Train L3 embeddings with bug-fixes

First install the latest version of `kapre` by running (in your conda environment) `pip install --upgrade kapre`.

For all jobs, use [jobs/l3embedding-train-melspec2-03202018.sbatch](https://github.com/marl/l3embedding/blob/master/jobs/l3embedding-train-melspec2-03202018.sbatch) as a template, modifying it in the specified ways.

## MusAudioSet
* [x] linear - @jtcramer
- Set `#SBATCH --job-name=l3embedding-train-mus-linear`
- Set `#SBATCH --mail-user=`
- Set `#SBATCH --output="l3embedding-train-mus-linear-%j.out"`
- Set `#SBATCH --err="l3embedding-train-mus-linear-%j.err"`
- Change `source activate l3embedding` to your conda environment
- Set `TRAIN_DATA_DIR=/beegfs/work/AudioSetSamples/music_train`
- Set `VAL_DATA_DIR=/beegfs/work/AudioSetSamples/music_valid`
- Set `GSHEET_ID` to the appropriate string
- Set `OUTPUT_DIR=/scratch//l3_output`
- Change `--model-type` to `--model-type cnn_L3_kapredbinputbn`
* [x] melspec1 - @hohsiangwu
- Set `#SBATCH --job-name=l3embedding-train-mus-melspec1`
- Set `#SBATCH --mail-user=`
- Set `#SBATCH --output="l3embedding-train-mus-melspec1-%j.out"`
- Set `#SBATCH --err="l3embedding-train-mus-melspec1-%j.err"`
- Change `source activate l3embedding` to your conda environment
- Set `TRAIN_DATA_DIR=/beegfs/work/AudioSetSamples/music_train`
- Set `VAL_DATA_DIR=/beegfs/work/AudioSetSamples/music_valid`
- Set `GSHEET_ID` to the appropriate string
- Set `OUTPUT_DIR=/scratch//l3_output`
- Change `--model-type` to `--model-type cnn_L3_melspec1`
* [ ] melspec2 - @justinsalamon
- Set `#SBATCH --job-name=l3embedding-train-mus-melspec2`
- Set `#SBATCH --mail-user=`
- Set `#SBATCH --output="l3embedding-train-mus-melspec2-%j.out"`
- Set `#SBATCH --err="l3embedding-train-mus-melspec2-%j.err"`
- Change `source activate l3embedding` to your conda environment
- Set `TRAIN_DATA_DIR=/beegfs/work/AudioSetSamples/music_train`
- Set `VAL_DATA_DIR=/beegfs/work/AudioSetSamples/music_valid`
- Set `GSHEET_ID` to the appropriate string
- Set `OUTPUT_DIR=/scratch//l3_output`
- Change `--model-type` to `--model-type cnn_L3_melspec2`

## EnvAudioSet
* [x] linear - @jtcramer
- Set `#SBATCH --job-name=l3embedding-train-env-linear`
- Set `#SBATCH --mail-user=`
- Set `#SBATCH --output="l3embedding-train-env-linear-%j.out"`
- Set `#SBATCH --err="l3embedding-train-env-linear-%j.err"`
- Change `source activate l3embedding` to your conda environment
- Set `TRAIN_DATA_DIR=/beegfs/work/AudioSetSamples_environmental/environmental_train`
- Set `VAL_DATA_DIR=/beegfs/work/AudioSetSamples_environmental/environmental_valid`
- Set `GSHEET_ID` to the appropriate string
- Set `OUTPUT_DIR=/scratch//l3_output`
- Change `--model-type` to `--model-type cnn_L3_kapredbinputbn`
* [x] melspec1 - @hohsiangwu
- Set `#SBATCH --job-name=l3embedding-train-env-melspec1`
- Set `#SBATCH --mail-user=`
- Set `#SBATCH --output="l3embedding-train-env-melspec1-%j.out"`
- Set `#SBATCH --err="l3embedding-train-env-melspec1-%j.err"`
- Change `source activate l3embedding` to your conda environment
- Set `TRAIN_DATA_DIR=/beegfs/work/AudioSetSamples_environmental/environmental_train`
- Set `VAL_DATA_DIR=/beegfs/work/AudioSetSamples_environmental/environmental_valid`
- Set `GSHEET_ID` to the appropriate string
- Set `OUTPUT_DIR=/scratch//l3_output`
- Change `--model-type` to `--model-type cnn_L3_melspec1`
* [ ] melspec2 - @justinsalamon
- Set `#SBATCH --job-name=l3embedding-train-env-melspec2`
- Set `#SBATCH --mail-user=`
- Set `#SBATCH --output="l3embedding-train-env-melspec2-%j.out"`
- Set `#SBATCH --err="l3embedding-train-env-melspec2-%j.err"`
- Change `source activate l3embedding` to your conda environment
- Set `TRAIN_DATA_DIR=/beegfs/work/AudioSetSamples_environmental/environmental_train`
- Set `VAL_DATA_DIR=/beegfs/work/AudioSetSamples_environmental/environmental_valid`
- Set `GSHEET_ID` to the appropriate string
- Set `OUTPUT_DIR=/scratch//l3_output`
- Change `--model-type` to `--model-type cnn_L3_melspec2`

Run `sbatch l3embedding-train.sbatch`.

Use jobs/l3embedding-train-melspec2-03202018.sbatch as the template and review the unchecked MusAudioSet and EnvAudioSet melspec2 entries. Configure the two sbatch jobs with their listed paths, model types, environment, and output settings, then run sbatch l3embedding-train.sbatch; completion means both melspec2 training jobs have been run.

Written by the indexing model from the issue text.