def test_packing_weights():
il_nn = 2
hl_nn = [2, 3]
ol_nn = 2
nn = NeuralNetwork(il_nn, hl_nn, ol_nn)
weights = [[[1.9911426, -1.2063], [1.9911426, -1.2063]],\
[[0.23531701, 2.9445081], [0.23531701, 2.9445081],\
[0.23531701, 2.9445081]], [[0.065952107, 0.82053429, -0.21171442],\
[0.065952107, 0.82053429, -0.21171442]]]
expected_packed_weights = [1.9911426, -1.2063, 1.9911426, -1.2063,\
0.23531701, 2.9445081, 0.23531701, 2.9445081, 0.23531701, 2.9445081,\
0.065952107, 0.82053429, -0.21171442, 0.065952107, 0.82053429,\
-0.21171442]
packed_weights = nn.pack_weights(weights)
for w in range(len(packed_weights)):
assert (packed_weights[w] == expected_packed_weights[w])
def test_demo(web3, chain):
_hashed_data_groups = []
accuracy_criteria = 5000 # 50.00%
total_gas_used = 0
timeout = 180
w_scale = 1000 # Scale up weights by 1000x
b_scale = 1000 # Scale up biases by 1000x
danku, deploy_tx = chain.provider.get_or_deploy_contract('Danku_demo')
deploy_receipt = wait_for_transaction_receipt(web3,
deploy_tx,
timeout=timeout)
total_gas_used += deploy_receipt["gasUsed"]
dbg.dprint("Deploy gas: " + str(deploy_receipt["gasUsed"]))
offer_account = web3.eth.accounts[1]
solver_account = web3.eth.accounts[2]
# Fund contract
fund_tx = web3.eth.sendTransaction({
'from': offer_account,
'to': danku.address,
'value': web3.toWei(1, "ether")
})
fund_receipt = wait_for_transaction_receipt(web3, fund_tx, timeout=timeout)
total_gas_used += fund_receipt["gasUsed"]
dbg.dprint("Fund gas: " + str(fund_receipt["gasUsed"]))
# Check that offerer was deducted
bal = web3.eth.getBalance(offer_account)
# Deduct reward amount (1 ETH) and gas cost (21040 wei)
assert bal == 999998999999999999978960
wallet_amount = 1000000000000000000000000 # minus the reward amount
scd = DemoDataset(training_percentage=0.8, partition_size=25)
scd.generate_nonce()
scd.sha_all_data_groups()
dbg.dprint("All data groups: " + str(scd.data))
dbg.dprint("All nonces: " + str(scd.nonce))
# Initialization step 1
dbg.dprint("Hashed data groups: " + str(scd.hashed_data_group))
dbg.dprint("Hashed Hex data groups: " +
str(list(map(lambda x: "0x" + x.hex(), scd.hashed_data_group))))
# Keep track of all block numbers, so we can send them in time
# Start at a random block between 0-1000
chain.wait.for_block(randbelow(1000))
dbg.dprint("Starting block: " + str(web3.eth.blockNumber))
init1_tx = danku.transact().init1(scd.hashed_data_group, accuracy_criteria,
offer_account)
init1_receipt = wait_for_transaction_receipt(web3,
init1_tx,
timeout=timeout)
total_gas_used += init1_receipt["gasUsed"]
dbg.dprint("Init1 gas: " + str(init1_receipt["gasUsed"]))
chain.wait.for_receipt(init1_tx)
init1_block_number = web3.eth.blockNumber
dbg.dprint("Init1 block: " + str(init1_block_number))
submission_t = danku.call().submission_stage_block_size(
) # get submission timeframe
evaluation_t = danku.call().evaluation_stage_block_size(
) # get evaluation timeframe
test_reveal_t = danku.call().reveal_test_data_groups_block_size(
) # get revealing testing dataset timeframe
# Initialization step 2
# Get data group indexes
chain.wait.for_block(init1_block_number + 1)
dgi = []
init2_block_number = web3.eth.blockNumber
dbg.dprint("Init2 block: " + str(init2_block_number))
for i in range(scd.num_data_groups):
dgi.append(i)
dbg.dprint("Data group indexes: " + str(dgi))
init2_tx = danku.transact().init2()
init2_receipt = wait_for_transaction_receipt(web3,
init2_tx,
timeout=timeout)
total_gas_used += init2_receipt["gasUsed"]
dbg.dprint("Init2 gas: " + str(init2_receipt["gasUsed"]))
chain.wait.for_receipt(init2_tx)
# Can only access one element of a public array at a time
training_partition = list(map(lambda x: danku.call().training_partition(x),\
range(scd.num_train_data_groups)))
testing_partition = list(map(lambda x: danku.call().testing_partition(x),\
range(scd.num_test_data_groups)))
# get partitions
dbg.dprint("Training partition: " + str(training_partition))
dbg.dprint("Testing partition: " + str(testing_partition))
scd.partition_dataset(training_partition, testing_partition)
# Initialization step 3
# Time to reveal the training dataset
training_nonces = []
training_data = []
for i in training_partition:
training_nonces.append(scd.nonce[i])
# Pack data into a 1-dimension array
# Since the data array is too large, we're going to send them in single data group chunks
train_data = scd.pack_data(scd.train_data)
test_data = scd.pack_data(scd.test_data)
init3_tx = []
for i in range(len(training_partition)):
start = i * scd.dps * scd.partition_size
end = start + scd.dps * scd.partition_size
dbg.dprint("(" + str(training_partition[i]) + ") Train data,nonce: " +
str(train_data[start:end]) + "," + str(scd.train_nonce[i]))
iter_tx = danku.transact().init3(train_data[start:end],
scd.train_nonce[i])
iter_receipt = wait_for_transaction_receipt(web3,
iter_tx,
timeout=timeout)
total_gas_used += iter_receipt["gasUsed"]
dbg.dprint("Reveal train data iter " + str(i) + " gas: " +
str(iter_receipt["gasUsed"]))
init3_tx.append(iter_tx)
chain.wait.for_receipt(init3_tx[i])
init3_block_number = web3.eth.blockNumber
dbg.dprint("Init3 block: " + str(init3_block_number))
# Get the training data from the contract
contract_train_data_length = danku.call().get_train_data_length()
contract_train_data = []
for i in range(contract_train_data_length):
for j in range(scd.dps):
contract_train_data.append(danku.call().train_data(i, j))
contract_train_data = scd.unpack_data(contract_train_data)
dbg.dprint("Contract training data: " + str(contract_train_data))
il_nn = 2
hl_nn = [4, 4]
ol_nn = 2
# Train a neural network with contract data
nn = NeuralNetwork(il_nn, hl_nn, ol_nn, 0.001, 1000000, 5, 100000)
contract_train_data = nn.binary_2_one_hot(contract_train_data)
nn.load_train_data(contract_train_data)
nn.init_network()
nn.train()
trained_weights = nn.weights
trained_biases = nn.bias
dbg.dprint("Trained weights: " + str(trained_weights))
dbg.dprint("Trained biases: " + str(trained_biases))
packed_trained_weights = nn.pack_weights(trained_weights)
dbg.dprint("Packed weights: " + str(packed_trained_weights))
packed_trained_biases = nn.pack_biases(trained_biases)
dbg.dprint("Packed biases: " + str(packed_trained_biases))
int_packed_trained_weights = scale_packed_data(packed_trained_weights,\
w_scale)
dbg.dprint("Packed integer weights: " + str(int_packed_trained_weights))
int_packed_trained_biases = scale_packed_data(packed_trained_biases,\
b_scale)
dbg.dprint("Packed integer biases: " + str(int_packed_trained_biases))
dbg.dprint("Solver address: " + str(solver_account))
# Submit the solution to the contract
submit_tx = danku.transact().submit_model(solver_account, il_nn, ol_nn, hl_nn,\
int_packed_trained_weights, int_packed_trained_biases)
submit_receipt = wait_for_transaction_receipt(web3,
submit_tx,
timeout=timeout)
total_gas_used += submit_receipt["gasUsed"]
dbg.dprint("Submit gas: " + str(submit_receipt["gasUsed"]))
chain.wait.for_receipt(submit_tx)
# Get submission index ID
submission_id = danku.call().get_submission_id(solver_account, il_nn,\
ol_nn, hl_nn, int_packed_trained_weights, int_packed_trained_biases)
dbg.dprint("Submission ID: " + str(submission_id))
# Wait until the submission period ends
chain.wait.for_block(init3_block_number + submission_t)
# Reveal the testing dataset after the submission period ends
reveal_tx = []
for i in range(len(testing_partition)):
start = i * scd.dps * scd.partition_size
end = start + scd.dps * scd.partition_size
dbg.dprint("(" + str(testing_partition[i]) + ") Test data,nonce: " +
str(test_data[start:end]) + "," + str(scd.test_nonce[i]))
iter_tx = danku.transact().reveal_test_data(test_data[start:end],
scd.test_nonce[i])
iter_receipt = wait_for_transaction_receipt(web3,
iter_tx,
timeout=timeout)
total_gas_used += iter_receipt["gasUsed"]
dbg.dprint("Reveal test data iter " + str(i) + " gas: " +
str(iter_receipt["gasUsed"]))
reveal_tx.append(iter_tx)
chain.wait.for_receipt(reveal_tx[i])
# Wait until the test reveal period ends
chain.wait.for_block(init3_block_number + submission_t + test_reveal_t)
# Evaluate the submitted solution
eval_tx = danku.transact().evaluate_model(submission_id)
eval_receipt = wait_for_transaction_receipt(web3, eval_tx, timeout=timeout)
total_gas_used += eval_receipt["gasUsed"]
dbg.dprint("Eval gas: " + str(eval_receipt["gasUsed"]))
# Wait until the evaluation period ends
chain.wait.for_block(init3_block_number + submission_t + test_reveal_t +
evaluation_t)
bal2 = web3.eth.getBalance(offer_account)
# Finalize the contract
final_tx = danku.transact().finalize_contract()
final_receipt = wait_for_transaction_receipt(web3,
final_tx,
timeout=timeout)
total_gas_used += final_receipt["gasUsed"]
dbg.dprint("Final gas: " + str(final_receipt["gasUsed"]))
contract_finalized = danku.call().contract_terminated()
dbg.dprint("Contract finalized: " + str(contract_finalized))
assert contract_finalized == True
# Get best submission accuracy & ID
best_submission_accuracy = danku.call().best_submission_accuracy()
best_submission_index = danku.call().best_submission_index()
dbg.dprint("Best submission ID: " + str(best_submission_index))
dbg.dprint("Best submission accuracy: " + str(best_submission_accuracy))
l_nn = [il_nn] + hl_nn + [ol_nn]
input_layer = train_data[:2]
hidden_layers = [0] * sum(hl_nn)
output_layer = [0] * ol_nn
weights = int_packed_trained_weights
biases = int_packed_trained_biases
# Test forward
fwd_pass2 = danku.call().forward_pass2(l_nn, input_layer, hidden_layers,
output_layer, weights, biases)
dbg.dprint("Test input: " + str(train_data[:2]))
dbg.dprint("Expected output: " + str(train_data[2]))
dbg.dprint("local nn prediction: " + str(nn.predict([train_data[:2]])))
dbg.dprint("forward_pass2: " + str(fwd_pass2))
dbg.dprint("Total gas used: " + str(total_gas_used))
scatter_x = np.array(list(map(lambda x: x[1:2][0], scd.data)))
scatter_y = np.array(list(map(lambda x: x[:1][0], scd.data)))
group = np.array(list(map(lambda x: x[2:3][0], scd.data)))
cdict = {0: "blue", 1: "red"}
fig, ax = plt.subplots()
for g in np.unique(group):
ix = np.where(group == g)
ax.scatter(scatter_x[ix], scatter_y[ix], c=cdict[g], label=g, s=4)
ax.legend()
plt.show()
bal = web3.eth.getBalance(solver_account)
# Verify that the solver account received the reward amount
assert bal == 1000001000000000000000000
bal = web3.eth.getBalance(offer_account)
# Verify the offer account balance
assert bal == 999998999999999999978960
assert (False)
hl_nn = [2, 3] # 2 hidden layers with 2 and 5 neurons respectively
ol_nn = 2 # 2 output neurons for binary classification
# Train a neural network with contract data
print("Training a neural network with the following:\n\
configuration: " + str(il_nn) + " x " + str(hl_nn) + " x " + str(ol_nn) +
"\n\
total iteration: 100000\n\
learning rate: 0.001\n")
nn = NeuralNetwork(il_nn, hl_nn, ol_nn, 0.001, 100000, 5, 10000)
nn.load_train_data(nn.binary_2_one_hot(contract_train_data))
nn.init_network()
nn.train()
trained_weights = nn.weights
trained_biases = nn.bias
packed_trained_weights = nn.pack_weights(trained_weights)
packed_trained_biases = nn.pack_biases(trained_biases)
int_packed_trained_weights = scale_packed_data(packed_trained_weights,\
w_scale)
int_packed_trained_biases = scale_packed_data(packed_trained_biases,\
b_scale)
print("Neural network trained!\n")
# Print submission data
print("Weights: " + str(int_packed_trained_weights) + "\n")
print("Biases: " + str(int_packed_trained_biases) + "\n")
print("Num_neurons_input_layer: " + str(il_nn) + "\n")
print("Num_neurons_output_layer: " + str(ol_nn) + "\n")
def test_single_solver_refunded_contract(web3, chain):
_hashed_data_groups = []
accuracy_criteria = 9950 # 99.50%
w_scale = 1000 # Scale up weights by 1000x
b_scale = 1000 # Scale up biases by 1000x
dbg.dprint("Start amount bal[0]: " + str(web3.eth.getBalance(web3.eth.accounts[0])))
cyphai, _ = chain.provider.get_or_deploy_contract('Cyphai')
offer_account = web3.eth.accounts[1]
solver_account = web3.eth.accounts[2]
# Fund contract
web3.eth.sendTransaction({
'from': offer_account,
'to': cyphai.address,
'value': web3.toWei(1, "ether")
})
# Check that offerer was deducted
bal = web3.eth.getBalance(offer_account)
# Deduct reward amount (1 ETH) and gas cost (21040 wei)
assert bal == 999998999999999999978960
wallet_amount = 1000000000000000000000000 # minus the reward amount
scd = SampleHalfDividedDataset(training_percentage=0.8)
scd.generate_nonce()
scd.sha_all_data_groups()
dbg.dprint("All data groups: " + str(scd.data))
dbg.dprint("All nonces: " + str(scd.nonce))
# Initialization step 1
dbg.dprint("Hashed data groups: " + str(scd.hashed_data_group))
dbg.dprint("Hashed Hex data groups: " +
str(list(map(lambda x: "0x" + x.hex(), scd.hashed_data_group))))
# Keep track of all block numbers, so we can send them in time
# Start at a random block between 0-1000
chain.wait.for_block(randbelow(1000))
dbg.dprint("Starting block: " + str(web3.eth.blockNumber))
init1_tx = cyphai.transact().init1(scd.hashed_data_group, accuracy_criteria,
offer_account)
chain.wait.for_receipt(init1_tx)
init1_block_number = web3.eth.blockNumber
dbg.dprint("Init1 block: " + str(init1_block_number))
submission_t = cyphai.call().submission_stage_block_size() # get submission timeframe
evaluation_t = cyphai.call().evaluation_stage_block_size() # get evaluation timeframe
test_reveal_t = cyphai.call().reveal_test_data_groups_block_size() # get revealing testing dataset timeframe
# Initialization step 2
# Get data group indexes
chain.wait.for_block(init1_block_number + 1)
dgi = []
init2_block_number = web3.eth.blockNumber
dbg.dprint("Init2 block: " + str(init2_block_number))
for i in range(scd.num_data_groups):
dgi.append(i)
dbg.dprint("Data group indexes: " + str(dgi))
init2_tx = cyphai.transact().init2()
chain.wait.for_receipt(init2_tx)
# Can only access one element of a public array at a time
training_partition = list(map(lambda x: cyphai.call().training_partition(x),\
range(scd.num_train_data_groups)))
testing_partition = list(map(lambda x: cyphai.call().testing_partition(x),\
range(scd.num_test_data_groups)))
# get partitions
dbg.dprint("Training partition: " + str(training_partition))
dbg.dprint("Testing partition: " + str(testing_partition))
scd.partition_dataset(training_partition, testing_partition)
# Initialization step 3
# Time to reveal the training dataset
training_nonces = []
training_data = []
for i in training_partition:
training_nonces.append(scd.nonce[i])
# Pack data into a 1-dimension array
# Since the data array is too large, we're going to send them in single data group chunks
train_data = scd.pack_data(scd.train_data)
test_data = scd.pack_data(scd.test_data)
init3_tx = []
for i in range(len(training_partition)):
start = i*scd.dps*scd.partition_size
end = start + scd.dps*scd.partition_size
dbg.dprint("(" + str(training_partition[i]) + ") Train data,nonce: " + str(train_data[start:end]) + "," + str(scd.train_nonce[i]))
init3_tx.append(cyphai.transact().init3(train_data[start:end], scd.train_nonce[i]))
chain.wait.for_receipt(init3_tx[i])
init3_block_number = web3.eth.blockNumber
dbg.dprint("Init3 block: " + str(init3_block_number))
# Get the training data from the contract
contract_train_data_length = cyphai.call().get_train_data_length()
contract_train_data = []
for i in range(contract_train_data_length):
for j in range(scd.dps):
contract_train_data.append(cyphai.call().train_data(i,j))
contract_train_data = scd.unpack_data(contract_train_data)
dbg.dprint("Contract training data: " + str(contract_train_data))
il_nn = 2
hl_nn = []
ol_nn = 2
# Train a neural network with contract data
nn = NeuralNetwork(il_nn, hl_nn, ol_nn)
contract_train_data = nn.binary_2_one_hot(contract_train_data)
nn.load_train_data(contract_train_data)
nn.init_network()
nn.train()
trained_weights = nn.weights
trained_biases = nn.bias
dbg.dprint("Trained weights: " + str(trained_weights))
dbg.dprint("Trained biases: " + str(trained_biases))
packed_trained_weights = nn.pack_weights(trained_weights)
dbg.dprint("Packed weights: " + str(packed_trained_weights))
packed_trained_biases = nn.pack_biases(trained_biases)
dbg.dprint("Packed biases: " + str(packed_trained_biases))
int_packed_trained_weights = scale_packed_data(packed_trained_weights,\
w_scale)
dbg.dprint("Packed integer weights: " + str(int_packed_trained_weights))
int_packed_trained_biases = scale_packed_data(packed_trained_biases,\
b_scale)
dbg.dprint("Packed integer biases: " + str(int_packed_trained_biases))
dbg.dprint("Solver address: " + str(solver_account))
# Submit the solution to the contract
submit_tx = cyphai.transact().submit_model(solver_account, il_nn, ol_nn, hl_nn,\
int_packed_trained_weights, int_packed_trained_biases)
chain.wait.for_receipt(submit_tx)
# Get submission index ID
submission_id = cyphai.call().get_submission_id(solver_account, il_nn,\
ol_nn, hl_nn, int_packed_trained_weights, int_packed_trained_biases)
dbg.dprint("Submission ID: " + str(submission_id))
# Wait until the submission period ends
chain.wait.for_block(init3_block_number + submission_t)
# Reveal the testing dataset after the submission period ends
reveal_tx = []
for i in range(len(testing_partition)):
start = i*scd.dps*scd.partition_size
end = start + scd.dps*scd.partition_size
dbg.dprint("(" + str(testing_partition[i]) + ") Test data,nonce: " + str(test_data[start:end]) + "," + str(scd.test_nonce[i]))
reveal_tx.append(cyphai.transact().reveal_test_data(test_data[start:end], scd.test_nonce[i]))
chain.wait.for_receipt(reveal_tx[i])
# Wait until the test reveal period ends
chain.wait.for_block(init3_block_number + submission_t + test_reveal_t)
# Evaluate the submitted solution
eval_tx = cyphai.transact().evaluate_model(submission_id)
# Wait until the evaluation period ends
chain.wait.for_block(init3_block_number + submission_t + test_reveal_t + evaluation_t)
bal2 = web3.eth.getBalance(offer_account)
# Finalize the contract
final_tx = cyphai.transact().finalize_contract()
contract_finalized = cyphai.call().contract_terminated()
dbg.dprint("Contract finalized: " + str(contract_finalized))
assert contract_finalized == True
# Get best submission accuracy & ID
best_submission_accuracy = cyphai.call().best_submission_accuracy()
best_submission_index = cyphai.call().best_submission_index()
dbg.dprint("Best submission ID: " + str(best_submission_index))
dbg.dprint("Best submission accuracy: " + str(best_submission_accuracy))
l_nn = [il_nn] + hl_nn + [ol_nn]
input_layer = train_data[:2]
hidden_layers = [0] * sum(hl_nn)
output_layer = [0] * ol_nn
weights = int_packed_trained_weights
biases = int_packed_trained_biases
# Test forward
fwd_pass2 = cyphai.call().forward_pass2(l_nn, input_layer, hidden_layers, output_layer, weights, biases)
dbg.dprint("Test input: " + str(train_data[:2]))
dbg.dprint("Expected output: " + str(train_data[2]))
dbg.dprint("local nn prediction: " + str(nn.predict([train_data[:2]])))
dbg.dprint("forward_pass2: " + str(fwd_pass2))
bal = web3.eth.getBalance(solver_account)
# Verify that the solver account didn't receive the reward amount
assert bal == 1000000000000000000000000
bal = web3.eth.getBalance(offer_account)
# Verify the offer account got refunded the reward amount
assert bal == 999999999999999999978960