Deep Learning Practice · End Term · Jan 2026 · FN

**1. Read the configuration:** nc is the number of detection classes. The file states nc: 5. **2. Verify against the names:** helmet, no_helmet, person, car and truck are five distinct class names. Their zero-based class IDs are 0 through 4. **3. Conclude:** The model is trained for five object classes. The class ID 2 in one annotation identifies person; it is not the total number of classes. **Answer: 5.**

**1. Identify the coordinate format:** YOLO annotation coordinates are normalized by the image dimensions. The supplied x-center is 0.25 and the image width is 640 pixels. **2. Convert to pixels:** \[x_c=x_{c,\mathrm{norm}}W=0.25\times640=160\text{ pixels}.\] **3. Check:** One quarter of a 640-pixel width is 160 pixels from the left edge. Image height is not used for an x-coordinate. **Answer: 160.**

**1. Convert center and width:** The x-center is \(0.25\times640=160\) pixels. The normalized box width is 0.40, giving \(w=0.40\times640=256\) pixels. **2. Move half a width left:** The top-left corner has \[x_{\min}=x_c-\frac{w}{2}=160-\frac{256}{2}=160-128=32\text{ pixels}.\] **3. Check:** The right edge is \(160+128=288\); the width \(288-32=256\) agrees with the annotation. **Answer: 32.**

**1. Use the convolution size formula:** For input size \(n\), kernel \(k\), stride \(s\), padding \(p\) and unit dilation, \[n_{\mathrm{out}}=\left\lfloor\frac{n+2p-k}{s}\right\rfloor+1.\] **2. First convolution:** \(\lfloor(84-4)/2\rfloor+1=41\). Sixteen filters produce \(41\times41\times16\). ReLU changes values but not dimensions. **3. Second convolution:** \(\lfloor(41-2)/2\rfloor+1=19+1=20\). Thirty-two filters produce \(20\times20\times32\). **4. Flatten:** Multiplying all non-batch dimensions gives \(20\times20\times32=12{,}800\) features per image. **Answer: 12800.**

**1. Understand a 1-by-1 convolution:** At one spatial position, one output filter multiplies each of the 1,024 input-channel values by a weight. Thus each output value needs 1,024 multiplications. **2. Count output values:** There are \(14\times14=196\) spatial positions and 256 output channels, giving \(196\times256=50{,}176\) output values. **3. Multiply:** \[196\times256\times1{,}024=51{,}380{,}224\text{ multiplications}.\] In millions this is 51.380224, rounded to 51.38. Bias additions are not multiplications. **Answer: 51.38.**

**1. Identify the intended YOLO output convention:** In the original YOLO-style grid formulation, each cell predicts B bounding boxes with five values per box: x, y, width, height and confidence. It also predicts one set of C class probabilities shared by the cell. **2. Compute channels per cell:** \[D=5B+C=5(3)+80=95.\] **3. Distinguish depth from total size:** The tensor is \(13\times13\times95\); its depth is 95, rather than the product of all three dimensions. **Convention note:** This uses cell-shared class probabilities. A detector predicting separate class scores for every box would use a different formula, such as \(B(5+C)\). The question’s intended formulation is the original grid convention. **Answer: 95.**

**1. Count one filter’s weights:** A 3-by-3 filter spans all 64 input channels, so it contains \(3\times3\times64=576\) weights. **2. Count all filters:** There are 64 output filters: \[N=64\times576=64\times64\times3\times3=36{,}864.\] **3. Exclude biases:** The question asks only for weights. Including one bias per filter would add 64, but those are not counted. **Answer: 36864.**

**1. Reduce training storage/work:** Mixed precision uses lower-precision tensors where appropriate. Freezing a backbone avoids storing its parameter gradients and optimizer states, although its forward computation is still required. **2. Keep the GPU supplied:** Background data workers and prefetching can overlap loading/augmentation with GPU work, reducing idle time when input loading is a bottleneck. **3. Evaluate batch size carefully:** A larger batch may improve throughput if it fits, but normally uses more activation memory. It does not lower memory usage as claimed in C. **Answer: A, B, D — Selected: Mixed Precision Training (FP16): Using 16-bit floating-point numbers instead of 32-bit for certain operations to speed up computation and reduce GPU memory consumption.; Freeze the Backbone: Keeping the early layers of the pre-trained model non- trainable (requires grad=False) and only training the final classification head.; Data Prefetching: Using multiple CPU workers (e.g., num workers > 0 in PyTorch) to load and augment data in the background while the GPU processes the current batch.** A: **Correct:** Suitable lower-precision operations can reduce memory and improve speed on compatible hardware. B: **Correct:** Training only the head avoids updating and storing optimizer state for the frozen backbone. C: **Incorrect:** Larger batches normally increase memory demand; throughput gains do not imply lower memory use. D: **Correct:** Background workers can overlap input preparation with GPU computation.

**1. Match device placement:** PyTorch operations require compatible tensors and model parameters on the same device. Move both to the selected GPU for GPU execution. **2. Save model state:** A state_dict stores parameter tensors and persistent buffers, such as batch-normalization statistics. It does not store the architecture’s Python source code. Saving this state is generally more portable than serializing a whole model object. **3. Restore appropriately:** Instantiate a compatible model class, load the state dictionary and use evaluation mode for inference. To resume training faithfully, also restore optimizer/scheduler states and other required training state. **Precision note:** B says “only learnable parameters”, but state_dict also includes persistent buffers. Its intended comparison with a whole-model pickle remains the standard recommended practice. **Answer: A, B, C — Selected: To utilize a GPU in PyTorch, you must explicitly move both the model parameters and the input data tensors to the same device (e.g.,.to(’cuda’)).; Saving a model using torch.save(model.state dict(), PATH) is generally preferred over saving the entire model object because it only stores the learnable parameters (weights and biases).; When resuming training or performing inference, you must first instantiate the model architecture and then load the weights using model.load state dict(torch.load(PATH, map location=device)).** A: **Correct:** Both the network and its tensor inputs must be placed on the selected compatible device. B: **Correct:** Saving a state dictionary is preferred here; it includes weights and persistent buffers, rather than architecture source. C: **Correct:** Loading a state dictionary requires a compatible instantiated architecture; full training resumption also needs optimizer state. D: **Incorrect:** A state dictionary does not contain source code and cannot generally load into an unrelated architecture.

**1. Identify the detector:** YOLOv8 is a one-stage detector. Its detection head uses an anchor-free design instead of predefined anchor-box templates. **2. Separate predictions:** The detection head has separate paths/components for class prediction and bounding-box regression. Both operate on shared feature maps within the detector forward pass. **3. Contrast two-stage detection:** An RPN first proposing regions is characteristic of Faster R-CNN, rather than standard YOLOv8. **Answer: A, B, C — Selected: The model predicts object locations directly without using predefined anchor boxes.; The model uses separate components within its detection head for classification and bounding box regression.; The model performs detection in a single forward pass without a separate proposal stage.** A: **Correct:** YOLOv8 uses anchor-free detection, although feature-map reference points are still used. B: **Correct:** Its decoupled head separates classification and localization predictions. C: **Correct:** It predicts detections through a one-stage network without an RPN proposal stage. D: **Incorrect:** A separate RPN is not part of the standard YOLOv8 pipeline.

**1. Explain the goal:** Super-resolution reconstructs a high-resolution image from a low-resolution input. Matching average pixel values alone can produce overly smooth images. **2. Use perceptual content loss:** Compare generated and target images through feature maps of a pretrained network such as VGG. This encourages perceptually meaningful structure. **3. Add adversarial loss:** A discriminator teaches the generator to produce images that look like natural high-resolution examples. SRGAN combines this with content loss; it is not merely a pixel-MSE system or an ImageNet classifier. **Answer: A, B — Selected: Adversarial Loss: To encourage the generator to produce solutions on the natural image manifold.; Content (Perceptual) Loss: Based on feature maps from a pre-trained VGG network.** A: **Correct:** Adversarial training encourages natural-looking high-resolution textures. B: **Correct:** VGG-feature content loss compares perceptual structure rather than only individual pixels. C: **Incorrect:** Pixel MSE alone does not describe SRGAN’s perceptual/adversarial objective. D: **Incorrect:** The pretrained network supplies features; the task is not 1,000-way classification.

**1. Follow the encoder:** U-Net’s contracting path reduces resolution while learning increasingly rich features. **2. Follow the decoder:** Its expansive path upsamples the features to recover a dense spatial prediction, such as a segmentation mask. The original architecture uses up-convolutions/transposed convolutions. **3. Recover local detail:** Skip connections concatenate matching-resolution encoder features with decoder features, helping retain boundaries and localization. Discarding spatial information at every layer would undermine pixel-level prediction. **Answer: A, B, C — Selected: It features a symmetric architecture consisting of a contracting path (encoder) and an expansive path (decoder).; Skip connections concatenate high-resolution features from the contracting path directly to the up sampled features in the expansive path.; The expansive path uses transposed convolutions (or up-convolutions) to increase the spatial resolution of the feature maps.** A: **Correct:** The encoder-decoder arrangement contracts and then expands spatial resolution. B: **Correct:** Concatenated skip features restore fine localization details. C: **Correct:** Upsampling operations increase the decoder feature-map resolution. D: **Incorrect:** Repeated global pooling would remove the spatial layout needed for segmentation.

**1. Identify the two stages:** Faster R-CNN generates candidate object regions and then classifies/refines them. Proposal and detection filtering depend on image content and scores. **2. Filter candidates:** Confidence thresholds, top-k selection and non-maximum suppression remove low-confidence and duplicate boxes. The remaining count can differ between images, even though the network imposes maximum proposal/detection limits. **3. Interpret the code as inference:** To run the example, call model.eval() before passing only images. The default training mode expects targets. The question is describing its inference outputs. **Important distinction:** YOLO inference can also return a variable number of final detections after filtering. This behavior is not exclusive to Faster R-CNN. **Answer: A, B — Selected: The model generates a variable number of region proposals based on the image content.; The model filters region proposals using confidence scores and non maximum suppression.** A: **Correct:** Content-dependent proposal scoring/filtering leads to a variable set of useful regions within configured limits. B: **Correct:** Score filtering and non-maximum suppression change the final detection count. C: **Incorrect:** A fixed-grid prediction scheme does not describe Faster R-CNN’s region-based architecture. D: **Incorrect:** Its returned final box count need not be constant.

**1. Capture multiple scales:** Inception modules apply parallel branches with different filter sizes and pooling. Their outputs are combined, allowing the network to capture patterns at multiple scales. **2. Control computation:** A 1-by-1 convolution reduces channel depth before expensive 3-by-3 or 5-by-5 operations, reducing their multiplication and parameter counts. **3. Support training:** Original GoogLeNet includes intermediate auxiliary classifiers that provide extra training signal and regularization. Its parallel multi-scale design differs from VGG’s primarily stacked 3-by-3 filters. **Answer: A, B, C — Selected: It uses Inception modules that apply multiple filter sizes (1×1, 3×3, 5×5) in parallel to capture multi-scale features.; It utilizes 1×1 convolutions as bottleneck layers to reduce dimensionality before computationally expensive operations.; It incorporates Auxiliary Classifiers in intermediate layers to inject additional gradient signal and combat the vanishing gradient problem.** A: **Correct:** Parallel filters of several sizes capture features at different spatial scales. B: **Correct:** Channel bottlenecks reduce the cost of subsequent larger filters. C: **Correct:** Auxiliary training heads provide additional gradient signal during training. D: **Incorrect:** GoogLeNet is not exclusively a chain of 3-by-3 convolutions.

**1. Describe a convolutional filter:** A small filter connects to a local patch and then slides across many image positions. **2. Explain parameter sharing:** The same filter weights are reused at every location. A convolution has \(k_hk_wC_{\mathrm{in}}C_{\mathrm{out}}\) weights, independent of image width and height. Increasing spatial size therefore increases output computation, rather than the filter parameter count. **3. State the scope:** Local connectivity and shared parameters allow convolutional feature extractors to process different spatial sizes. A later flatten-and-dense head can still impose a fixed input size; adaptive/global pooling is often used to remove that restriction. **Answer: A, C — Selected: Neurons are connected only to a local region of the input.; Parameter Sharing: Using the same set of weights (filters) across different spatial locations of the input.** A: **Correct:** Local receptive fields avoid separate dense connections to every image pixel. B: **Incorrect:** Max pooling does not normally reduce the image to one pixel before the first hidden layer. C: **Correct:** Reusing a filter across all positions is the central reason parameter count is independent of image size. D: **Incorrect:** Flattened fully connected layers generally depend on a fixed number of input values.

**1. Read the first VGG16 layer:** It has 64 filters, a 3-by-3 kernel, stride 1 and padding 1. The input shape is batch 1, channels 3, height 224, width 224. **2. Compute spatial output:** \[\left\lfloor\frac{224+2(1)-3}{1}\right\rfloor+1=224.\] Both height and width remain 224. **3. Set output channels:** Sixty-four filters produce 64 channels, so the output is torch.Size([1, 64, 224, 224]). No pooling layer is applied by model.features[0]. **Answer: C — Selected: torch.Size([1, 64, 224, 224]).** A: **Incorrect:** The output has 64 channels, not the original three. B: **Incorrect:** A stride-one padded convolution does not halve the spatial dimensions. C: **Correct:** Batch size stays one, filter count gives 64 channels and padding preserves 224-by-224 size. D: **Incorrect:** A 222-by-222 output would correspond to no padding.

**1. Describe the operation:** A 1-by-1 convolution mixes input channels independently at each spatial position. **2. Add expressive power efficiently:** Placing an activation such as ReLU after it introduces another nonlinear transformation without directly enlarging the spatial receptive field. Compared with a larger kernel at the same channel widths, the 1-by-1 convolution is relatively inexpensive. **3. Keep the residual design:** It can reduce or restore channel width in a bottleneck, while the skip connection still provides the residual learning path. The convolution alone is linear; the following activation supplies the nonlinearity. **Answer: B — Selected: They allow for the addition of non-linearity (via activation functions) without increasing the receptive field or computational cost excessively.** A: **Incorrect:** A standard stride-one 1-by-1 convolution does not upsample spatial dimensions. B: **Correct:** A cheap channel-mixing layer followed by an activation adds nonlinear modeling capacity. C: **Incorrect:** Convolution is a learned weighted operation, not max pooling. D: **Incorrect:** Bottleneck blocks still rely on residual skip connections.

**1. Use the original YOLO training rule:** When several box predictors in a cell could represent an object, one predictor is assigned responsibility for that ground-truth object. **2. Compare overlap:** Intersection over Union is \[\mathrm{IoU}=\frac{\text{intersection area}}{\text{union area}}.\] The predictor with the highest IoU with the ground-truth box receives responsibility under this original rule. **3. Separate the extra sentence:** The printed question also mentions varying output counts. That relates to inference filtering, not this training-assignment rule. Modern YOLO versions can use more elaborate assignment strategies; C answers the stated classic responsibility question. **Answer: C — Selected: The bounding box that has the highest Intersection over Union (IoU) with the ground truth.** A: **Incorrect:** The classic rule assigns one responsible predictor for an object, rather than every box in the cell. B: **Incorrect:** Class confidence alone does not measure localization overlap with the true box. C: **Correct:** Maximum overlap identifies the responsible predictor in the original YOLO rule. D: **Incorrect:** Distance from the image center is not the assignment criterion.

**1. Separate raw candidates from returned detections:** The network evaluates many candidate locations. results[0].boxes contains the detections retained after postprocessing, rather than every raw network output. **2. Filter predictions:** Low-confidence predictions are discarded. Non-maximum suppression removes redundant overlapping boxes for the same object/class according to the configured settings. **3. Interpret shape:** boxes.xyxy has shape \((N,4)\), with four corner coordinates per retained box. Different images can contain different numbers of confidently detected objects, so N varies. **Answer: B — Selected: The model filters predictions based on confidence scores and non maximum suppression, resulting in a variable number of final detections.** A: **Incorrect:** A fixed candidate layout does not imply a fixed count after postprocessing. B: **Correct:** Confidence filtering and suppression make the final detection count image-dependent. C: **Incorrect:** YOLOv8 is not a two-stage RPN detector. D: **Incorrect:** A class can have zero, one or many detected instances; class count does not fix N.

**1. Describe the rearrangement:** Pixel Shuffle moves values from channel positions into spatial positions, increasing image resolution. **2. State the shape rule:** For scale factor \(r\), it maps \[(B,Cr^2,H,W)\longrightarrow(B,C,rH,rW).\] For example, four channel groups can be rearranged into a two-times increase in both height and width. **3. Apply it to super-resolution:** SRResNet and SRGAN use sub-pixel convolution blocks, with learned convolutional features followed by Pixel Shuffle, to produce high-resolution outputs. **Answer: C — Selected: The Pixel Shuffle (Sub-pixel Convolution) layer.** A: **Incorrect:** Max pooling downsamples rather than rearranging channels for upsampling. B: **Incorrect:** Bilinear interpolation computes interpolated values; it is a different operation. C: **Correct:** Pixel Shuffle is the channel-to-spatial rearrangement used in sub-pixel upsampling. D: **Incorrect:** Global average pooling collapses spatial dimensions.

**1. Identify the size problem:** Region proposals cover different-sized rectangles in a shared convolutional feature map. A dense classifier needs a consistent input size for every proposal. **2. Apply RoI pooling:** The proposal region is divided into a fixed grid of bins. Pooling within each bin produces the same spatial output size regardless of the proposal’s original dimensions. **3. Feed the detection head:** The fixed-size region features support classification and bounding-box regression. Fast R-CNN uses supplied proposals; the RPN belongs to Faster R-CNN. **Answer: B — Selected: The Region of Interest (RoI) Pooling layer.** A: **Incorrect:** An RPN generates proposals in Faster R-CNN; it is not the fixed-size feature extraction step. B: **Correct:** RoI pooling converts each variable-size proposal region into a fixed-size feature representation. C: **Incorrect:** A generic convolution stack alone does not normalize arbitrary proposal sizes. D: **Incorrect:** Global average pooling over the whole image would not preserve separate proposal-specific features.

**1. Identify the expensive operation:** A k-by-k convolution costs roughly \(HWk^2C_{\mathrm{in}}C_{\mathrm{out}}\) multiplications. Large input channel depth makes 3-by-3 and 5-by-5 branches expensive. **2. Insert a bottleneck:** A 1-by-1 convolution first reduces channel count from \(C_{\mathrm{in}}\) to a smaller value. The larger filter then operates on this narrower representation. **3. Explain the benefit:** This makes parallel multi-scale branches practical within a manageable compute and parameter budget. The layer can also be followed by a nonlinearity, but channel reduction is the primary motivation here. **Answer: B — Selected: To reduce the dimensionality (depth) of feature maps to manage computational complexity.** A: **Incorrect:** A standard 1-by-1 bottleneck is not a transposed convolution. B: **Correct:** Reducing channel depth lowers the cost of the subsequent larger convolutions. C: **Incorrect:** Residual skip connections address a different design concern. D: **Incorrect:** A learned convolution does not enforce zero mean/unit variance; normalization layers serve that role.