def predict(self, X):
"""Return the probability of x belonging to either class"""
hidden = sigmoid(self.Wh @ X + self.bh)
scores = self.Ws @ hidden + self.bs
probs = softmax_vectorized(scores)
return probs.argmax(axis=0)
def forward(self, scores, ys):
self.ys = ys
self.probs = softmax_vectorized(np.array(scores))
y_hats = self.probs[self.ys, range(len(self.ys))]
# Loss
losses = -np.log(y_hats)
loss = losses.sum()
return loss, losses, self.probs
def forward_backward_prop(self, X=None, ys=None, rollout=None,
Whh=None, bhh=None, Wxh=None, bxh=None, Ws=None, bs=None,
hidden=None, predict=False):
"""Perform forward and backward prop over a single training example
Returns loss and gradients
"""
# Hidden and input weights
Whh = self.Whh if not type(Whh) == np.ndarray else Whh
bhh = self.bhh if not type(bhh) == np.ndarray else bhh
Wxh = self.Wxh if not type(Wxh) == np.ndarray else Wxh
bxh = self.bxh if not type(bxh) == np.ndarray else bxh
# Softmax weights
Ws = self.Ws if not type(Ws) == np.ndarray else Ws
bs = self.bs if not type(bs) == np.ndarray else bs
# Initial hidden state
hidden = self.hidden if not type(hidden) == np.ndarray else hidden
# Where to start in the sequence and how far to go
rollout = self.rollout if not rollout else rollout
# Convert X and ys to dictionaries?
if predict:
X = {t+1:X[:, [t]] for t in range(rollout)}
ys = {t+1:1 for t in range(rollout)}
else:
X, ys = {}, {}
# Forward pass!
dWhh, dbhh = np.zeros_like(Whh), np.zeros_like(bhh)
dWxh, dbxh = np.zeros_like(Wxh), np.zeros_like(bxh)
dWs, dbs = np.zeros_like(Ws), np.zeros_like(bs)
loss = 0.
hiddens = {0: hidden}
dhiddens, dhiddens_downstream, dhiddens_local = {}, {rollout:np.zeros((self.H, 1))}, {}
scores, probs = {}, {}
for t in range(1, rollout+1):
# Get the next input in the sequence
X[t], ys[t] = next(self.input)
# Previous hidden layer and input at time t
Z = (Whh @ hiddens[t-1] + bhh) + (Wxh @ X[t] + bxh)
hiddens[t] = np.tanh(Z)
# Softmax
scores[t] = Ws @ hiddens[t] + bs
probs[t] = softmax_vectorized(scores[t])
y_hat = probs[t][ys[t]]
# Loss
loss += -np.log(y_hat).sum()
# Add regularization
loss += self.regularizer * 0.5*(np.sum(Whh**2) + np.sum(bhh**2) +
np.sum(Wxh**2) + np.sum(bxh**2) +
np.sum(Ws**2) + np.sum(bs**2))
if predict:
return hiddens[rollout], scores
# Backpropagate!
backwards = list(reversed(range(rollout+1)))
for t in backwards[:-1]:
# Scores
dscores = probs[t]
dscores[ys[t], 0] -= 1
# Softmax weights
dbs += dscores
dWs += dscores @ hiddens[t].T
dhiddens_local[t] = Ws.T @ dscores
dhiddens[t] = dhiddens_local[t] + dhiddens_downstream[t] # Karpathy optimization
dZ = tanh_grad(hiddens[t]) * dhiddens[t]
# Input and hidden weights
dbxh += dZ
dWxh += dZ @ X[t].T
dbhh += dZ
dWhh += dZ @ hiddens[t-1].T
# Set up incoming hidden weight gradient for previous time step
dhiddens_downstream[t-1] = Whh.T @ dZ
# Regularization
#
# Hidden and input weights
dWhh += (self.regularizer*Whh)
dbhh += (self.regularizer*bhh)
dWxh += (self.regularizer*Wxh)
dbxh += (self.regularizer*bxh)
# Softmax weights
dWs += (self.regularizer*Ws)
dbs += (self.regularizer*bs)
# Log additional info?
if self.inspect:
self.xs, self.ys = str(X), str(ys)
self.scores, self.probs = scores, probs
self.loss = loss
self.dWhh, self.dbhh, self.dWxh, self.dbxh = dWhh, dbhh, dWxh, dbxh
self.dWs, self.dbs = dWs, dbs
self.hiddens = hiddens
self.dhiddens = dhiddens
self.dhiddens_local, self.dhiddens_downstream = dhiddens_local, dhiddens_downstream
return State(loss, Gradients(dWhh, dbhh, dWxh, dbxh, dWs, dbs), hiddens[rollout])
def forward_backward_prop(self, X=None, ys=None, rollout=None, train_index=None,
Whh=None, bhh=None, Wxh=None, bxh=None, Ws=None, bs=None,
hidden=None, predict=False):
"""Perform forward and backward prop over a single training example
Returns loss and gradients
"""
# Hidden and input weights
Whh = self.Whh if not type(Whh) == np.ndarray else Whh
bhh = self.bhh if not type(bhh) == np.ndarray else bhh
Wxh = self.Wxh if not type(Wxh) == np.ndarray else Wxh
bxh = self.bxh if not type(bxh) == np.ndarray else bxh
# Softmax weights
Ws = self.Ws if not type(Ws) == np.ndarray else Ws
bs = self.bs if not type(bs) == np.ndarray else bs
# Initial hidden state
hidden = self.hidden if not type(hidden) == np.ndarray else hidden
# Where to start in the sequence and how far to go
rollout = self.rollout if not rollout else rollout
train_index = self.train_index if not train_index else train_index
# Get next portion of sequence to train on
if not type(X) == np.ndarray:
X = self.X_train[:, train_index:train_index+rollout]
ys = self.ys_train[train_index:train_index+rollout]
# Got to the end and need to wrap around?
if train_index+rollout > self.T:
rollover_index = (train_index+rollout) % self.T
X = np.hstack([X, self.X_train[:, :rollover_index]])
ys = np.hstack([ys, self.ys_train[:rollover_index]])
# Append column of zeros to align X and Y with natural time
X, ys = np.hstack([np.zeros((self.N, 1)), X]), np.hstack([np.zeros(1, dtype=np.int), ys])
# Forward pass!
dWhh, dbhh = np.zeros_like(Whh), np.zeros_like(bhh)
dWxh, dbxh = np.zeros_like(Wxh), np.zeros_like(bxh)
dWs, dbs = np.zeros_like(Ws), np.zeros_like(bs)
loss = 0.
hiddens = {0: hidden}
dhiddens, dhiddens_downstream, dhiddens_local = {}, {rollout:np.zeros((self.H, 1))}, {}
scores, probs = {}, {}
for t in range(1, rollout+1):
# Previous hidden layer and input at time t
Z = (Whh @ hiddens[t-1] + bhh) + (Wxh @ X[:,[t]] + bxh)
hiddens[t] = np.tanh(Z)
# Softmax
scores[t] = Ws @ hiddens[t] + bs
probs[t] = softmax_vectorized(scores[t])
y_hat = probs[t][ys[t]]
# Loss
loss += -np.log(y_hat).sum()
# Add regularization
loss += self.regularizer * 0.5*(np.sum(Whh**2) + np.sum(bhh**2) +
np.sum(Wxh**2) + np.sum(bxh**2) +
np.sum(Ws**2) + np.sum(bs**2))
if predict:
return hiddens[rollout], scores
# Backpropagate!
backwards = list(reversed(range(rollout+1)))
for t in backwards[:-1]:
# Scores
dscores = probs[t]
dscores[ys[t], 0] -= 1
# Softmax weights
dbs += dscores
dWs += dscores @ hiddens[t].T
dhiddens_local[t] = Ws.T @ dscores
dhiddens[t] = dhiddens_local[t] + dhiddens_downstream[t] # Karpathy optimization
dZ = tanh_grad(hiddens[t]) * dhiddens[t]
# Input and hidden weights
dbxh += dZ
dWxh += dZ @ X[:,[t]].T
dbhh += dZ
dWhh += dZ @ hiddens[t-1].T
# Set up incoming hidden weight gradient for previous time step
dhiddens_downstream[t-1] = Whh.T @ dZ
# Regularization
#
# Hidden and input weights
dWhh += (self.regularizer*Whh)
dbhh += (self.regularizer*bhh)
dWxh += (self.regularizer*Wxh)
dbxh += (self.regularizer*bxh)
# Softmax weights
dWs += (self.regularizer*Ws)
dbs += (self.regularizer*bs)
# Log additional info?
if self.inspect:
self.xs, self.ys = str(X[:, 1:]), str(ys[1:])
self.scores, self.probs = scores, probs
self.loss = loss
self.dWhh, self.dbhh, self.dWxh, self.dbxh = dWhh, dbhh, dWxh, dbxh
self.dWs, self.dbs = dWs, dbs
self.hiddens = hiddens
self.dhiddens = dhiddens
self.dhiddens_local, self.dhiddens_downstream = dhiddens_local, dhiddens_downstream
return State(loss, Gradients(dWhh, dbhh, dWxh, dbxh, dWs, dbs), hiddens[rollout])