Class: LSTM_CELL
- Inherits:
-
Object
- Object
- LSTM_CELL
- Defined in:
- lib/CELL.rb
Overview
██╗░░░░░░██████╗████████╗███╗░░░███╗ ░█████╗░███████╗██╗░░░░░██╗░░░░░██║░░░░░██╔════╝╚══██╔══╝████╗░████║ ██╔══██╗██╔════╝██║░░░░░██║░░░░░██║░░░░░╚█████╗░░░░██║░░░██╔████╔██║ ██║░░╚═╝█████╗░░██║░░░░░██║░░░░░██║░░░░░░╚═══██╗░░░██║░░░██║╚██╔╝██║ ██║░░██╗██╔══╝░░██║░░░░░██║░░░░░███████╗██████╔╝░░░██║░░░██║░╚═╝░██║ ╚█████╔╝███████╗███████╗███████╗╚══════╝╚═════╝░░░░╚═╝░░░╚═╝░░░░░╚═╝ ░╚════╝░╚══════╝╚══════╝╚══════╝
Instance Method Summary collapse
- #applyWeightChange ⇒ Object
-
#backwardPropagation(top_diff_h, top_diff_c) ⇒ Object
Back propagation.
-
#forwardPropagation ⇒ Object
Forward Propagation.
- #getBottomDeltaCt ⇒ Object
- #getBottomDeltaHt ⇒ Object
- #getCt ⇒ Object
- #getHt ⇒ Object
- #getYt ⇒ Object
-
#hp(nArr1, nArr2) ⇒ Object
Hadamard product.
-
#init(alpha, sz, terminal_output = nil) ⇒ Object
constructing the required arrays (Numo::NArrays).
- #setCprev(cp) ⇒ Object
- #setHprev(hp) ⇒ Object
-
#setXt(xt) ⇒ Object
Getters and Setters.
- #setYt(yt) ⇒ Object
- #sigmoid(value_) ⇒ Object
- #sigmoidDer(v) ⇒ Object
- #sigmoidVector(v) ⇒ Object
-
#tanhDer(v) ⇒ Object
Transfer functions.
Instance Method Details
#applyWeightChange ⇒ Object
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 |
# File 'lib/CELL.rb', line 144 def applyWeightChange() #puts ">> " + @dWg[0,0].to_s ### Apply input weight changes @Wi += @@Alpha * @dWi #@@Alpha @Wf += @@Alpha * @dWf @Wo += @@Alpha * @dWo @Wg += @@Alpha * @dWg ### Apply Cell output weight changes @Ui += @@Alpha * @dUi @Uf += @@Alpha * @dUf @Uo += @@Alpha * @dUo @Ug += @@Alpha * @dUg ### Apply bias changes @Bi += @@Alpha * @dBi @Bf += @@Alpha * @dBf @Bo += @@Alpha * @dBo @Bg += @@Alpha * @dBg #puts "<< " + @Wg[0,0].to_s ### Change related to cell input @dWi = DFloat.zeros(1, @sz) @dWf = DFloat.zeros(1, @sz) @dWo = DFloat.zeros(1, @sz) @dWg = DFloat.zeros(1, @sz) ### Change related to cell state @dUi = DFloat.zeros(1, @sz) @dUf = DFloat.zeros(1, @sz) @dUo = DFloat.zeros(1, @sz) @dUg = DFloat.zeros(1, @sz) ### Change required to Bias @dBi = DFloat.zeros(1, @sz) @dBf = DFloat.zeros(1, @sz) @dBo = DFloat.zeros(1, @sz) @dBg = DFloat.zeros(1, @sz) end |
#backwardPropagation(top_diff_h, top_diff_c) ⇒ Object
Back propagation
107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 |
# File 'lib/CELL.rb', line 107 def backwardPropagation(top_diff_h, top_diff_c) dC = hp(@O, top_diff_h) + top_diff_c dO = hp(@Ct, top_diff_h) dI = hp(@G, dC) dG = hp(@I, dC) dF = hp(@Cprev, dC) dIinput = hp(sigmoidDer(@I), dI) dFinput = hp(sigmoidDer(@F), dF) dOinput = hp(sigmoidDer(@O), dO) dGinput = hp(tanhDer(@G), dG) ### Calculating change required to input weight matrices @dWi -= hp(dIinput, @Xt) @dWf -= hp(dFinput, @Xt) @dWo -= hp(dOinput, @Xt) @dWg -= hp(dGinput, @Xt) #p @dWi.to_a ### Calculating change required to cell state weight matrices @dUi += hp(dIinput, @Hprev) @dUf += hp(dFinput, @Hprev) @dUo += hp(dOinput, @Hprev) @dUg += hp(dGinput, @Hprev) ### Calculating change required to bias vectors @dBi += dIinput @dBf += dFinput @dBo += dOinput @dBg += dGinput ### Calculating change required to HPrev dHprev = DFloat.zeros(1, @sz) dHprev += @Wi.transpose.dot(dIinput) dHprev += @Wf.transpose.dot(dFinput) dHprev += @Wo.transpose.dot(dOinput) dHprev += @Wg.transpose.dot(dGinput) ### Calculating and returning bottoms @bottom_diff_c = hp(dC, @F) @bottom_diff_h = dHprev end |
#forwardPropagation ⇒ Object
Forward Propagation
92 93 94 95 96 97 98 99 100 101 102 103 104 105 |
# File 'lib/CELL.rb', line 92 def forwardPropagation() ### Calculating gate values @Zf = @Wf.dot(@Xt) + @Uf.dot(@Hprev) + @Bf @F = sigmoidVector(@Zf) @Zi = @Wi.dot(@Xt) + @Ui.dot(@Hprev) + @Bi @I = sigmoidVector(@Zi) @Zo = @Wo.dot(@Xt) + @Uo.dot(@Hprev) + @Bo @O = sigmoidVector(@Zo) @Zg = @Wg.dot(@Xt) + @Ug.dot(@Hprev) + @Bg @G = NMath.tanh(@Zg) ### Calculating cell states @Ct = hp(@F, @Cprev) + hp(@I, @G) @Ht = hp(@O, NMath.tanh(@Ct)) end |
#getBottomDeltaCt ⇒ Object
249 250 251 |
# File 'lib/CELL.rb', line 249 def getBottomDeltaCt() return @bottom_diff_c end |
#getBottomDeltaHt ⇒ Object
246 247 248 |
# File 'lib/CELL.rb', line 246 def getBottomDeltaHt() return @bottom_diff_h end |
#getCt ⇒ Object
258 259 260 |
# File 'lib/CELL.rb', line 258 def getCt return @Ct end |
#getHt ⇒ Object
255 256 257 |
# File 'lib/CELL.rb', line 255 def getHt() return @Ht end |
#getYt ⇒ Object
252 253 254 |
# File 'lib/CELL.rb', line 252 def getYt() return @Yt end |
#hp(nArr1, nArr2) ⇒ Object
Hadamard product
81 82 83 84 85 86 87 88 89 90 |
# File 'lib/CELL.rb', line 81 def hp(nArr1, nArr2) nArr3 = DFloat.zeros(nArr1.shape()) dims = nArr1.shape() for i in 0...(dims[0]) for j in 0...(dims[1]) nArr3[i,j] = (nArr1[i,j] * nArr2[i,j]) end end return nArr3 end |
#init(alpha, sz, terminal_output = nil) ⇒ Object
constructing the required arrays (Numo::NArrays)
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 |
# File 'lib/CELL.rb', line 32 def init(alpha, sz, terminal_output=nil) @sz = sz @@Alpha = alpha ### Weight matrices with regard to inputs @Wg = DFloat.new(1, @sz).rand @Wi = DFloat.new(1, @sz).rand @Wf = DFloat.new(1, @sz).rand @Wo = DFloat.new(1, @sz).rand ### Weight matrices with regard to cell previous cell output @Ug = DFloat.new(1, @sz).rand @Ui = DFloat.new(1, @sz).rand @Uf = DFloat.new(1, @sz).rand @Uo = DFloat.new(1, @sz).rand ### Bias @Bg = DFloat.new(1, @sz).rand @Bi = DFloat.new(1, @sz).rand @Bf = DFloat.new(1, @sz).rand @Bo = DFloat.new(1, @sz).rand ### Cell input @Xt = DFloat.zeros(1, @sz) @Yt = DFloat.zeros(1, @sz) ### Cell state @Hprev = DFloat.zeros(1, @sz) # Previous cell output @Cprev = DFloat.zeros(1, @sz) # Previous cell state @Ct = DFloat.zeros(1, @sz) # Cell state @Ht = DFloat.zeros(1, @sz) # Cell output ### Internal Cell gates @F = DFloat.new(1, @sz).rand @I = DFloat.new(1, @sz).rand @O = DFloat.new(1, @sz).rand @G = DFloat.new(1, @sz).rand ### BPTT variables ### Change related to cell input @dWi = DFloat.zeros(1, @sz) @dWf = DFloat.zeros(1, @sz) @dWo = DFloat.zeros(1, @sz) @dWg = DFloat.zeros(1, @sz) ### Change related to cell state @dUi = DFloat.zeros(1, @sz) @dUf = DFloat.zeros(1, @sz) @dUo = DFloat.zeros(1, @sz) @dUg = DFloat.zeros(1, @sz) ### Change required to Bias @dBi = DFloat.zeros(1, @sz) @dBf = DFloat.zeros(1, @sz) @dBo = DFloat.zeros(1, @sz) @dBg = DFloat.zeros(1, @sz) end |
#setCprev(cp) ⇒ Object
238 239 240 241 242 243 244 245 |
# File 'lib/CELL.rb', line 238 def setCprev(cp) @Cprev = cp if cp.shape()[0] != 1 puts "Size of cp: " + cp.shape().to_s raise "Setting @Cprev of LSTM cell raised an incorrect dimension error" return end end |
#setHprev(hp) ⇒ Object
230 231 232 233 234 235 236 237 |
# File 'lib/CELL.rb', line 230 def setHprev(hp) if hp.shape()[0] != 1 puts "Size of hp: " + hp.shape().to_s raise "Setting @Hprev of LSTM cell raised an incorrect dimension error" return end @Hprev = hp end |
#setXt(xt) ⇒ Object
Getters and Setters
214 215 216 217 218 219 220 221 |
# File 'lib/CELL.rb', line 214 def setXt(xt) if xt.shape()[0] != 1 puts "Size of xt: " + xt.shape().to_s raise "Setting @Xt of LSTM cell raised an incorrect dimension error" return end @Xt = xt end |
#setYt(yt) ⇒ Object
222 223 224 225 226 227 228 229 |
# File 'lib/CELL.rb', line 222 def setYt(yt) if yt.shape()[0] != 1 puts "Size of yt: " + yt.shape().to_s raise "Setting @Yt of LSTM cell raised an incorrect dimension error" return end @Yt = yt end |
#sigmoid(value_) ⇒ Object
209 210 211 212 |
# File 'lib/CELL.rb', line 209 def sigmoid(value_) output_value = 1 / (1 + Math.exp(-1 * value_)) return output_value end |
#sigmoidDer(v) ⇒ Object
189 190 191 192 193 194 195 196 197 198 |
# File 'lib/CELL.rb', line 189 def sigmoidDer(v) #input vector must be a horisontal vector output_vector = DFloat.zeros(v.shape[0], v.shape[1]) for i in 0...v.shape[0] for j in 0...v.shape[1] output_vector[i,j] = v[i,j] * (1 - v[i,j]) end end return output_vector end |
#sigmoidVector(v) ⇒ Object
199 200 201 202 203 204 205 206 207 208 |
# File 'lib/CELL.rb', line 199 def sigmoidVector(v) #input vector must be a horisontal vector output_vector = DFloat.zeros(v.shape[0], v.shape[1]) for i in 0...v.shape[0] for j in 0...v.shape[1] output_vector[i,j] = sigmoid(v[i,j]) end end return output_vector end |
#tanhDer(v) ⇒ Object
Transfer functions
179 180 181 182 183 184 185 186 187 188 |
# File 'lib/CELL.rb', line 179 def tanhDer(v) #input vector must be a horisontal vector output_vector = DFloat.zeros(v.shape[0], v.shape[1]) for i in 0...v.shape[0] for j in 0...v.shape[1] output_vector[i,j] = 1 - (v[i,j]**2) end end return output_vector end |