Torchkeras，一个源码不足300行的深度学习框架

Torchkeras

了解过深度学习框架的都知道，Tensorflow是早期的主流框架，而后又出现了Keras,keras对Tensorflow进行了封装，使得搭建深度学模型的过程简化到了几个简单的步骤：summary、compile、fit、evaluate、 predict。Pytorch虽然比Tensorflow出现的晚，但是其在框架的实现方式上，更为优雅，可以很好的与Python的原生编程思维结合起来，因此越来越多的人开始转向Pytorch。开发的过程就是不断模块化的过程，Torchkeras便是这一原则下的产物，它将Pytorch进行了封装，使得利用Pytorch搭建深度学模型的过程也可以像Keras那样，并且非常的灵活。

torchkeras

https://github.com/lyhue1991/torchkeras

源码

# -*- coding: utf-8 -*-
import os
import datetime
import numpy as np 
import pandas as pd 
import torch
from collections import OrderedDict
from prettytable import PrettyTable

__version__ = "1.5.3"

#On macOs, run pytorch and matplotlib at the same time in jupyter should set this.
os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" 

# Some modules do the computation themselves using parameters or the parameters of children, treat these as layers
layer_modules = (torch.nn.MultiheadAttention, )

def summary(model, input_shape, input_dtype = torch.FloatTensor, batch_size=-1,
            layer_modules = layer_modules,*args, **kwargs):
    def register_hook(module):
        def hook(module, inputs, outputs):
            
            class_name = str(module.__class__).split(".")[-1].split("'")[0]
            module_idx = len(summary)

            key = "%s-%i" % (class_name, module_idx + 1)

            info = OrderedDict()
            info["id"] = id(module)
            if isinstance(outputs, (list, tuple)):
                try:
                    info["out"] = [batch_size] + list(outputs[0].size())[1:]
                except AttributeError:
                    # pack_padded_seq and pad_packed_seq store feature into data attribute
                    info["out"] = [batch_size] + list(outputs[0].data.size())[1:]
            else:
                info["out"] = [batch_size] + list(outputs.size())[1:]

            info["params_nt"], info["params"] = 0, 0 
            for name, param in module.named_parameters():
                info["params"] += param.nelement() * param.requires_grad
                info["params_nt"] += param.nelement() * (not param.requires_grad)

            # if the current module is already-used, mark as "(recursive)"
            # check if this module has params
            if list(module.named_parameters()):
                for v in summary.values():
                    if info["id"] == v["id"]:
                        info["params"] = "(recursive)"

            summary[key] = info

        # ignore Sequential and ModuleList and other containers
        if isinstance(module, layer_modules) or not module._modules:
            hooks.append(module.register_forward_hook(hook))

    hooks = []
    summary = OrderedDict()

    model.apply(register_hook)
    
    # multiple inputs to the network
    if isinstance(input_shape, tuple):
        input_shape = [input_shape]
        
    # batch_size of 2 for batchnorm
    x = [torch.rand(2, *size).type(input_dtype) for size in input_shape]
    # print(type(x[0]))
    
    try:
        with torch.no_grad():
            model(*x) if not (kwargs or args) else model(*x, *args, **kwargs)
    except Exception:
        # This can be usefull for debugging
        print("Failed to run torchkeras.summary...")
        raise
    finally:
        for hook in hooks:
            hook.remove()
            
    print("----------------------------------------------------------------")
    line_new = "{:>20}  {:>25} {:>15}".format("Layer (type)", "Output Shape", "Param #")
    print(line_new)
    print("================================================================")
    total_params = 0
    total_output = 0
    trainable_params = 0
    for layer in summary:
        # layer, output_shape, params
        line_new = "{:>20}  {:>25} {:>15}".format(
            layer,
            str(summary[layer]["out"]),
            "{0:,}".format(summary[layer]["params"]+summary[layer]["params_nt"])
        )
        total_params += (summary[layer]["params"]+summary[layer]["params_nt"])
        total_output += np.prod(summary[layer]["out"])
        trainable_params += summary[layer]["params"]
        print(line_new)

    # assume 4 bytes/number
    total_input_size = abs(np.prod(input_shape) * batch_size * 4. / (1024 ** 2.))
    total_output_size = abs(2. * total_output * 4. / (1024 ** 2.))  # x2 for gradients
    total_params_size = abs(total_params * 4. / (1024 ** 2.))
    total_size = total_params_size + total_output_size + total_input_size

    print("================================================================")
    print("Total params: {0:,}".format(total_params))
    print("Trainable params: {0:,}".format(trainable_params))
    print("Non-trainable params: {0:,}".format(total_params - trainable_params))
    print("----------------------------------------------------------------")
    print("Input size (MB): %0.6f" % total_input_size)
    print("Forward/backward pass size (MB): %0.6f" % total_output_size)
    print("Params size (MB): %0.6f" % total_params_size)
    print("Estimated Total Size (MB): %0.6f" % total_size)
    print("----------------------------------------------------------------")


class Model(torch.nn.Module):
    
    # print time bar...
    @staticmethod
    def print_bar(): 
        nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
        print("\n"+"="*80 + "%s"%nowtime)
        
    def __init__(self,net = None):
        super(Model, self).__init__()
        self.net = net

    def forward(self,x):
        if self.net:
            return self.net.forward(x)
        else:
            raise NotImplementedError
    
    def compile(self, loss_func, 
               optimizer=None, metrics_dict=None,device = None):
        self.loss_func = loss_func
        self.optimizer = optimizer if optimizer else torch.optim.Adam(self.parameters(),lr = 0.001)
        self.metrics_dict = metrics_dict if metrics_dict else {}
        self.history = {}
        self.device = device if torch.cuda.is_available() else None
        if self.device:
            self.to(self.device)
        

    def summary(self,input_shape,input_dtype = torch.FloatTensor, batch_size=-1 ):
        summary(self,input_shape,input_dtype,batch_size)
    
    def train_step(self, features, labels):  
           
        self.train()
        self.optimizer.zero_grad()
        if self.device:
            features = features.to(self.device)
            labels = labels.to(self.device)

        # forward
        predictions = self.forward(features)
        loss = self.loss_func(predictions,labels)
        
        # evaluate metrics
        train_metrics = {"loss":loss.item()}
        for name,metric_func in self.metrics_dict.items():
            train_metrics[name] = metric_func(predictions,labels).item()

        # backward
        loss.backward()

        # update parameters
        self.optimizer.step()
        self.optimizer.zero_grad()
        
        return train_metrics
    
    @torch.no_grad()
    def evaluate_step(self, features,labels):
        
        self.eval()
        
        if self.device:
            features = features.to(self.device)
            labels = labels.to(self.device)
            
        with torch.no_grad():
            predictions = self.forward(features)
            loss = self.loss_func(predictions,labels)
        
        val_metrics = {"val_loss":loss.item()}
        for name,metric_func in self.metrics_dict.items():
            val_metrics["val_"+name] = metric_func(predictions,labels).item()
            
        return val_metrics
        
    
    def fit(self,epochs,dl_train,dl_val = None,log_step_freq = 1):
        
        print("Start Training ...")
        Model.print_bar()
        
        dl_val = dl_val if dl_val else []
        
        for epoch in range(1,epochs+1):
            
            # 1，training loop -------------------------------------------------
            
            train_metrics_sum, step = {}, 0
            for features,labels in dl_train:
                step = step + 1
                train_metrics = self.train_step(features,labels)
    
                for name,metric in train_metrics.items():
                    train_metrics_sum[name] = train_metrics_sum.get(name,0.0)+metric
                    
                if step%log_step_freq == 0:   
                    logs = {"step":step}
                    logs.update({k:round(v/step,3) for k,v in train_metrics_sum.items()})
                    print(logs)
                
            for name,metric_sum in train_metrics_sum.items():
                self.history[name] = self.history.get(name,[])+[metric_sum/step]
                
                
            # 2，validate loop -------------------------------------------------

            val_metrics_sum, step = {}, 0
            for features,labels in dl_val:
                step = step + 1
                val_metrics = self.evaluate_step(features,labels)
                for name,metric in val_metrics.items():
                    val_metrics_sum[name] = val_metrics_sum.get(name,0.0)+metric
            for name,metric_sum in val_metrics_sum.items():
                self.history[name] = self.history.get(name,[])+[metric_sum/step]
                
            # 3，print logs -------------------------------------------------
            
            infos = {"epoch":epoch}
            infos.update({k:round(self.history[k][-1],3) for k in self.history})
            tb = PrettyTable()
            tb.field_names = infos.keys()
            tb.add_row(infos.values())
            print("\n",tb)
            Model.print_bar()
        
        print("Finished Training...")
            
        return pd.DataFrame(self.history)
    
    @torch.no_grad()
    def evaluate(self,dl_val):
        self.eval()
        val_metrics_list = {}
        for features,labels in dl_val:
            val_metrics = self.evaluate_step(features,labels)
            for name,metric in val_metrics.items():
                val_metrics_list[name] = val_metrics_list.get(name,[])+[metric]
        
        return {name:np.mean(metric_list) for name,metric_list in val_metrics_list.items()}
    
    @torch.no_grad()
    def predict(self,dl):
        self.eval()
        if self.device:
            result = torch.cat([self.forward(t[0].to(self.device)) for t in dl])
        else:
            result = torch.cat([self.forward(t[0]) for t in dl])
        return(result.data)

使用教程

torchkeras，像Keras一样训练Pytorch模型！

https://zhuanlan.zhihu.com/p/149903272