fix bugs

demo1_config.yaml → config_demo1.yaml

@@ -12,9 +12,10 @@ do_lower_case: True
 max_seq_len: 512
 
 batch_size: 5
-num_epochs: 2
+num_epochs: 3
 optimizer: "adam"
 learning_rate: 3e-5
 warmup_proportion: 0.1
 weight_decay: 0.1
 
+print_every_n_steps: 1

demo1.py

 import paddlepalm as palm
 
 if __name__ == '__main__':
-    controller = palm.Controller('demo1_config.yaml', task_dir='demo1_tasks')
+    controller = palm.Controller('config_demo1.yaml', task_dir='demo1_tasks')
     controller.load_pretrain('pretrain_model/ernie/params')
     controller.train()
 
-    controller = palm.Controller(config='demo1_config.yaml', task_dir='demo1_tasks', for_train=False)
+    controller = palm.Controller(config='config_demo1.yaml', task_dir='demo1_tasks', for_train=False)
     controller.pred('mrqa', inference_model_dir='output_model/firstrun/infer_model')
 

demo2.py

@@ -3,8 +3,8 @@ import paddlepalm as palm
 if __name__ == '__main__':
     controller = palm.Controller('config_demo2.yaml', task_dir='demo2_tasks')
     controller.load_pretrain('pretrain_model/ernie/params')
-    controller.train()
+    # controller.train()
 
-    controller = palm.Controller(config='config_demo2.yaml', task_dir='demo2_tasks', for_train=False)
-    controller.pred('mrqa', inference_model_dir='output_model/secondrun/infer_model')
+    # controller = palm.Controller(config='config_demo2.yaml', task_dir='demo2_tasks', for_train=False)
+    # controller.pred('mrqa', inference_model_dir='output_model/secondrun/infer_model')
 

paddlepalm/backbone/ernie.py

@@ -76,7 +76,7 @@ class Model(backbone):
                 "sentence_embedding": [[-1, self._emb_size], 'float32'],
                 "sentence_pair_embedding": [[-1, self._emb_size], 'float32']}
 
-    def build(self, inputs):
+    def build(self, inputs, scope_name=""):
 
         src_ids = inputs['token_ids']
         pos_ids = inputs['position_ids']
@@ -90,25 +90,25 @@ class Model(backbone):
             size=[self._voc_size, self._emb_size],
             dtype=self._emb_dtype,
             param_attr=fluid.ParamAttr(
-                name=self._word_emb_name, initializer=self._param_initializer),
+                name=scope_name+self._word_emb_name, initializer=self._param_initializer),
             is_sparse=False)
 
         # fluid.global_scope().find_var('backbone-word_embedding').get_tensor()
-        embedding_table = fluid.default_main_program().global_block().var(self._word_emb_name)
+        embedding_table = fluid.default_main_program().global_block().var(scope_name+self._word_emb_name)
         
         position_emb_out = fluid.layers.embedding(
             input=pos_ids,
             size=[self._max_position_seq_len, self._emb_size],
             dtype=self._emb_dtype,
             param_attr=fluid.ParamAttr(
-                name=self._pos_emb_name, initializer=self._param_initializer))
+                name=scope_name+self._pos_emb_name, initializer=self._param_initializer))
 
         sent_emb_out = fluid.layers.embedding(
             sent_ids,
             size=[self._sent_types, self._emb_size],
             dtype=self._emb_dtype,
             param_attr=fluid.ParamAttr(
-                name=self._sent_emb_name, initializer=self._param_initializer))
+                name=scope_name+self._sent_emb_name, initializer=self._param_initializer))
 
         emb_out = emb_out + position_emb_out
         emb_out = emb_out + sent_emb_out
@@ -118,13 +118,13 @@ class Model(backbone):
             size=[self._task_types, self._emb_size],
             dtype=self._emb_dtype,
             param_attr=fluid.ParamAttr(
-                name=self._task_emb_name,
+                name=scope_name+self._task_emb_name,
                 initializer=self._param_initializer))
 
         emb_out = emb_out + task_emb_out
 
         emb_out = pre_process_layer(
-            emb_out, 'nd', self._prepostprocess_dropout, name='pre_encoder')
+            emb_out, 'nd', self._prepostprocess_dropout, name=scope_name+'pre_encoder')
 
         self_attn_mask = fluid.layers.matmul(
             x=input_mask, y=input_mask, transpose_y=True)
@@ -151,7 +151,7 @@ class Model(backbone):
             preprocess_cmd="",
             postprocess_cmd="dan",
             param_initializer=self._param_initializer,
-            name='encoder')
+            name=scope_name+'encoder')
 
         
         next_sent_feat = fluid.layers.slice(
@@ -162,8 +162,8 @@ class Model(backbone):
             size=self._emb_size,
             act="tanh",
             param_attr=fluid.ParamAttr(
-                name="pooled_fc.w_0", initializer=self._param_initializer),
-            bias_attr="pooled_fc.b_0")
+                name=scope_name+"pooled_fc.w_0", initializer=self._param_initializer),
+            bias_attr=scope_name+"pooled_fc.b_0")
 
         return {'embedding_table': embedding_table,
                 'word_embedding': emb_out,

paddlepalm/mtl_controller.py

@@ -430,23 +430,25 @@ class Controller(object):
 
         # build backbone and task layers
         # 不指定scope名字会挂，框架有坑
+        train_prog = fluid.default_main_program()
+        train_init_prog = fluid.default_startup_program()
+        # 别用unique_name.guard了，没用的，无法作用到param_attr里的name上
         with fluid.unique_name.guard("backbone-"):
-            bb_output_vars = train_backbone.build(net_inputs)
-        # bb_output_vars = train_backbone.build(net_inputs)
+            bb_output_vars = train_backbone.build(net_inputs, scope_name='__paddlepalm_')
         assert sorted(bb_output_vars.keys()) == sorted(train_backbone.outputs_attr.keys())
+        #for var in train_init_prog.blocks[0].vars:
+        #    print(var)
         
         # 会挂
         # 这里是否有必要新建一个program？是的，被坑死了
         pred_prog = fluid.Program()
         pred_init_prog = fluid.Program()
 
-        train_prog = fluid.default_main_program()
-        train_init_prog = fluid.default_startup_program()
-
         with fluid.program_guard(main_program = pred_prog, startup_program = pred_init_prog):
             pred_net_inputs = create_net_inputs(pred_input_attrs)
-            with fluid.unique_name.guard("backbone-"):
-                pred_bb_output_vars = pred_backbone.build(pred_net_inputs)
+            # 别用unique_name.guard了，没用的，无法作用到param_attr里的name上
+            # with fluid.unique_name.guard("backbone-"):
+            pred_bb_output_vars = pred_backbone.build(pred_net_inputs, scope_name='__paddlepalm_')
 
         fluid.framework.switch_main_program(train_prog)
         fluid.framework.switch_startup_program(train_init_prog)
@@ -503,13 +505,13 @@ class Controller(object):
 
         num_examples = main_reader.num_examples
         for inst in instances:
-            max_train_steps = int(main_conf['num_epochs']* inst.mix_ratio * num_examples) // main_conf['batch_size']  // dev_count
+            max_train_steps = int(main_conf['num_epochs']* inst.mix_ratio * (num_examples // main_conf['batch_size']  // dev_count))
             if inst.is_target:
                 print('{}: expected train steps {}.'.format(inst.name, max_train_steps))
             inst.steps_pur_epoch = inst.reader['train'].num_examples // main_conf['batch_size']  // dev_count
             inst.expected_train_steps = max_train_steps
 
-        global_max_train_steps = int(main_conf['num_epochs'] * num_examples * sum(mrs)) // main_conf['batch_size']  // dev_count
+        global_max_train_steps = int(main_conf['num_epochs'] * sum(mrs) * (num_examples // main_conf['batch_size']  // dev_count))
         print('Estimated overall train steps {}.'.format(global_max_train_steps))
 
         if 'warmup_proportion' in main_conf and main_conf['warmup_proportion'] > 0:

paddlepalm/optimizer/adam.py

@@ -90,11 +90,6 @@ def optimize(loss, config, max_train_steps=None, warmup_steps=0, train_program=N
 
     _, param_grads = optimizer.minimize(loss)
 
-    for block in fluid.default_main_program().blocks:
-        for var_name in block.vars:
-            if var_name.startswith("embedding"):
-                print(block.vars[var_name])
-            
 
     if config.get('weight_decay', 0) > 0:
         for param, grad in param_grads:

paddlepalm/utils/saver.py

@@ -19,6 +19,8 @@ import os
 import six
 import ast
 import copy
+import tarfile
+import shutil
 
 import numpy as np
 import paddle.fluid as fluid
@@ -48,18 +50,30 @@ def init_pretraining_params(exe,
     assert os.path.exists(pretraining_params_path
                           ), "[%s] cann't be found." % pretraining_params_path
 
+
+    assert os.path.exists(os.path.join(pretraining_params_path, '__palmmodel__')), "__palmmodel__ not found."
+    print("Loading pretraining parameters from {}...".format(
+        pretraining_params_path))
+
+    with tarfile.open(os.path.join(pretraining_params_path, '__palmmodel__'), 'r:') as f:
+        f.extractall(os.path.join(pretraining_params_path, '.temp'))
+
+    pretraining_params_path = os.path.join(pretraining_params_path, '.temp')
+
     def existed_params(var):
         if not isinstance(var, fluid.framework.Parameter):
             return False
+        if not os.path.exists(os.path.join(pretraining_params_path, var.name)):
+            print('Warning: {} not found in {}.'.format(var.name, pretraining_params_path))
         return os.path.exists(os.path.join(pretraining_params_path, var.name))
 
-    print("Load pretraining parameters from {}...\n".format(
-        pretraining_params_path))
-
     fluid.io.load_vars(
         exe,
         pretraining_params_path,
         main_program=main_program,
         predicate=existed_params)
 
+    shutil.rmtree(pretraining_params_path)
+    print('')
+
 

run_demo1.sh

-export CUDA_VISIBLE_DEVICES=0
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 
 python demo1.py
 

run_demo2.sh

-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+export CUDA_VISIBLE_DEVICES=0
 
 python demo2.py
 

script/convert_params.sh

@@ -5,13 +5,32 @@ if [[ $# != 1 ]]; then
     exit 1
 fi
 
+if [[ -f $1/__palminfo__ ]]; then
+    echo "already converted."
+    exit 0
+fi
+
 echo "converting..."
-cd $1
+if [[ -d $1/params ]]; then
+    cd $1/params
+else
+    cd $1
+fi
+
 mkdir .palm.backup
 
 for file in $(ls *)
-    do cp $file "backbone-"$file; mv $file .palm.backup
+    do cp $file .palm.backup; mv $file "__paddlepalm_"$file
 done
+tar -cf __rawmodel__ .palm.backup/*
+rm .palm.backup/*
+mv __rawmodel__ .palm.backup
+# find . ! -name '__rawmodel__' -exec rm {} +
+tar -cf __palmmodel__ __paddlepalm_*
+touch __palminfo__
+ls __paddlepalm_* > __palminfo__
+rm __paddlepalm_*
+
 cd - >/dev/null
 
 echo "done!"

script/recover_params.sh

@@ -5,7 +5,29 @@ if [[ $# != 1 ]]; then
     exit 1
 fi
 
-rm $1/backbone-*
-mv $1/.palm.backup/* $1
-rm -rf $1/.palm.backup
+if [[ ! -d $1 ]]; then
+    echo "$1 not found."
+    exit 1
+fi
+
+if [[ ! -f $1/__palmmodel__ ]]; then
+    echo "paddlepalm model not found."
+    exit 1
+fi
+
+echo "recovering..."
+if [[ -d $1/params ]]; then
+    cd $1/params
+else
+    cd $1
+fi
+rm __palm*
+mv .palm.backup/__rawmodel__ .
+rm -rf .palm.backup
+tar -xf __rawmodel__
+mv .palm.backup/* .
+rm __rawmodel__
+
+rm -rf .palm.backup
+cd - >/dev/null