print("Spatial stream")
spatial_model_restored = legacy_load_model(
filepath="spatial.h5",
custom_objects={
'sparse_categorical_cross_entropy_loss':
sparse_categorical_cross_entropy_loss,
"acc_top_1": acc_top_1,
"acc_top_5": acc_top_5
})
spatial_model_restored.summary()
# evaluate
_, spatial_test_loader, test_video_level_label = frame_dataloader.SpatialDataLoader(
width=int(spatial_model_restored.inputs[0].shape[1]),
height=int(spatial_model_restored.inputs[0].shape[2]),
batch_size=get_batch_size(spatial_model_restored, spatial=True),
testing_samples_per_video=19).run()
video_level_loss, video_level_accuracy_1, video_level_accuracy_5, test_video_level_preds = eval_model(
spatial_model_restored, spatial_test_loader, test_video_level_label, 19)
print("Spatial Model validation", "prec@1", video_level_accuracy_1, "prec@5",
video_level_accuracy_5, "loss", video_level_loss)
"""
Evaluate motion stream
"""
# download
# drive_manager = DriveManager("heavy-mot-xception-adam-1e-05-imnet")
# drive_manager.download_file('1kvslNL8zmZYaHRmhgAM6-l_pNDDA0EKZ', "motion.zip") # the id of the zip file contains my network
# load into ram
print("Motion stream")
pred_file=pred_file, h5py_file=h5py_file, drive_manager=drive_manager, log_file=log_file)
data_loader = partial(frame_dataloader.SpatialDataLoader,
testing_samples_per_video=testing_samples_per_video,
augmenter_level=augmenter_level,
log_stream=log_stream)
if checkpoint_found:
# restore the model
print("Model restored")
eval_globals.best_video_level_accuracy_1 = float(zip_file_name.split("-")[1])
print("Current Best", eval_globals.best_video_level_accuracy_1)
spatial_model_restored = legacy_load_model(filepath=h5py_file, custom_objects={'sparse_categorical_cross_entropy_loss': sparse_categorical_cross_entropy_loss, "acc_top_1": acc_top_1, "acc_top_5": acc_top_5})
# init data loader
train_loader, test_loader, test_video_level_label = data_loader(width=int(spatial_model_restored.inputs[0].shape[1]), height=int(spatial_model_restored.inputs[0].shape[2]), batch_size=get_batch_size(spatial_model_restored, spatial=True)).run()
# training
spatial_model_restored.fit_generator(train_loader,
steps_per_epoch=len(train_loader), # generates a batch per step
epochs=epochs,
use_multiprocessing=True, workers=workers,
# validation_data=gen_test(), validation_steps=len(test_loader.dataset)
callbacks=[SpatialValidationCallback(model=spatial_model_restored, test_loader=test_loader, test_video_level_label=test_video_level_label), # returns callback instance
keras.callbacks.ReduceLROnPlateau(monitor='val_loss', patience=validate_every, verbose=1)],
initial_epoch=int(zip_file_name.split("-")[0])) # get epoch number
else:
# init the model
print("Starting from scratch")
data_loader = partial(frame_dataloader.MotionDataLoader,
testing_samples_per_video=testing_samples_per_video,
augmenter_level=augmenter_level,
log_stream=log_stream, stacked_frames=stacked_frames)
if checkpoint_found:
# restore the model from the checkpoint
log("Model restored")
eval_globals.best_video_level_accuracy_1 = float(zip_file_name.split("-")[1])
log("Current Best", eval_globals.best_video_level_accuracy_1)
motion_model_restored = legacy_load_model(filepath=h5py_file, custom_objects={'sparse_categorical_cross_entropy_loss': sparse_categorical_cross_entropy_loss, "acc_top_1": acc_top_1, "acc_top_5": acc_top_5})
# init data loader
train_loader, test_loader, test_video_level_label = data_loader(width=int(motion_model_restored.inputs[0].shape[1]),
height=int(motion_model_restored.inputs[0].shape[2]),
batch_size=get_batch_size(motion_model_restored,
spatial=False)).run()
# training
motion_model_restored.fit_generator(train_loader,
steps_per_epoch=len(train_loader), # generates a batch per step
epochs=epochs,
use_multiprocessing=True, workers=workers,
# validation_data=gen_test(), validation_steps=len(test_loader.dataset)
callbacks=[ MotionValidationCallback(model=motion_model_restored, test_loader=test_loader, test_video_level_label=test_video_level_label), # returns callback instance
keras.callbacks.ReduceLROnPlateau(monitor='val_loss', patience=validate_every, verbose=1)],
initial_epoch=int(zip_file_name.split("-")[0])) # get epoch number
else:
# init the model from scratch
log("Starting from scratch")
zip_file_name.split("-")[1])
print("Current Best", eval_globals.best_video_level_accuracy_1)
spatial_model_restored = legacy_load_model(
filepath=h5py_file,
custom_objects={
'sparse_categorical_cross_entropy_loss':
sparse_categorical_cross_entropy_loss,
"acc_top_1": acc_top_1,
"acc_top_5": acc_top_5
})
# init data loader
train_loader, test_loader, test_video_level_label = data_loader(
width=int(spatial_model_restored.inputs[0].shape[1]),
height=int(spatial_model_restored.inputs[0].shape[2]),
batch_size=get_batch_size(spatial_model_restored, spatial=True)).run()
# training
spatial_model_restored.fit_generator(
train_loader,
steps_per_epoch=len(train_loader), # generates a batch per step
epochs=epochs,
use_multiprocessing=False,
workers=workers,
# validation_data=gen_test(), validation_steps=len(test_loader.dataset)
callbacks=[
SpatialValidationCallback(
model=spatial_model_restored,
test_loader=test_loader,
test_video_level_label=test_video_level_label
), # returns callback instance