def get_2dcnn_wahl_model(feature_set_training, label_set_training, param):
'''
Get 2D convolution based neural network based model
:param feature_set_training:
:param label_set_training:
:param param:
:return:
'''
input_feature = feature_set_training
descriptor_size = input_feature.shape[1]
N_classes = label_set_training.shape[1]
assert (N_classes > 1) # TOP HAT NOTATION REQUIRED!
# input layer
x = tf.placeholder(tf.float32, shape=[None, descriptor_size])
y_ = tf.placeholder(tf.float32, shape=[None, N_classes])
x_image = tf.reshape(
x, [-1, 6, 6, 6, 6, 1]) # be careful of shape consistency
###### Reshaping
rows_num, cols_num = common_utils_nn.find_best_square_reshape_from_array(
descriptor_size)
print('Using desc size %d, reshaping into rows %d x cols %d' %
(descriptor_size, rows_num, cols_num))
h_pool1_reshape = tf.reshape(x_image, [-1, rows_num, cols_num, 1])
###### conv layer 1
W_conv1 = common_utils_nn.weight_variable(
[5, 5, 1, param.layer1_feature_maps_num], param.init_stddev)
b_conv1 = common_utils_nn.bias_variable([param.layer1_feature_maps_num])
h_conv1 = tf.nn.relu(
common_utils_nn.conv2d(h_pool1_reshape, W_conv1) + b_conv1)
###### conv layer 2
W_conv2 = common_utils_nn.weight_variable(
[5, 5, param.layer1_feature_maps_num, param.layer2_feature_maps_num],
param.init_stddev)
b_conv2 = common_utils_nn.bias_variable([param.layer2_feature_maps_num])
h_conv2 = tf.nn.relu(common_utils_nn.conv2d(h_conv1, W_conv2) + b_conv2)
###### conv layer 3
W_conv3 = common_utils_nn.weight_variable(
[5, 5, param.layer2_feature_maps_num, param.layer3_feature_maps_num],
param.init_stddev)
b_conv3 = common_utils_nn.bias_variable([param.layer3_feature_maps_num])
h_conv3 = tf.nn.relu(common_utils_nn.conv2d(h_conv2, W_conv3) + b_conv3)
# pooling
h_pool3 = common_utils_nn.max_pool_2x2(h_conv3)
# pooling again
h_pool_final_flat = tf.reshape(h_pool3, [
-1, h_pool3.shape[1].value * h_pool3.shape[2].value *
param.layer3_feature_maps_num
]) # be careful of shape consistency
###### fully connected layer 1
W_fc1 = common_utils_nn.weight_variable([
h_pool3.shape[1].value * h_pool3.shape[2].value *
param.layer3_feature_maps_num, param.layer4_fc_num
], param.init_stddev) # be careful of shape consistency
b_fc1 = common_utils_nn.bias_variable([param.layer4_fc_num])
h_fc1 = tf.nn.relu(tf.matmul(h_pool_final_flat, W_fc1) + b_fc1)
###### dropout layer
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = common_utils_nn.weight_variable([param.layer4_fc_num, N_classes],
param.init_stddev)
b_fc2 = common_utils_nn.bias_variable([N_classes])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
layers_var = {'h_conv1': h_conv1, 'h_conv2': h_conv2, 'h_conv3': h_conv3}
return x, y_conv, y_, keep_prob, layers_var
def get_3dcnn_our_model(feature_set_training, label_set_training, param):
input_feature = feature_set_training
desc_dims = np.array(([20, 4, 5, 3]))
descriptor_size = input_feature.shape[1]
N_classes = label_set_training.shape[1]
assert (N_classes > 1) # TOP HAT NOTATION REQUIRED!
# input layer
x = tf.placeholder(tf.float32, shape=[None, descriptor_size])
y_ = tf.placeholder(tf.float32, shape=[None, N_classes])
x_image = tf.reshape(
x, [-1, 20, 4, 5, 3, 1]) # be careful of shape consistency
###### Reshape into 3d
h_pool1_reshape = tf.reshape(x_image, [-1, 20, 4, 15, 1])
###### conv3d layer 1
W_conv1 = common_utils_nn.weight_variable(
[5, 5, 1, 1, param.layer1_feature_maps_num], param.init_stddev)
b_conv1 = common_utils_nn.bias_variable([param.layer1_feature_maps_num])
h_conv1 = tf.nn.relu(
common_utils_nn.conv3d(h_pool1_reshape, W_conv1) + b_conv1)
###### conv3d layer 2
W_conv2 = common_utils_nn.weight_variable([
5, 5, 1, param.layer1_feature_maps_num, param.layer2_feature_maps_num
], param.init_stddev)
b_conv2 = common_utils_nn.bias_variable([param.layer2_feature_maps_num])
h_conv2 = tf.nn.relu(common_utils_nn.conv3d(h_conv1, W_conv2) + b_conv2)
###### conv3d layer 3
W_conv3 = common_utils_nn.weight_variable([
5, 5, 1, param.layer2_feature_maps_num, param.layer3_feature_maps_num
], param.init_stddev)
b_conv3 = common_utils_nn.bias_variable([param.layer3_feature_maps_num])
h_conv3 = tf.nn.relu(common_utils_nn.conv3d(h_conv2, W_conv3) + b_conv3)
###### Reshape into 3d
rows_num, cols_num = common_utils_nn.find_best_square_reshape_from_array(
descriptor_size)
h_conv3_reshape = tf.reshape(h_conv3, [-1, rows_num, cols_num, 1])
# pooling
h_pool3 = common_utils_nn.max_pool_2x2(h_conv3_reshape)
# pooling again
h_pool_final_flat = tf.reshape(h_pool3, [
-1, h_pool3.shape[1].value * h_pool3.shape[2].value *
param.layer3_feature_maps_num
]) # be careful of shape consistency
###### fully connected layer 1
W_fc1 = common_utils_nn.weight_variable([
h_pool3.shape[1].value * h_pool3.shape[2].value *
param.layer3_feature_maps_num, param.layer4_fc_num
], param.init_stddev) # be careful of shape consistency
b_fc1 = common_utils_nn.bias_variable([param.layer4_fc_num])
h_fc1 = tf.nn.relu(tf.matmul(h_pool_final_flat, W_fc1) + b_fc1)
###### dropout layer
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = common_utils_nn.weight_variable([param.layer4_fc_num, N_classes],
param.init_stddev)
b_fc2 = common_utils_nn.bias_variable([N_classes])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
layers_var = {'h_conv1': h_conv1, 'h_conv2': h_conv2, 'h_conv3': h_conv3}
return x, y_conv, y_, keep_prob, layers_var
def get_4dcnn_our_model(feature_set_training, label_set_training, param):
input_feature = feature_set_training
descriptor_size = input_feature.shape[1]
N_classes = label_set_training.shape[1]
assert (N_classes > 1) # TOP HAT NOTATION REQUIRED!
# input layer
x = tf.placeholder(tf.float32, shape=[None, descriptor_size])
y_ = tf.placeholder(tf.float32, shape=[None, N_classes])
x_image = tf.reshape(
x, [-1, 20, 4, 5, 3, 1]) # be careful of shape consistency
###### conv layer 1 4d
W_conv1 = common_utils_nn.weight_variable(
[5, 2, 2, 1, 1, param.layer1_feature_maps_num], param.init_stddev)
b_conv1 = common_utils_nn.bias_variable([param.layer1_feature_maps_num])
h_conv1 = tf.nn.relu(
common_utils_nn.conv4d_stacked(x_image, W_conv1, stack_axis=1) +
b_conv1)
###### conv layer 2 4d
W_conv2 = common_utils_nn.weight_variable([
5, 2, 2, 1, param.layer1_feature_maps_num,
param.layer2_feature_maps_num
], param.init_stddev)
b_conv2 = common_utils_nn.bias_variable([param.layer2_feature_maps_num])
h_conv2 = tf.nn.relu(
common_utils_nn.conv4d_stacked(h_conv1, W_conv2, stack_axis=1) +
b_conv2)
###### Reshaping
rows_num, cols_num = common_utils_nn.find_best_square_reshape_from_array(
descriptor_size) #is this correct? # was 1200
h_pool1_reshape = tf.reshape(
h_conv2, [-1, rows_num, cols_num, param.layer2_feature_maps_num])
###### conv layer 3python wrapper_stanford_train.py --batch_size 88 --log_dir stanford10_wahl/ --dataset stanford10OurSplitNew --gpu 0 --max_epoch 1500
W_conv3 = common_utils_nn.weight_variable(
[5, 5, param.layer2_feature_maps_num, param.layer3_feature_maps_num],
param.init_stddev)
b_conv3 = common_utils_nn.bias_variable([param.layer3_feature_maps_num])
h_conv3 = tf.nn.relu(
common_utils_nn.conv2d(h_pool1_reshape, W_conv3) + b_conv3)
# pooling
h_pool3 = common_utils_nn.max_pool_2x2(h_conv3)
# pooling again
h_pool_final_flat = tf.reshape(h_pool3, [
-1, h_pool3.shape[1].value * h_pool3.shape[2].value *
param.layer3_feature_maps_num
]) # be careful of shape consistency
###### fully connected layer 1
W_fc1 = common_utils_nn.weight_variable([
h_pool3.shape[1].value * h_pool3.shape[2].value *
param.layer3_feature_maps_num, param.layer4_fc_num
], param.init_stddev) # be careful of shape consistency
b_fc1 = common_utils_nn.bias_variable([param.layer4_fc_num])
h_fc1 = tf.nn.relu(tf.matmul(h_pool_final_flat, W_fc1) + b_fc1)
###### dropout layer
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = common_utils_nn.weight_variable([param.layer4_fc_num, N_classes],
param.init_stddev)
b_fc2 = common_utils_nn.bias_variable([N_classes])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
layers_var = {'h_conv1': h_conv1, 'h_conv2': h_conv2, 'h_conv3': h_conv3}
return x, y_conv, y_, keep_prob, layers_var