('backbeam_hole_1', 'backbeam_hole_2',
'frontbeam_hole_1', 'frontbeam_hole_2'),
'beam_top': ('backbeam_hole_3', 'frontbeam_hole_3'),
'backrest': ('backrest_hole_1', 'backrest_hole_2')
}
data_dir = os.path.expanduser(data_dir)
out_dir = os.path.expanduser(out_dir)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
logger = utils.setupRootLogger(filename=os.path.join(out_dir, 'log.txt'))
if results_file is None:
results_file = os.path.join(out_dir, 'results.csv')
else:
results_file = os.path.expanduser(results_file)
fig_dir = os.path.join(out_dir, 'figures')
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
out_data_dir = os.path.join(out_dir, 'data')
if not os.path.exists(out_data_dir):
os.makedirs(out_data_dir)
def saveVariable(var, var_name):
joblib.dump(var, os.path.join(out_data_dir, f'{var_name}.pkl'))
def loadAll(seq_ids, var_name, data_dir):
def loadOne(seq_id):
fn = os.path.join(data_dir, f'trial={seq_id}_{var_name}')
return joblib.load(fn)
return tuple(map(loadOne, seq_ids))
# Load vocab
with open(os.path.join(data_dir, "part-vocab.yaml"), 'rt') as f:
link_vocab = yaml.safe_load(f)
assembly_vocab = joblib.load(os.path.join(data_dir, 'assembly-vocab.pkl'))
# Load data
trial_ids = utils.getUniqueIds(data_dir, prefix='trial=')
feature_seqs = loadAll(trial_ids, 'feature-seq.pkl', data_dir)
label_seqs = loadAll(trial_ids, 'label-seq.pkl', data_dir)
if part_symmetries:
# Construct equivalence classes from vocab
eq_classes, assembly_eq_classes, eq_class_vocab = makeEqClasses(
assembly_vocab, part_symmetries)
lib_assembly.writeAssemblies(
os.path.join(fig_dir, 'eq-class-vocab.txt'), eq_class_vocab)
label_seqs = tuple(assembly_eq_classes[label_seq]
for label_seq in label_seqs)
saveVariable(eq_class_vocab, 'assembly-vocab')
else:
eq_classes = None
def impute_nan(input_seq):
input_is_nan = np.isnan(input_seq)
logger.info(f"{input_is_nan.sum()} NaN elements")
input_seq[input_is_nan] = 0 # np.nanmean(input_seq)
return input_seq
# feature_seqs = tuple(map(impute_nan, feature_seqs))
for trial_id, label_seq, feat_seq in zip(trial_ids, label_seqs,
feature_seqs):
saveVariable(feat_seq, f"trial={trial_id}_feature-seq")
saveVariable(label_seq, f"trial={trial_id}_label-seq")
device = torchutils.selectDevice(gpu_dev_id)
# Define cross-validation folds
dataset_size = len(trial_ids)
cv_folds = utils.makeDataSplits(dataset_size, **cv_params)
def getSplit(split_idxs):
split_data = tuple(
tuple(s[i] for i in split_idxs)
for s in (feature_seqs, label_seqs, trial_ids))
return split_data
for cv_index, cv_splits in enumerate(cv_folds):
train_data, val_data, test_data = tuple(map(getSplit, cv_splits))
train_feats, train_labels, train_ids = train_data
train_set = torchutils.SequenceDataset(train_feats,
train_labels,
device=device,
labels_dtype=torch.long,
seq_ids=train_ids,
transpose_data=True)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=batch_size,
shuffle=True)
test_feats, test_labels, test_ids = test_data
test_set = torchutils.SequenceDataset(test_feats,
test_labels,
device=device,
labels_dtype=torch.long,
seq_ids=test_ids,
transpose_data=True)
test_loader = torch.utils.data.DataLoader(test_set,
batch_size=batch_size,
shuffle=False)
val_feats, val_labels, val_ids = val_data
val_set = torchutils.SequenceDataset(val_feats,
val_labels,
device=device,
labels_dtype=torch.long,
seq_ids=val_ids,
transpose_data=True)
val_loader = torch.utils.data.DataLoader(val_set,
batch_size=batch_size,
shuffle=True)
logger.info(
f'CV fold {cv_index + 1} / {len(cv_folds)}: {len(trial_ids)} total '
f'({len(train_ids)} train, {len(val_ids)} val, {len(test_ids)} test)'
)
input_dim = train_set.num_obsv_dims
output_dim = train_set.num_label_types
if model_name == 'linear':
model = torchutils.LinearClassifier(
input_dim, output_dim, **model_params).to(device=device)
elif model_name == 'dummy':
model = DummyClassifier(input_dim, output_dim, **model_params)
elif model_name == 'AssemblyClassifier':
model = AssemblyClassifier(assembly_vocab,
link_vocab,
eq_classes=eq_classes,
**model_params)
else:
raise AssertionError()
criterion = torch.nn.CrossEntropyLoss()
if model_name != 'dummy':
train_epoch_log = collections.defaultdict(list)
val_epoch_log = collections.defaultdict(list)
metric_dict = {
'Avg Loss': metrics.AverageLoss(),
'Accuracy': metrics.Accuracy(),
'Precision': metrics.Precision(),
'Recall': metrics.Recall(),
'F1': metrics.Fmeasure()
}
optimizer_ft = torch.optim.Adam(model.parameters(),
lr=learning_rate,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=0,
amsgrad=False)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer_ft,
step_size=1,
gamma=1.00)
model, last_model_wts = torchutils.trainModel(
model,
criterion,
optimizer_ft,
lr_scheduler,
train_loader,
val_loader,
device=device,
metrics=metric_dict,
train_epoch_log=train_epoch_log,
val_epoch_log=val_epoch_log,
**train_params)
logger.info(f'scale={float(model._scale)}')
logger.info(f'alpha={float(model._alpha)}')
# Test model
metric_dict = {
'Avg Loss': metrics.AverageLoss(),
'Accuracy': metrics.Accuracy(),
'Precision': metrics.Precision(),
'Recall': metrics.Recall(),
'F1': metrics.Fmeasure()
}
test_io_history = torchutils.predictSamples(
model.to(device=device),
test_loader,
criterion=criterion,
device=device,
metrics=metric_dict,
data_labeled=True,
update_model=False,
seq_as_batch=train_params['seq_as_batch'],
return_io_history=True)
metric_str = ' '.join(str(m) for m in metric_dict.values())
logger.info('[TST] ' + metric_str)
d = {k: v.value for k, v in metric_dict.items()}
utils.writeResults(results_file, d, sweep_param_name, model_params)
if plot_predictions:
io_fig_dir = os.path.join(fig_dir, 'model-io')
if not os.path.exists(io_fig_dir):
os.makedirs(io_fig_dir)
label_names = ('gt', 'pred')
preds, scores, inputs, gt_labels, ids = zip(*test_io_history)
for batch in test_io_history:
batch = tuple(
x.cpu().numpy() if isinstance(x, torch.Tensor) else x
for x in batch)
for preds, _, inputs, gt_labels, seq_id in zip(*batch):
fn = os.path.join(io_fig_dir,
f"trial={seq_id}_model-io.png")
utils.plot_array(inputs.sum(axis=-1), (gt_labels, preds),
label_names,
fn=fn,
**viz_params)
def saveTrialData(pred_seq, score_seq, feat_seq, label_seq, trial_id):
saveVariable(pred_seq, f'trial={trial_id}_pred-label-seq')
saveVariable(score_seq, f'trial={trial_id}_score-seq')
saveVariable(label_seq, f'trial={trial_id}_true-label-seq')
for batch in test_io_history:
batch = tuple(x.cpu().numpy() if isinstance(x, torch.Tensor) else x
for x in batch)
for io in zip(*batch):
saveTrialData(*io)
saveVariable(train_ids, f'cvfold={cv_index}_train-ids')
saveVariable(test_ids, f'cvfold={cv_index}_test-ids')
saveVariable(val_ids, f'cvfold={cv_index}_val-ids')
saveVariable(train_epoch_log,
f'cvfold={cv_index}_{model_name}-train-epoch-log')
saveVariable(val_epoch_log,
f'cvfold={cv_index}_{model_name}-val-epoch-log')
saveVariable(metric_dict,
f'cvfold={cv_index}_{model_name}-metric-dict')
saveVariable(model, f'cvfold={cv_index}_{model_name}-best')
train_fig_dir = os.path.join(fig_dir, 'train-plots')
if not os.path.exists(train_fig_dir):
os.makedirs(train_fig_dir)
if train_epoch_log:
torchutils.plotEpochLog(train_epoch_log,
subfig_size=(10, 2.5),
title='Training performance',
fn=os.path.join(
train_fig_dir,
f'cvfold={cv_index}_train-plot.png'))
if val_epoch_log:
torchutils.plotEpochLog(val_epoch_log,
subfig_size=(10, 2.5),
title='Heldout performance',
fn=os.path.join(
train_fig_dir,
f'cvfold={cv_index}_val-plot.png'))