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, 24, 16]  # Corresponding layers in DenseNet121

        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 DenseNet121 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
import torch.nn.functional as F
from torch.utils.cpp_extension import load_inline

_cpp_src = r"""
#include <torch/extension.h>
torch::Tensor bn_train_relu_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias, torch::Tensor running_mean, torch::Tensor running_var, double momentum, double eps);
torch::Tensor bn_eval_relu_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias, torch::Tensor running_mean, torch::Tensor running_var, double eps);
"""
_cuda_src = r"""
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <math.h>

__global__ void bn_train_relu_kernel(const float* __restrict__ x, float* __restrict__ y,
                                     const float* __restrict__ w, const float* __restrict__ b,
                                     float* __restrict__ rm, float* __restrict__ rv,
                                     int N, int C, int HW, float momentum, float eps) {
    int c = blockIdx.x;
    int tid = threadIdx.x;
    int M = N * HW;
    float sum = 0.0f, sum2 = 0.0f;
    for (int n = 0; n < N; ++n) {
        const float* xp = x + (n * C + c) * HW;
        for (int i = tid; i < HW; i += blockDim.x) {
            float v = xp[i];
            sum += v;
            sum2 += v * v;
        }
    }
    __shared__ float s[256];
    __shared__ float q[256];
    s[tid] = sum; q[tid] = sum2;
    __syncthreads();
    for (int stride = 128; stride > 0; stride >>= 1) {
        if (tid < stride) { s[tid] += s[tid + stride]; q[tid] += q[tid + stride]; }
        __syncthreads();
    }
    float mean = s[0] / (float)M;
    float var = q[0] / (float)M - mean * mean;
    if (var < 0.0f) var = 0.0f;
    if (tid == 0) {
        rm[c] = rm[c] * (1.0f - momentum) + mean * momentum;
        rv[c] = rv[c] * (1.0f - momentum) + var * ((float)M / ((float)M - 1.0f)) * momentum;
    }
    float inv = rsqrtf(var + eps);
    float ww = w[c], bb = b[c];
    for (int n = 0; n < N; ++n) {
        const float* xp = x + (n * C + c) * HW;
        float* yp = y + (n * C + c) * HW;
        for (int i = tid; i < HW; i += blockDim.x) {
            float v = (xp[i] - mean) * inv * ww + bb;
            yp[i] = v > 0.0f ? v : 0.0f;
        }
    }
}

__global__ void bn_eval_relu_kernel(const float* __restrict__ x, float* __restrict__ y,
                                    const float* __restrict__ w, const float* __restrict__ b,
                                    const float* __restrict__ rm, const float* __restrict__ rv,
                                    int total, int C, int HW, float eps) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < total) {
        int c = (idx / HW) % C;
        float v = (x[idx] - rm[c]) * rsqrtf(rv[c] + eps) * w[c] + b[c];
        y[idx] = v > 0.0f ? v : 0.0f;
    }
}

torch::Tensor bn_train_relu_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias,
                                    torch::Tensor running_mean, torch::Tensor running_var,
                                    double momentum, double eps) {
    x = x.contiguous();
    auto y = torch::empty_like(x);
    int N = (int)x.size(0), C = (int)x.size(1), HW = (int)(x.size(2) * x.size(3));
    bn_train_relu_kernel<<<C, 256>>>(x.data_ptr<float>(), y.data_ptr<float>(),
        weight.data_ptr<float>(), bias.data_ptr<float>(), running_mean.data_ptr<float>(), running_var.data_ptr<float>(),
        N, C, HW, (float)momentum, (float)eps);
    return y;
}

torch::Tensor bn_eval_relu_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias,
                                   torch::Tensor running_mean, torch::Tensor running_var, double eps) {
    x = x.contiguous();
    auto y = torch::empty_like(x);
    int N = (int)x.size(0), C = (int)x.size(1), HW = (int)(x.size(2) * x.size(3));
    int total = N * C * HW;
    int threads = 256;
    bn_eval_relu_kernel<<<(total + threads - 1) / threads, threads>>>(x.data_ptr<float>(), y.data_ptr<float>(),
        weight.data_ptr<float>(), bias.data_ptr<float>(), running_mean.data_ptr<float>(), running_var.data_ptr<float>(),
        total, C, HW, (float)eps);
    return y;
}
"""

_ext = load_inline(name="densenet_bnrelu_opt_final", cpp_sources=_cpp_src, cuda_sources=_cuda_src,
                   functions=["bn_train_relu_forward", "bn_eval_relu_forward"],
                   extra_cflags=["-O3"], extra_cuda_cflags=["-O3"], verbose=False)

def _bn_relu(bn, x, training: bool):
    if x.is_cuda and x.dtype is torch.float32:
        if training and x.size(2) <= 14:
            return _ext.bn_train_relu_forward(x, bn.weight, bn.bias, bn.running_mean, bn.running_var, bn.momentum, bn.eps)
        if not training:
            return _ext.bn_eval_relu_forward(x, bn.weight, bn.bias, bn.running_mean, bn.running_var, bn.eps)
    return F.relu(bn(x), inplace=True)

class DenseBlock(nn.Module):
    def __init__(self, num_layers: int, num_input_features: int, growth_rate: int):
        super().__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[2](_bn_relu(layer[0], x, self.training))
            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().__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):
        t = self.transition
        return t[3](t[2](_bn_relu(t[0], x, self.training)))

class ModelNew(nn.Module):
    def __init__(self, growth_rate: int = 32, num_classes: int = 1000):
        super().__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, 24, 16]
        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[0](x)
        x = _bn_relu(self.features[1], x, self.training)
        x = self.features[3](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 = _bn_relu(self.final_bn, x, self.training)
        x = F.adaptive_avg_pool2d(x, (1, 1)).view(x.size(0), -1)
        return self.classifier(x)