Skip to content

Commit 3706720

Browse files
JPLAY0JPLAY0
committed
A3Q1: almost finished 🎉
1 parent b54862d commit 3706720

File tree

3 files changed

+73
-22
lines changed

3 files changed

+73
-22
lines changed

‎assignment3/cs231n/classifiers/rnn.py‎

Lines changed: 21 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,3 @@
1-
from builtins import range
2-
from builtins import object
31
import numpy as np
42

53
from cs231n.layers import *
@@ -74,7 +72,6 @@ def __init__(self, word_to_idx, input_dim=512, wordvec_dim=128,
7472
for k, v in self.params.items():
7573
self.params[k] = v.astype(self.dtype)
7674

77-
7875
def loss(self, features, captions):
7976
"""
8077
Compute training-time loss for the RNN. We input image features and
@@ -142,7 +139,27 @@ def loss(self, features, captions):
142139
############################################################################
143140
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
144141

145-
pass
142+
caches = []
143+
out, c = affine_forward(features, W_proj, b_proj)
144+
caches.append(c)
145+
word_in, c = word_embedding_forward(captions_in, W_embed)
146+
caches.append(c)
147+
if self.cell_type == 'rnn':
148+
out, c = rnn_forward(word_in, out, Wx, Wh, b)
149+
else:
150+
pass
151+
caches.append(c)
152+
out, c = temporal_affine_forward(out, W_vocab, b_vocab)
153+
caches.append(c)
154+
loss, dx = temporal_softmax_loss(out, captions_out, mask)
155+
156+
dx, grads['W_vocab'], grads['b_vocab'] = temporal_affine_backward(dx, caches.pop())
157+
if self.cell_type == 'rnn':
158+
d_caption, dx, grads['Wx'], grads['Wh'], grads['b'] = rnn_backward(dx, caches.pop())
159+
else:
160+
pass
161+
grads['W_embed'] = word_embedding_backward(d_caption, caches.pop())
162+
_, grads['W_proj'], grads['b_proj'] = affine_backward(dx, caches.pop())
146163

147164
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
148165
############################################################################
@@ -151,7 +168,6 @@ def loss(self, features, captions):
151168

152169
return loss, grads
153170

154-
155171
def sample(self, features, max_length=30):
156172
"""
157173
Run a test-time forward pass for the model, sampling captions for input

‎assignment3/cs231n/rnn_layers.py‎

Lines changed: 43 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,5 @@
1-
from __future__ import print_function, division
2-
from builtins import range
31
import numpy as np
42

5-
63
"""
74
This file defines layer types that are commonly used for recurrent neural
85
networks.
@@ -30,13 +27,14 @@ def rnn_step_forward(x, prev_h, Wx, Wh, b):
3027
"""
3128
next_h, cache = None, None
3229
##############################################################################
33-
# TODO: Implement a single forward step for the vanilla RNN. Store the next #
30+
# Implement a single forward step for the vanilla RNN. Store the next #
3431
# hidden state and any values you need for the backward pass in the next_h #
3532
# and cache variables respectively. #
3633
##############################################################################
3734
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
3835

39-
pass
36+
next_h = np.tanh(prev_h @ Wh + x @ Wx + b)
37+
cache = x, prev_h, Wx, Wh, b, next_h
4038

4139
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
4240
##############################################################################
@@ -62,14 +60,20 @@ def rnn_step_backward(dnext_h, cache):
6260
"""
6361
dx, dprev_h, dWx, dWh, db = None, None, None, None, None
6462
##############################################################################
65-
# TODO: Implement the backward pass for a single step of a vanilla RNN. #
63+
# Implement the backward pass for a single step of a vanilla RNN. #
6664
# #
6765
# HINT: For the tanh function, you can compute the local derivative in terms #
6866
# of the output value from tanh. #
6967
##############################################################################
7068
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
7169

72-
pass
70+
x, prev_h, Wx, Wh, b, next_h = cache
71+
dnext_h: np.ndarray = (1 - next_h ** 2) * dnext_h
72+
dx = dnext_h @ Wx.T
73+
dprev_h = dnext_h @ Wh.T
74+
dWx = x.T @ dnext_h
75+
dWh = prev_h.T @ dnext_h
76+
db = dnext_h.sum(0)
7377

7478
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
7579
##############################################################################
@@ -78,7 +82,7 @@ def rnn_step_backward(dnext_h, cache):
7882
return dx, dprev_h, dWx, dWh, db
7983

8084

81-
def rnn_forward(x, h0, Wx, Wh, b):
85+
def rnn_forward(x: np.ndarray, h0, Wx, Wh, b):
8286
"""
8387
Run a vanilla RNN forward on an entire sequence of data. We assume an input
8488
sequence composed of T vectors, each of dimension D. The RNN uses a hidden
@@ -98,13 +102,19 @@ def rnn_forward(x, h0, Wx, Wh, b):
98102
"""
99103
h, cache = None, None
100104
##############################################################################
101-
# TODO: Implement forward pass for a vanilla RNN running on a sequence of #
105+
# Implement forward pass for a vanilla RNN running on a sequence of #
102106
# input data. You should use the rnn_step_forward function that you defined #
103107
# above. You can use a for loop to help compute the forward pass. #
104108
##############################################################################
105109
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
106110

107-
pass
111+
N, T, D = x.shape
112+
cache = []
113+
h = np.zeros((N, T, h0.shape[1]))
114+
for i in range(T):
115+
h0, c = rnn_step_forward(x[:, i], h0, Wx, Wh, b)
116+
h[:, i] += h0
117+
cache.append(c)
108118

109119
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
110120
##############################################################################
@@ -113,7 +123,7 @@ def rnn_forward(x, h0, Wx, Wh, b):
113123
return h, cache
114124

115125

116-
def rnn_backward(dh, cache):
126+
def rnn_backward(dh, cache: list):
117127
"""
118128
Compute the backward pass for a vanilla RNN over an entire sequence of data.
119129
@@ -134,13 +144,25 @@ def rnn_backward(dh, cache):
134144
"""
135145
dx, dh0, dWx, dWh, db = None, None, None, None, None
136146
##############################################################################
137-
# TODO: Implement the backward pass for a vanilla RNN running an entire #
147+
# Implement the backward pass for a vanilla RNN running an entire #
138148
# sequence of data. You should use the rnn_step_backward function that you #
139149
# defined above. You can use a for loop to help compute the backward pass. #
140150
##############################################################################
141151
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
142152

143-
pass
153+
N, T, H = dh.shape
154+
D = cache[0][0].shape[1]
155+
dx = np.zeros((N, T, D))
156+
dh0 = np.zeros((N, H))
157+
dWx = np.zeros((D, H))
158+
dWh = np.zeros((H, H))
159+
db = np.zeros((H,))
160+
161+
for i in reversed(range(T)):
162+
dx[:, i], dh0, dWx_, dWh_, db_ = rnn_step_backward(dh[:, i] + dh0, cache.pop())
163+
dWx += dWx_
164+
dWh += dWh_
165+
db += db_
144166

145167
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
146168
##############################################################################
@@ -166,13 +188,14 @@ def word_embedding_forward(x, W):
166188
"""
167189
out, cache = None, None
168190
##############################################################################
169-
# TODO: Implement the forward pass for word embeddings. #
191+
# Implement the forward pass for word embeddings. #
170192
# #
171193
# HINT: This can be done in one line using NumPy's array indexing. #
172194
##############################################################################
173195
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
174196

175-
pass
197+
out = W[x]
198+
cache = x, W
176199

177200
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
178201
##############################################################################
@@ -198,14 +221,17 @@ def word_embedding_backward(dout, cache):
198221
"""
199222
dW = None
200223
##############################################################################
201-
# TODO: Implement the backward pass for word embeddings. #
224+
# Implement the backward pass for word embeddings. #
202225
# #
203226
# Note that words can appear more than once in a sequence. #
204227
# HINT: Look up the function np.add.at #
205228
##############################################################################
206229
# *****START OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
207230

208-
pass
231+
x, w = cache
232+
dW = np.zeros_like(w)
233+
N, T, D = dout.shape
234+
np.add.at(dW, x.flatten(), dout.reshape(-1, D))
209235

210236
# *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****
211237
##############################################################################

‎assignment3/requirements.txt‎

Lines changed: 9 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -65,3 +65,12 @@ tornado==4.3
6565
traitlets==4.0.0
6666
Werkzeug==0.15.3
6767
yarl==1.2.4
68+
69+
# pytorch
70+
# jupyterlab
71+
# h5py
72+
# scipy==1.2.1
73+
# cython
74+
# matplotlib
75+
# future
76+
# install all this package is enough,pillow and scipy should at the same chanel

0 commit comments

Comments
 (0)