regularizer = reg.Regularizer()
regularizer.fixresize(200, 200)
if createdataset:
create_dataset(savepath=path, im_regularizer=regularizer)
formatting = confidence_filtered_pb_format(minconf)
generator = DatasetManager(train_samples=train_samples,
valid_samples=valid_samples,
batch_size=batch_size,
dataset_dir=path,
formatting=formatting)
gt = generator.train()
cw = create_weights(gt)
data_processing_plan = ProcessingPlan()
model1 = train_model(model_generator=lambda: simple_classifier_rgb(weight_decay=weight_decay),
dataset_manager=generator,
loss='binary_crossentropy',
learning_rate=learning_rate,
patience=patience,
data_processing_plan=data_processing_plan,
tb_path='palm_back/',
model_name=name,
model_path=resources_path(os.path.join("models", "palm_back", name)),
epochs=epochs,
enable_telegram_log=False,
class_weight=cw)
# Plan the processing needed before providing inputs and outputs for training and validation
data_processing_plan = ProcessingPlan(
augmenter=Augmenter().shift_hue(augmentation_prob).shift_sat(
augmentation_prob).shift_val(augmentation_prob),
regularizer=Regularizer().normalize() if normalize else None,
keyset={IN('img')})
data_processing_plan.add_outer(key=IN('img'),
fun=lambda x: preprocess_mobile(255 * x))
model = train_model(
model_generator=lambda: regressor_2(input_shape=np.shape(dm.train()[0][
IN('img')][0])),
dataset_manager=dm,
loss={OUT('heats'): loss},
learning_rate=learning_rate,
patience=patience,
data_processing_plan=data_processing_plan,
tb_path="joints",
tb_plots={
'target':
lambda feed: joint_skeleton_regressor(
feed, img_key=IN('img'), joints_key=OUT('heats')),
'output':
lambda feed: joint_skeleton_regressor(
feed, img_key=IN('img'), joints_key=NET_OUT('heats'))
},
model_name=model,
model_path=joint_locators_path(),
epochs=epochs,
enable_telegram_log=False)
.shift_sat(augmentation_prob)
.shift_val(augmentation_prob),
regularizer=Regularizer().normalize() if normalize else None,
keyset={IN('img')})
data_processing_plan.add_outer(key=IN('img'), fun=lambda x: preprocess_mobile(255*x))
model = train_model(model_generator=lambda: transfer_mobile_net_joints(# input_shape=np.shape(dm.train()[0][IN('img')][0]),
dropout_rate=drate,
train_mobilenet=retrain,
# activation=K.layers.LeakyReLU(alpha=leaky_slope)
),
dataset_manager=dm,
loss={OUT('heats'): heatmap_loss,
OUT('vis'): 'binary_crossentropy'},
learning_rate=learning_rate,
patience=patience,
data_processing_plan=data_processing_plan,
tb_path="joints",
tb_plots={'target': lambda feed: joint_skeleton_impression(feed,
img_key=IN('img'),
heats_key=OUT('heats'),
vis_key=OUT('vis')),
'output': lambda feed: joint_skeleton_impression(feed,
img_key=IN('img'),
heats_key=NET_OUT('heats'),
vis_key=NET_OUT('vis')),
},
model_name=model,
model_path=joint_locators_path(),
epochs=epochs,
enable_telegram_log=True)
# Plan the processing needed before providing inputs and outputs for training and validation
data_processing_plan = ProcessingPlan(augmenter=Augmenter().shift_hue(augmentation_prob)
.shift_sat(augmentation_prob)
.shift_val(augmentation_prob),
regularizer=Regularizer().normalize() if normalize else None,
keyset={IN(0)}) # Today we just need to augment one input...
model1 = train_model(model_generator=lambda: normalized_convs(input_shape=np.shape(generator.train()[0][IN(0)])[1:],
dropout_rate=drate,
weight_decay=weight_decay,
activation=lambda: K.layers.LeakyReLU(alpha=leaky_slope)),
dataset_manager=generator,
loss={OUT(0): lambda x, y: prop_heatmap_penalized_fp_loss(x, y,
white_priority=white_priority,
delta=delta)
},
learning_rate=learning_rate,
patience=patience,
data_processing_plan=data_processing_plan,
tb_path="heat_maps/",
tb_plots={'plain_input': lambda feed: feed[IN(0)],
'plain_target': lambda feed: feed[OUT(0)],
'plain_output': lambda feed: feed[NET_OUT(0)],
'combined_mask': lambda feed: get_image_with_mask(feed[IN(0)],
feed[NET_OUT(0)]),
'crops': crop_sprite},
model_name=model,
model_path=croppers_path(),
epochs=epochs,
enable_telegram_log=True)
model1 = train_model(
model_generator=lambda:
transfer_mobile_net( # input_shape=np.shape(generator.train()[0][IN(0)])[1:],
# weight_decay=weight_decay,
# train_vgg=retrain_vgg_model,
train_mobilenet=retrain_vgg_model,
dropout_rate=drate),
dataset_manager=generator,
loss={
OUT(0):
lambda x, y: prop_heatmap_penalized_fp_loss(
x, y, white_priority=white_priority, delta=delta)
},
learning_rate=learning_rate,
patience=patience,
data_processing_plan=data_processing_plan,
tb_path="heat_maps/",
tb_plots={
'plain_input':
lambda feed: feed[IN(0)],
'plain_target':
lambda feed: feed[OUT(0)],
'plain_output':
lambda feed: feed[NET_OUT(0)],
'combined_mask':
lambda feed: get_image_with_mask(feed[IN(0)], feed[NET_OUT(0)]),
'crops':
crop_sprite
},
model_name=model,
model_path=croppers_path(),
epochs=epochs,
enable_telegram_log=True)