Update bert based DARTS demo

paddleslim/common/meter.py

@@ -16,8 +16,9 @@ __all__ = ['AvgrageMeter']
 
 
 class AvgrageMeter(object):
-    def __init__(self):
+    def __init__(self, format="{}"):
         self.reset()
+        self._format = format
 
     def reset(self):
         self.avg = 0
@@ -28,3 +29,6 @@ class AvgrageMeter(object):
         self.sum += val * n
         self.cnt += n
         self.avg = self.sum / self.cnt
+
+    def __repr__(self):
+        return self._format.format(self.avg)

paddleslim/nas/darts/architect.py

@@ -56,7 +56,10 @@ class Architect(object):
         else:
             loss = self._backward_step(valid_data)
             self.optimizer.minimize(loss)
+
+#        print("alphas gradient: {}".format(self.model.arch_parameters()[0].gradient()))
         self.optimizer.clear_gradients()
+        return self.model.arch_parameters()[0].gradient()
 
     def _backward_step(self, valid_data):
         loss = self.model.loss(valid_data)

paddleslim/nas/darts/search_space/conv_bert/cls.py

@@ -50,7 +50,9 @@ class AdaBERTClassifier(Layer):
                  beta=4,
                  conv_type="conv_bn",
                  search_layer=True,
-                 teacher_model=None):
+                 teacher_model=None,
+                 alphas=None,
+                 k=None):
         super(AdaBERTClassifier, self).__init__()
         self._n_layer = n_layer
         self._num_labels = num_labels
@@ -60,6 +62,8 @@ class AdaBERTClassifier(Layer):
         self._beta = beta
         self._conv_type = conv_type
         self._search_layer = search_layer
+        self._alphas = alphas
+        self._k = k
         print(
             "----------------------load teacher model and test----------------------------------------"
         )
@@ -86,20 +90,36 @@ class AdaBERTClassifier(Layer):
                 bias_attr=fluid.ParamAttr(
                     name="s_cls_out_%d_b" % i,
                     initializer=fluid.initializer.Constant(0.)))
-            fc = self.add_sublayer("cls_fc_%d" % i, fc)
+            fc = self.add_sublayer("s_cls_fc_%d" % i, fc)
             self.cls_fc.append(fc)
 
-    def forward(self, data_ids):
+    def forward(self, data_ids, alphas=None, k=None):
         src_ids = data_ids[0]
         position_ids = data_ids[1]
         sentence_ids = data_ids[2]
-        return self.student(src_ids, position_ids, sentence_ids)
+        return self.student(
+            src_ids,
+            position_ids,
+            sentence_ids,
+            alphas=self._alphas,
+            k=self._k)
 
     def arch_parameters(self):
         return self.student.arch_parameters()
 
+    def model_parameters(self):
+
+        model_parameters = [
+            p for p in self.student.parameters()
+            if p.name not in [a.name for a in self.arch_parameters()]
+        ]
+        return model_parameters
+
     def genotype(self):
-        return self.arch_parameters()
+        alphas = self.arch_parameters()[0].numpy()
+        alphas = [np.argmax(edge) for edge in alphas]
+        k = np.argmax(self.arch_parameters()[1].numpy())
+        return "layers: {}; edges: {} ".format(k, alphas)
 
     def new(self):
         model_new = AdaBERTClassifier(
@@ -108,8 +128,7 @@ class AdaBERTClassifier(Layer):
         )
         return model_new
 
-    def loss(self, data_ids):
-        T = 1.0
+    def valid(self, data_ids):
         src_ids = data_ids[0]
         position_ids = data_ids[1]
         sentence_ids = data_ids[2]
@@ -117,16 +136,57 @@ class AdaBERTClassifier(Layer):
         labels = data_ids[4]
         flops = []
         model_size = []
+        alphas = self.arch_parameters()[0].numpy(
+        ) if self._alphas is None else self._alphas
+        k = self.arch_parameters()[1].numpy() if self._k is None else self._k
+
+        print(alphas.shape)
+        print(k.shape)
+
         enc_outputs, next_sent_feats, k_i = self.student(
             src_ids,
             position_ids,
             sentence_ids,
             flops=flops,
-            model_size=model_size)
+            model_size=model_size,
+            alphas=alphas,
+            k=k)
+
+        logits = self.cls_fc[-1](next_sent_feats[-1])
+        probs = fluid.layers.softmax(logits)
+        accuracy = fluid.layers.accuracy(input=probs, label=labels)
+
+        model_size = np.sum(model_size)
+        flops = np.sum(flops)
+        ret = {
+            "accuracy": accuracy.numpy(),
+            "model_size(MB)": model_size / 1e6,
+            "FLOPs(M)": flops / 1e6
+        }
+        return ret
 
+    def loss(self, data_ids):
+        T = 1.0
+        src_ids = data_ids[0]
+        position_ids = data_ids[1]
+        sentence_ids = data_ids[2]
+        input_mask = data_ids[3]
+        labels = data_ids[4]
+        flops = []
+        model_size = []
         self.teacher.eval()
         total_loss, t_logits, t_losses, accuracys, num_seqs = self.teacher(
             data_ids)
+        self.teacher.train()
+
+        enc_outputs, next_sent_feats, k_i = self.student(
+            src_ids,
+            position_ids,
+            sentence_ids,
+            flops=flops,
+            model_size=model_size,
+            alphas=self._alphas,
+            k=self._k)
 
         # define kd loss
         kd_losses = []
@@ -140,21 +200,16 @@ class AdaBERTClassifier(Layer):
         kd_weights = np.exp(kd_weights - np.max(kd_weights))
 
         kd_weights = kd_weights / kd_weights.sum(axis=0)
-
+        s_probs = None
         for i in range(len(next_sent_feats)):
             j = int(np.ceil(i * (float(len(t_logits)) / len(next_sent_feats))))
             t_logit = t_logits[j]
             s_sent_feat = next_sent_feats[i]
             fc = self.cls_fc[i]
-            s_sent_feat = fluid.layers.dropout(
-                x=s_sent_feat,
-                dropout_prob=0.1,
-                dropout_implementation="upscale_in_train")
             s_logits = fc(s_sent_feat)
-
+            t_logit.stop_gradient = True
             t_probs = fluid.layers.softmax(t_logit)
             s_probs = fluid.layers.softmax(s_logits)
-            t_probs.stop_gradient = True
             kd_loss = t_probs * fluid.layers.log(s_probs / T)
             kd_loss = fluid.layers.reduce_sum(kd_loss, dim=1)
             kd_loss = kd_loss * kd_weights[i]
@@ -167,17 +222,28 @@ class AdaBERTClassifier(Layer):
         ce_loss = fluid.layers.cross_entropy(s_probs, labels)
         ce_loss = fluid.layers.reduce_mean(ce_loss) * k_i
 
+        len_model_size = len(model_size)
         # define e loss
+        if self._alphas is not None:
+            flops = np.sum(flops)
+            model_size = np.sum(model_size)
+        else:
+            flops = fluid.layers.sum(flops)
             model_size = fluid.layers.sum(model_size)
-        #        print("model_size: {}".format(model_size.numpy()/1e6))
         model_size = model_size / self.student.max_model_size()
-        flops = fluid.layers.sum(flops) / self.student.max_flops()
-        e_loss = (len(next_sent_feats) * k_i / self._n_layer) * (
-            flops + model_size)
+        flops = flops / self.student.max_flops()
+        e_loss = (flops + model_size) * (len(next_sent_feats) * k_i /
+                                         self._n_layer)
+        print(
+            "len(next_sent_feats): {}; k_i: {}; flops: {}; model_size: {}; len: {}".
+            format(
+                len(next_sent_feats), k_i,
+                flops.numpy(), model_size.numpy(), len_model_size))
         # define total loss
         loss = (1 - self._gamma
                 ) * ce_loss - self._gamma * kd_loss + self._beta * e_loss
-        #        print("ce_loss: {}; kd_loss: {}; e_loss: {}".format((
-        #            1 - gamma) * ce_loss.numpy(), -gamma * kd_loss.numpy(), beta *
-        #                                                            e_loss.numpy()))
         return loss, ce_loss, kd_loss, e_loss
+
+
+#        loss = ce_loss + self._beta * e_loss 
+#        return loss, ce_loss, ce_loss, e_loss

paddleslim/nas/darts/search_space/conv_bert/model/bert.py

@@ -110,10 +110,9 @@ class BertModelLayer(Layer):
                 position_ids,
                 sentence_ids,
                 flops=[],
-                model_size=[]):
-        """
-        forward
-        """
+                model_size=[],
+                alphas=None,
+                k=None):
         src_emb = self._src_emb(src_ids)
         pos_emb = self._pos_emb(position_ids)
         sent_emb = self._sent_emb(sentence_ids)
@@ -122,9 +121,8 @@ class BertModelLayer(Layer):
         emb_out = emb_out + sent_emb
 
         emb_out = self._emb_fac(emb_out)
-
         enc_outputs, k_i = self._encoder(
-            emb_out, flops=flops, model_size=model_size)
+            emb_out, flops=flops, model_size=model_size, alphas=alphas, k=k)
 
         if not self.return_pooled_out:
             return enc_outputs

paddleslim/nas/darts/search_space/conv_bert/model/cls.py

@@ -57,12 +57,12 @@ class ClsModelLayer(Layer):
             fc = Linear(
                 input_dim=self.config["hidden_size"],
                 output_dim=num_labels,
-                param_attr=fluid.ParamAttr(
+                param_attr=fluid.paramattr(
                     name="cls_out_%d_w" % i,
-                    initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
-                bias_attr=fluid.ParamAttr(
+                    initializer=fluid.initializer.truncatednormal(scale=0.02)),
+                bias_attr=fluid.paramattr(
                     name="cls_out_%d_b" % i,
-                    initializer=fluid.initializer.Constant(0.)))
+                    initializer=fluid.initializer.constant(0.)))
             fc = self.add_sublayer("cls_fc_%d" % i, fc)
             self.cls_fc.append(fc)
 

paddleslim/nas/darts/search_space/conv_bert/model/transformer_encoder.py

@@ -18,6 +18,7 @@ from __future__ import division
 from __future__ import print_function
 
 import numpy as np
+from collections import Iterable
 
 import paddle
 import paddle.fluid as fluid
@@ -25,13 +26,13 @@ from paddle.fluid.dygraph import Embedding, LayerNorm, Linear, Layer, Conv2D, Ba
 from paddle.fluid.initializer import NormalInitializer
 
 GConv_PRIMITIVES = [
-    'std_gconv_3', 'std_gconv_5', 'std_gconv_7', 'dil_gconv_3', 'dil_gconv_5',
-    'dil_gconv_7', 'avg_pool_3', 'max_pool_3', 'none', 'skip_connect'
+    'none', 'std_gconv_3', 'std_gconv_5', 'std_gconv_7', 'dil_gconv_3',
+    'dil_gconv_5', 'dil_gconv_7', 'avg_pool_3', 'max_pool_3', 'skip_connect'
 ]
 
 ConvBN_PRIMITIVES = [
-    'std_conv_bn_3', 'std_conv_bn_5', 'std_conv_bn_7', 'dil_conv_bn_3',
-    'dil_conv_bn_5', 'dil_conv_bn_7', 'avg_pool_3', 'max_pool_3', 'none',
+    'none', 'std_conv_bn_3', 'std_conv_bn_5', 'std_conv_bn_7', 'dil_conv_bn_3',
+    'dil_conv_bn_5', 'dil_conv_bn_7', 'avg_pool_3', 'max_pool_3',
     'skip_connect'
 ]
 
@@ -117,7 +118,11 @@ class MixedOp(fluid.dygraph.Layer):
 
     def forward(self, x, weights, flops=[], model_size=[]):
         for i in range(len(self._ops)):
-            if weights[i].numpy() != 0:
+            if isinstance(weights, Iterable):
+                weights_i = weights[i]
+            else:
+                weights_i = weights[i].numpy()
+            if weights_i != 0:
                 flops.append(FLOPs.values()[i] * weights[i])
                 model_size.append(ModelSize.values()[i] * weights[i])
                 return self._ops[i](x) * weights[i]
@@ -166,6 +171,7 @@ class ConvBNRelu(fluid.dygraph.Layer):
                  use_cudnn=True,
                  name=None):
         super(ConvBNRelu, self).__init__()
+        self._name = name
         conv_std = (2.0 /
                     (filter_size[0] * filter_size[1] * out_c * in_c))**0.5
         conv_param = fluid.ParamAttr(
@@ -187,7 +193,7 @@ class ConvBNRelu(fluid.dygraph.Layer):
     def forward(self, inputs):
         conv = self.conv(inputs)
         bn = self.bn(conv)
-        return bn
+        return conv
 
 
 class GateConv(fluid.dygraph.Layer):
@@ -261,24 +267,30 @@ class Cell(fluid.dygraph.Layer):
         states = [s0, s1]
         offset = 0
         for i in range(self._steps):
-            s = fluid.layers.sums([
-                self._ops[offset + j](h,
+            edges = []
+            for j, h in enumerate(states):
+                edge = self._ops[offset + j](h,
                                              weights[offset + j],
                                              flops=flops,
                                              model_size=model_size)
-                for j, h in enumerate(states)
-            ])
+                edges.append(edge)
+            s = edges[0]
+            for n in range(1, len(edges)):
+                s = s + edges[n]
+#            s = fluid.layers.sums(edges)
             offset += len(states)
             states.append(s)
-        out = fluid.layers.sum(states[-self._steps:])
+
+        states = states[-self._steps:]
+        out = states[0]
+        for n in range(1, len(states)):
+            out = out + states[n]
+
+#        out = fluid.layers.sums(states[-self._steps:])
         return out
 
 
 class EncoderLayer(Layer):
-    """
-    encoder
-    """
-
     def __init__(self,
                  n_layer,
                  hidden_size=768,
@@ -342,14 +354,16 @@ class EncoderLayer(Layer):
             default_initializer=NormalInitializer(
                 loc=0.0, scale=1e-3))
 
-    def forward(self, enc_input, flops=[], model_size=[]):
+    def forward(self, enc_input, flops=[], model_size=[], alphas=None, k=None):
         tmp = fluid.layers.reshape(
             enc_input, [-1, 1, enc_input.shape[1],
                         self._hidden_size])  #(bs, 1, seq_len, hidden_size)
 
         tmp = self.conv0(tmp)  # (bs, hidden_size, seq_len, 1)
 
+        if alphas is None:
             alphas = gumbel_softmax(self.alphas)
+        if k is None:
             k = fluid.layers.reshape(gumbel_softmax(self.k), [-1])
 
         outputs = []
@@ -364,7 +378,11 @@ class EncoderLayer(Layer):
                 enc_output, [-1, enc_output.shape[1],
                              self._hidden_size])  # (bs, seq_len, hidden_size)
             outputs.append(enc_output)
-            if self._search_layer and k[i].numpy() != 0:
+            if isinstance(k, Iterable):
+                k_i = k[i]
+            else:
+                k_i = k[i].numpy()
+            if k_i != 0:
                 outputs[-1] = outputs[-1] * k[i]
                 return outputs, k[i]
         return outputs, 1.0

paddleslim/nas/darts/train_search.py

@@ -18,10 +18,12 @@ from __future__ import print_function
 
 __all__ = ['DARTSearch']
 
+import math
 import logging
 from itertools import izip
 import numpy as np
 import paddle.fluid as fluid
+from paddle.fluid.framework import Variable
 from paddle.fluid.dygraph.base import to_variable
 from ...common import AvgrageMeter, get_logger
 from .architect import Architect
@@ -41,6 +43,7 @@ class DARTSearch(object):
                  model,
                  train_reader,
                  valid_reader,
+                 test_reader=None,
                  learning_rate=0.025,
                  batchsize=64,
                  num_imgs=50000,
@@ -54,6 +57,7 @@ class DARTSearch(object):
         self.model = model
         self.train_reader = train_reader
         self.valid_reader = valid_reader
+        self.test_reader = test_reader
         self.learning_rate = learning_rate
         self.batchsize = batchsize
         self.num_imgs = num_imgs
@@ -82,10 +86,13 @@ class DARTSearch(object):
         step_id = 0
         for train_data, valid_data in izip(train_loader(), valid_loader()):
             if epoch >= self.epochs_no_archopt:
-                architect.step(train_data, valid_data)
+                alphas_grad = architect.step(train_data, valid_data)
 
             loss, ce_loss, kd_loss, e_loss = self.model.loss(train_data)
-
+            if math.isnan(e_loss.numpy()):
+                print("alphas_grad: {}".format(alphas_grad))
+                print("alphas: {}".format(self.model.arch_parameters()[0]
+                                          .numpy()))
             if self.use_data_parallel:
                 loss = self.model.scale_loss(loss)
                 loss.backward()
@@ -93,67 +100,55 @@ class DARTSearch(object):
             else:
                 loss.backward()
 
-            grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(5)
-            optimizer.minimize(loss, grad_clip)
+#            grad_clip = fluid.dygraph_grad_clip.GradClipByGlobalNorm(5)
+#            optimizer.minimize(loss, grad_clip)
+            optimizer.minimize(loss)
+
             self.model.clear_gradients()
 
             batch_size = train_data[0].shape[0]
             objs.update(loss.numpy(), batch_size)
+
+            e_loss = e_loss.numpy() if isinstance(e_loss, Variable) else e_loss
             ce_losses.update(ce_loss.numpy(), batch_size)
             kd_losses.update(kd_loss.numpy(), batch_size)
-            e_losses.update(e_loss.numpy(), batch_size)
+            e_losses.update(e_loss, batch_size)
 
             if step_id % self.log_freq == 0:
-                #logger.info("Train Epoch {}, Step {}, loss {:.6f}; ce: {:.6f}; kd: {:.6f}; e: {:.6f}".format(
-                #    epoch, step_id, objs.avg[0], ce_losses.avg[0], kd_losses.avg[0], e_losses.avg[0]))
                 logger.info(
                     "Train Epoch {}, Step {}, loss {}; ce: {}; kd: {}; e: {}".
                     format(epoch, step_id,
                            loss.numpy(),
-                           ce_loss.numpy(), kd_loss.numpy(), e_loss.numpy()))
+                           ce_loss.numpy(), kd_loss.numpy(), e_loss))
             step_id += 1
         return objs.avg[0]
 
     def valid_one_epoch(self, valid_loader, epoch):
-        objs = AvgrageMeter()
-        top1 = AvgrageMeter()
-        top5 = AvgrageMeter()
         self.model.eval()
-
+        meters = {}
         for step_id, valid_data in enumerate(valid_loader):
-            image = to_variable(image)
-            label = to_variable(label)
-            n = image.shape[0]
-            logits = self.model(image)
-            prec1 = fluid.layers.accuracy(input=logits, label=label, k=1)
-            prec5 = fluid.layers.accuracy(input=logits, label=label, k=5)
-            loss = fluid.layers.reduce_mean(
-                fluid.layers.softmax_with_cross_entropy(logits, label))
-            objs.update(loss.numpy(), n)
-            top1.update(prec1.numpy(), n)
-            top5.update(prec5.numpy(), n)
+            ret = self.model.valid(valid_data)
+            for key, value in ret.items():
+                if key not in meters:
+                    meters[key] = AvgrageMeter()
+                meters[key].update(value, 1)
 
             if step_id % self.log_freq == 0:
-                logger.info(
-                    "Valid Epoch {}, Step {}, loss {:.6f}, acc_1 {:.6f}, acc_5 {:.6f}".
-                    format(epoch, step_id, objs.avg[0], top1.avg[0], top5.avg[
-                        0]))
-        return top1.avg[0]
+                logger.info("Valid Epoch {}, Step {}, {}".format(
+                    epoch, step_id, meters))
 
     def train(self):
         if self.use_data_parallel:
             strategy = fluid.dygraph.parallel.prepare_context()
-        model_parameters = [
-            p for p in self.model.parameters()
-            if p.name not in [a.name for a in self.model.arch_parameters()]
-        ]
-        logger.info("param size = {:.6f}MB".format(
+        model_parameters = self.model.model_parameters()
+        logger.info("parameter size in super net: {:.6f}M".format(
             count_parameters_in_MB(model_parameters)))
         step_per_epoch = int(self.num_imgs * 0.5 / self.batchsize)
         if self.unrolled:
             step_per_epoch *= 2
         learning_rate = fluid.dygraph.CosineDecay(
             self.learning_rate, step_per_epoch, self.num_epochs)
+
         optimizer = fluid.optimizer.MomentumOptimizer(
             learning_rate,
             0.9,
@@ -167,6 +162,9 @@ class DARTSearch(object):
                 self.train_reader)
             self.valid_reader = fluid.contrib.reader.distributed_batch_reader(
                 self.valid_reader)
+            if self.test_reader is not None:
+                self.test_reader = fluid.contrib.reader.distributed_batch_reader(
+                    self.test_reader)
 
         train_loader = fluid.io.DataLoader.from_generator(
             capacity=64,
@@ -182,6 +180,17 @@ class DARTSearch(object):
         train_loader.set_batch_generator(self.train_reader, places=self.place)
         valid_loader.set_batch_generator(self.valid_reader, places=self.place)
 
+        if self.test_reader is not None:
+            test_loader = fluid.io.DataLoader.from_generator(
+                capacity=64,
+                use_double_buffer=True,
+                iterable=True,
+                return_list=True)
+            test_loader.set_batch_generator(
+                self.test_reader, places=self.place)
+        else:
+            test_loader = valid_loader
+
         architect = Architect(self.model, learning_rate,
                               self.arch_learning_rate, self.place,
                               self.unrolled)
@@ -199,8 +208,8 @@ class DARTSearch(object):
             self.train_one_epoch(train_loader, valid_loader, architect,
                                  optimizer, epoch)
 
-            if epoch == self.num_epochs - 1:
-                #                valid_top1 = self.valid_one_epoch(valid_loader, epoch)
-                logger.info("Epoch {}, valid_acc {:.6f}".format(epoch, 1))
-            if save_parameters:
-                fluid.save_dygraph(self.model.state_dict(), "./weights")
+
+#            if epoch == self.num_epochs - 1:
+#                self.valid_one_epoch(test_loader, epoch)
+#            if save_parameters:
+#                fluid.save_dygraph(self.model.state_dict(), "./weights")

paddleslim/teachers/bert/model/transformer_encoder.py

@@ -57,7 +57,7 @@ class PrePostProcessLayer(Layer):
             elif cmd == "d":  # add dropout
                 if dropout_rate:
                     self.functors.append(lambda x: fluid.layers.dropout(
-                        x, dropout_prob=dropout_rate, is_test=False))
+                        x, dropout_prob=dropout_rate, is_test=True))
                     self.exec_order += "d"
 
     def forward(self, x, residual=None):
@@ -111,8 +111,8 @@ class PositionwiseFeedForwardLayer(Layer):
             hidden = fluid.layers.dropout(
                 hidden,
                 dropout_prob=self._dropout_rate,
-                upscale_in_train="upscale_in_train",
-                is_test=False)
+                #                upscale_in_train="upscale_in_train",
+                is_test=True)
         out = self._h2o(hidden)
         return out
 
@@ -218,13 +218,13 @@ class MultiHeadAttentionLayer(Layer):
         #alpha=self._d_model**-0.5)
         if attn_bias is not None:
             product += attn_bias
-        weights = fluid.layers.softmax(product)
+        weights = fluid.layers.softmax(product)  # 48
         if self._dropout_rate:
             weights_droped = fluid.layers.dropout(
                 weights,
                 dropout_prob=self._dropout_rate,
-                dropout_implementation="upscale_in_train",
-                is_test=False)
+                #                dropout_implementation="upscale_in_train",
+                is_test=True)
             out = fluid.layers.matmul(weights_droped, transpose_v)
         else:
             out = fluid.layers.matmul(weights, transpose_v)