def cnn2mlp(model_path, model_params=CNN_MODEL_PARAMS, verbose=False):
cnn_model = load_model2(model_path)
cnn_weights, cnn_biases = extract_weights(cnn_model,
with_bias=True)
mlp_weights, _, mlp_biases = extract_cnn_weights(cnn_weights,
biases=cnn_biases,
verbose=verbose,
as_sparse=True)
return SimpleMLP(mlp_weights, mlp_biases, model_params)
def cnn_tensors_to_flat_weights_and_graph(weights_array, max_weight_convention, as_sparse):
flat_weights_array, _ = extract_cnn_weights(weights_array,
with_avg=(max_weight_convention=='one_on_n'),
as_sparse=as_sparse)
for w_tens in flat_weights_array:
assert len(w_tens.shape) == 2, f"weight tensor should have been flattened but has shape {w_tens.shape}"
adj_mat = weights_to_graph(flat_weights_array)
return (flat_weights_array, adj_mat)
def save_weights(model, model_path, network_type, _log):
weight_path = str(model_path) + '-weights.pckl'
weights = extract_weights(model)
picklify(weight_path, weights)
ex.add_artifact(weight_path)
if network_type == 'cnn':
_log.info('Expanding CNN layers...')
expanded_weights, constraints = extract_cnn_weights(weights,
as_sparse=True,
verbose=True)
expanded_weight_path = str(model_path) + '-weights-expanded.pckl'
constraintst_path = str(model_path) + '-constraints-expanded.pckl'
picklify(expanded_weight_path, expanded_weights)
ex.add_artifact(expanded_weight_path)
picklify(constraintst_path, constraints)
ex.add_artifact(constraintst_path)
def save_weights(model, model_path, network_type, unroll_cnns, _log):
weight_path = str(model_path) + '-weights.pckl'
weights = extract_weights(model)
picklify(weight_path, weights)
ex.add_artifact(weight_path)
if network_type == 'cnn':
_log.info('Expanding CNN layers...')
if unroll_cnns: # if extracting cnn weights via unrolling into mlp
expanded_weights, constraints = extract_cnn_weights(weights,
as_sparse=True,
verbose=True)
expanded_weight_path = str(model_path) + '-weights-expanded.pckl'
constraintst_path = str(model_path) + '-constraints-expanded.pckl'
picklify(expanded_weight_path, expanded_weights)
ex.add_artifact(expanded_weight_path)
picklify(constraintst_path, constraints)
ex.add_artifact(constraintst_path)
else: # if extracting cn weights via treating each filter as a unit
_log.info(
f'Raw CNN model weight shapes: {[wgt.shape for wgt in weights]}'
)
extracted_weights_l1 = extract_cnn_weights_filters_as_units(
weights, norm=1)
extracted_weights_l2 = extract_cnn_weights_filters_as_units(
weights, norm=2)
_log.info(
f'Extracted CNN model weight shapes: {[ew.shape for ew in extracted_weights_l1]}'
)
extracted_weight_path_l1 = str(
model_path) + '-weights-filter-units_l1.pckl'
extracted_weight_path_l2 = str(
model_path) + '-weights-filter-units_l2.pckl'
picklify(extracted_weight_path_l1, extracted_weights_l1)
ex.add_artifact(extracted_weight_path_l1)
picklify(extracted_weight_path_l2, extracted_weights_l2)
ex.add_artifact(extracted_weight_path_l2)
if network_type == 'cnn_vgg':
_log.info('Expanding CNN_VGG layers...')
weights = extract_weights(model, with_batch_norm=True)
extracted_weights_l1 = extract_cnn_weights_filters_as_units(
weights, norm=1, with_batch_norm=True)
_log.info(
f'Extracted CNN model weight shapes: {[ew.shape for ew in extracted_weights_l1]}'
)
extracted_weight_path_l1 = str(
model_path) + '-weights-filter-units_l1.pckl'
picklify(extracted_weight_path_l1, extracted_weights_l1)
ex.add_artifact(extracted_weight_path_l1)
def plot_eigenvalues(weights_path, n_eigenvalues=None, ax=None, **kwargs):
weights = load_weights(weights_path)
if 'cnn' in str(weights_path):
weights, _ = extract_cnn_weights(
weights, with_avg=True) #(max_weight_convention=='one_on_n'))
print([w.shape for w in weights])
# TODO: take simpler solution from delete_isolated_ccs_refactored
adj_mat = weights_to_graph(weights)
_, components = sparse.csgraph.connected_components(adj_mat)
most_common_component_counts = Counter(components).most_common(2)
main_component_id = most_common_component_counts[0][0]
assert (len(most_common_component_counts) == 1
or most_common_component_counts[1][1] == 1)
main_component_mask = (components == main_component_id)
selected_adj_mat = adj_mat[main_component_mask, :][:, main_component_mask]
nrom_laplacian_matrix = sparse.csgraph.laplacian(selected_adj_mat,
normed=True)
if n_eigenvalues == None:
start, end = 0, selected_adj_mat.shape[0] - 2
elif isinstance(n_eigenvalues, int):
start, end = 0, n_eigenvalues
elif isinstance(n_eigenvalues, tuple):
start, end = n_eigenvalues
else:
raise TypeError(
'n_eigenvalues should be either None or int or tuple or slice.')
"""
eigen_values, _ = sparse.linalg.eigs(nrom_laplacian_matrix, k=end,
which='SM')
"""
sigma = 1
OP = nrom_laplacian_matrix - sigma * sparse.eye(
nrom_laplacian_matrix.shape[0])
OPinv = sparse.linalg.LinearOperator(
matvec=lambda v: sparse.linalg.minres(OP, v, tol=1e-5)[0],
shape=nrom_laplacian_matrix.shape,
dtype=nrom_laplacian_matrix.dtype)
eigen_values, _ = sparse.linalg.eigsh(nrom_laplacian_matrix,
sigma=sigma,
k=end,
which='LM',
tol=1e-5,
OPinv=OPinv)
eigen_values = np.sort(eigen_values)
eigen_values = eigen_values[start:end]
if ax is None:
_, ax = plt.subplots(1)
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
if 'linestyle' not in kwargs:
kwargs['linestyle'] = 'none'
kwargs['marker'] = '*'
kwargs['markersize'] = 5
return ax.plot(range(start + 1, end + 1), eigen_values, **kwargs)