def _get_nonseq_folded_model(x_shape):
x = x_in = layers.Input(x_shape, name="input")
x1 = layers.Conv2D(filters=1,
kernel_size=(1, 1),
strides=(1, 1),
name="conv2d_1")(x)
x2 = layers.Conv2D(filters=1,
kernel_size=(1, 1),
strides=(1, 1),
name="conv2d_2")(x)
x = layers.Maximum()([x1, x2])
x = QConv2DBatchnorm(filters=2,
kernel_size=(2, 2),
strides=(4, 4),
kernel_initializer="ones",
bias_initializer="zeros",
use_bias=False,
kernel_quantizer=kernel_quantizer,
beta_initializer="zeros",
gamma_initializer="ones",
moving_mean_initializer="zeros",
moving_variance_initializer="ones",
folding_mode=folding_mode,
ema_freeze_delay=ema_freeze_delay,
name="foldconv2d")(x)
x = layers.Flatten(name="flatten")(x)
x = layers.Dense(2,
use_bias=False,
kernel_initializer="ones",
name="dense")(x)
model = Model(inputs=[x_in], outputs=[x])
model.layers[4].set_weights(
[kernel, gamma, beta, iteration, moving_mean, moving_variance])
return model
def test_populate_bias_quantizer_from_accumulator():
"""Test populate_bias_quantizer_from_accumulator function.
Define a qkeras model with a QConv2DBatchnorm layer. Set bias quantizer in the
layer as None. Call populate_bias_quantizer_from_accumulator function
to automatically generate bias quantizer type from the MAC accumulator type.
Set the bias quantizer accordingly in the model.
Call populate_bias_quantizer_from_accumulator again in this model. This time
since bias quantizer is already set, populate_bias_quantizer_from_accumulator
function should not change the bias quantizer.
"""
x_shape = (2, 2, 1)
# get a qkeras model with QConv2DBatchnorm layer. Set bias quantizer in the
# layer as None.
x = x_in = layers.Input(x_shape, name="input")
x1 = QConv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), use_bias=False,
kernel_quantizer="quantized_bits(4, 0, 1)", name="conv2d_1")(x)
x2 = QConv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), use_bias=False,
kernel_quantizer="quantized_bits(4, 0, 1)", name="conv2d_2")(x)
x = layers.Maximum()([x1, x2])
x = QActivation("quantized_relu(4, 1)")(x)
x = QConv2DBatchnorm(
filters=2, kernel_size=(2, 2), strides=(4, 4),
kernel_initializer="ones", bias_initializer="zeros", use_bias=False,
kernel_quantizer="quantized_bits(4, 0, 1)", bias_quantizer=None,
beta_initializer="zeros",
gamma_initializer="ones", moving_mean_initializer="zeros",
moving_variance_initializer="ones", folding_mode="batch_stats_folding",
ema_freeze_delay=10,
name="foldconv2d")(x)
x1 = x
x2 = layers.Flatten(name="flatten")(x)
x2 = QDense(2, use_bias=False, kernel_initializer="ones",
kernel_quantizer="quantized_bits(6, 2, 1)", name="dense")(x2)
model = Model(inputs=[x_in], outputs=[x1, x2])
assert_equal(model.layers[5].get_quantizers()[1], None)
# Call populate_bias_quantizer_from_accumulator function
# to automatically generate bias quantizer from the MAC accumulator type.
_ = bn_folding_utils.populate_bias_quantizer_from_accumulator(
model, ["quantized_bits(8, 0, 1)"])
q = model.layers[5].get_quantizers()[1]
assert_equal(q.__str__(), "quantized_bits(10,3,1)")
# Call populate_bias_quantizer_from_accumulator function again
# bias quantizer should not change
_ = bn_folding_utils.populate_bias_quantizer_from_accumulator(
model, ["quantized_bits(8, 0, 1)"])
q = model.layers[5].get_quantizers()[1]
assert_equal(q.__str__(), "quantized_bits(10,3,1)")
def squeeze_excite_block(input, ratio=0.25):
cse = channel_squeeze_excite_block(input, ratio)
sse = spatial_squeeze_excite_block(input)
output = layers.Maximum()([cse, sse])
return output
def ecnn(input_shape=(320,320,3), base_model_input_shape=(40,40,3), name='ECNN', num_classes=10, augment=False, auxiliary=False, sampling=True, scales='all', pooling=None, dropout=False, gaze=None, scale4_freeze=False, return_logits=False):
#ImageNet cortical sampling model
#check args
if input_shape != (320, 320, 3):
raise ValueError
if base_model_input_shape != (40, 40, 3):
raise ValueError
if scales not in ['all', 'scale4']:
raise ValueError
if scales != 'all' and auxiliary:
raise ValueError
if pooling is not None:
if pooling not in ['max', 'avg']:
raise ValueError
if pooling is not None and dropout:
raise NotImplementedError
if scales == 'scale4':
if dropout:
raise NotImplementedError
if pooling is not None:
raise ValueError
#base models
if scales == 'all':
scale1_network = ResNet18(include_top=False, input_shape=base_model_input_shape, conv1_stride=1, max_pool_stride=1, filters=45, subnetwork_name='scale1-{}'.format(name), pooling='avg')
scale2_network = ResNet18(include_top=False, input_shape=base_model_input_shape, conv1_stride=1, max_pool_stride=1, filters=45, subnetwork_name='scale2-{}'.format(name), pooling='avg')
scale3_network = ResNet18(include_top=False, input_shape=base_model_input_shape, conv1_stride=1, max_pool_stride=1, filters=45, subnetwork_name='scale3-{}'.format(name), pooling='avg')
scale4_network = ResNet18(include_top=False, input_shape=base_model_input_shape, conv1_stride=1, max_pool_stride=1, filters=45, subnetwork_name='scale4-{}'.format(name), pooling='avg')
model_input = layers.Input(shape=input_shape)
#data augmentation
if augment:
x = layers.Lambda(lambda tensor: glimpse.image_augmentation(tensor, dataset='imagenet10'), name='image_augmentation')(model_input)
else:
x = model_input
#preprocess
if gaze is not None:
gaze_x = tf.constant(gaze[0], tf.int32)
gaze_y = tf.constant(gaze[1], tf.int32)
gaze = [gaze_x, gaze_y]
else:
#img shape (320, 320, 3)
if not scale4_freeze:
gaze = 40
else:
gaze = 80
if not scale4_freeze:
scale_sizes = [40, 80, 160, 240]
scale_radii = [1, 2, 4, 6]
else:
scale_sizes = [40, 80, 160, 320]
scale_radii = [1, 2, 4, 8]
scale_center = [input_shape[0] // 2, input_shape[0] // 2]
if not sampling:
scales_x = layers.Lambda(lambda tensor: glimpse.image_scales(tensor, scale_center, scale_radii, scale_sizes, gaze, scale4_freeze), name='scale_sampling')(x)
else:
scales_x = layers.Lambda(lambda tensor: glimpse.warp_image_and_image_scales(tensor, input_shape[0], input_shape[0], scale_center, scale_radii, scale_sizes, gaze, scale4_freeze), name='nonuniform_and_scale_sampling')(x)
#unpack scales
scale1_x = scales_x[0]
scale2_x = scales_x[1]
scale3_x = scales_x[2]
scale4_x = scales_x[3]
if scales == 'all':
scale1_x = scale1_network(scale1_x)
scale2_x = scale2_network(scale2_x)
scale3_x = scale3_network(scale3_x)
scale4_x = scale4_network(scale4_x)
if scales == 'all':
if pooling is None:
x = layers.concatenate([scale1_x, scale2_x, scale3_x, scale4_x])
elif pooling == 'avg':
x = layers.Average()([scale1_x, scale2_x, scale3_x, scale4_x])
elif pooling == 'max':
x = layers.Maximum()([scale1_x, scale2_x, scale3_x, scale4_x])
else:
raise ValueError
if dropout:
x = layers.Dropout(0.75)(x)
elif scales == 'scale4':
x = scale4_x
else:
raise ValueError
if not return_logits:
model_output = layers.Dense(num_classes, activation='softmax', name='probs')(x)
else:
model_output = layers.Dense(num_classes, activation=None, name='probs')(x)
if auxiliary:
#aux output
if return_logits:
raise NotImplementedError
scale1_aux_out = layers.Dense(num_classes, activation='softmax', name='scale1_aux_probs')(scale1_x)
scale2_aux_out = layers.Dense(num_classes, activation='softmax', name='scale2_aux_probs')(scale2_x)
scale3_aux_out = layers.Dense(num_classes, activation='softmax', name='scale3_aux_probs')(scale3_x)
scale4_aux_out = layers.Dense(num_classes, activation='softmax', name='scale4_aux_probs')(scale4_x)
model = tf.keras.models.Model(inputs=model_input, outputs=[model_output, scale1_aux_out, scale2_aux_out, scale3_aux_out, scale4_aux_out])
else:
model = tf.keras.models.Model(inputs=model_input, outputs=model_output)
return model