def pruning(arch_name, masks):
masks_file = ensure_dir(root_dir() / "masks" / (arch_name + ".pkl"))
if masks_file.exists():
with open(masks_file, "rb") as file:
mask_values = pickle.load(file)
else:
mask_values = {}
pretrained_model_dir = root_dir() / "downloads" / "model" / arch_name
model_dir = root_dir() / "train" / arch_name
warm_start_hook = WarmStartHook(pretrained_model_dir, model_dir)
pruning_hook = PruningHook(masks, mask_values)
estimator = tf.estimator.Estimator(
model_fn=partial(
model_fn,
pruning_hook=pruning_hook,
pretrained_model_dir=pretrained_model_dir,
),
model_dir=model_dir,
)
input_fn = tf.estimator.inputs.numpy_input_fn(
np.random.rand(*input_shapes[arch_name]).astype("f"),
np.random.randint(1000, size=(1, 1)),
batch_size=1,
shuffle=False,
)
estimator.train(input_fn=input_fn, hooks=[warm_start_hook, pruning_hook])
with open(masks_file, "wb") as file:
pickle.dump(pruning_hook.mask_values, file)
return
def execute(self, note):
main_panel = note.getBody()
# first running the webserver
# self.facade.registerProxy(model.WebServerProxy(main_panel,
# config={'host': '127.0.0.1', 'port': 12345,
# #'path': '%s/ui/html' % utils.root_dir()
# 'path': None,
# },
# http_root_obj=http_root.MainHttpRoot(),
# orm_base_obj=vo.Base,
# db_plugin=utils.db_str_conn('mainapp.db')),)
self.facade.registerProxy(model.WebServerProxy(main_panel))
with open("%s/plugins.yaml" % utils.root_dir()) as stream:
cfg = yaml.safe_load(stream)
if cfg:
# register startup class
for key, val in cfg.iteritems():
try:
plugin_facade_cls = utils.class_for_name(val["module"] + ".main", "PluginFacade")
facade = plugin_facade_cls.getInstance(key=key)
if facade:
logging.info("%s: plugin loaded" % key)
facade.PLUGIN_NAME = key
# calling startup
facade.sendNotification(facade.STARTUP, main_panel)
except Exception, e:
logging.info("%s: %s" % (key, e.message))
def load_form(self, params):
current_path = utils.root_dir()
if 'postfix_path' in params:
current_path += params["postfix_path"]
variables = params["vars"] if "vars" in params else {}
content = pages.retrieve(params["name"], BASE_PATH=current_path, ASIDE_JS=None, TITLE=params["title"],
**variables)
self.viewComponent.webView.setHtml(*content)
def execute(self, note):
main_panel = note.getBody()
main_panel.on_shutdown.signal.connect(StartupCommand._on_shutdown)
main_facade = AsideFacade.getInstance(key=MAIN_APP_KEY)
web_proxy = main_facade.retrieveProxy(model.WebServerProxy.NAME)
path = "%s/%s" % (utils.root_dir(), self.facade.PLUGIN_DIR)
static_dir = '%s/html' % (path,)
web_proxy.tree_mount(self.facade.PLUGIN_NAME, http_root.QueueHTTPRoot(), static_dir,
orm_base_obj=vo.Base,
db_plugin=utils.db_str_conn("queue.db", path=path))
from utils import root_dir, nice_json
from flask import Flask
from werkzeug.exceptions import NotFound
import json
app = Flask(__name__)
with open("{}/database/showtimes.json".format(root_dir()), "r") as f:
showtimes = json.load(f)
# no argument was passed, so this is a dummy placeholder
@app.route("/", methods=['GET'])
def hello():
return nice_json({
"uri": "/",
"subresource_uris": {
"showtimes": "/showtimes",
"showtime": "/showtimes/<date>"
}
})
# if no date arugment is passed, then return all possible showtimes
@app.route("/showtimes", methods=['GET'])
def showtimes_list():
return nice_json(showtimes)
# if a date argument is passed, then use it to fetch the approprate key/value pair
@app.route("/showtimes/<date>", methods=['GET'])
def test_root_dir(self):
# FIXME: How to test this without either hardcoding the path or
# doing the same work as root_dir() ?
self.assertEqual(utils.root_dir(),
"/Users/stain/Documents/i/src/forgetmp3")
def download_tf_model(arch_name, model_dir):
full_model_dir = root_dir() / "downloads" / model_dir
if not full_model_dir.exists():
full_model_dir.mkdir(mode=0o755, parents=True, exist_ok=True)
if not (full_model_dir / arch_name / "checkpoint").exists():
download_arch(arch_name, str(full_model_dir / arch_name) + "/")
def main(_):
now = time.strftime("%Y-%m-%d-%H_%M_%S", time.localtime(time.time()))
arch_name = FLAGS.arch_name
if FLAGS.init_checkpoint == None or FLAGS.init_checkpoint == "":
pretrained_model_dir = root_dir() / "downloads" / "model" / arch_name
else:
pretrained_model_dir = FLAGS.init_checkpoint
if FLAGS.pre_masks_dir == None or FLAGS.init_checkpoint == "":
all_mask_values = {}
else:
with open(FLAGS.pre_masks_dir, "rb") as file:
all_mask_values = pickle.load(file)
if FLAGS.output_dir == None:
model_dir = root_dir() / "train" / arch_name / now
else:
model_dir = root_dir() / "train" / FLAGS.output_dir
if FLAGS.data_dir != None:
data_dir = FLAGS.data_dir
else:
raise ValueError("data_dir must set")
if FLAGS.pruning_type != "":
pruning_type = FLAGS.pruning_type
else:
raise ValueError("pruning_type must set")
tf.logging.info("Pruning_type : %s\n" % pruning_type)
if "tw" in pruning_type or "bw" in pruning_type:
masks = [
"vgg_16/conv1/conv1_1/weights:0",
"vgg_16/conv1/conv1_2/weights:0",
"vgg_16/conv2/conv2_1/weights:0",
"vgg_16/conv2/conv2_2/weights:0",
"vgg_16/conv3/conv3_1/weights:0",
"vgg_16/conv3/conv3_2/weights:0",
"vgg_16/conv3/conv3_3/weights:0",
"vgg_16/conv4/conv4_1/weights:0",
"vgg_16/conv4/conv4_2/weights:0",
"vgg_16/conv4/conv4_3/weights:0",
"vgg_16/conv5/conv5_1/weights:0",
"vgg_16/conv5/conv5_2/weights:0",
"vgg_16/conv5/conv5_3/weights:0",
# "vgg_16/fc6/weights:0",
# "vgg_16/fc7/weights:0",
# "vgg_16/fc8/weights:0",
]
else:
masks = [
"vgg_16/conv1/conv1_1/weights:0",
"vgg_16/conv1/conv1_2/weights:0",
"vgg_16/conv2/conv2_1/weights:0",
"vgg_16/conv2/conv2_2/weights:0",
"vgg_16/conv3/conv3_1/weights:0",
"vgg_16/conv3/conv3_2/weights:0",
"vgg_16/conv3/conv3_3/weights:0",
"vgg_16/conv4/conv4_1/weights:0",
"vgg_16/conv4/conv4_2/weights:0",
"vgg_16/conv4/conv4_3/weights:0",
"vgg_16/conv5/conv5_1/weights:0",
"vgg_16/conv5/conv5_2/weights:0",
"vgg_16/conv5/conv5_3/weights:0",
# "vgg_16/fc6/weights:0",
# "vgg_16/fc7/weights:0",
# "vgg_16/fc8/weights:0",
]
if arch_name == "alexnet":
masks = [
"conv2d/kernel:0",
"conv2d_1/kernel:0",
"conv2d_2/kernel:0",
"conv2d_3/kernel:0",
"conv2d_4/kernel:0",
# "dense/kernel:0",
# "dense_1/kernel:0",
# "dense_2/kernel:0",
]
batch_size = int(FLAGS.batch_size)
num_epochs = int(FLAGS.num_train_epochs)
if FLAGS.score_type == "weight":
is_masked_weight = True
else:
is_masked_weight = False
pruning_steps = int(FLAGS.pruning_steps)
finetune_steps = int(FLAGS.finetune_steps)
mini_finetune_steps = int(FLAGS.mini_finetune_steps)
if is_masked_weight:
tf.logging.info("Pruning score : Weight ! \n")
else:
tf.logging.info("Pruning score : Taylor ! \n")
if "tw" in pruning_type:
pruning_layers = [[0, 1], [2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]]
sparsity_stages = [25, 40, 50, 60, 65, 70, 75, 80, 85, 90, 92, 94, 96]
if "ew" in pruning_type:
pruning_layers = [[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]
sparsity_stages = [25, 40, 50, 60, 65, 70, 75, 80, 85, 90, 92, 94, 96]
if "vw" in pruning_type:
pruning_layers = [[0, 1], [2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]]
sparsity_stages = [25, 40, 50, 60, 65, 70, 75, 80, 85, 90, 92, 94, 96]
if "bw" in pruning_type:
pruning_layers = [[0, 1], [2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]]
sparsity_stages = [25, 40, 50, 60, 65, 70, 75, 80, 85, 90, 92, 94, 96]
early_stop = [100, 100, 100, 100, 100, 100]
previous_checkpoint = tf.train.latest_checkpoint(pretrained_model_dir)
previous_mask_values = copy.deepcopy(all_mask_values)
# Set hooks
pruning_hook = PruningHook(masks, pruning_type, is_masked_weight,
previous_mask_values)
# Set up training hook that logs the training accuracy every 100 steps.
tensors_to_log = {
"train_accuracy": "train_accuracy",
"train_accuracy_top_5": "train_accuracy_top_5",
}
logging_hook = tf.train.LoggingTensorHook(tensors=tensors_to_log,
every_n_iter=100)
session_config = tf.ConfigProto(allow_soft_placement=True)
session_config.gpu_options.allow_growth = True
estimator_config = tf.estimator.RunConfig(
session_config=session_config,
save_checkpoints_secs=60 * 60 * 24 * 5, # 5 days
save_checkpoints_steps=None,
keep_checkpoint_max=len(sparsity_stages) + 3,
)
if arch_name == "vgg16":
pre_fn = preprocess_image
if arch_name == "alexnet":
pre_fn = alexnet_preprocess_image
# Set input_fn
def eval_input_fn():
return imagenet.test(
data_dir,
batch_size,
num_parallel_calls=40,
transform_fn=lambda ds: ds.map(lambda image, label:
(image, label - 1)),
preprocessing_fn=pre_fn)
def train_input_fn():
return imagenet.train(
data_dir,
batch_size,
is_shuffle=True,
num_epochs=num_epochs,
num_parallel_calls=40,
transform_fn=lambda ds: ds.map(lambda image, label:
(image, label - 1)),
preprocessing_fn=pre_fn)
warm_start_settings = tf.estimator.WarmStartSettings(previous_checkpoint)
estimator = tf.estimator.Estimator(
model_fn=partial(model_fn, pruning_hook=None, is_pruning=False),
config=estimator_config,
warm_start_from=warm_start_settings,
)
tf.logging.info('First Evaluation !!!')
eval_results = estimator.evaluate(input_fn=eval_input_fn)
tf.logging.info('Done First Evaluation !!!\n')
tf.logging.info("First Evaluation results:\n\n\t%s\n" % eval_results)
## Multi stages to prune the CNN model
for stage in range(len(sparsity_stages)):
stage_dir = model_dir / ("sparsity_stage_" +
str(sparsity_stages[stage]))
sparsity = sparsity_stages[stage]
pruning_hook.sparsity = sparsity
tf.logging.info("\n\n\n\n\n\n\n\n\n\n\n\n")
tf.logging.info("Stage : %d, Sparsity : %f, Finetune_steps : %d" %
(stage, sparsity, finetune_steps))
tf.logging.info("early_stop:")
tf.logging.info(early_stop)
is_early_stop = True
for layer in range(len(pruning_layers)):
# if "ew" in pruning_type:
# sparsity = sparsity_layers[layer]
pruning_hook.sparsity = sparsity
tf.logging.info("\n\n\n\n\n")
if early_stop[layer] < sparsity:
continue
is_early_stop = False
layer_dir = stage_dir / ("layer_" + str(layer))
masks_dir = layer_dir / "masks"
os.makedirs(masks_dir)
masks_now = pruning_layers[layer]
assert (previous_checkpoint)
warm_start_settings = tf.estimator.WarmStartSettings(
previous_checkpoint)
pruning_hook.masks_now = masks_now
pruning_hook.mask_values = copy.deepcopy(previous_mask_values)
# Set estimator
pruning_estimator = tf.estimator.Estimator(
model_fn=partial(model_fn,
pruning_hook=pruning_hook,
is_pruning=True),
model_dir=layer_dir,
warm_start_from=warm_start_settings,
config=estimator_config,
)
tf.logging.info('Prune !!!')
pruning_hook.initialize(pruning_steps)
pruning_estimator.train(input_fn=train_input_fn,
hooks=[logging_hook, pruning_hook],
steps=pruning_steps)
tf.logging.info('Done Prune !!!\n')
# Set fine-tune estimator
# The mask_values must be updated by estimator_pruning.train().
# The mask_values in estimator is constant,
# then we build the estimator in every iteration for fine-tune.
estimator = tf.estimator.Estimator(
model_fn=partial(model_fn,
pruning_hook=pruning_hook,
is_pruning=False),
model_dir=layer_dir,
config=estimator_config,
)
mini_train_times = 0
while (True):
tf.logging.info('Mini Fine Tune !!!')
estimator.train(input_fn=train_input_fn,
hooks=[logging_hook],
steps=mini_finetune_steps)
tf.logging.info('Done Mini Fine Tune !!!\n')
tf.logging.info('Evaluation !!!')
eval_results = estimator.evaluate(input_fn=eval_input_fn, )
tf.logging.info(
"Pruning Sparsity %d Layer %d Evaluation results:\n\n\t%s\n"
% (sparsity, layer, eval_results))
tf.logging.info('Done Evaluation !!!\n')
threshold = 0.8885
if sparsity > 75:
threshold = 0.88818
if sparsity > 85:
threshold = 0.0
if arch_name == "alexnet":
threshold = 0.775
if eval_results["accuracy_top_5"] > threshold:
previous_checkpoint = tf.train.latest_checkpoint(layer_dir)
previous_mask_values = copy.deepcopy(
pruning_hook.mask_values)
tf.logging.info('GOOD CHECKPOINT!!!\n')
with open(
masks_dir /
("goog_mask_" + str(sparsity) + ".pkl"), "wb") as file:
pickle.dump(pruning_hook.mask_values, file)
break
else:
tf.logging.info('BAD CHECKPOINT!!!\n')
with open(
masks_dir / ("bad_mask_" + str(sparsity) + ".pkl"),
"wb") as file:
pickle.dump(pruning_hook.mask_values, file)
mini_train_times += 1
if (mini_train_times > 8):
early_stop[layer] = sparsity
break
if is_early_stop:
break
tf.logging.info("\n\n\n\n\n")
all_layer_dir = stage_dir / "all_layers"
masks_dir = all_layer_dir / "masks"
os.makedirs(masks_dir)
assert (previous_checkpoint)
warm_start_settings = tf.estimator.WarmStartSettings(
previous_checkpoint)
pruning_hook.mask_values = copy.deepcopy(previous_mask_values)
# Set fine-tune estimator
# The mask_values must be updated by estimator_pruning.train().
# The mask_values in estimator is constant,
# then we build the estimator in every iteration for fine-tune.
estimator = tf.estimator.Estimator(
model_fn=partial(
model_fn,
pruning_hook=pruning_hook,
is_pruning=False,
),
model_dir=all_layer_dir,
config=estimator_config,
warm_start_from=warm_start_settings,
)
tf.logging.info('Fine Tune !!!')
estimator.train(input_fn=train_input_fn,
hooks=[logging_hook],
steps=finetune_steps)
tf.logging.info('Done Fine Tune !!!\n')
tf.logging.info('Last Evaluation !!!')
eval_results = estimator.evaluate(input_fn=eval_input_fn)
tf.logging.info('Done Evaluation !!!\n')
tf.logging.info("After Pruning Evaluation results:\n\n\t%s\n" %
eval_results)
previous_checkpoint = tf.train.latest_checkpoint(all_layer_dir)
previous_mask_values = copy.deepcopy(pruning_hook.mask_values)
with open(masks_dir / ("mask_" + str(sparsity) + ".pkl"),
"wb") as file:
pickle.dump(pruning_hook.mask_values, file)
return
from utils import root_dir, nice_json
from flask import Flask
from werkzeug.exceptions import NotFound
import json
app = Flask(__name__)
# load data into memory;
with open("{}/database/movies.json".format(root_dir()), "r") as f:
movies = json.load(f)
# app.route binds url to function; a good example of an decorator
@app.route("/", methods=['GET'])
def hello():
return nice_json({
"uri": "/",
"subresource_uris": {
"movies": "/movies",
"movie": "/movies/<id>"
}
})
# triggered by http://127.0.0.1:5001/movies/7daf7208-be4d-4944-a3ae-c1c2f516f3e6
@app.route("/movies/<movieid>", methods=['GET'])
def movie_info(movieid):
if movieid not in movies:
raise NotFound
result = movies[movieid]
result["uri"] = "/movies/{}".format(movieid)
