-
Notifications
You must be signed in to change notification settings - Fork 3
/
lstm_models_copy.py
219 lines (174 loc) · 11.4 KB
/
lstm_models_copy.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
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
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
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
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
#author : Suhas Pillai
import numpy as np
import pdb
from lstm_layer import *
from lstm_net import *
from ctc_loss import *
from lstm_net_cythonic import *
class Models:
'''
This class is the deep network, created using many sub modules from lstm_layer_net.py and lstm_layer.py files
'''
def __init__(self):
pass
def MDLSTM_model_graves(self,X,model,dict_conv_param,grd_truth_seq=None):
'''
The function is the forward pass for Deep Network
'''
lstm_layer_obj = Layer()
lstm_net_obj = LSTM_net()
lstm_net_obj_cython = LSTM_net_cythonic()
ctc_obj = CTC()
conv_param_1 = dict_conv_param['conv_1']
width = conv_param_1['width']
height = conv_param_1['height']
X = lstm_layer_obj.get_input_block_cython(X,width,height) # Get input blocks
#----------------------------------------------MDLSTM lstm_conv_feedforward block----------------------------------------------------#
model_2 = model['model_2']
conv_param_2 = dict_conv_param['conv_2']
# call to cython
out_2, cache_2 = lstm_net_obj_cython.MDLSTM_lstm_conv_feed_layer_forward(X,model_2,conv_param_2)
_,_,_,_,_,_,_,_,dout_conv_frwd_2 = cache_2
# pass through tanh
out_tanh_2, cache_tanh_2 = lstm_layer_obj.tanh_forward(out_2)
#----------------------------------------------MDLSTM lstm_conv_feedward block----------------------------------------------------#
out_tanh_2 = out_tanh_2.T
N_conv_2,C_conv_2,W_conv_2,H_conv_2 = dout_conv_frwd_2.shape
out_tanh_2_re = out_tanh_2.reshape(out_tanh_2.shape[0],W_conv_2,H_conv_2)
model_3 = model['model_3']
conv_param_3 = dict_conv_param['conv_3']
# call to cython
out_3, cache_3 = lstm_net_obj_cython.MDLSTM_lstm_conv_feed_layer_forward(out_tanh_2_re,model_3,conv_param_3)
_,_,_,_,_,_,_,_,dout_conv_frwd_3 = cache_3
# pass through tanh
out_tanh_3, cache_tanh_3 = lstm_layer_obj.tanh_forward(out_3)
#-----------------------------------------------------------------last lstm layer---------------------------------------------#
out_tanh_3 = out_tanh_3.T
N_conv_3,C_conv_3,W_conv_3,H_conv_3= dout_conv_frwd_3.shape
out_tanh_3_re = out_tanh_3.reshape(out_tanh_3.shape[0],W_conv_3,H_conv_3)
model_4 = model['model_4']
h_frwd_4,h_bckd_4,h_frwd_unflip_4,h_bckd_unflip_align_4,cache_lstm_4 = lstm_net_obj.MDLSTM_lstm_forward_layer(out_tanh_3_re,model_4)
# calling cython
# Some reshaping required
C= h_frwd_4.shape[0]
W = h_frwd_4.shape[1]-1
H = h_frwd_4.shape[2]-1
h_frwd_4 = h_frwd_4[:,1:,1:].reshape(C,W*H)
h_frwd_4 = h_frwd_4 .T
h_bckd_4 = h_bckd_4[:,1:,1:].reshape(C,W*H)
h_bckd_4 = h_bckd_4 .T
h_frwd_unflip_4 = h_frwd_unflip_4[:,1:,1:].reshape(C,W*H)
h_frwd_unflip_4 = h_frwd_unflip_4.T
h_bckd_unflip_align_4 = h_bckd_unflip_align_4[:,1:,1:].reshape(C,W*H)
h_bckd_unflip_align_4 =h_bckd_unflip_align_4.T
#------------------------------------------------------------Fully Connnected---------------------------------------------------#
W_h_to_o_4_frwd = model_4['ff']['W_h_to_o_frwd']
b_h_to_o_4_frwd = model_4['ff']['b_h_to_o_frwd']
W_h_to_o_4_bckd = model_4['ff']['W_h_to_o_bckd']
b_h_to_o_4_bckd = model_4['ff']['b_h_to_o_bckd']
W_h_to_o_4_frwd_flip = model_4['ff']['W_h_to_o_frwd_flip']
b_h_to_o_4_frwd_flip = model_4['ff']['b_h_to_o_frwd_flip']
W_h_to_o_4_bckd_flip = model_4['ff']['W_h_to_o_bckd_flip']
b_h_to_o_4_bckd_flip = model_4['ff']['b_h_to_o_bckd_flip']
out_4_frwd,cache_affine_forward_4_frwd = lstm_layer_obj.affine_forward(h_frwd_4,W_h_to_o_4_frwd, b_h_to_o_4_frwd)
out_4_bckd,cache_affine_forward_4_bckd = lstm_layer_obj.affine_forward(h_bckd_4,W_h_to_o_4_bckd, b_h_to_o_4_bckd)
out_4_frwd_flip,cache_affine_forward_4_frwd_flip = lstm_layer_obj.affine_forward(h_frwd_unflip_4,W_h_to_o_4_frwd_flip, b_h_to_o_4_frwd_flip)
out_4_bckd_flip,cache_affine_forward_4_bckd_flip = lstm_layer_obj.affine_forward(h_bckd_unflip_align_4, W_h_to_o_4_bckd_flip, b_h_to_o_4_bckd_flip)
cache_affine_forward_4 = (cache_affine_forward_4_frwd, cache_affine_forward_4_bckd, cache_affine_forward_4_frwd_flip,cache_affine_forward_4_bckd_flip)
#----------------------------------------------------Sum across fully connected--------------------------------------------------#
out_4_sum= out_4_frwd + out_4_bckd + out_4_frwd_flip + out_4_bckd_flip
out_4_sum = out_4_sum.T
out_4 = out_4_sum.reshape(out_4_sum.shape[0],W,H)
out_4=out_4.sum(1) #sum across vertical dimensions.
if grd_truth_seq==None:
out_4= out_4- np.max(out_4,axis=0)
out_4 = np.exp(out_4)
out_4 = out_4 / np.sum(out_4,axis=0)
return out_4
check_prob = out_4.copy()
check_prob = check_prob-np.max(check_prob,axis=0)
check_prob = np.exp(check_prob)
check_prob = check_prob / np.sum(check_prob,axis=0)
check_prob = np.sum(check_prob,1)/check_prob.shape[1]
#----------------------------------------------------Calling CTC cost function-------------------------------------------------------#
grd_truth_seq = np.asarray(grd_truth_seq)
loss,dscores,check = ctc_obj.ctc_loss_mass(out_4,grd_truth_seq,0,False)
dscores = dscores.T
cache = (cache_2 ,cache_tanh_2,cache_3,cache_tanh_3,cache_lstm_4,cache_affine_forward_4)
return loss,dscores,check,cache,check_prob
def MDLSTM_model_backward_graves(self,dscores,cache):
'''
The function is used for backward propagation of Deep network.
'''
lstm_net_obj_cython = LSTM_net_cythonic()
lstm_layer_obj = Layer()
lstm_net_obj = LSTM_net()
cache_2 ,cache_tanh_2,cache_3,cache_tanh_3,cache_lstm_4,cache_affine_forward_4 = cache
cache_lstm_frwd,cache_lstm_bckd,cache_lstm_frwd_flip,cache_lstm_bckd_flip = cache_lstm_4
X,model,h,cell_state,arr_i,arr_f_d1,arr_f_d2,arr_o,arr_g = cache_lstm_frwd
C = h.shape[0]
W= h.shape[1]-1
H=h.shape[2]-1
dout_sum = np.zeros((dscores.shape[1],W, H))
# giving error with respect to every pixel, which was summed.
for iter_i in xrange(H):
for iter_j in xrange(W):
dout_sum[:,iter_j,iter_i] = dscores[iter_i,:]
cache_affine_forward_4_frwd, cache_affine_forward_4_bckd, cache_affine_forward_4_frwd_flip,cache_affine_forward_4_bckd_flip = cache_affine_forward_4
dout_sum = dout_sum.reshape(dout_sum.shape[0],dout_sum.shape[1]*dout_sum.shape[2])
dout_sum = dout_sum.T
#-----------------------------------------------Calling affine Backward-------------------------------------------------------------#
dout_frwd_4, dW_h_to_o_4_frwd,db_h_to_o_4_frwd = lstm_layer_obj.affine_backward(dout_sum, cache_affine_forward_4_frwd)
dout_frwd_4 = dout_frwd_4.T
dout_frwd_4 = dout_frwd_4.reshape(C,W,H)
dout_bckd_4, dW_h_to_o_4_bckd,db_h_to_o_4_bckd = lstm_layer_obj.affine_backward(dout_sum, cache_affine_forward_4_bckd)
dout_bckd_4 =dout_bckd_4.T
dout_bckd_4 = dout_bckd_4.reshape(C,W,H)
dout_frwd_flip_4, dW_h_to_o_4_frwd_flip,db_h_to_o_4_frwd_flip = lstm_layer_obj.affine_backward(dout_sum, cache_affine_forward_4_frwd_flip)
dout_frwd_flip_4 = dout_frwd_flip_4.T
dout_frwd_flip_4 = dout_frwd_flip_4.reshape(C,W,H)
dout_bckd_flip_4, dW_h_to_o_4_bckd_flip,db_h_to_o_4_bckd_flip = lstm_layer_obj.affine_backward(dout_sum, cache_affine_forward_4_bckd_flip)
dout_bckd_flip_4 = dout_bckd_flip_4.T
dout_bckd_flip_4 = dout_bckd_flip_4.reshape(C,W,H)
dx_4,grads_frwd_4,grads_bckd_4,grads_frwd_flip_4,grads_bckd_flip_4 = lstm_net_obj.MDLSTM_backward_layer(dout_frwd_4,dout_bckd_4,dout_frwd_flip_4,dout_bckd_flip_4,cache_lstm_4)
# calling cython
dx_4 = dx_4.reshape(dx_4.shape[0],dx_4.shape[1] * dx_4.shape[2])
dx_4 = dx_4.T
# passing through tanh
dx_tanh_3 = lstm_layer_obj.tanh_backward(dx_4,cache_tanh_3)
#-----------------------------------------------Backward Propagation of lstm_conv_feed_layer----------------------------------------#
dx_tanh_3 = dx_tanh_3.T #(W*H,C) --->(C,W*H)
dx_3, grads_conv_3, grads_frwd_3, grads_bckd_3, grads_frwd_flip_3, grads_bckd_flip_3 = lstm_net_obj_cython. MDLSTM_lstm_conv_feed_layer_backward(dx_tanh_3,cache_3)
dx_3 = dx_3.reshape(dx_3.shape[0],dx_3.shape[1] * dx_3.shape[2])
dx_3 = dx_3.T
#pass through tanh
dx_tanh_2 = lstm_layer_obj.tanh_backward(dx_3,cache_tanh_2)
#-----------------------------------------------Backward Propagation of lstm_conv_feed_layer----------------------------------------#
# calling cython
dx_tanh_2 = dx_tanh_2.T
dx_2,grads_conv_2, grads_frwd_2, grads_bckd_2, grads_frwd_flip_2, grads_bckd_flip_2 = lstm_net_obj_cython. MDLSTM_lstm_conv_feed_layer_backward(dx_tanh_2,cache_2)
dx_2 = dx_2.reshape(1,dx_2.shape[0],dx_2.shape[1],dx_2.shape[2])
# --------------------------------------------------Storing Gradients----------------------------------------------------------------#
grads_model_2 = {'conv':grads_conv_2,'forward':grads_frwd_2,'backward':grads_bckd_2,'forward_flip':grads_frwd_flip_2,'backward_flip':grads_bckd_flip_2}
grads_model_3 = {'conv':grads_conv_3,'forward':grads_frwd_3,'backward':grads_bckd_3,'forward_flip':grads_frwd_flip_3,'backward_flip':grads_bckd_flip_3}
grads_model_4 = {'ff':{'W_h_to_o_frwd':dW_h_to_o_4_frwd,'b_h_to_o_frwd':db_h_to_o_4_frwd,'W_h_to_o_bckd':dW_h_to_o_4_bckd,'b_h_to_o_bckd':db_h_to_o_4_bckd,\
'W_h_to_o_frwd_flip':dW_h_to_o_4_frwd_flip,'b_h_to_o_frwd_flip':db_h_to_o_4_frwd_flip,'W_h_to_o_bckd_flip':dW_h_to_o_4_bckd_flip, \
'b_h_to_o_bckd_flip':db_h_to_o_4_bckd_flip},'forward':grads_frwd_4,'backward':grads_bckd_4,'forward_flip':grads_frwd_flip_4,'backward_flip':grads_bckd_flip_4}
return dx_2,grads_model_2, grads_model_3, grads_model_4
def Model_MDLSTM(self,X,model,dict_conv_param,grd_truth_seq,reg):
'''
The method calls forward and backward pass, i.e the above 2 methods.
'''
loss_forward,dscores,check,cache,check_prob = self.MDLSTM_model_graves(X,model,dict_conv_param,grd_truth_seq)
dx_2,grads_model_2, grads_model_3, grads_model_4 = self.MDLSTM_model_backward_graves(dscores,cache)
grads={'model_2':grads_model_2,'model_3':grads_model_3,'model_4':grads_model_4}
# L2 regularization
reg_loss = 0.0
for model_name in model:
for field in model[model_name]:
for key in model[model_name][field]:
grads[model_name][field][key] += reg * model[model_name][field][key]
reg_loss += np.sum(model[model_name][field][key]**2)
reg_loss = 0.5 * reg * reg_loss
loss_forward += reg_loss
return loss_forward, grads,check,check_prob