def load_intermediate_outputs(input_filename,
embedding_json,
fc_json,
lstm_hidden_json,
lstm_cell_json,
layer_name=None):
#layer_name is currently not needed - for networks with more layers we will need it, as the json structure will change
keys_hidden, data_hidden = data_format.get_data(lstm_hidden_json)
keys_cell, data_cell = data_format.get_data(lstm_cell_json)
keys_fc, data_fc = data_format.get_data(fc_json)
keys_ebd, data_ebd = data_format.get_data(embedding_json)
lstm_hidden = data_hidden[input_filename]
lstm_cell = data_cell[input_filename]
fc_out = data_fc[input_filename]
embedding_output_data = data_ebd[input_filename]
T = embedding_output_data.shape
d = lstm_cell.shape[1]
return fc_out, lstm_hidden, lstm_cell, embedding_output_data, d, T
def get_DstMatrix_singleReview(review_MaxAct_json, final_embed_mat,
embedding_size, dictionary_w):
# Get similarity matrix between neurons based on the custom distance function defined above, calculated based on a single review.
keys, data = data_format.get_data(review_MaxAct_json)
kkeys = list(keys)
dstMat = np.zeros((len(kkeys), len(kkeys)))
for i in range((len(kkeys))):
for j in range(i, len(kkeys)):
dstMat[i, j] = neuron_distance(final_embed_mat=final_embed_mat,
embedding_size=embedding_size,
neuron1=list(data[kkeys[i]]),
neuron2=list(data[kkeys[j]]),
dictionary_w=dictionary_w)
return dstMat
def get_NeuronExcitingWords_dict(lstm_hidden_json, kkeys, k, save_dir, topN=5):
# Get the N words that excite each LSTM cell maximum, i.e. neuron of output has maximum value during forward pass
d = collections.OrderedDict()
keys_hidden, data_hidden = data_format.get_data(lstm_hidden_json)
kdata = data_hidden[k]
for i in range(min(len(kkeys), kdata.shape[0])):
ord_cells = np.argsort(kdata[i, :], axis=0, kind='quicksort')
d[kkeys[i]] = ord_cells[-(topN + 1):-1].tolist()
NtoW = invert_dict_nonunique(d, topN)
NtoW_keys = map(int, list(NtoW.keys()))
for i in range(kdata.shape[1]):
if i not in NtoW_keys:
NtoW[str(i)] = []
with open(save_dir + str(k) + "_ActCells.json", 'w') as f:
json.dump(NtoW, f)
return str(k) + "_ActCells.json", NtoW
def get_MostExcitingWords_allReviews(save_dir, neuronWords_jsons, topN=5):
#Get list of top-N exciting words for each neuron based on the whole dataset of reviews
#neuron-word data dictionary
nw_data = dict()
done = []
for i in neuronWords_jsons:
keys, data = data_format.get_data(save_dir + i)
kkeys = list(keys)
for j in kkeys:
if j not in done:
if j in list(nw_data.keys()):
vals = list(data[j]) + list(
set(list(nw_data[j])) - set(list(data[j])))
nw_data[j] = vals
else:
nw_data[j] = data[j]
if len(list(nw_data[j])) >= topN:
done.append(j)
return nw_data
x=0,
y=0,
anchor="bottom_left"))
"""
------------------------------------------------------------------------------------------------------------------------
MAIN APP CODE
------------------------------------------------------------------------------------------------------------------------
"""
# Provide data paths and files
load_dir = "./bokeh_vis/static/"
lstm_layer_name = "lstm"
#Get trained model parameters: weights and gate values
keys, data = data_format.get_data(load_dir + "model.json")
#Get raw input
keys_raw, data_raw = data_format.get_data(load_dir + "test_data_text.pickle")
#Load auxiliary data
with open(load_dir + "lstm_predictions.pickle", "rb") as handle:
predicted_tgs = pickle.load(handle)
with open(load_dir + "exploratoryDataFull.pickle", 'rb') as f:
(testX, embed_mat, excitingWords_fullSet, similarityMatrix_AllReviews,
similarityMatrix_PerReview, neuron_types, totalLRP, LRP) = pickle.load(f)
_, lstm_hidden = data_format.get_data(
load_dir + "test_model_internals_lstm_hidden.pickle")
#Get preset buttons' selections