Dogscats ensemble method - mismatching layer weights

MTAU · December 13, 2017, 2:54am

I’ll admit to having had some pain using the methods learned on new datasets or applying the methods taught in the lectures.

In this workbook I tried using the dogscats ensemble methods on the Kaggle Invasive species dataset. Was happy to work through debugging methods to expose and check mismatched layers, currently stumped at this layer where
a layer from fc_model.layers has weights of shape [(4096, 4096), (4096,)]
while
a layer from fc_layers has weights of shape [(4096, 2), (2,)]

(refer def train_dense_layers)

github.com

aspiringguru/fastai_bmt_reworking/blob/master/Kaggle-Invasive/Kaggle_Invasive_3_ensemble_broken_debugging_unresolved.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "#https://www.kaggle.com/c/invasive-species-monitoring/data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from datetime import datetime\n",
    "HOMEPATH = \"/home/ubuntu/fastai/\"\n",
    "DATA_HOME_DIR = HOMEPATH + \"data/invasive/\"\n",

This file has been truncated. show original

Taking a step back to redo some fundamentals. Broader question of work method to debug code in Keras remains.