import torch
import torch.nn as nn
import torch.nn.functional as F

class DenseBlock(nn.Module):
    def __init__(self, num_layers: int, num_input_features: int, growth_rate: int):
        """
        :param num_layers: The number of layers in the dense block
        :param num_input_features: The number of input feature maps
        :param growth_rate: The growth rate for the dense block (new features added per layer)
        """
        super(DenseBlock, self).__init__()
        layers = []
        for i in range(num_layers):
            layers.append(self._make_layer(num_input_features + i * growth_rate, growth_rate))
        self.layers = nn.ModuleList(layers)

    def _make_layer(self, in_features: int, growth_rate: int):
        """
        Creates a single layer with BatchNorm, ReLU, Conv2D, and Dropout.
        """
        return nn.Sequential(
            nn.BatchNorm2d(in_features),
            nn.ReLU(inplace=True),
            nn.Conv2d(in_features, growth_rate, kernel_size=3, padding=1, bias=False),
            nn.Dropout(0.0)
        )

    def forward(self, x):
        """
        :param x: Input tensor of shape (batch_size, num_input_features, height, width)
        :return: Concatenated output tensor with shape (batch_size, num_output_features, height, width)
        """
        features = [x]
        for layer in self.layers:
            new_feature = layer(x)
            features.append(new_feature)
            x = torch.cat(features, 1)  # Concatenate along channel axis
        return x

class TransitionLayer(nn.Module):
    def __init__(self, num_input_features: int, num_output_features: int):
        """
        :param num_input_features: The number of input feature maps
        :param num_output_features: The number of output feature maps
        """
        super(TransitionLayer, self).__init__()
        self.transition = nn.Sequential(
            nn.BatchNorm2d(num_input_features),
            nn.ReLU(inplace=True),
            nn.Conv2d(num_input_features, num_output_features, kernel_size=1, bias=False),
            nn.AvgPool2d(kernel_size=2, stride=2)
        )

    def forward(self, x):
        """
        :param x: Input tensor of shape (batch_size, num_input_features, height, width)
        :return: Downsampled tensor with reduced number of feature maps
        """
        return self.transition(x)

class Model(nn.Module):
    def __init__(self, growth_rate: int = 32, num_classes: int = 1000):
        """
        :param growth_rate: The growth rate of the DenseNet (new features added per layer)
        :param num_classes: The number of output classes for classification
        """
        super(Model, self).__init__()

        # Initial convolution and pooling
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        )

        # Each dense block is followed by a transition layer, except the last one
        num_features = 64
        block_layers = [6, 12, 48, 32]  # Corresponding layers in DenseNet201

        self.dense_blocks = nn.ModuleList()
        self.transition_layers = nn.ModuleList()

        for i, num_layers in enumerate(block_layers):
            block = DenseBlock(num_layers=num_layers, num_input_features=num_features, growth_rate=growth_rate)
            self.dense_blocks.append(block)
            num_features = num_features + num_layers * growth_rate

            if i != len(block_layers) - 1:
                transition = TransitionLayer(num_input_features=num_features, num_output_features=num_features // 2)
                self.transition_layers.append(transition)
                num_features = num_features // 2

        # Final batch norm and classifier
        self.final_bn = nn.BatchNorm2d(num_features)
        self.classifier = nn.Linear(num_features, num_classes)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        :param x: Input tensor of shape (batch_size, 3, height, width)
        :return: Output tensor of shape (batch_size, num_classes)
        """
        x = self.features(x)

        for i, block in enumerate(self.dense_blocks):
            x = block(x)
            if i != len(self.dense_blocks) - 1:
                x = self.transition_layers[i](x)

        x = self.final_bn(x)
        x = F.relu(x, inplace=True)
        x = F.adaptive_avg_pool2d(x, (1, 1)).view(x.size(0), -1)
        x = self.classifier(x)
        return x

# Testing the DenseNet201 model
batch_size = 10
num_classes = 10
height, width = 224, 224  # Standard input size for DenseNet

def get_inputs():
    return [torch.rand(batch_size, 3, height, width)]

def get_init_inputs():
    return [32, num_classes]

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

_cuda_src = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda.h>
#include <cuda_runtime.h>

__global__ void relu_avg_linear_kernel(const float* __restrict__ x,
                                       const float* __restrict__ w,
                                       const float* __restrict__ b,
                                       float* __restrict__ y,
                                       int N, int C, int H, int Wd, int K) {
    int nk = blockIdx.x;
    int n = nk / K;
    int k = nk - n * K;
    int HW = H * Wd;
    int total = C * HW;
    float sum = 0.0f;
    for (int idx = threadIdx.x; idx < total; idx += 256) {
        int c = idx / HW;
        float v = x[((n * C + c) * H * Wd) + (idx - c * HW)];
        v = fmaxf(v, 0.0f);
        sum += v * w[k * C + c];
    }
    __shared__ float sm[256];
    sm[threadIdx.x] = sum;
    __syncthreads();
    for (int s = 128; s > 0; s >>= 1) {
        if (threadIdx.x < s) sm[threadIdx.x] += sm[threadIdx.x + s];
        __syncthreads();
    }
    if (threadIdx.x == 0) y[n * K + k] = sm[0] / (float)HW + b[k];
}

torch::Tensor relu_avg_linear(torch::Tensor x, torch::Tensor w, torch::Tensor b) {
    int N = (int)x.size(0), C = (int)x.size(1), H = (int)x.size(2), Wd = (int)x.size(3), K = (int)w.size(0);
    auto y = torch::empty({N, K}, x.options());
    relu_avg_linear_kernel<<<N * K, 256, 0, at::cuda::getCurrentCUDAStream()>>>(x.data_ptr<float>(), w.data_ptr<float>(), b.data_ptr<float>(), y.data_ptr<float>(), N, C, H, Wd, K);
    return y;
}
'''

_ext = load_inline(
    name='densenet201_final_fused_ext_submit',
    cpp_sources='torch::Tensor relu_avg_linear(torch::Tensor x, torch::Tensor w, torch::Tensor b);',
    cuda_sources=_cuda_src,
    functions=['relu_avg_linear'],
    extra_cuda_cflags=['-O3'],
    verbose=False,
)

class DenseBlock(nn.Module):
    def __init__(self, num_layers: int, num_input_features: int, growth_rate: int):
        super(DenseBlock, self).__init__()
        self.layers = nn.ModuleList([self._make_layer(num_input_features + i * growth_rate, growth_rate) for i in range(num_layers)])

    def _make_layer(self, in_features: int, growth_rate: int):
        return nn.Sequential(
            nn.BatchNorm2d(in_features),
            nn.ReLU(inplace=True),
            nn.Conv2d(in_features, growth_rate, kernel_size=3, padding=1, bias=False),
            nn.Dropout(0.0),
        )

    def forward(self, x):
        features = [x]
        for layer in self.layers:
            new_feature = layer(x)
            features.append(new_feature)
            x = torch.cat(features, 1)
        return x

class TransitionLayer(nn.Module):
    def __init__(self, num_input_features: int, num_output_features: int):
        super(TransitionLayer, self).__init__()
        self.transition = nn.Sequential(
            nn.BatchNorm2d(num_input_features),
            nn.ReLU(inplace=True),
            nn.Conv2d(num_input_features, num_output_features, kernel_size=1, bias=False),
            nn.AvgPool2d(kernel_size=2, stride=2),
        )

    def forward(self, x):
        return self.transition(x)

class ModelNew(nn.Module):
    def __init__(self, growth_rate: int = 32, num_classes: int = 1000):
        super(ModelNew, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
        )
        num_features = 64
        block_layers = [6, 12, 48, 32]
        self.dense_blocks = nn.ModuleList()
        self.transition_layers = nn.ModuleList()
        for i, num_layers in enumerate(block_layers):
            self.dense_blocks.append(DenseBlock(num_layers, num_features, growth_rate))
            num_features += num_layers * growth_rate
            if i != len(block_layers) - 1:
                self.transition_layers.append(TransitionLayer(num_features, num_features // 2))
                num_features //= 2
        self.final_bn = nn.BatchNorm2d(num_features)
        self.classifier = nn.Linear(num_features, num_classes)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.features(x)
        for i, block in enumerate(self.dense_blocks):
            x = block(x)
            if i != len(self.dense_blocks) - 1:
                x = self.transition_layers[i](x)
        x = self.final_bn(x)
        return _ext.relu_avg_linear(x.contiguous(), self.classifier.weight.contiguous(), self.classifier.bias.contiguous())