Skip to content

Commit

Permalink
Removed casting to np array of layer_types + layer_params because of …
Browse files Browse the repository at this point in the history 
…error
  • Loading branch information
@konegen
konegen committed Nov 6, 2024
1 parent 0aa0ecb commit 275f768
Show file tree
Hide file tree
Showing 5 changed files with 341 additions and 199 deletions.
161 changes: 103 additions & 58 deletions src/pruning.py
View file
Original file line number Diff line number Diff line change
@@ -1,25 +1,33 @@
'''Copyright [2020] Hahn-Schickard-Gesellschaft fuer angewandte Forschung e.V.,
"""Copyright [2020] Hahn-Schickard-Gesellschaft fuer angewandte Forschung e.V.,
Daniel Konegen + Marcus Rueb
Copyright [2021] Karlsruhe Institute of Technology, Daniel Konegen
Copyright [2022] Hahn-Schickard-Gesellschaft fuer angewandte Forschung e.V.,
Daniel Konegen + Marcus Rueb
SPDX-License-Identifier: Apache-2.0
============================================================================'''
============================================================================"""

import os

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import load_model
from sklearn.model_selection import train_test_split
from tensorflow.keras.models import load_model

from pruning_helper_classes import *
from pruning_helper_functions import *
from pruning_helper_functions_dense import *
from pruning_helper_functions_conv import *
from pruning_helper_functions_dense import *


def factor_pruning(keras_model, prun_factor_dense=10, prun_factor_conv=10,
metric='L1', comp=None, num_classes=None, label_one_hot=None):
def factor_pruning(
keras_model,
prun_factor_dense=10,
prun_factor_conv=10,
metric="L1",
comp=None,
num_classes=None,
label_one_hot=None,
):
"""
A given keras model get pruned. The factor for dense and conv says how
many percent of the dense and conv layers should be deleted.
Expand Down Expand Up @@ -48,45 +56,67 @@ def factor_pruning(keras_model, prun_factor_dense=10, prun_factor_conv=10,

if num_classes <= 2 and comp is None:
comp = {
"optimizer": 'adam',
"optimizer": "adam",
"loss": tf.keras.losses.BinaryCrossentropy(),
"metrics": 'accuracy'}
"metrics": "accuracy",
}
elif num_classes > 3 and comp is None:
if label_one_hot:
comp = {
"optimizer": 'adam',
"optimizer": "adam",
"loss": tf.keras.losses.CategoricalCrossentropy(),
"metrics": 'accuracy'}
"metrics": "accuracy",
}
else:
comp = {
"optimizer": 'adam',
"optimizer": "adam",
"loss": tf.keras.losses.SparseCategoricalCrossentropy(),
"metrics": 'accuracy'}
"metrics": "accuracy",
}

layer_types, layer_params, layer_output_shape, layer_bias, netstr = (
load_model_param(model))
load_model_param(model)
)
num_new_neurons = np.zeros(shape=len(layer_params), dtype=np.int16)
num_new_filters = np.zeros(shape=len(layer_params), dtype=np.int16)

layer_params, num_new_neurons, num_new_filters, layer_output_shape = (
model_pruning(layer_types, layer_params, layer_output_shape,
layer_bias, netstr, num_new_neurons, num_new_filters,
prun_factor_dense, prun_factor_conv, metric))
layer_params, num_new_neurons, num_new_filters, layer_output_shape = model_pruning(
layer_types,
layer_params,
layer_output_shape,
layer_bias,
netstr,
num_new_neurons,
num_new_filters,
prun_factor_dense,
prun_factor_conv,
metric,
)

print("Finish with pruning")

pruned_model = build_pruned_model(model, layer_params, layer_types,
num_new_neurons, num_new_filters, comp)
pruned_model = build_pruned_model(
model, layer_params, layer_types, num_new_neurons, num_new_filters, comp
)

print("Model built")

return pruned_model


def accuracy_pruning(keras_model, comp, x_train, y_train=None, x_val=None,
y_val=None, pruning_acc=None, max_acc_loss=5,
num_classes=None, label_one_hot=None,
data_loader_path=None):
def accuracy_pruning(
keras_model,
comp,
x_train,
y_train=None,
x_val=None,
y_val=None,
pruning_acc=None,
max_acc_loss=5,
num_classes=None,
label_one_hot=None,
data_loader_path=None,
):
"""
A given keras model gets pruned. Either an accuracy value (in %) can be
specified, which the minimized model has to still achieve. Or the maximum
Expand Down Expand Up @@ -135,16 +165,14 @@ def accuracy_pruning(keras_model, comp, x_train, y_train=None, x_val=None,
else:
if data_loader_path == None or os.path.isfile(data_loader_path):
if x_val is not None and y_val is not None:
original_model_acc = original_model.evaluate(
x_val, y_val)[-1]
original_model_acc = original_model.evaluate(x_val, y_val)[-1]
else:
x_train, x_val, y_train, y_val = train_test_split(
x_train, y_train, test_size=0.2)
original_model_acc = original_model.evaluate(
x_val, y_val)[-1]
x_train, y_train, test_size=0.2
)
original_model_acc = original_model.evaluate(x_val, y_val)[-1]
elif os.path.isdir(data_loader_path):
original_model_acc = original_model.evaluate(
x_val)[-1]
original_model_acc = original_model.evaluate(x_val)[-1]
print("Start model accuracy: " + str(original_model_acc * 100) + "%")
req_acc = original_model_acc - (max_acc_loss / 100)

Expand All @@ -156,23 +184,36 @@ def accuracy_pruning(keras_model, comp, x_train, y_train=None, x_val=None,

print("Next pruning factors: " + str(pruning_factor))

model = factor_pruning(original_model, prun_factor_dense=pruning_factor,
prun_factor_conv=pruning_factor, metric='L1',
comp=comp, num_classes=num_classes,
label_one_hot=label_one_hot)
model = factor_pruning(
original_model,
prun_factor_dense=pruning_factor,
prun_factor_conv=pruning_factor,
metric="L1",
comp=comp,
num_classes=num_classes,
label_one_hot=label_one_hot,
)

if data_loader_path == None or os.path.isfile(data_loader_path):
history = model.fit(x=x_train, y=y_train, batch_size=64,
validation_data=(x_val, y_val),
epochs=train_epochs, callbacks=callbacks)
history = model.fit(
x=x_train,
y=y_train,
batch_size=64,
validation_data=(x_val, y_val),
epochs=train_epochs,
callbacks=callbacks,
)
elif os.path.isdir(data_loader_path):
history = model.fit_generator(
x_train, steps_per_epoch=len(x_train),
x_train,
steps_per_epoch=len(x_train),
validation_data=x_val,
validation_steps=len(x_val),
epochs=train_epochs, callbacks=callbacks)
epochs=train_epochs,
callbacks=callbacks,
)

if history.history['val_accuracy'][-1] < req_acc:
if history.history["val_accuracy"][-1] < req_acc:
# Required accuracy is not reached
if lowest_pruning_factor_not_working > pruning_factor:
lowest_pruning_factor_not_working = pruning_factor
Expand All @@ -182,8 +223,10 @@ def accuracy_pruning(keras_model, comp, x_train, y_train=None, x_val=None,
return original_model

if last_pruning_step == 2:
print("Pruningfactor dense and conv: " +
str(pruning_factor - last_pruning_step))
print(
"Pruningfactor dense and conv: "
+ str(pruning_factor - last_pruning_step)
)
return pruned_model
elif last_pruning_step == 5:
pruning_factor -= 3
Expand All @@ -199,40 +242,42 @@ def accuracy_pruning(keras_model, comp, x_train, y_train=None, x_val=None,
# Required accuracy is reached
pruned_model = model
# Set pruning factor for next pruning step
if (len(history.history['val_accuracy']) <=
int(0.3 * train_epochs)):
if len(history.history["val_accuracy"]) <= int(0.3 * train_epochs):
pruning_factor += 15
last_pruning_step = 15
elif (len(history.history['val_accuracy']) <=
int(0.5 * train_epochs)):
elif len(history.history["val_accuracy"]) <= int(0.5 * train_epochs):
pruning_factor += 10
last_pruning_step = 10
elif (len(history.history['val_accuracy']) <=
int(0.7 * train_epochs)):
elif len(history.history["val_accuracy"]) <= int(0.7 * train_epochs):
pruning_factor += 5
last_pruning_step = 5
elif (len(history.history['val_accuracy']) >
int(0.7 * train_epochs)):
elif len(history.history["val_accuracy"]) > int(0.7 * train_epochs):
pruning_factor += 2
last_pruning_step = 2

if lowest_pruning_factor_not_working < pruning_factor:
# Check if pruning factor is higher than the lowest one which
# didn't work and adjust the pruning factor if it's true
if (lowest_pruning_factor_not_working -
(pruning_factor - last_pruning_step) <= 2):
print("Pruningfactor dense and conv: " +
str(pruning_factor - last_pruning_step))
if (
lowest_pruning_factor_not_working - (pruning_factor - last_pruning_step)
<= 2
):
print(
"Pruningfactor dense and conv: "
+ str(pruning_factor - last_pruning_step)
)
return pruned_model
elif (lowest_pruning_factor_not_working -
(pruning_factor - last_pruning_step) <= 5):
elif (
lowest_pruning_factor_not_working - (pruning_factor - last_pruning_step)
<= 5
):
pruning_factor = (pruning_factor - last_pruning_step) + 2
last_pruning_step = 2

if all_pruning_factors.count(pruning_factor) >= 1:
# Check if the pruning factor for next iteration was already
# applied
if history.history['val_accuracy'][-1] < req_acc:
if history.history["val_accuracy"][-1] < req_acc:
# If required accuracy wasn't reached, the pruning factor is
# lowered in the step before. If the new pruning factor was
# already applied, this is one which worked, so you increase
Expand Down Expand Up @@ -263,4 +308,4 @@ def accuracy_pruning(keras_model, comp, x_train, y_train=None, x_val=None,

all_pruning_factors.append(pruning_factor)

return pruned_model
return pruned_model
7 changes: 4 additions & 3 deletions src/pruning_helper_classes.py
View file
Original file line number Diff line number Diff line change
@@ -1,14 +1,15 @@
'''Copyright [2020] Hahn-Schickard-Gesellschaft fuer angewandte Forschung e.V.,
"""Copyright [2020] Hahn-Schickard-Gesellschaft fuer angewandte Forschung e.V.,
Daniel Konegen + Marcus Rueb
Copyright [2021] Karlsruhe Institute of Technology, Daniel Konegen
Copyright [2022] Hahn-Schickard-Gesellschaft fuer angewandte Forschung e.V.,
Daniel Konegen + Marcus Rueb
SPDX-License-Identifier: Apache-2.0
============================================================================'''
============================================================================"""

import tensorflow as tf
from typing import NamedTuple

import tensorflow as tf


class NetStructure(NamedTuple):
parents: list
Expand Down
Loading

0 comments on commit 275f768

Please sign in to comment.