From bd680f157fb41177b1f2c3325879d5850505357b Mon Sep 17 00:00:00 2001 From: dangqingqing Date: Thu, 26 Oct 2017 19:13:24 +0800 Subject: [PATCH] fix compiling warning. --- paddle/operators/lstm_op.h | 4 +- paddle/operators/math/sequence2batch.h | 7 +-- .../paddle/v2/framework/tests/test_lstm_op.py | 46 +++++++------------ 3 files changed, 23 insertions(+), 34 deletions(-) diff --git a/paddle/operators/lstm_op.h b/paddle/operators/lstm_op.h index f910e3bc340..d147b84aefe 100644 --- a/paddle/operators/lstm_op.h +++ b/paddle/operators/lstm_op.h @@ -155,7 +155,7 @@ class LSTMGradKernel : public framework::OpKernel { auto* batch_cell_pre_act = ctx.Input("BatchCellPreAct"); auto* hidden_g = ctx.Input(framework::GradVarName("Hidden")); - auto* cell_g = ctx.Input(framework::GradVarName("Cell")); + // auto* cell_g = ctx.Input(framework::GradVarName("Cell")); auto* in_g = ctx.Output(framework::GradVarName("Input")); auto* weight_g = ctx.Output(framework::GradVarName("Weight")); @@ -219,8 +219,8 @@ class LSTMGradKernel : public framework::OpKernel { LoDTensor batch_cell_g; batch_cell_g.mutable_data(out_dims, ctx.GetPlace()); batch_cell_g.set_lod(batch_gate->lod()); - to_batch(device_ctx, *cell_g, batch_cell_g, false); // TODO(qingqing) support the case output cell has gradient. + // to_batch(device_ctx, *cell_g, batch_cell_g, false); zero(device_ctx, &batch_cell_g, static_cast(0.0)); LoDTensor batch_gate_g; diff --git a/paddle/operators/math/sequence2batch.h b/paddle/operators/math/sequence2batch.h index b833a326c89..b1ba35a6d4a 100644 --- a/paddle/operators/math/sequence2batch.h +++ b/paddle/operators/math/sequence2batch.h @@ -58,7 +58,8 @@ class LoDTensor2BatchFunctor { if (!is_cal_batch_lod) { auto lods = batch.lod(); PADDLE_ENFORCE_EQ(lods.size(), 2UL); - PADDLE_ENFORCE_EQ(lods[1].size(), lod_tensor.dims()[0]); + PADDLE_ENFORCE_EQ(lods[1].size(), + static_cast(lod_tensor.dims()[0])); CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, lods[1].data(), batch, true); return; @@ -111,10 +112,10 @@ class LoDTensor2BatchFunctor { size_t* batch_starts = batch_lods[0].data(); size_t* seq2batch_idx = batch_lods[1].data(); batch_starts[0] = 0; - for (size_t n = 0; n < num_batch; n++) { + for (int n = 0; n < num_batch; n++) { auto batch_id = static_cast(batch_starts[n]); for (size_t i = 0; i < seq_info.size(); ++i) { - size_t seq_len = seq_info[i].length; + int seq_len = seq_info[i].length; int start = seq_info[i].start; if (n < seq_len) { seq2batch_idx[batch_id] = diff --git a/python/paddle/v2/framework/tests/test_lstm_op.py b/python/paddle/v2/framework/tests/test_lstm_op.py index e10972bb3af..7f428cd617c 100644 --- a/python/paddle/v2/framework/tests/test_lstm_op.py +++ b/python/paddle/v2/framework/tests/test_lstm_op.py @@ -52,7 +52,7 @@ def lstm( g = np.dot(h_pre, w_h) # 1 x 4D g = g + x g = np.reshape(g, (1, g.size)) - c_tmp, g_i, g_f, g_o = np.split(g, 4, axis=1) + c, g_i, g_f, g_o = np.split(g, 4, axis=1) if w_c is None: g_i = act_gate(g_i) # 1 x D g_f = act_gate(g_f) # 1 x D @@ -60,7 +60,7 @@ def lstm( w_ic, w_fc, w_oc = np.split(w_c, 3, axis=1) g_i = act_gate(g_i + w_ic * c_pre) # 1 x D g_f = act_gate(g_f + w_fc * c_pre) # 1 x D - c = g_f * c_pre + g_i * act_cand(c_tmp) # 1 x D + c = g_f * c_pre + g_i * act_cand(c) # 1 x D if w_c is None: g_o = act_gate(g_o) # 1 x D @@ -68,8 +68,7 @@ def lstm( _, _, w_oc = np.split(w_c, 3, axis=1) g_o = act_gate(g_o + w_oc * c) # 1 x D h = g_o * act_cell(c) - bg = np.concatenate((act_cand(c_tmp), g_i, g_f, g_o), axis=1) - return h, c, bg + return h, c def _reverse(x, lod): y = np.zeros_like(x) @@ -82,7 +81,6 @@ def lstm( batch_size = len(offset) - 1 hidden = [] cell = [] - gate = [] input = _reverse(input, offset) if is_reverse else input if w_b is not None: input = input + np.tile(w_b, (offset[-1], 1)) @@ -94,30 +92,26 @@ def lstm( c_pre = c0[i] # 1 x D for j in range(seq_len): # compute one step - h_pre, c_pre, g_pre = _step(x[j], w_h, w_c, h_pre, c_pre, act_gate, - act_cell, act_cand) + h_pre, c_pre = _step(x[j], w_h, w_c, h_pre, c_pre, act_gate, + act_cell, act_cand) hidden.append(h_pre.flatten()) cell.append(c_pre.flatten()) - gate.append(g_pre.flatten()) hidden = np.array(hidden).astype('float64') cell = np.array(cell).astype('float64') - gate = np.array(gate).astype('float64') hidden = _reverse(hidden, offset) if is_reverse else hidden cell = _reverse(cell, offset) if is_reverse else cell - assert gate.shape == input.shape assert hidden.shape == (input.shape[0], input.shape[1] / 4) assert cell.shape == (input.shape[0], input.shape[1] / 4) - return hidden, cell, gate + return hidden, cell class TestLstmOp(OpTest): def set_argument(self): - self.lod = [[0, 2, 6, 9]] + self.lod = [[0, 2, 6]] self.D = 16 - self.sort_idx = [2, 6, 0, 3, 7, 1, 4, 8, 5] self.act_gate = 'sigmoid' self.act_cell = 'tanh' @@ -141,22 +135,18 @@ class TestLstmOp(OpTest): w_b = b[:, 0:4 * self.D] w_c = b[:, 4 * self.D:] - h, c, g = lstm(x, self.lod, h0, c0, w, w_b, w_c, self.is_reverse, - ACTVATION[self.act_gate], ACTVATION[self.act_cell], - ACTVATION[self.act_cand]) - - g_sort = np.zeros_like(x) - for i, j in enumerate(self.sort_idx): - g_sort[i, :] = g[j, :] + h, c = lstm(x, self.lod, h0, c0, w, w_b, w_c, self.is_reverse, + ACTVATION[self.act_gate], ACTVATION[self.act_cell], + ACTVATION[self.act_cand]) self.inputs = {'Input': (x, self.lod), 'Weight': w, 'Bias': b} - self.inputs['H0'] = h0 - self.inputs['C0'] = c0 + if self.has_initial_state: + self.inputs['H0'] = h0 + self.inputs['C0'] = c0 self.outputs = { 'Hidden': (h, self.lod), 'Cell': (c, self.lod), - 'BatchGate': g_sort, } self.attrs = { 'usePeepholes': True, @@ -179,9 +169,8 @@ class TestLstmOp(OpTest): class TestLstmOpHasNoInitial(TestLstmOp): def set_argument(self): - self.lod = [[0, 2, 6, 9]] - self.D = 64 - self.sort_idx = [2, 6, 0, 3, 7, 1, 4, 8, 5] + self.lod = [[0, 2, 6]] + self.D = 16 self.act_gate = 'sigmoid' self.act_cell = 'tanh' @@ -193,9 +182,8 @@ class TestLstmOpHasNoInitial(TestLstmOp): class TestLstmOpRerverse(TestLstmOp): def set_argument(self): - self.lod = [[0, 2, 6, 9]] - self.D = 64 - self.sort_idx = [2, 6, 0, 3, 7, 1, 4, 8, 5] + self.lod = [[0, 2, 6]] + self.D = 16 self.act_gate = 'sigmoid' self.act_cell = 'tanh' -- GitLab