Github user njayaram2 commented on a diff in the pull request:
https://github.com/apache/madlib/pull/243#discussion_r175949965
--- Diff: src/ports/postgres/modules/convex/mlp_igd.py_in ---
@@ -222,67 +243,83 @@ def mlp(schema_madlib, source_table, output_table,
independent_varname,
it_args.update({
'group_by_clause': group_by_clause,
'using_clause': using_clause,
- 'grouping_str_comma': grouping_str_comma
+ 'grouping_str_comma': grouping_str_comma,
})
first_try = True
temp_output_table = unique_string(desp='temp_output_table')
+
+ layer_sizes = [num_input_nodes] + hidden_layer_sizes +
[num_output_nodes]
+
for _ in range(n_tries):
+ prev_state = None
if not warm_start:
coeff = []
- for i in range(len(layer_sizes) - 1):
- fan_in = layer_sizes[i]
- fan_out = layer_sizes[i + 1]
+ for fan_in, fan_out in zip(layer_sizes, layer_sizes[1:]):
# Initalize according to Glorot and Bengio (2010)
# See design doc for more info
span = math.sqrt(6.0 / (fan_in + fan_out))
- dim = (layer_sizes[i] + 1) * layer_sizes[i + 1]
- rand = plpy.execute("""SELECT
array_agg({span}*2*(random()-0.5))
- AS random
- FROM generate_series(0,{dim})
- """.format(span=span, dim=dim))[0]["random"]
+ dim = (fan_in + 1) * fan_out
+ rand = [span * (random() - 0.5) for _ in range(dim)]
--- End diff --
Its supposed to be explained in the design doc as per the comment. I think
these formulae are taken from a research paper.
---
