from collections import Counter
from keras.models import model_from_json
from keras.models import Sequential
from keras.layers.embeddings import Embedding
from keras.layers.core import Dense, Activation, Dropout, TimeDistributedDense,RepeatVector,Merge
from keras.layers.recurrent import GRU,LSTM
from keras.models import model_from_json
import helper
"""
script to train a char-level LSTM or GRU to predict the next word based on
the previous words, previous chars, and topic vector
"""
#load our ModelHelper, pickled when training
mh = helper.loadHelper("../models/recipeRNN3noState")
temp = mh.word_indices
maxlen = mh.get("maxlen")
maxWord = mh.get("maxWord")
vecSize = mh.get("vecSize")
batchSize = 4
names = mh.get("recipeNames")
conVecs = mh.get("contextVectors")
charModel = Sequential()
charModel.add(LSTM(128, return_sequences=True,input_shape=(maxlen, mh.numChars)))
charModel.add(Dropout(.2))
#charModel.add(TimeDistributedDense(128))
#print(contextVecs[:3])
#stop=raw_input("done transforming docs")
#these will be useful for sampling randomly as a seed
mh.add("contextVectors",contextVecs)
mh.add("recipeNames",names)
mh.save("../models/recipeRNN3noState")
#define how we go from document
#define model
if len(sys.argv) > 1:
model = helper.loadThatModel(sys.argv[1])
mh = helper.loadHelper(sys.argv[1])
else:
charModel = Sequential()
charModel.add(LSTM(128, return_sequences=True,input_shape=(maxlen, mh.numChars)))
charModel.add(Dropout(.2))
#charModel.add(TimeDistributedDense(128))
wordModel = Sequential()
wordModel.add(Embedding(mh.vocSize+1, 256, input_length=maxWord))
wordModel.add(LSTM(512, return_sequences=True))
wordModel.add(Dropout(.2))
wordModel.add(LSTM(512, return_sequences=False))
wordModel.add(RepeatVector(maxlen))
contextModel = Sequential()
contextModel.add(Dense(256,input_shape=(vecSize,)))
contextVecs = contextVecs - np.mean(contextVecs,axis=0)
contextVecs = contextVecs / (np.std(contextVecs,axis=0)+0.00000001)
#these will be useful for sampling randomly as a seed
mh.add("contextVectors",contextVecs)
mh.add("recipeNames",names)
mh.add("recipes",recipes)
mh.save("../models/recipeRNN3noState")
#define model
if len(sys.argv) > 1:
model = helper.loadThatModel(sys.argv[1])
mh = helper.loadHelper(sys.argv[1])
print("model loaded")
contextVecs = mh.get("contextVectors")
recipes = mh.get("recipes")
else:
print("creating model")
charModel = Sequential()
charModel.add(LSTM(128, return_sequences=True,input_shape=(maxlen, mh.numChars)))
charModel.add(Dropout(.2))
#charModel.add(TimeDistributedDense(128))
wordModel = Sequential()
wordModel.add(Embedding(mh.vocSize, 512, input_length=maxWord))
wordModel.add(Dropout(0.2))
wordModel.add(LSTM(1024, return_sequences=True))
wordModel.add(Dropout(.2))