update 3.9.3

Hhhhhhand · Jul 25, 2023 · 4f3bdf5 · 4f3bdf5
1 parent 357f8ba
commit 4f3bdf5
Show file tree

Hide file tree

Showing 9 changed files with 311 additions and 399 deletions.
diff --git a/01，kerasmodel_example.ipynb b/01，kerasmodel_example.ipynb
diff --git a/README_en.md b/README_en.md
@@ -125,7 +125,6 @@ KerasModel. Here are some examples.
 |**NLP**|||
 |TextClassification——BERT 🔥🔥| transformers |  [BERT](./examples/BERT——transformers.ipynb) |
 |TokenClassification——BERT | transformers |  [BERT_NER](./examples/BERT_NER——transformers.ipynb) |
-|FinetuneLLM——ChatGLM2 🔥🔥🔥| transformers |  [ChatGLM2](./examples/ChatGLM2——transformers.ipynb) |
 |FinetuneLLM——ChatGLM2_LoRA 🔥🔥🔥| transformers,peft |  [ChatGLM2_LoRA](./examples/ChatGLM2_LoRA——transformers.ipynb) |
 |FinetuneLLM——ChatGLM2_AdaLoRA 🔥🔥🔥| transformers,peft |  [ChatGLM2_AdaLoRA](./examples/ChatGLM2_AdaLoRA——transformers.ipynb) |
 |FinetuneLLM——ChatGLM2_QLoRA 🔥🔥🔥| transformers |  [ChatGLM2_QLoRA_Kaggle](./examples/ChatGLM2_QLoRA_Kaggle——transformers.ipynb) |

diff --git a/kerasmodel_example.py b/kerasmodel_example.py
@@ -97,22 +97,19 @@ def reset(self):
 #model.load_ckpt(ckpt_path) #load trained ckpt and continue training
 dfhistory=model.fit(train_data=dl_train, 
                     val_data=dl_val, 
-                    epochs=10, 
-                    patience=5, 
+                    epochs=100, 
+                    patience=10, 
                     monitor="val_acc",
                     mode="max",
                     ckpt_path=ckpt_path,
                     plot=True,
                     wandb=False,
                     mixed_precision='fp16',
-                    cpu=True   
+                    cpu=True
                    )
 
 model.evaluate(dl_val,quiet=False)
 
-
-
-
 # ### 5, use the model
 
 net = model.net

diff --git a/push2github.ipynb b/push2github.ipynb
diff --git a/setup.py b/setup.py
@@ -21,7 +21,7 @@ def get_version():
     description="pytorch❤️keras",
     long_description=README,
     install_requires=[           
-         'accelerate',
+         'accelerate>=0.20.3',
          'tqdm',
          #'torch',
          #'torchmetrics',

diff --git a/torchkeras/__init__.py b/torchkeras/__init__.py
@@ -1,4 +1,4 @@
-__version__="3.9.2"
+__version__="3.9.3"
 
 import sys
 from .kerasmodel import KerasModel

diff --git a/torchkeras/kerascallbacks.py b/torchkeras/kerascallbacks.py
@@ -147,7 +147,6 @@ def on_validation_epoch_end(self, model:"KerasModel"):
         dfhistory = pd.DataFrame(model.history)
         self.progress.update(dfhistory['epoch'].iloc[-1])
 
-
     def on_fit_end(self,  model:"KerasModel"):
         dfhistory = pd.DataFrame(model.history)
         if dfhistory['epoch'].max()<model.epochs:
@@ -160,6 +159,8 @@ def __init__(self,figsize = (6,4),
         self.figsize = (6,4)
         self.save_path = save_path
         self.in_jupyter = is_jupyter()
+        if not self.in_jupyter:
+            print('\nView dynamic loss/metric plot: \n'+os.path.abspath(self.save_path))
 
     def on_fit_start(self,model: 'KerasModel'):
         self.metric =  model.monitor.replace('val_','')

diff --git a/torchkeras/kerasmodel.py b/torchkeras/kerasmodel.py
@@ -156,11 +156,8 @@ def fit(self, train_data, val_data=None, epochs=10, ckpt_path='checkpoint',
         callbacks = callbacks if callbacks is not None else []
 
         if bool(plot):
-            if is_jupyter():
-                from torchkeras.kerascallbacks import VisProgress
-                callbacks = [VisProgress()]+callbacks
-            from torchkeras.kerascallbacks import VisMetric
-            callbacks = [VisMetric()]+callbacks
+            from torchkeras.kerascallbacks import VisProgress,VisMetric
+            callbacks = [VisMetric(),VisProgress()]+callbacks
 
         if wandb!=False:
             from torchkeras.kerascallbacks import WandbCallback
@@ -174,16 +171,12 @@ def fit(self, train_data, val_data=None, epochs=10, ckpt_path='checkpoint',
 
         start_epoch = 1 if self.from_scratch else 0
 
-        if quiet is None:
-            if is_jupyter():
-                quiet = True
-            else:
-                quiet = False
+        if is_jupyter() or bool(plot) or quiet is None:
+            quiet = True
 
         quiet_fn = (lambda epoch:quiet) if isinstance(quiet,bool) else (
             (lambda epoch:epoch>quiet) if isinstance(quiet,int) else quiet)
 
-
         for epoch in range(start_epoch,epochs+1):
             should_quiet = quiet_fn(epoch)
 
@@ -246,15 +239,16 @@ def fit(self, train_data, val_data=None, epochs=10, ckpt_path='checkpoint',
                         monitor,arr_scores[best_score_idx])))
 
             if len(arr_scores)-best_score_idx>patience:
-                self.accelerator.print(colorful(
-                    "<<<<<< {} without improvement in {} epoch,""early stopping >>>>>>"
-                ).format(monitor,patience))
-                break; 
+                break
 
         if self.accelerator.is_local_main_process:   
             dfhistory = pd.DataFrame(self.history)
             [cb.on_fit_end(model = self) for cb in self.callbacks 
                  if hasattr(cb,'on_fit_end')]
+            if epoch<epochs:
+                self.accelerator.print(colorful(
+                        "<<<<<< {} without improvement in {} epoch,""early stopping >>>>>> \n"
+                    ).format(monitor,patience))
             self.net = self.accelerator.unwrap_model(self.net)
             self.net.cpu()
             self.load_ckpt(ckpt_path)

diff --git a/torchkeras/pbar.py b/torchkeras/pbar.py
@@ -2,9 +2,12 @@
 '''
 reference from https://github.com/fastai/fastprogress/
 '''
-import time
+import time,sys
 from IPython.display import clear_output, display, HTML
+from torchkeras.utils import is_jupyter
 
+from tqdm.utils import _term_move_up
+move_up = _term_move_up()
 
 def format_time(t):
     "Format `t` (in seconds) to (h):mm:ss"
@@ -50,18 +53,27 @@ def html_progress_bar(value, total, label='', postfix='', interrupted=False):
     </div>
     """
 
+def text_progress_bar(value, total, label='', postfix='', interrupted=False):
+    bar_style = "🟥" if interrupted else "⬜️"  #"🟥","⬜️" "○", "*"
+    percentage = round(value / total * 50)
+    finished = "🟩" * (percentage)  #"🟩","●"
+    unfinished = bar_style * (50 - percentage)
+    bar = "\r{}{} {}".format(finished,unfinished,label)+" "*20+"\t"*50+f"{postfix}"+" "*20+"\t"*50
+    return bar
+
 class ProgressBar:
     update_every,first_its,lt = 0.2,5,'<'
-    def __init__(self, gen, total=None, 
-                 display=True, comment=''):
+    def __init__(self, gen, total=None, comment=''):
         self.gen,self.comment = gen,comment
         self.postfix = ''
         self.total = None if total=='noinfer' else len(gen) if total is None else total
         self.last_v = 0
-        self.display = display
         self.last_v = None
+        self.display = True
+        self.in_jupyter = is_jupyter()
         self.update(0)
 
+
     def update(self, val):
         if self.last_v is None:
             self.on_iter_begin()
@@ -76,30 +88,44 @@ def update(self, val):
             self.wait_for = max(int(self.update_every / (avg_t+1e-8)),1)
             self.pred_t = None if self.total is None else avg_t * self.total
             self.last_v,self.last_t = val,cur_t
-            self.update_bar(val)
             if self.total is not None and val >= self.total:
                 self.on_iter_end()
-                self.last_v = None
+                self.last_v = self.total
+            else:
+                self.update_bar(val)
 
     def on_iter_begin(self):
-        self.html_code = '\n'.join([html_progress_bar(0, self.total, ""), ""])
-        display(HTML(html_progress_bar_styles))
-        self.out = display(HTML(self.html_code), display_id=True)
+        if self.in_jupyter:
+            self.html_code = '\n'.join([html_progress_bar(0, self.total, ""), ""])
+            display(HTML(html_progress_bar_styles))
+            self.out = display(HTML(self.html_code), display_id=True)
+        else:
+            print('\n')
 
     def on_iter_end(self):
         total_time = format_time(time.time() - self.start_t)
-        self.comment = f'[{total_time}]' 
-        if hasattr(self, 'out'): 
-            self.on_update(self.total,self.comment,self.postfix)
+        self.comment = f'100% [{self.total}/{self.total}] [{total_time}]'   
+        self.on_update(self.total,self.comment,self.postfix,False,1)
+        self.display = False
+        if not self.in_jupyter:
+            print('\n')
 
-    def on_update(self, val, comment='', postfix='', interrupted=False): 
-        self.progress = html_progress_bar(val, self.total,comment,postfix,interrupted)
-        if self.display: 
-            self.out.update(HTML(self.progress))
-
+    def on_update(self, val, comment='', postfix='', interrupted=False, up=1): 
+        if not self.display:
+            return 
+        if self.in_jupyter:
+                self.progress = html_progress_bar(val, self.total,comment,postfix,interrupted)
+                self.out.update(HTML(self.progress))
+        else:
+            progress = text_progress_bar(val, self.total, comment, postfix, interrupted)
+            print(move_up*up+progress,end='')
+
     def on_interrupt(self,msg='interrupted'):
-        self.on_update(self.last_v,self.comment+f'[{msg}]',interrupted=True)
-
+        self.on_update(self.last_v,self.comment+f' [{msg}]',self.postfix,interrupted=True,up=1)
+        self.display = False
+        if not self.in_jupyter:
+            print('\n')
+
     def __iter__(self):
         if self.total != 0: self.update(0)
         try:
@@ -121,15 +147,17 @@ def update_bar(self, val):
         if val ==0:
             self.comment = f'0% [0/{self.total}]'
             return self.on_update(0, self.comment)
-        pct = '' if self.total is None else f'{100 * val/self.total:.2f}% '
+        pct = '' if self.total is None else f'{100 * val/self.total:.2f}%'
         tot = '?' if self.total is None else str(self.total)
         elapsed_t = self.last_t - self.start_t
         remaining_t = '?' if self.pred_t is None else format_time(self.pred_t - elapsed_t)
         elapsed_t = format_time(elapsed_t)
-        self.comment = f'{pct}[{val}/{tot} {elapsed_t}{self.lt}{remaining_t}]'
+        self.comment = f'{pct} [{val}/{tot}] [{elapsed_t}{self.lt}{remaining_t}]'
         self.on_update(val, self.comment, self.postfix)
 
     def set_postfix(self,**kwargs):
+        if not self.display:
+            return 
         postfix = ''
         if 'i' in kwargs and 'n' in kwargs:
             from tqdm.std import Bar 
@@ -139,14 +167,14 @@ def set_postfix(self,**kwargs):
             ratio = i/n
             postfix+=format(Bar(ratio,default_len=20))
             postfix+=f'{100 * i/n:.2f}%'
-            postfix+=f' [{i}/{n}] '
+            postfix+=f' [{i}/{n}]'
         if kwargs:
-            postfix+='['
+            postfix+=' ['
             for i,(key,value) in enumerate(kwargs.items()):
                 if isinstance(value,float):
-                    postfix = postfix+f'{key}={format_number(value)},'
+                    postfix = postfix+f'{key}={format_number(value)}, '
                 else:
-                    postfix = postfix+f'{key}={value},'
-            postfix = postfix[:-1]+']'
+                    postfix = postfix+f'{key}={value}, '
+            postfix = postfix[:-2]+']'
         self.postfix = postfix
         self.on_update(self.last_v,self.comment,self.postfix)