def _get_layer_string(self, f, scale):
assert isinstance(f, Layer)
result = f.abbreviation
if scale and f.neuron_scale is not None:
self._output_scale = shape_string(
f.neuron_scale if f.output_scale is None else f.output_scale)
result += '_{}'.format(f.neuron_scale)
return result
def structure_string(self, detail=True, scale=True):
# Get functions to be added to structure string
assert isinstance(self.children, list)
fs = [
f for f in self.children
if isinstance(f, Net) or detail or f.is_nucleus
]
# Add input layer
result = ('' if self.input_ is None else 'input[{}] => '.format(
shape_string(self.input_.sample_shape)))
# Check interconnection type
next_net, next_layer = ' => ', ' -> '
if self._inter_type not in (pedia.cascade,
self.RECURRENT) or self.is_branch:
if self._inter_type in [pedia.sum, pedia.prod, pedia.concat]:
result += self._inter_type
if self.is_branch: result += 'branch'
else: next_layer, next_net = ', ', ', '
result += '('
# Add children
str_list, next_token = [], None
for f in fs:
if isinstance(f, Net):
if next_token is None: next_token = next_net
assert next_token == next_net
str_list.append(f.structure_string(detail, scale))
else:
assert isinstance(f, Layer)
if next_token is None: next_token = next_layer
assert next_token == next_layer
str_list.append(self._get_layer_string(f, scale))
str_list = merger(str_list)
result += next_token.join(str_list)
# Check is_branch flag
if self.is_branch:
result += ' -> output'
# Check interconnection type
if self._inter_type not in (pedia.cascade,
self.RECURRENT) or self.is_branch:
result += ')'
# Add output scale
if self.is_root and not self._inter_type == pedia.fork:
result += ' => output[{}]'.format(self.output_shape_str)
# Return
return result
def structure_string(self, detail=True, scale=True):
# Get functions to be added to structure string
assert isinstance(self.children, list)
fs = [
f for f in self.children
if isinstance(f, Net) or detail or f.is_nucleus
]
# Add input layer
result = ('' if self.input_ is None else 'input_{} => '.format(
shape_string(self.input_.sample_shape)))
# Check interconnection type
next_net, next_layer = ' => ', ' -> '
if self._inter_type != pedia.cascade or self.is_branch:
if self._inter_type in [pedia.sum, pedia.prod, pedia.concat]:
result += self._inter_type
if self.is_branch: result += 'branch'
else: next_layer, next_net = ', ', ', '
result += '('
# Add children
for (i, f) in zip(range(len(self.children)), fs):
if isinstance(f, Net):
result += next_net if i != 0 else ''
result += f.structure_string(detail, scale)
else:
assert isinstance(f, Layer)
result += next_layer if i != 0 else ''
result += self._get_layer_string(f, scale)
# Check is_branch flag
if self.is_branch:
result += ' -> output'
# Check interconnection type
if self._inter_type != pedia.cascade or self.is_branch: result += ')'
# Add output scale
if self.is_root and not self._inter_type == pedia.fork:
result += ' => output_{}'.format(self.children[-1]._output_scale)
# Return
return result
def structure_detail(self):
"""A list of structure strings with format
Layer (type) Output Shape Params #
Currently only work for sequential model
TODO: refactoring is badly needed
"""
from tframe.nets.rnet import RNet
from tframe.nets.customized_net import CustomizedNet
widths = hub.structure_detail_widths
indent = 3
# rows is a list of lists of 3 cols
rows = []
# TODO: the line below should be removed is things are settled down
# add_to_rows = lambda cols: rows.append(fs.table_row(cols, widths))
# Dense total will be used when model weights are pruned
total_params, dense_total = 0, 0
if self.is_root:
rows.append(['input', shape_string(self.input_.sample_shape), ''])
for child in self.children:
if isinstance(child, Layer):
_row, num, dense_num = self._get_layer_detail(child)
rows.append(_row)
elif isinstance(child, (RNet, CustomizedNet)):
num, dense_num = child.params_num
cols = [child.structure_string(), child.output_shape_str,
stark.get_num_string(num, dense_num)]
rows.append(cols)
elif isinstance(child, Net):
_rows, num, dense_num = child.structure_detail
# TODO
rows += _rows
else:
raise TypeError('!! unknown child type {}'.format(type(child)))
# Accumulate total_params and dense_total_params
total_params += num
dense_total += dense_num
# Check total params
if not (hub.prune_on or hub.etch_on):
assert total_params == sum([np.prod(v.shape) for v in self.var_list])
if self.is_root:
headers = ['Layers', 'Output Shape', 'Params #']
# Decide cell widths
widths = [max(len(h), max([len(r[i]) for r in rows]))
for i, h in enumerate(headers)]
# Put all these stuff into a table
t = Table(*widths, margin=0, tab=9, buffered=True, indent=indent)
t.specify_format(align='llr')
t.print_header(*headers)
for i, row in enumerate(rows):
t.print_row(*row)
# Draw line
t.hline() if i != len(rows) - 1 else t.dhline()
t.print_with_margin('Total params: {}'.format(
stark.get_num_string(total_params, dense_total)))
t.hline()
return t.content, total_params, dense_total
else: return rows, total_params, dense_total
def structure_detail(self):
"""A list of structure strings with format
Layer (type) Output Shape Params #
Currently only work for sequential model
"""
from tframe.nets.rnet import RNet
widths = [33, 24, 20]
indent = 3
rows = []
add_to_rows = lambda cols: rows.append(fs.table_row(cols, widths))
# Dense total will be used when model weights are pruned
total_params, dense_total = 0, 0
if self.is_root:
add_to_rows(['input', shape_string(self.input_.sample_shape), ''])
def get_num_string(num, dense_num):
if num == 0: num_str = ''
elif hub.prune_on or hub.etch_on:
num_str = '{} ({:.1f}%)'.format(num, 100.0 * num / dense_num)
else: num_str = str(num)
return num_str
for child in self.children:
if isinstance(child, Layer):
# Try to find variable in child
variables = [v for v in self.var_list if child.group_name in v.name]
num, dense_num = stark.get_params_num(variables, consider_prune=True)
# Generate a row
cols = [self._get_layer_string(child, True, True),
child.output_shape_str, get_num_string(num, dense_num)]
add_to_rows(cols)
elif isinstance(child, RNet):
num, dense_num = child.params_num
cols = [child.structure_string(), child.output_shape_str,
get_num_string(num, dense_num)]
add_to_rows(cols)
elif isinstance(child, Net):
_rows, num, dense_num = child.structure_detail
rows += _rows
else:
raise TypeError('!! unknown child type {}'.format(type(child)))
# Accumulate total_params and dense_total_params
total_params += num
dense_total += dense_num
if self.is_root:
# Head
detail = ''
add_with_indent = lambda d, c: d + ' ' * indent + c + '\n'
width = sum(widths)
detail = add_with_indent(detail, '-' * width)
detail = add_with_indent(
detail, fs.table_row(['Layers', 'Output Shape', 'Params #'], widths))
detail = add_with_indent(detail, '=' * width)
# Content
for i, row in enumerate(rows):
if i > 0:
detail = add_with_indent(detail, '-' * width)
detail = add_with_indent(detail, row)
# Summary
detail = add_with_indent(detail, '=' * width)
detail = add_with_indent(
detail, 'Total params: {}'.format(
get_num_string(total_params, dense_total)))
detail += ' ' * indent + '-' * width
return detail, total_params, dense_total
else: return rows, total_params, dense_total