def __init__(self, dataset_path, frame=None, loop=True):
self._dataset_path = dataset_path
self._frame = frame
self._color_frame = frame
self._ir_frame = frame
self._im_index = 0
self._running = False
from dexnet.learning import TensorDataset
self._dataset = TensorDataset.open(self._dataset_path)
self._num_images = self._dataset.num_datapoints
self._image_rescale_factor = 1.0
if 'image_rescale_factor' in self._dataset.metadata.keys():
self._image_rescale_factor = 1.0 / self._dataset.metadata[
'image_rescale_factor']
datapoint = self._dataset.datapoint(
0, [TensorDatasetVirtualSensor.CAMERA_INTR_FIELD])
camera_intr_vec = datapoint[
TensorDatasetVirtualSensor.CAMERA_INTR_FIELD]
self._color_intr = CameraIntrinsics.from_vec(
camera_intr_vec,
frame=self._color_frame).resize(self._image_rescale_factor)
self._ir_intr = CameraIntrinsics.from_vec(
camera_intr_vec,
frame=self._ir_frame).resize(self._image_rescale_factor)
def analyze_classification_performance(model_dir, config, dataset_path=None):
# read params
predict_batch_size = config['batch_size']
randomize = config['randomize']
plotting_config = config['plotting']
figsize = plotting_config['figsize']
font_size = plotting_config['font_size']
legend_font_size = plotting_config['legend_font_size']
line_width = plotting_config['line_width']
colors = plotting_config['colors']
dpi = plotting_config['dpi']
style = '-'
class_remapping = None
if 'class_remapping' in config.keys():
class_remapping = config['class_remapping']
# read training config
training_config_filename = os.path.join(model_dir, 'training_config.yaml')
training_config = YamlConfig(training_config_filename)
# read training params
indices_filename = None
if dataset_path is None:
dataset_path = training_config['dataset']
indices_filename = os.path.join(model_dir, 'splits.npz')
dataset_prefix, dataset_name = os.path.split(dataset_path)
if dataset_name == '':
_, dataset_name = os.path.split(dataset_prefix)
x_names = training_config['x_names']
y_name = training_config['y_name']
batch_size = training_config['training']['batch_size']
iterator_config = training_config['data_iteration']
x_name = x_names[0]
# set analysis dir
analysis_dir = os.path.join(model_dir, 'analysis')
if not os.path.exists(analysis_dir):
os.mkdir(analysis_dir)
# setup log file
experiment_log_filename = os.path.join(analysis_dir, '%s_analysis.log' %(dataset_name))
if os.path.exists(experiment_log_filename):
os.remove(experiment_log_filename)
formatter = logging.Formatter('%(asctime)s %(levelname)s: %(message)s')
hdlr = logging.FileHandler(experiment_log_filename)
hdlr.setFormatter(formatter)
logging.getLogger().addHandler(hdlr)
# setup plotting
plt.figure(figsize=(figsize, figsize))
# read dataset
dataset = TensorDataset.open(dataset_path)
# read dataset splits
if indices_filename is None:
splits = {dataset_name: np.arange(dataset.num_datapoints)}
else:
splits = np.load(indices_filename)['arr_0'].tolist()
# load cnn
logging.info('Loading model %s' %(model_dir))
cnn = ClassificationCNN.open(model_dir)
# save examples
logging.info('Saving examples of each class')
all_labels = np.arange(cnn.num_classes)
label_counts = {}
[label_counts.update({l:0}) for l in all_labels]
for tensor_ind in range(dataset.num_tensors):
tensor = dataset.tensor(y_name, tensor_ind)
for label in tensor:
label_counts[label] += 1
d = utils.sqrt_ceil(cnn.num_classes)
plt.clf()
for i, label in enumerate(all_labels):
tensor_ind = 0
label_found = False
while not label_found and tensor_ind < dataset.num_tensors:
tensor = dataset.tensor(y_name, tensor_ind)
ind = np.where(tensor.arr == label)[0]
if ind.shape[0] > 0:
ind = ind[0] + dataset.datapoints_per_tensor * (tensor_ind)
label_found = True
tensor_ind += 1
if not label_found:
continue
datapoint = dataset[ind]
example_im = datapoint[x_name]
plt.subplot(d,d,i+1)
plt.imshow(example_im[:,:,:3].astype(np.uint8))
plt.title('Class %03d: %.3f%%' %(label, float(label_counts[label]) / dataset.num_datapoints), fontsize=3)
plt.axis('off')
plt.savefig(os.path.join(analysis_dir, '%s_classes.pdf' %(dataset_name)))
# evaluate on dataset
results = {}
for split_name, indices in splits.iteritems():
logging.info('Evaluating performance on split: %s' %(split_name))
# predict
if randomize:
pred_probs, true_labels = cnn.evaluate_on_dataset(dataset, indices=indices, batch_size=predict_batch_size)
pred_labels = np.argmax(pred_probs, axis=1)
else:
true_labels = []
pred_labels = []
pred_probs = []
for datapoint in dataset:
im = ColorImage(datapoint['color_ims'].astype(np.uint8)[:,:,:3])
true_label = datapoint['stp_labels']
pred_prob = cnn.predict(im)
pred_label = np.argmax(pred_prob, axis=1)
true_labels.append(true_label)
pred_labels.append(pred_label)
pred_probs.append(pred_prob.ravel())
"""
if class_remapping is not None:
true_label = class_remapping[true_label]
plt.figure()
plt.imshow(im.raw_data)
plt.title('T: %d, P: %d' %(true_label, pred_label))
plt.show()
"""
true_labels = np.array(true_labels)
pred_labels = np.array(pred_labels)
pred_probs = np.array(pred_probs)
# apply optional class re-mapping
if class_remapping is not None:
new_true_labels = np.zeros(true_labels.shape)
for orig_label, new_label in class_remapping.iteritems():
new_true_labels[true_labels==orig_label] = new_label
true_labels = new_true_labels
# compute classification results
result = ClassificationResult([pred_probs], [true_labels])
results[split_name] = result
# print stats
logging.info('SPLIT: %s' %(split_name))
logging.info('Acc: %.3f' %(result.accuracy))
logging.info('AP: %.3f' %(result.ap_score))
logging.info('AUC: %.3f' %(result.auc_score))
# save confusion matrix
confusion = result.confusion_matrix.data
plt.clf()
plt.imshow(confusion, cmap=plt.cm.gray, interpolation='none')
plt.locator_params(nticks=cnn.num_classes)
plt.xlabel('Predicted', fontsize=font_size)
plt.ylabel('Actual', fontsize=font_size)
plt.savefig(os.path.join(analysis_dir, '%s_confusion.pdf' %(split_name)), dpi=dpi)
# save analysis
result_filename = os.path.join(analysis_dir, '%s.cres' %(split_name))
result.save(result_filename)
# plot
colormap = plt.get_cmap('tab10')
num_colors = 9
plt.clf()
for i, split_name in enumerate(splits.keys()):
result = results[split_name]
precision, recall, taus = result.precision_recall_curve()
color = colormap(float(colors[i%num_colors]) / num_colors)
plt.plot(recall, precision, linewidth=line_width, color=color, linestyle=style, label=split_name)
plt.xlabel('Recall', fontsize=font_size)
plt.ylabel('Precision', fontsize=font_size)
plt.title('Precision-Recall Curve', fontsize=font_size)
handles, plt_labels = plt.gca().get_legend_handles_labels()
plt.legend(handles, plt_labels, loc='best', fontsize=legend_font_size)
plt.savefig(os.path.join(analysis_dir, '%s_precision_recall.pdf' %(dataset_name)), dpi=dpi)
plt.clf()
for i, split_name in enumerate(splits.keys()):
result = results[split_name]
fpr, tpr, taus = result.roc_curve()
color = colormap(float(colors[i%num_colors]) / num_colors)
plt.plot(fpr, tpr, linewidth=line_width, color=color, linestyle=style, label=split_name)
plt.xlabel('FPR', fontsize=font_size)
plt.ylabel('TPR', fontsize=font_size)
plt.title('Receiver Operating Characteristic', fontsize=font_size)
handles, plt_labels = plt.gca().get_legend_handles_labels()
plt.legend(handles, plt_labels, loc='best', fontsize=legend_font_size)
plt.savefig(os.path.join(analysis_dir, '%s_roc.pdf' %(dataset_name)), dpi=dpi)
######################################################################
if config_filename is None:
config_filename = os.path.join(
os.path.dirname(os.path.realpath(__file__)), '..',
'cfg/tools/generate_gqcnn_dataset.yaml')
config = YamlConfig(config_filename)
# set tensor dataset config
tensor_config = config['tensors']
tensor_config['fields']['depth_ims_tf_table_96'][
'height'] = im_final_height
tensor_config['fields']['depth_ims_tf_table_96']['width'] = im_final_width
# tensor_dataset = TensorDataset(output_dir, tensor_config)
tensor_dataset = TensorDataset.open(output_dir)
tensor_datapoint = tensor_dataset.datapoint_template
######################################################################
# open tensor dirs
if not os.path.exists(dataset_path):
raise ValueError('Dataset %s not found!' % (dataset_path))
# create subdirectories
image_dir = os.path.join(dataset_path, 'images')
if not os.path.exists(image_dir):
raise ValueError('Image folder %s not found!' % (image_dir))
grasp_dir = os.path.join(dataset_path, 'grasps')
if not os.path.exists(grasp_dir):
'--config_filename',
type=str,
default=None,
help=
'Yaml filename containing configuration parameters for the visualization'
)
args = parser.parse_args()
dataset_path = args.dataset_path
config_filename = args.config_filename
# handle config filename
if config_filename is None:
config_filename = os.path.join(
os.path.dirname(os.path.realpath(__file__)), '..',
'cfg/tools/visualize_gqcnn_dataset.yaml')
# turn relative paths absolute
if not os.path.isabs(dataset_path):
dataset_path = os.path.join(os.getcwd(), dataset_path)
if not os.path.isabs(config_filename):
config_filename = os.path.join(os.getcwd(), config_filename)
# read config
config = YamlConfig(config_filename)
# open tensor dataset
dataset = TensorDataset.open(dataset_path)
# visualize a tensor dataset
visualize_tensor_dataset(dataset, config)
def finetune_classification_cnn(config):
""" Main function. """
# read params
dataset = config['dataset']
x_names = config['x_names']
y_name = config['y_name']
model_dir = config['model_dir']
debug = config['debug']
num_classes = None
if 'num_classes' in config.keys():
num_classes = config['num_classes']
batch_size = config['training']['batch_size']
train_pct = config['training']['train_pct']
model_save_period = config['training']['model_save_period']
data_aug_config = config['data_augmentation']
preproc_config = config['preprocessing']
iterator_config = config['data_iteration']
model_config = config['model']
base_model_config = model_config['base']
optimization_config = config['optimization']
train_config = config['training']
generator_image_shape = None
if 'image_shape' in data_aug_config.keys():
generator_image_shape = data_aug_config['image_shape']
optimizer_name = optimization_config['optimizer']
model_params = {}
if 'params' in model_config.keys():
model_params = model_config['params']
base_model_params = {}
if 'params' in base_model_config.keys():
base_model_params = base_model_config['params']
if debug:
seed = 108
random.seed(seed)
np.random.seed(seed)
# generate model dir
if not os.path.exists(model_dir):
os.mkdir(model_dir)
model_id = utils.gen_experiment_id()
model_dir = os.path.join(model_dir, 'model_%s' % (model_id))
if not os.path.exists(model_dir):
os.mkdir(model_dir)
logging.info('Saving model to %s' % (model_dir))
latest_model_filename = os.path.join(model_dir, 'weights_{epoch:05d}.h5')
best_model_filename = os.path.join(model_dir, 'weights.h5')
# save config
training_config_filename = os.path.join(model_dir, 'training_config.yaml')
config.save(training_config_filename)
# open dataset
dataset = TensorDataset.open(dataset)
# split dataset
indices_filename = os.path.join(model_dir, 'splits.npz')
if os.path.exists(indices_filename):
indices = np.load(indices_filename)['arr_0'].tolist()
train_indices = indices['train']
val_indices = indices['val']
else:
train_indices, val_indices = dataset.split(train_pct)
indices = np.array({'train': train_indices, 'val': val_indices})
np.savez_compressed(indices_filename, indices)
num_train = train_indices.shape[0]
num_val = val_indices.shape[0]
val_steps = int(np.ceil(float(num_val) / batch_size))
# init generator
train_generator_filename = os.path.join(model_dir,
'train_preprocessor.pkl')
val_generator_filename = os.path.join(model_dir, 'val_preprocessor.pkl')
if os.path.exists(train_generator_filename):
logging.info('Loading generators')
train_generator = pkl.load(open(train_generator_filename, 'rb'))
val_generator = pkl.load(open(val_generator_filename, 'rb'))
else:
logging.info('Fitting generator')
train_generator = TensorDataGenerator(num_classes=num_classes,
**data_aug_config)
val_generator = TensorDataGenerator(
featurewise_center=data_aug_config['featurewise_center'],
featurewise_std_normalization=data_aug_config[
'featurewise_std_normalization'],
image_shape=generator_image_shape,
num_classes=num_classes)
fit_start = time.time()
train_generator.fit(dataset,
x_names,
y_name,
indices=train_indices,
**preproc_config)
val_generator.mean = train_generator.mean
val_generator.std = train_generator.std
val_generator.min_output = train_generator.min_output
val_generator.max_output = train_generator.max_output
val_generator.num_classes = train_generator.num_classes
fit_stop = time.time()
logging.info('Generator fit took %.3f sec' % (fit_stop - fit_start))
pkl.dump(train_generator, open(train_generator_filename, 'wb'))
pkl.dump(val_generator, open(val_generator_filename, 'wb'))
if num_classes is None:
num_classes = int(train_generator.num_classes)
# init iterator
train_iterator = train_generator.flow_from_dataset(dataset,
x_names,
y_name,
indices=train_indices,
batch_size=batch_size,
**iterator_config)
val_iterator = val_generator.flow_from_dataset(dataset,
x_names,
y_name,
indices=val_indices,
batch_size=batch_size,
**iterator_config)
# setup model
base_cnn = ClassificationCNN.open(base_model_config['model'],
base_model_config['type'],
input_name=x_names[0],
**base_model_params)
cnn = FinetunedClassificationCNN(base_cnn=base_cnn,
name='dexresnet',
num_classes=num_classes,
output_name=y_name,
im_preprocessor=val_generator,
**model_params)
# setup training
cnn.freeze_base_cnn()
if optimizer_name == 'sgd':
optimizer = SGD(lr=optimization_config['lr'],
momentum=optimization_config['momentum'])
elif optimizer_name == 'adam':
optimizer = Adam(lr=optimization_config['lr'])
else:
raise ValueError('Optimizer %s not supported!' % (optimizer_name))
model = cnn.model
model.compile(optimizer=optimizer,
loss=optimization_config['loss'],
metrics=optimization_config['metrics'])
# train
steps_per_epoch = int(np.ceil(float(num_train) / batch_size))
latest_model_ckpt = ModelCheckpoint(latest_model_filename,
period=model_save_period)
best_model_ckpt = ModelCheckpoint(best_model_filename,
save_best_only=True,
period=model_save_period)
train_history_cb = TrainHistory(model_dir)
callbacks = [latest_model_ckpt, best_model_ckpt, train_history_cb]
history = model.fit_generator(
train_iterator,
steps_per_epoch=steps_per_epoch,
epochs=train_config['epochs'],
callbacks=callbacks,
validation_data=val_iterator,
validation_steps=val_steps,
class_weight=train_config['class_weight'],
use_multiprocessing=train_config['use_multiprocessing'])
# save model
cnn.save(model_dir)
# save history
history_filename = os.path.join(model_dir, 'history.pkl')
pkl.dump(history.history, open(history_filename, 'wb'))