def forward(self, X, dump_chunks=-1):
A = self.A
b = self.b
""" Layer 1: linear activation """
layer1 = linear_forward(X, A, b)
if dump_chunks > 0:
dump_big_matrix(layer1, "lin_l1_mat", dump_chunks)
return [layer1]
def forward(self, X, dump_chunks = -1):
A = self.A
b = self.b
""" Layer 1: linear activation """
layer1 = linear_forward(X, A, b)
if dump_chunks > 0:
dump_big_matrix(layer1, "lin_l1_mat", dump_chunks)
return [layer1]
def backward(self, X, layers, Y, dump_chunks = -1):
A = self.A
b = self.b
layer1 = layers[-1]
""" softmax classification """
L, dLdl1 = softmax_loss(layer1, Y)
if dump_chunks > 0:
dump_big_matrix(dLdl1, "lin_dLdl1_mat", dump_chunks)
""" regularization: loss = 1/2 * lam * sum_nk(A_nk * A_nk) """
L += 0.5 * self.lam * np.sum(A * A)
""" backpropagation for Layer 1 """
dLdX, dLdA, dLdb = linear_backward(dLdl1, X, A)
if dump_chunks > 0:
dump_big_matrix(dLdX, "lin_dLdX_mat", dump_chunks)
dump_big_matrix(dLdA, "lin_dLdA_mat", 1)
dump_big_matrix(dLdb, "lin_dLdb_mat", 1)
""" regularization gradient """
dLdA = dLdA.reshape(A.shape)
dLdA += self.lam * A
""" tune the parameter """
self.v = self.mu * self.v - self.rho * dLdA
self.A += self.v
self.b += -self.rho * dLdb
return L
def backward(self, X, layers, Y, dump_chunks=-1):
A = self.A
b = self.b
layer1 = layers[-1]
""" softmax classification """
L, dLdl1 = softmax_loss(layer1, Y)
if dump_chunks > 0:
dump_big_matrix(dLdl1, "lin_dLdl1_mat", dump_chunks)
""" regularization: loss = 1/2 * lam * sum_nk(A_nk * A_nk) """
L += 0.5 * self.lam * np.sum(A * A)
""" backpropagation for Layer 1 """
dLdX, dLdA, dLdb = linear_backward(dLdl1, X, A)
if dump_chunks > 0:
dump_big_matrix(dLdX, "lin_dLdX_mat", dump_chunks)
dump_big_matrix(dLdA, "lin_dLdA_mat", 1)
dump_big_matrix(dLdb, "lin_dLdb_mat", 1)
""" regularization gradient """
dLdA = dLdA.reshape(A.shape)
dLdA += self.lam * A
""" tune the parameter """
self.v = self.mu * self.v - self.rho * dLdA
self.A += self.v
self.b += -self.rho * dLdb
return L
def forward(self, X, dump_chunks=-1):
A1 = self.A1
b1 = self.b1
A3 = self.A3
b3 = self.b3
"""
Layer 1 : linear
Layer 2 : ReLU
Layer 3 : linear
"""
layer1 = linear_forward(X, A1, b1)
layer2 = ReLU_forward(layer1)
layer3 = linear_forward(layer2, A3, b3)
if dump_chunks > 0:
dump_big_matrix(layer1, "nn_l1_mat", dump_chunks)
dump_big_matrix(layer2, "nn_l2_mat", dump_chunks)
dump_big_matrix(layer3, "nn_l3_mat", dump_chunks)
return [layer1, layer2, layer3]
def forward(self, X, dump_chunks = -1):
A1 = self.A1
b1 = self.b1
A3 = self.A3
b3 = self.b3
"""
Layer 1 : linear
Layer 2 : ReLU
Layer 3 : linear
"""
layer1 = linear_forward(X, A1, b1)
layer2 = ReLU_forward(layer1)
layer3 = linear_forward(layer2, A3, b3)
if dump_chunks > 0:
dump_big_matrix(layer1, "nn_l1_mat", dump_chunks)
dump_big_matrix(layer2, "nn_l2_mat", dump_chunks)
dump_big_matrix(layer3, "nn_l3_mat", dump_chunks)
return [layer1, layer2, layer3]
def backward(self, X, layers, Y, dump_chunks=-1):
A1 = self.A1
b1 = self.b1
A3 = self.A3
b3 = self.b3
layer1, layer2, layer3 = layers
""" softmax classification """
L, dLdl3 = softmax_loss(layer3, Y)
""" backpropagation for Layer 3 """
dLdl2, dLdA3, dLdb3 = linear_backward(dLdl3, layer2, A3)
""" backpropagation for Layer 2 """
dLdl1 = ReLU_backward(dLdl2, layer1)
""" backpropagation for Layer 1 """
dLdX, dLdA1, dLdb1 = linear_backward(dLdl1, X, A1)
""" regularization """
L += 0.5 * self.lam * (np.sum(A1 * A1) + np.sum(A3 * A3))
""" regularization gradient """
dLdA3 = dLdA3.reshape(A3.shape)
dLdA1 = dLdA1.reshape(A1.shape)
dLdA3 += self.lam * A3
dLdA1 += self.lam * A1
""" tune the parameter """
self.v1 = self.mu * self.v1 - self.rho * dLdA1
self.v3 = self.mu * self.v3 - self.rho * dLdA3
self.A1 += self.v1
self.A3 += self.v3
self.b1 += -self.rho * dLdb1
self.b3 += -self.rho * dLdb3
""" dump """
if dump_chunks > 0:
dump_big_matrix(dLdl3, "nn_dLdl3_mat", dump_chunks)
dump_big_matrix(dLdl2, "nn_dLdl2_mat", dump_chunks)
dump_big_matrix(dLdl1, "nn_dLdl1_mat", dump_chunks)
dump_big_matrix(dLdX, "nn_dLdX_mat", dump_chunks)
dump_big_matrix(dLdA3, "nn_dLdA3_mat", 1)
dump_big_matrix(dLdb3, "nn_dLdb3_mat", 1)
dump_big_matrix(dLdA1, "nn_dLdA1_mat", 1)
dump_big_matrix(dLdb1, "nn_dLdb1_mat", 1)
return L
def backward(self, X, layers, Y, dump_chunks=-1):
A1 = self.A1
b1 = self.b1
S1 = self.S1
P1 = self.P1
F3 = self.F3
S3 = self.S3
A4 = self.A4
b4 = self.b4
S4 = self.S4
P4 = self.P4
F6 = self.F6
S6 = self.S6
A7 = self.A7
b7 = self.b7
S7 = self.S7
P7 = self.P7
F9 = self.F9
S9 = self.S9
A10 = self.A10
b10 = self.b10
(layer1, X_col1), layer2, (layer3, X_idx3), \
(layer4, X_col4), layer5, (layer6, X_idx6), \
(layer7, X_col7), layer8, (layer9, X_idx9), layer10 = layers
s = time()
L, dLdl10 = softmax_loss(layer10, Y)
e = time()
if self.verbose:
print """ Softmax loss calc done : %.2f sec """ % (e - s)
s = time()
dLdl9, dLdA10, dLdb10 = linear_backward(dLdl10, layer9, A10)
e = time()
if self.verbose:
print """ Layer10: FC (1 x 1 x 10) backward done: %.2f sec """ % (
e - s)
""" Pool (4 x 4 x 20) Backward """
s = time()
dLdl8 = max_pool_backward(dLdl9, layer8, X_idx9, F9, S9)
e = time()
if self.verbose:
print """ Layer9: Pool (4 x 4 x 20) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl7 = ReLU_backward(dLdl8, layer7)
e = time()
if self.verbose:
print """ Layer8: ReLU (8 x 8 x 20) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl6, dLdA7, dLdb7 = conv_backward(dLdl7, layer6, X_col7, A7, S7, P7)
e = time()
if self.verbose:
print """ Layer7: Conv (8 x 8 x 20) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl5 = max_pool_backward(dLdl6, layer5, X_idx6, F6, S6)
e = time()
if self.verbose:
print """ Layer6: Pool (8 x 8 x 20) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl4 = ReLU_backward(dLdl5, layer4)
e = time()
if self.verbose:
print """ Layer5: ReLU (16 x 16 x 20) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl3, dLdA4, dLdb4 = conv_backward(dLdl4, layer3, X_col4, A4, S4, P4)
e = time()
if self.verbose:
print """ Layer4: Conv (16 x 16 x 20) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl2 = max_pool_backward(dLdl3, layer2, X_idx3, F3, S3)
e = time()
if self.verbose:
print """ Layer3: Pool (16 x 16 x 16) backward done: %.2f sec """ % (
e - s)
s = time()
dLdl1 = ReLU_backward(dLdl2, layer1)
e = time()
if self.verbose:
print """ Layer2: ReLU (32 x 32 x 16) backward done: %.2f sec """ % (
e - s)
s = time()
dLdX, dLdA1, dLdb1 = conv_backward(dLdl1, X, X_col1, A1, S1, P1)
e = time()
if self.verbose:
print """ Layer1: Conv (32 x 32 x 16) backward done: %.2f sec """ % (
e - s)
""" regularization """
L += 0.5 * self.lam * np.sum(A1 * A1)
L += 0.5 * self.lam * np.sum(A4 * A4)
L += 0.5 * self.lam * np.sum(A7 * A7)
L += 0.5 * self.lam * np.sum(A10 * A10)
""" regularization gradient """
dLdA10 = dLdA10.reshape(A10.shape)
dLdA7 = dLdA7.reshape(A7.shape)
dLdA4 = dLdA4.reshape(A4.shape)
dLdA1 = dLdA1.reshape(A1.shape)
dLdA10 += self.lam * A10
dLdA7 += self.lam * A7
dLdA4 += self.lam * A4
dLdA1 += self.lam * A1
""" tune the parameter """
self.v1 = self.mu * self.v1 - self.rho * dLdA1
self.v4 = self.mu * self.v4 - self.rho * dLdA4
self.v7 = self.mu * self.v7 - self.rho * dLdA7
self.v10 = self.mu * self.v10 - self.rho * dLdA10
self.A1 += self.v1
self.A4 += self.v4
self.A7 += self.v7
self.A10 += self.v10
self.b1 += -self.rho * dLdb1
self.b4 += -self.rho * dLdb4
self.b7 += -self.rho * dLdb7
self.b10 += -self.rho * dLdb10
""" dump """
if dump_chunks > 0:
dump_big_matrix(dLdl10, "cnn_dLdl10_mat", dump_chunks)
dump_big_matrix(dLdl9, "cnn_dLdl9_mat", dump_chunks)
dump_big_matrix(dLdl8, "cnn_dLdl8_mat", dump_chunks)
dump_big_matrix(dLdl7, "cnn_dLdl7_mat", dump_chunks)
dump_big_matrix(dLdl6, "cnn_dLdl6_mat", dump_chunks)
dump_big_matrix(dLdl5, "cnn_dLdl5_mat", dump_chunks)
dump_big_matrix(dLdl4, "cnn_dLdl4_mat", dump_chunks)
dump_big_matrix(dLdl3, "cnn_dLdl3_mat", dump_chunks)
dump_big_matrix(dLdl2, "cnn_dLdl2_mat", dump_chunks)
dump_big_matrix(dLdl1, "cnn_dLdl1_mat", dump_chunks)
dump_big_matrix(dLdX, "cnn_dLdX_mat", dump_chunks)
dump_big_matrix(dLdA10, "cnn_dLdA10_mat", 1)
dump_big_matrix(dLdb10, "cnn_dLdb10_mat", 1)
dump_big_matrix(dLdA7, "cnn_dLdA7_mat", 1)
dump_big_matrix(dLdb7, "cnn_dLdb7_mat", 1)
dump_big_matrix(dLdA4, "cnn_dLdA4_mat", 1)
dump_big_matrix(dLdb4, "cnn_dLdb4_mat", 1)
dump_big_matrix(dLdA1, "cnn_dLdA1_mat", 1)
dump_big_matrix(dLdb1, "cnn_dLdb1_mat", 1)
return L
def forward(self, X, dump_chunks=-1):
A1 = self.A1
b1 = self.b1
S1 = self.S1
P1 = self.P1
F3 = self.F3
S3 = self.S3
A4 = self.A4
b4 = self.b4
S4 = self.S4
P4 = self.P4
F6 = self.F6
S6 = self.S6
A7 = self.A7
b7 = self.b7
S7 = self.S7
P7 = self.P7
F9 = self.F9
S9 = self.S9
A10 = self.A10
b10 = self.b10
s = time()
layer1, X_col1 = conv_forward(X, A1, b1, S1, P1)
e = time()
if self.verbose:
print """ Layer1: Conv (32 x 32 x 16) forward doen: %.2f sec """ % (
e - s)
s = time()
layer2 = ReLU_forward(layer1)
e = time()
if self.verbose:
print """ Layer2: ReLU (32 x 32 x 16) forward done: %.2f sec """ % (
e - s)
s = time()
layer3, X_idx3 = max_pool_forward(layer2, F3, S3)
e = time()
if self.verbose:
print """ Layer3: Pool (16 x 16 x 16) forward done: %.2f sec """ % (
e - s)
s = time()
layer4, X_col4 = conv_forward(layer3, A4, b4, S4, P4)
e = time()
if self.verbose:
print """ Layer4: Conv (16 x 16 x 20) forward done: %.2f sec """ % (
e - s)
s = time()
layer5 = ReLU_forward(layer4)
e = time()
if self.verbose:
print """ Layer5: ReLU (16 x 16 x 20) forward done: %.2f sec """ % (
e - s)
s = time()
layer6, X_idx6 = max_pool_forward(layer5, F6, S6)
e = time()
if self.verbose:
print """ Layer6: Pool (8 x 8 x 20) forward done: %.2f sec """ % (
e - s)
s = time()
layer7, X_col7 = conv_forward(layer6, A7, b7, S7, P7)
e = time()
if self.verbose:
print """ Layer7: Conv (8 x 8 x 20) forward: %.2f sec """ % (e - s)
s = time()
layer8 = ReLU_forward(layer7)
e = time()
if self.verbose:
print """ Layer8: ReLU (8 x 8 x 20) forward done: %.2f sec """ % (
e - s)
s = time()
layer9, X_idx9 = max_pool_forward(layer8, F9, S9)
e = time()
if self.verbose:
print """ Layer9: Pool (4 x 4 x 20) forward done: %.2f sec """ % (
e - s)
s = time()
layer10 = linear_forward(layer9, A10, b10)
e = time()
if self.verbose:
print """ Layer10: FC (1 x 1 x 10) forward done: %.2f sec """ % (
e - s)
if dump_chunks > 0:
dump_big_matrix(layer1, "cnn_l1_mat", dump_chunks)
dump_big_matrix(layer2, "cnn_l2_mat", dump_chunks)
dump_big_matrix(layer3, "cnn_l3_mat", dump_chunks)
dump_big_matrix(layer4, "cnn_l4_mat", dump_chunks)
dump_big_matrix(layer5, "cnn_l5_mat", dump_chunks)
dump_big_matrix(layer6, "cnn_l6_mat", dump_chunks)
dump_big_matrix(layer7, "cnn_l7_mat", dump_chunks)
dump_big_matrix(layer8, "cnn_l8_mat", dump_chunks)
dump_big_matrix(layer9, "cnn_l9_mat", dump_chunks)
dump_big_matrix(layer10, "cnn_l10_mat", dump_chunks)
return [(layer1, X_col1), layer2, (layer3, X_idx3), (layer4, X_col4),
layer5, (layer6, X_idx6), (layer7, X_col7), layer8,
(layer9, X_idx9), layer10]
def backward(self, X, layers, Y, dump_chunks = -1):
A1 = self.A1
b1 = self.b1
A3 = self.A3
b3 = self.b3
layer1, layer2, layer3 = layers
""" softmax classification """
L, dLdl3 = softmax_loss(layer3, Y)
""" backpropagation for Layer 3 """
dLdl2, dLdA3, dLdb3 = linear_backward(dLdl3, layer2, A3)
""" backpropagation for Layer 2 """
dLdl1 = ReLU_backward(dLdl2, layer1)
""" backpropagation for Layer 1 """
dLdX, dLdA1, dLdb1 = linear_backward(dLdl1, X, A1)
""" regularization """
L += 0.5 * self.lam * (np.sum(A1*A1) + np.sum(A3*A3))
""" regularization gradient """
dLdA3 = dLdA3.reshape(A3.shape)
dLdA1 = dLdA1.reshape(A1.shape)
dLdA3 += self.lam * A3
dLdA1 += self.lam * A1
""" tune the parameter """
self.v1 = self.mu * self.v1 - self.rho * dLdA1
self.v3 = self.mu * self.v3 - self.rho * dLdA3
self.A1 += self.v1
self.A3 += self.v3
self.b1 += - self.rho * dLdb1
self.b3 += - self.rho * dLdb3
""" dump """
if dump_chunks > 0:
dump_big_matrix(dLdl3, "nn_dLdl3_mat", dump_chunks)
dump_big_matrix(dLdl2, "nn_dLdl2_mat", dump_chunks)
dump_big_matrix(dLdl1, "nn_dLdl1_mat", dump_chunks)
dump_big_matrix(dLdX, "nn_dLdX_mat", dump_chunks)
dump_big_matrix(dLdA3, "nn_dLdA3_mat", 1)
dump_big_matrix(dLdb3, "nn_dLdb3_mat", 1)
dump_big_matrix(dLdA1, "nn_dLdA1_mat", 1)
dump_big_matrix(dLdb1, "nn_dLdb1_mat", 1)
return L
def backward(self, X, layers, Y, dump_chunks = -1):
A1 = self.A1
b1 = self.b1
S1 = self.S1
P1 = self.P1
F3 = self.F3
S3 = self.S3
A4 = self.A4
b4 = self.b4
S4 = self.S4
P4 = self.P4
F6 = self.F6
S6 = self.S6
A7 = self.A7
b7 = self.b7
S7 = self.S7
P7 = self.P7
F9 = self.F9
S9 = self.S9
A10 = self.A10
b10 = self.b10
(layer1, X_col1), layer2, (layer3, X_idx3), \
(layer4, X_col4), layer5, (layer6, X_idx6), \
(layer7, X_col7), layer8, (layer9, X_idx9), layer10 = layers
s = time()
L, dLdl10 = softmax_loss(layer10, Y)
e = time()
if self.verbose:
print """ Softmax loss calc done : %.2f sec """ % (e - s)
s = time()
dLdl9, dLdA10, dLdb10 = linear_backward(dLdl10, layer9 , A10)
e = time()
if self.verbose:
print """ Layer10: FC (1 x 1 x 10) backward done: %.2f sec """ % (e - s)
""" Pool (4 x 4 x 20) Backward """
s = time()
dLdl8 = max_pool_backward(dLdl9, layer8, X_idx9, F9, S9)
e = time()
if self.verbose:
print """ Layer9: Pool (4 x 4 x 20) backward done: %.2f sec """ % (e - s)
s = time()
dLdl7 = ReLU_backward(dLdl8, layer7)
e = time()
if self.verbose:
print """ Layer8: ReLU (8 x 8 x 20) backward done: %.2f sec """ % (e - s)
s = time()
dLdl6, dLdA7, dLdb7 = conv_backward(dLdl7, layer6, X_col7, A7, S7, P7)
e = time()
if self.verbose:
print """ Layer7: Conv (8 x 8 x 20) backward done: %.2f sec """ % (e - s)
s = time()
dLdl5 = max_pool_backward(dLdl6, layer5, X_idx6, F6, S6)
e = time()
if self.verbose:
print """ Layer6: Pool (8 x 8 x 20) backward done: %.2f sec """ % (e - s)
s = time()
dLdl4 = ReLU_backward(dLdl5, layer4)
e = time()
if self.verbose:
print """ Layer5: ReLU (16 x 16 x 20) backward done: %.2f sec """ % (e - s)
s = time()
dLdl3, dLdA4, dLdb4 = conv_backward(dLdl4, layer3, X_col4, A4, S4, P4)
e = time()
if self.verbose:
print """ Layer4: Conv (16 x 16 x 20) backward done: %.2f sec """ % (e - s)
s = time()
dLdl2 = max_pool_backward(dLdl3, layer2, X_idx3, F3, S3)
e = time()
if self.verbose:
print """ Layer3: Pool (16 x 16 x 16) backward done: %.2f sec """ % (e - s)
s = time()
dLdl1 = ReLU_backward(dLdl2, layer1)
e = time()
if self.verbose:
print """ Layer2: ReLU (32 x 32 x 16) backward done: %.2f sec """ % (e - s)
s = time()
dLdX, dLdA1, dLdb1 = conv_backward(dLdl1, X, X_col1, A1, S1, P1)
e = time()
if self.verbose:
print """ Layer1: Conv (32 x 32 x 16) backward done: %.2f sec """ % (e - s)
""" regularization """
L += 0.5 * self.lam * np.sum(A1*A1)
L += 0.5 * self.lam * np.sum(A4*A4)
L += 0.5 * self.lam * np.sum(A7*A7)
L += 0.5 * self.lam * np.sum(A10*A10)
""" regularization gradient """
dLdA10 = dLdA10.reshape(A10.shape)
dLdA7 = dLdA7.reshape(A7.shape)
dLdA4 = dLdA4.reshape(A4.shape)
dLdA1 = dLdA1.reshape(A1.shape)
dLdA10 += self.lam * A10
dLdA7 += self.lam * A7
dLdA4 += self.lam * A4
dLdA1 += self.lam * A1
""" tune the parameter """
self.v1 = self.mu * self.v1 - self.rho * dLdA1
self.v4 = self.mu * self.v4 - self.rho * dLdA4
self.v7 = self.mu * self.v7 - self.rho * dLdA7
self.v10 = self.mu * self.v10 - self.rho * dLdA10
self.A1 += self.v1
self.A4 += self.v4
self.A7 += self.v7
self.A10 += self.v10
self.b1 += -self.rho * dLdb1
self.b4 += -self.rho * dLdb4
self.b7 += -self.rho * dLdb7
self.b10 += -self.rho * dLdb10
""" dump """
if dump_chunks > 0:
dump_big_matrix(dLdl10, "cnn_dLdl10_mat", dump_chunks)
dump_big_matrix(dLdl9, "cnn_dLdl9_mat", dump_chunks)
dump_big_matrix(dLdl8, "cnn_dLdl8_mat", dump_chunks)
dump_big_matrix(dLdl7, "cnn_dLdl7_mat", dump_chunks)
dump_big_matrix(dLdl6, "cnn_dLdl6_mat", dump_chunks)
dump_big_matrix(dLdl5, "cnn_dLdl5_mat", dump_chunks)
dump_big_matrix(dLdl4, "cnn_dLdl4_mat", dump_chunks)
dump_big_matrix(dLdl3, "cnn_dLdl3_mat", dump_chunks)
dump_big_matrix(dLdl2, "cnn_dLdl2_mat", dump_chunks)
dump_big_matrix(dLdl1, "cnn_dLdl1_mat", dump_chunks)
dump_big_matrix(dLdX, "cnn_dLdX_mat", dump_chunks)
dump_big_matrix(dLdA10, "cnn_dLdA10_mat", 1)
dump_big_matrix(dLdb10, "cnn_dLdb10_mat", 1)
dump_big_matrix(dLdA7, "cnn_dLdA7_mat", 1)
dump_big_matrix(dLdb7, "cnn_dLdb7_mat", 1)
dump_big_matrix(dLdA4, "cnn_dLdA4_mat", 1)
dump_big_matrix(dLdb4, "cnn_dLdb4_mat", 1)
dump_big_matrix(dLdA1, "cnn_dLdA1_mat", 1)
dump_big_matrix(dLdb1, "cnn_dLdb1_mat", 1)
return L
def forward(self, X, dump_chunks = -1):
A1 = self.A1
b1 = self.b1
S1 = self.S1
P1 = self.P1
F3 = self.F3
S3 = self.S3
A4 = self.A4
b4 = self.b4
S4 = self.S4
P4 = self.P4
F6 = self.F6
S6 = self.S6
A7 = self.A7
b7 = self.b7
S7 = self.S7
P7 = self.P7
F9 = self.F9
S9 = self.S9
A10 = self.A10
b10 = self.b10
s = time()
layer1, X_col1 = conv_forward(X, A1, b1, S1, P1)
e = time()
if self.verbose:
print """ Layer1: Conv (32 x 32 x 16) forward doen: %.2f sec """ % (e - s)
s = time()
layer2 = ReLU_forward(layer1)
e = time()
if self.verbose:
print """ Layer2: ReLU (32 x 32 x 16) forward done: %.2f sec """ % (e - s)
s = time()
layer3, X_idx3 = max_pool_forward(layer2, F3, S3)
e = time()
if self.verbose:
print """ Layer3: Pool (16 x 16 x 16) forward done: %.2f sec """ % (e - s)
s = time()
layer4, X_col4 = conv_forward(layer3, A4, b4, S4, P4)
e = time()
if self.verbose:
print """ Layer4: Conv (16 x 16 x 20) forward done: %.2f sec """ % (e - s)
s = time()
layer5 = ReLU_forward(layer4)
e = time()
if self.verbose:
print """ Layer5: ReLU (16 x 16 x 20) forward done: %.2f sec """ % (e - s)
s = time()
layer6, X_idx6 = max_pool_forward(layer5, F6, S6)
e = time()
if self.verbose:
print """ Layer6: Pool (8 x 8 x 20) forward done: %.2f sec """ % (e - s)
s = time()
layer7, X_col7 = conv_forward(layer6, A7, b7, S7, P7)
e = time()
if self.verbose:
print """ Layer7: Conv (8 x 8 x 20) forward: %.2f sec """ % (e - s)
s = time()
layer8 = ReLU_forward(layer7)
e = time()
if self.verbose:
print """ Layer8: ReLU (8 x 8 x 20) forward done: %.2f sec """ % (e - s)
s = time()
layer9, X_idx9 = max_pool_forward(layer8, F9, S9)
e = time()
if self.verbose:
print """ Layer9: Pool (4 x 4 x 20) forward done: %.2f sec """ % (e - s)
s = time()
layer10 = linear_forward(layer9, A10, b10)
e = time()
if self.verbose:
print """ Layer10: FC (1 x 1 x 10) forward done: %.2f sec """ % (e - s)
if dump_chunks > 0:
dump_big_matrix(layer1, "cnn_l1_mat", dump_chunks)
dump_big_matrix(layer2, "cnn_l2_mat", dump_chunks)
dump_big_matrix(layer3, "cnn_l3_mat", dump_chunks)
dump_big_matrix(layer4, "cnn_l4_mat", dump_chunks)
dump_big_matrix(layer5, "cnn_l5_mat", dump_chunks)
dump_big_matrix(layer6, "cnn_l6_mat", dump_chunks)
dump_big_matrix(layer7, "cnn_l7_mat", dump_chunks)
dump_big_matrix(layer8, "cnn_l8_mat", dump_chunks)
dump_big_matrix(layer9, "cnn_l9_mat", dump_chunks)
dump_big_matrix(layer10, "cnn_l10_mat", dump_chunks)
return [
(layer1, X_col1), layer2, (layer3, X_idx3),
(layer4, X_col4), layer5, (layer6, X_idx6),
(layer7, X_col7), layer8, (layer9, X_idx9), layer10]
