def test_same():
image = get_test_image(tile_factor=0)
rf = KerasReceptiveField(get_build_func(padding='same'), init_weights=True)
rf_params0 = rf.compute(input_shape=ImageShape(*image.shape),
input_layer='input_image',
output_layer='conv')
print(rf_params0)
image = get_test_image(tile_factor=1)
rf = KerasReceptiveField(get_build_func(padding='same'), init_weights=True)
rf_params1 = rf.compute(input_shape=ImageShape(*image.shape),
input_layer='input_image',
output_layer='conv')
print(rf_params1)
assert_allclose(rf_params0, rf_params1)
def test_same():
image = get_test_image(tile_factor=0)
rf = KerasReceptiveField(get_build_func(padding="same"), init_weights=True)
rf_params0 = rf.compute(
input_shape=ImageShape(*image.shape),
input_layer="input_image",
output_layers=["conv1"],
)
print(rf_params0)
image = get_test_image(tile_factor=1)
rf = KerasReceptiveField(get_build_func(padding="same"), init_weights=True)
rf_params1 = rf.compute(
input_shape=ImageShape(*image.shape),
input_layer="input_image",
output_layers=["conv1"],
)
print(rf_params1)
assert_allclose(rf_params0[0].rf, rf_params1[0].rf)
def test_expected_values():
image = get_test_image(tile_factor=0)
rf = KerasReceptiveField(get_build_func(padding="valid"),
init_weights=True)
rf_params0 = rf.compute(
input_shape=ImageShape(*image.shape),
input_layer="input_image",
output_layers=["conv1"],
)[0]
assert_allclose(rf_params0.rf.stride, (2, 2))
assert_allclose(rf_params0.rf.size,
(((2 + 1) * 2 + 2) * 2, ((2 + 1) * 2 + 2) * 2))
receptive_field_size = []
for model_file in model_files:
print(model_file)
# estimate receptive field of the model
def model_build_func(input_shape):
model = fcn_sherrah2016_regression_and_classifier(input_shape=input_shape, for_receptive_field=True)
model.load_weights(model_file)
return model
rf = KerasReceptiveField(model_build_func, init_weights=False)
rf_params = rf.compute(
input_shape=(500, 500, 3),
input_layer='input_image',
output_layers=['regression_output', 'classification_output'])
print(rf_params)
receptive_field_size.append(rf._rf_params[0].size)
for i, model_file in enumerate(model_files):
print(model_file)
print('Receptive field size: ' + str(receptive_field_size[i]))
'''Load model and visualise results
'''
# list of model files to inspect
model_files = glob.glob(os.path.join(saved_models_dir, model_name))
##########################################################################################
# fcn_sherrah2016_regression
##########################################################################################
# estimate receptive field of the model
def model_build_func(input_shape):
model = fcn_sherrah2016_regression(input_shape=input_shape,
for_receptive_field=True)
return model
rf = KerasReceptiveField(model_build_func, init_weights=True)
rf_params = rf.compute(input_shape=(512, 512, 3),
input_layer='input_image',
output_layers=['regression_output'])
print(rf_params)
print('Sherrah 2016 effective receptive field: ' + str(rf._rf_params[0].size))
##########################################################################################
# U-net
##########################################################################################
# estimate receptive field of the model
def model_build_func(input_shape):
model = unet(input_shape=input_shape, for_receptive_field=True)
return model
receptive_field_size = []
for model_file in model_files:
print(model_file)
# estimate receptive field of the model
def model_build_func(input_shape):
model = models.fcn_sherrah2016_regression(input_shape=input_shape,
for_receptive_field=True)
model.load_weights(model_file)
return model
rf = KerasReceptiveField(model_build_func, init_weights=False)
rf_params = rf.compute(input_shape=(500, 500, 3),
input_layer='input_image',
output_layers=['main_output'])
print(rf_params)
receptive_field_size.append(rf._rf_params[0].size)
for i, model_file in enumerate(model_files):
print(model_file)
print('Receptive field size: ' + str(receptive_field_size[i]))
'''Load model and visualise results
'''
for fold_i, model_file in enumerate(model_files):
# select test data (data not used for training)
idx_test = idx_test_all[fold_i]
for dilation_rate in dilation_rate_vec:
# estimate receptive field of the model
def model_build_func(input_shape):
return models.fcn_conv_bnorm_maxpool_regression(
input_shape=input_shape,
for_receptive_field=True,
nblocks=nblocks,
kernel_len=kernel_len,
dilation_rate=dilation_rate)
rf = KerasReceptiveField(model_build_func, init_weights=True)
try:
rf_params = rf.compute(input_shape=(1000, 1000, 1),
input_layer='input_image',
output_layer='main_output')
output['nblocks'].append(nblocks)
output['kernel_len'].append(kernel_len)
output['dilation_rate'].append(dilation_rate)
output['rf_size'].append(rf_params[2].w)
except ResourceExhaustedError: # if we run out of memory
output['nblocks'].append(nblocks)
output['kernel_len'].append(kernel_len)
output['dilation_rate'].append(dilation_rate)
output['rf_size'].append(float('nan'))
# convert to arrays, so that we can index with a vector
output['nblocks'] = np.array(output['nblocks'])
output['kernel_len'] = np.array(output['kernel_len'])
model = efn.EfficientNetB3(input_shape=shape, include_top=False, weights=None)
# model = C2D(shape)
# model.summary()
def model_build_func(input_shape):
return model
# compute receptive field
rf = KerasReceptiveField(model_build_func)
layers = list(filter(lambda y: (y.name.endswith("add") or y.name == "top_conv"), model.layers))
# layers = model.layers[1:]# list(filter(lambda y: y.name.endswith("project_BN") or y.name.endswith("add"), model.layers))
layers_names = [x.name for x in layers]
rf_params = rf.compute(shape, 'input_1', layers_names)
for i, x in enumerate(rf_params):
print("%-16s %-16s: %s" % (layers_names[i], layers[i].output_shape[1:], {"offset": x.rf.offset, "stride": x.rf.stride, "size": (x.rf.size.h, x.rf.size.w)}))
# debug receptive field
# rf.plot_rf_grids(get_default_image(shape, name='doge'))
# plt.show()
### EfficientNetB0 (224, 224, 3), RFs berechnet mit (528, 528, 3)
# block2b_add (56, 56, 24): {'offset': (3.5, 3.5), 'stride': (4.0, 4.0), 'size': (19, 19)} ### -> EfficientNetB0_Block2 (75264) ? Keine guten Ergebnisse zu erwarten
# block3b_add (28, 28, 40): {'offset': (7.5, 7.5), 'stride': (8.0, 8.0), 'size': (67, 67)} ### -> EfficientNetB0_Block3 (31360)
# block4b_add (14, 14, 80): {'offset': (15.5, 15.5), 'stride': (16.0, 16.0), 'size': (115, 115)}
# block4c_add (14, 14, 80): {'offset': (15.5, 15.5), 'stride': (16.0, 16.0), 'size': (147, 147)} ### -> EfficientNetB0_Block4 (15680)
# block5b_add (14, 14, 112): {'offset': (15.5, 15.5), 'stride': (16.0, 16.0), 'size': (275, 275)}
# block5c_add (14, 14, 112): {'offset': (15.5, 15.5), 'stride': (16.0, 16.0), 'size': (339, 339)} ### -> EfficientNetB0_Block5 (21952)