def matmul_opt(app, W, D, num_gpus):
# Section 1
cluster_state = ClusterState((num_gpus, 1), app.system)
W_ga: GraphArray = GraphArray.from_ba(W, cluster_state)
D_ga: GraphArray = GraphArray.from_ba(D, cluster_state)
initend = time.time()
# Section 2
Z_ga: GraphArray = opt.collapse_graph_array(app, W_ga @ D_ga)
endtime = time.time()
Z: BlockArray = opt.compute_graph_array(app, Z_ga)
Z.touch()
del Z
return initend, endtime
def forward(app, X, W):
Z = opt.collapse_graph_array(app, X @ W)
return Z
def one_step_fit_opt(app,
X,
y,
W_in_1,
W_1_2,
W_2_out,
num_gpus,
verbose=False):
# --forward propagation--
LR = app.one
cluster_state = ClusterState((num_gpus, 1), app.system)
one_ga: GraphArray = GraphArray.from_ba(app.one, cluster_state)
X_ga = GraphArray.from_ba(X, cluster_state)
y_ga = GraphArray.from_ba(y, cluster_state)
W_in_1_ga = GraphArray.from_ba(W_in_1, cluster_state)
W_1_2_ga = GraphArray.from_ba(W_1_2, cluster_state)
W_2_out_ga = GraphArray.from_ba(W_2_out, cluster_state)
if verbose:
print("forward Z_1_ga")
Z_1_ga: GraphArray = forward(app, X_ga, W_in_1_ga)
if verbose:
print("forward S_1_ga")
S_1_ga: GraphArray = opt.sigmoid(app, Z_1_ga, one_ga)
if verbose:
print("forward F_1_ga")
F_1_ga: GraphArray = opt.sigmoid_deriv(app, Z_1_ga, one_ga)
if verbose:
print("forward Z_2_ga")
Z_2_ga: GraphArray = forward(app, S_1_ga, W_1_2_ga)
S_2_ga: GraphArray = opt.sigmoid(app, Z_2_ga, one_ga)
F_2_ga: GraphArray = opt.sigmoid_deriv(app, Z_2_ga, one_ga)
if verbose:
print("forward Z_out_ga")
Z_out_ga: GraphArray = forward(app, S_2_ga, W_2_out_ga)
if verbose:
print("forward y_predict_ga")
y_predict_ga: GraphArray = opt.sigmoid(app, Z_out_ga, one_ga)
if verbose:
print("forward F_out_ga")
F_out_ga: GraphArray = opt.sigmoid_deriv(app, Z_out_ga, one_ga)
initend = time.time()
# --back propagation--
if verbose:
print("collapse D_out_ga")
D_out_ga = opt.collapse_graph_array(app,
F_out_ga.T * (y_predict_ga - y_ga).T)
if verbose:
print("collapse D_2_ga")
D_2_ga = opt.collapse_graph_array(app, F_2_ga.T * (W_2_out_ga @ D_out_ga))
if verbose:
print("collapse D_1_ga")
D_1_ga = opt.collapse_graph_array(app, F_1_ga.T * (W_1_2_ga @ D_2_ga))
distribute_graph_array(D_1_ga, cluster_state)
if verbose:
print("collapse_graph_array dW_in_1_ga")
dW_in_1_ga = opt.collapse_graph_array(app, (D_1_ga @ X_ga).T)
if verbose:
print("collapse_graph_array dW_1_2_ga")
dW_1_2_ga = opt.collapse_graph_array(app, (D_2_ga @ S_1_ga).T)
if verbose:
print("collapse_graph_array dW_2_out_ga")
dW_2_out_ga = opt.collapse_graph_array(app, (D_out_ga @ S_2_ga).T)
endtime = time.time()
dW_in_1_ga_ba: BlockArray = opt.compute_graph_array(app, dW_in_1_ga)
dW_1_2_ga_ba: BlockArray = opt.compute_graph_array(app, dW_1_2_ga)
dW_2_out_ga_ba: BlockArray = opt.compute_graph_array(app, dW_2_out_ga)
if verbose:
print("update W_in_1")
W_in_1 = W_in_1 - LR * dW_in_1_ga_ba
if verbose:
print("update W_1_2")
W_1_2 = W_1_2 - LR * dW_1_2_ga_ba
if verbose:
print("update W_2_out")
W_2_out = W_2_out - LR * dW_2_out_ga_ba
W_in_1.touch()
W_1_2.touch()
W_2_out.touch()
return initend, endtime
def one_step_fit_opt(app, X, y, W_in_1, W_1_2, W_2_out, B_1, B_2, B_out,
num_gpus):
# --forward proprogation--
# print("start forward proprogation")
LR = app.one
cluster_state = ClusterState((num_gpus, 1), app.system)
one_ga: GraphArray = GraphArray.from_ba(app.one, cluster_state)
X_ga = GraphArray.from_ba(X, cluster_state)
# print(f"X_ga block_shape {X_ga.block_shape}")
y_ga = GraphArray.from_ba(y, cluster_state)
W_in_1_ga = GraphArray.from_ba(W_in_1, cluster_state)
W_1_2_ga = GraphArray.from_ba(W_1_2, cluster_state)
W_2_out_ga = GraphArray.from_ba(W_2_out, cluster_state)
B_1_ga = GraphArray.from_ba(B_1, cluster_state)
B_2_ga = GraphArray.from_ba(B_2, cluster_state)
B_out_ga = GraphArray.from_ba(B_out, cluster_state)
Z_1_ga: GraphArray = forward(app, X_ga, W_in_1_ga, B_1_ga)
S_1_ga: GraphArray = opt.sigmoid(app, Z_1_ga, one_ga)
F_1_ga: GraphArray = opt.sigmoid_deriv(app, Z_1_ga, one_ga)
# print(f"S_1.shape {S_1.shape} S_1.block_shape {S_1.block_shape}")
# Z_1_ga: GraphArray = opt.relu(S_1_ga, zero_ga)
# print(f"Z_1.shape {Z_1.shape} Z_1.block_shape {Z_1.block_shape}")
# F_1_ga: GraphArray = opt.relu_deriv(S_1_ga, zero_ga, one_ga)
Z_2_ga: GraphArray = forward(app, S_1_ga, W_1_2_ga, B_2_ga)
S_2_ga: GraphArray = opt.sigmoid(app, Z_2_ga, one_ga)
F_2_ga: GraphArray = opt.sigmoid_deriv(app, Z_2_ga, one_ga)
# Z_2_ga: GraphArray = opt.relu(S_2_ga, zero_ga)
# print(f"S_2.shape {S_2.shape} S_2.block_shape {S_2.block_shape}")
# F_2_ga: GraphArray = opt.relu_deriv(S_2_ga, zero_ga, one_ga)
Z_out_ga: GraphArray = forward(app, S_2_ga, W_2_out_ga, B_out_ga)
y_predict_ga: GraphArray = opt.sigmoid(app, Z_out_ga, one_ga)
# y_predict_ga: GraphArray = opt.relu(S_out_ga, zero_ga)
# print("start back propogation")
# --back propogation--
D_out_ga = opt.collapse_graph_array(app, (y_predict_ga - y_ga).T)
D_2_ga = opt.collapse_graph_array(app, F_2_ga.T * (W_2_out_ga @ D_out_ga))
D_1_ga = opt.collapse_graph_array(app, F_1_ga.T * (W_1_2_ga @ D_2_ga))
D_out_ga_ba = opt.compute_graph_array(app, D_out_ga)
D_2_ga_ba = opt.compute_graph_array(app, D_2_ga)
D_1_ga_ba = opt.compute_graph_array(app, D_1_ga)
S_1_ga_ba = opt.compute_graph_array(app, S_1_ga)
S_2_ga_ba = opt.compute_graph_array(app, S_2_ga)
W_in_1: BlockArray = update_weight(app, LR, W_in_1, D_1_ga_ba, X)
W_1_2: BlockArray = update_weight(app, LR, W_1_2, D_2_ga_ba, S_1_ga_ba)
W_2_out: BlockArray = update_weight(app, LR, W_2_out, D_out_ga_ba,
S_2_ga_ba)
B_1: BlockArray = update_bias(app, LR, B_1, D_1_ga_ba)
B_2: BlockArray = update_bias(app, LR, B_2, D_2_ga_ba)
B_out: BlockArray = update_bias(app, LR, B_out, D_out_ga_ba)
# print("Start touching")
W_in_1.touch()
W_1_2.touch()
W_2_out.touch()
B_1.touch()
B_2.touch()
B_out.touch()
return W_in_1, W_1_2, W_2_out, B_1, B_2, B_out
def one_step_fit_opt(app,
X,
y,
W_in_1,
W_1_2,
W_2_out,
num_gpus,
verbose=False):
# --forward proprogation--
# print("start forward proprogation")
LR = app.one
cluster_state = ClusterState((num_gpus, 1), app.system)
one_ga: GraphArray = GraphArray.from_ba(app.one, cluster_state)
X_ga = GraphArray.from_ba(X, cluster_state)
# print(f"X_ga block_shape {X_ga.block_shape}")
y_ga = GraphArray.from_ba(y, cluster_state)
W_in_1_ga = GraphArray.from_ba(W_in_1, cluster_state)
W_1_2_ga = GraphArray.from_ba(W_1_2, cluster_state)
W_2_out_ga = GraphArray.from_ba(W_2_out, cluster_state)
# Distribute Weights
distribute_weights(app.one, cluster_state)
# distribute_weights(X, cluster_state)
# distribute_weights(y, cluster_state)
if verbose:
print("forward Z_1_ga")
Z_1_ga: GraphArray = forward(app, X_ga, W_in_1_ga) # --> 0/1
if verbose:
print("forward S_1_ga")
S_1_ga: GraphArray = opt.sigmoid(app, Z_1_ga, one_ga) # --> 0/1
# distribute_weights(S_1_ga, cluster_state)
if verbose:
print("forward F_1_ga")
F_1_ga: GraphArray = opt.sigmoid_deriv(app, Z_1_ga, one_ga) # --> 0/1
# print(f"S_1.shape {S_1.shape} S_1.block_shape {S_1.block_shape}")
# Z_1_ga: GraphArray = opt.relu(S_1_ga, zero_ga)
# print(f"Z_1.shape {Z_1.shape} Z_1.block_shape {Z_1.block_shape}")
# F_1_ga: GraphArray = opt.relu_deriv(S_1_ga, zero_ga, one_ga)
if verbose:
print("forward Z_2_ga")
Z_2_ga: GraphArray = forward(app, S_1_ga, W_1_2_ga)
S_2_ga: GraphArray = opt.sigmoid(app, Z_2_ga, one_ga)
F_2_ga: GraphArray = opt.sigmoid_deriv(app, Z_2_ga, one_ga)
# Z_2_ga: GraphArray = opt.relu(S_2_ga, zero_ga)
# print(f"S_2.shape {S_2.shape} S_2.block_shape {S_2.block_shape}")
# F_2_ga: GraphArray = opt.relu_deriv(S_2_ga, zero_ga, one_ga)
if verbose:
print("forward Z_out_ga")
Z_out_ga: GraphArray = forward(app, S_2_ga, W_2_out_ga) # --> 0/1
if verbose:
print("forward y_predict_ga")
y_predict_ga: GraphArray = opt.sigmoid(app, Z_out_ga, one_ga) # --> 0/1
if verbose:
print("forward F_out_ga")
F_out_ga: GraphArray = opt.sigmoid_deriv(app, Z_out_ga, one_ga) # --> 0/1
# print(F_out_ga.shape) -> (1000,)
# y_predict_ga: GraphArray = opt.relu(S_out_ga, zero_ga)
initend = time.time()
if verbose:
print(
"-----------------------------start back propogation-------------------------------"
)
print(
"-----------------------------start back propogation-------------------------------"
)
print(
"-----------------------------start back propogation-------------------------------"
)
# --back propogation--
if verbose:
print("collapse D_out_ga")
D_out_ga = opt.collapse_graph_array(app,
F_out_ga.T *
(y_predict_ga - y_ga).T) # --> 0/1
# D_out_ga = opt.collapse_graph_array(app, (y_predict_ga - y_ga) * F_out_ga)
if verbose:
print("collapse D_2_ga")
# print(f"W_2_out_ga shape {W_2_out_ga.shape}") -> (2048,)
# print(f"D_out_ga shape {D_out_ga.shape}") -> (1000,)
# F_2_ga.shape -> (1000, 2048)
D_2_ga = opt.collapse_graph_array(app, F_2_ga.T * (W_2_out_ga @ D_out_ga))
# D_2_ga = opt.collapse_graph_array(app, (D_out_ga @ W_2_out_ga.T) * F_2_ga)
if verbose:
print("collapse D_1_ga")
D_1_ga = opt.collapse_graph_array(app, F_1_ga.T *
(W_1_2_ga @ D_2_ga)) # --> 0/1
distribute_graph_array(D_1_ga, cluster_state)
# print(D_1_ga.shape)
# D_1_ga = opt.collapse_graph_array(app, (D_2_ga @ W_1_2_ga.T) * F_1_ga)
# print("-----------------------------start computing weights-------------------------------")
# print("-----------------------------start computing weights-------------------------------")
# print("-----------------------------start computing weights-------------------------------")
if verbose:
print("collapse_graph_array dW_in_1_ga")
dW_in_1_ga = opt.collapse_graph_array(
app, (D_1_ga @ X_ga).T) # --> now all exeucted on GPU 0
if verbose:
print("collapse_graph_array dW_1_2_ga")
dW_1_2_ga = opt.collapse_graph_array(app, (D_2_ga @ S_1_ga).T)
if verbose:
print("collapse_graph_array dW_2_out_ga")
dW_2_out_ga = opt.collapse_graph_array(app, (D_out_ga @ S_2_ga).T)
endtime = time.time()
dW_in_1_ga_ba: BlockArray = opt.compute_graph_array(app, dW_in_1_ga)
dW_1_2_ga_ba: BlockArray = opt.compute_graph_array(app, dW_1_2_ga)
dW_2_out_ga_ba: BlockArray = opt.compute_graph_array(app, dW_2_out_ga)
# W_in_1_ga = opt.collapse_graph_array(app, W_in_1_ga - one_ga * (D_1_ga @ X_ga).T)
# print("collapse_graph_array W_1_2_ga")
# W_1_2_ga = opt.collapse_graph_array(app, W_1_2_ga - one_ga * (D_2_ga @ S_1_ga).T)
# print("collapse_graph_array W_2_out_ga")
# W_2_out_ga = opt.collapse_graph_array(app, W_2_out_ga - one_ga * (D_out_ga @ S_2_ga).T)
# W_in_1: BlockArray = opt.compute_graph_array(app, W_in_1_ga)
# W_1_2: BlockArray = opt.compute_graph_array(app, W_1_2_ga)
# W_2_out: BlockArray = opt.compute_graph_array(app, W_2_out_ga)
if verbose:
print("update W_in_1")
W_in_1 -= dW_in_1_ga_ba
if verbose:
print("update W_1_2")
W_1_2 -= dW_1_2_ga_ba
if verbose:
print("update W_2_out")
W_2_out -= dW_2_out_ga_ba
# D_out_ga_ba = opt.compute_graph_array(app, D_out_ga)
# D_2_ga_ba = opt.compute_graph_array(app, D_2_ga)
# D_1_ga_ba = opt.compute_graph_array(app, D_1_ga)
# S_1_ga_ba = opt.compute_graph_array(app, S_1_ga)
# S_2_ga_ba = opt.compute_graph_array(app, S_2_ga)
# W_in_1: BlockArray = update_weight(app, LR, W_in_1, D_1_ga_ba, X)
# W_1_2: BlockArray = update_weight(app, LR, W_1_2, D_2_ga_ba, S_1_ga_ba)
# W_2_out: BlockArray = update_weight(app, LR, W_2_out, D_out_ga_ba, S_2_ga_ba)
# W - LR * (D @ X).T
# print("Start touching")
W_in_1.touch()
W_1_2.touch()
W_2_out.touch()
return initend, endtime