zeros = zeros.astype(X_mb.dtype)
mask_zeros = np.zeros((new_len, X_mb_mask.shape[1]))
mask_zeros = mask_zeros.astype(X_mb_mask.dtype)
X_mb = np.concatenate((X_mb, zeros), axis=0)
X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0)
assert len(X_mb[start:stop]) == cut_len
assert len(X_mb_mask[start:stop]) == cut_len
rval = function(X_mb[start:stop], X_mb_mask[start:stop], c_mb,
c_mb_mask, prev_h1, prev_h2, prev_h3, prev_kappa,
prev_w)
current_cost = rval[0]
prev_h1, prev_h2, prev_h3 = rval[1:4]
prev_h1 = prev_h1[-1]
prev_h2 = prev_h2[-1]
prev_h3 = prev_h3[-1]
prev_kappa = rval[4][-1]
prev_w = rval[5][-1]
partial_costs.append(current_cost)
return partial_costs
run_loop(_loop,
train_function,
train_itr,
cost_function,
valid_itr,
n_epochs=n_epochs,
checkpoint_dict=checkpoint_dict,
checkpoint_every_n=checkpoint_every_n,
skip_minimums=True)
partial_costs = []
for n in range(n_cuts):
start = n * cut_len
stop = (n + 1) * cut_len
if len(X_mb[start:stop]) < cut_len:
new_len = cut_len - len(X_mb) % cut_len
zeros = np.zeros((new_len, X_mb.shape[1],
X_mb.shape[2]))
zeros = zeros.astype(X_mb.dtype)
mask_zeros = np.zeros((new_len, X_mb_mask.shape[1]))
mask_zeros = mask_zeros.astype(X_mb_mask.dtype)
X_mb = np.concatenate((X_mb, zeros), axis=0)
X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0)
assert len(X_mb[start:stop]) == cut_len
assert len(X_mb_mask[start:stop]) == cut_len
rval = function(X_mb[start:stop],
X_mb_mask[start:stop],
c_mb, c_mb_mask,
prev_h1, prev_h2, prev_kappa, prev_w)
current_cost = rval[0]
prev_h1, prev_h2 = rval[1:3]
prev_h1 = prev_h1[-1]
prev_h2 = prev_h2[-1]
prev_kappa = rval[3][-1]
prev_w = rval[4][-1]
partial_costs.append(current_cost)
return partial_costs
run_loop(_loop, train_function, train_itr, cost_function, valid_itr,
n_epochs=n_epochs, checkpoint_dict=checkpoint_dict)
start = n * cut_len
stop = (n + 1) * cut_len
if len(X_mb[start:stop]) < cut_len:
new_len = cut_len - len(X_mb) % cut_len
zeros = np.zeros((new_len, X_mb.shape[1],
X_mb.shape[2]))
zeros = zeros.astype(X_mb.dtype)
mask_zeros = np.zeros((new_len, X_mb_mask.shape[1]))
mask_zeros = mask_zeros.astype(X_mb_mask.dtype)
X_mb = np.concatenate((X_mb, zeros), axis=0)
X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0)
assert len(X_mb[start:stop]) == cut_len
assert len(X_mb_mask[start:stop]) == cut_len
rval = function(X_mb[start:stop],
X_mb_mask[start:stop],
c_mb, c_mb_mask,
prev_h1, prev_h2, prev_h3, prev_kappa, prev_w)
current_cost = rval[0]
prev_h1, prev_h2, prev_h3 = rval[1:4]
prev_h1 = prev_h1[-1]
prev_h2 = prev_h2[-1]
prev_h3 = prev_h3[-1]
prev_kappa = rval[4][-1]
prev_w = rval[5][-1]
partial_costs.append(current_cost)
return partial_costs
run_loop(_loop, train_function, train_itr, cost_function, valid_itr,
n_epochs=n_epochs, checkpoint_dict=checkpoint_dict,
checkpoint_every_n=checkpoint_every_n, skip_minimums=True)
stop = (n + 1) * cut_len
if len(X_mb[start:stop]) < cut_len:
new_len = cut_len - len(X_mb) % cut_len
zeros = np.zeros((new_len, X_mb.shape[1], X_mb.shape[2]))
zeros = zeros.astype(X_mb.dtype)
mask_zeros = np.zeros((new_len, X_mb_mask.shape[1]))
mask_zeros = mask_zeros.astype(X_mb_mask.dtype)
X_mb = np.concatenate((X_mb, zeros), axis=0)
X_mb_mask = np.concatenate((X_mb_mask, mask_zeros), axis=0)
assert len(X_mb[start:stop]) == cut_len
assert len(X_mb_mask[start:stop]) == cut_len
rval = function(X_mb[start:stop], X_mb_mask[start:stop], c_mb,
c_mb_mask, prev_h1, prev_h2, prev_kappa, prev_w)
current_cost = rval[0]
prev_h1, prev_h2 = rval[1:3]
prev_h1 = prev_h1[-1]
prev_h2 = prev_h2[-1]
prev_kappa = rval[3][-1]
prev_w = rval[4][-1]
partial_costs.append(current_cost)
return partial_costs
run_loop(_loop,
train_function,
train_itr,
cost_function,
valid_itr,
n_epochs=n_epochs,
checkpoint_dict=checkpoint_dict)
train_loss = train_loss.mean()
valid_prediction = get_output(l_out, deterministic=True)[:, :, :width, :height]
valid_loss = squared_error(valid_prediction, target_var)
valid_loss = valid_loss.mean()
params = get_all_params(l_out, trainable=True)
# adam is the optimizer that is updating everything
updates = adam(train_loss, params, learning_rate=1E-4)
train_function = theano.function([input_var, target_var], train_loss,
updates=updates)
valid_function = theano.function([input_var, target_var], valid_loss)
predict_function = theano.function([input_var], prediction)
checkpoint_dict = {}
checkpoint_dict["train_function"] = train_function
checkpoint_dict["valid_function"] = valid_function
checkpoint_dict["predict_function"] = predict_function
def _loop(function, itr):
X_train, y_train = itr.next()
ret = function(X_train, y_train)
return [ret]
run_loop(_loop, train_function, train_itr, valid_function, valid_itr,
n_epochs=n_epochs, checkpoint_dict=checkpoint_dict,
checkpoint_every_n=100)
params = get_all_params(l_out, trainable=True)
# adam is the optimizer that is updating everything
updates = adam(train_loss, params, learning_rate=1E-4)
train_function = theano.function([input_var, target_var],
train_loss,
updates=updates)
valid_function = theano.function([input_var, target_var], valid_loss)
predict_function = theano.function([input_var], prediction)
checkpoint_dict = {}
checkpoint_dict["train_function"] = train_function
checkpoint_dict["valid_function"] = valid_function
checkpoint_dict["predict_function"] = predict_function
def _loop(function, itr):
X_train, y_train = itr.next()
ret = function(X_train, y_train)
return [ret]
run_loop(_loop,
train_function,
train_itr,
valid_function,
valid_itr,
n_epochs=n_epochs,
checkpoint_dict=checkpoint_dict,
checkpoint_every_n=100)