mbollmann

#!/usr/bin/env python3

"""Usage: conv_checkpoints_to_model.py MODFILE

Takes a trained model file with multiple saved checkpoints and converts these
checkpoints into standalone models.  This allows the different checkpoints to be
used, e.g., as parts of a model ensemble.

This script will:
- Analyze MODFILE to find all saved model components

mbollmann

--- theano/sandbox/cuda/opt.py	2017-05-31 23:26:09.972668647 +0200
+++ theano/sandbox/cuda/opt_patched.py	2017-06-01 00:49:43.818626738 +0200
@@ -38,10 +38,12 @@
     GpuElemwise, GpuDimShuffle, GpuReshape, GpuCAReduce,
     gpu_flatten,
     GpuSubtensor, GpuAdvancedSubtensor1,
-    GpuAdvancedIncSubtensor1, GpuAdvancedIncSubtensor1_dev20,
+    GpuAdvancedIncSubtensor1,
     GpuIncSubtensor, gpu_alloc, GpuAlloc, gpu_shape, GpuSplit, GpuAllocEmpty)
 from theano.sandbox.cuda.opt_util import pad_dims, unpad_dims

mbollmann

#!/usr/bin/python3
# -*- coding: utf-8 -*-

import argparse
import bibtexparser
from collections import Counter
import matplotlib.pyplot as plt
import seaborn as sns
import sys

mbollmann

class AttentionLSTM(LSTM):
    """LSTM with attention mechanism

    This is an LSTM incorporating an attention mechanism into its hidden states.
    Currently, the context vector calculated from the attended vector is fed
    into the model's internal states, closely following the model by Xu et al.
    (2016, Sec. 3.1.2), using a soft attention model following
    Bahdanau et al. (2014).

    The layer expects two inputs instead of the usual one:

mbollmann

from keras import backend as K
from keras.layers.recurrent import LSTM

class HiddenStateLSTM(LSTM):
    """LSTM with input/output capabilities for its hidden state.

    This layer behaves just like an LSTM, except that it accepts further inputs
    to be used as its initial states, and returns additional outputs,
    representing the layer's final states.

mbollmann

# Source:
# https://github.com/farizrahman4u/seq2seq/blob/master/seq2seq/layers/state_transfer_lstm.py

from keras import backend as K
from keras.layers.recurrent import LSTM

class StateTransferLSTM(LSTM):
    """LSTM with the ability to transfer its hidden state.

    This layer behaves just like an LSTM, except that it can transfer (or