Deep Learning Practice · End Term · Jan 2026 · AN

**1. Read nc:** In a YOLO dataset configuration, nc is the number of detection classes. Here nc: 3. **2. Check the names:** helmet, no_helmet and person give three distinct classes, with zero-based IDs 0, 1 and 2. **3. Conclude:** The model detects 3 classes. The leading 0 in the sample annotation labels that object as helmet; it does not mean that the dataset has zero classes. **Answer: 3.**

**1. Identify the normalized coordinate:** The annotation gives x-center 0.50. YOLO normalizes horizontal coordinates by image width, here 640 pixels. **2. Convert:** \[x_c=0.50(640)=320\text{ pixels}.\] **3. Check:** The object center is halfway across the image. Its y-center and image height are unnecessary for this x-coordinate. **Answer: 320.**

**1. Convert the box data:** The center is \(0.50(640)=320\) pixels and width is \(0.20(640)=128\) pixels. **2. Move half the width left:** \[x_{\min}=x_c-w/2=320-128/2=256\text{ pixels}.\] **3. Verify:** The right edge is 384, so the width is \(384-256=128\) and its midpoint is 320. **Answer: 256.**

**1. Use the convolution shape formula:** With unit dilation, output size is \(\lfloor(n+2p-k)/s\rfloor+1\). **2. Apply both layers:** First layer: \((64-3)/1+1=62\), giving \(62\times62\times32\). Second layer: \(\lfloor(62-3)/2\rfloor+1=30\), giving \(30\times30\times64\). **3. Flatten:** \[30(30)(64)=57{,}600.\] The batch dimension is not part of the single-image feature vector. **Answer: 57600.**

**1. Count output positions:** A stride-one 1-by-1 convolution preserves the 14-by-14 spatial size and produces 128 channels. **2. Count products per output:** Each output activation combines all 512 input channels, requiring 512 multiplications. **3. Multiply and convert units:** \[14(14)(128)(512)=12{,}845{,}056\text{ multiplications}=12.845056\text{ million}.\] Bias additions and other operations are not requested. **Answer: 12.845056.**

**1. Use the YOLO v1 convention:** Each cell predicts B boxes with five values each: x, y, width, height and confidence. The cell also predicts one shared vector of C class probabilities. **2. Compute depth:** \[D=5B+C=5(2)+20=30.\] **3. Distinguish depth and full shape:** The full output shape is \(7\times7\times30\). The requested channel depth is 30, rather than its total number of elements. **Answer: 30.**

**1. Count one filter:** Every 3-by-3 filter spans all 32 input channels, requiring \(3(3)(32)=288\) weights. **2. Count 64 filters:** \[N=64(288)=64(32)(3)(3)=18{,}432.\] **3. Exclude bias:** One bias per output channel would add 64 parameters, but the question explicitly asks for weights without bias. **Answer: 18432.**

**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, C — Selected: Mixed precision training reduces memory usage and speeds up computation.; Freezing early layers reduces the number of trainable parameters.; Using multiple data loading workers improves input pipeline efficiency.** A: **Correct:** Mixed precision reduces suitable tensor storage and can speed compatible GPU operations. B: **Correct:** Frozen early layers need no parameter gradients or optimizer updates, reducing trainable state. C: **Correct:** Background workers overlap loading/augmentation with computation when the input pipeline is a bottleneck. D: **Incorrect:** Larger batches normally require more activation memory; the word always is false.

**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: Both model parameters and input tensors must be moved to the same device for computation.; Saving only the state dict is preferred for flexibility and portability.; The model architecture must be defined before loading weights using load state dict.** 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. Stabilize activations:** Batch normalization uses mini-batch statistics during training, with learned affine scale/shift, to help control activation distributions. Evaluation normally uses accumulated statistics. **2. Regularize training:** Dropout randomly zeros selected activations during training. Data augmentation creates varied training examples through suitable transformations, such as crops or flips. **3. Keep nonlinear modeling:** These methods complement activation functions; batch normalization does not replace their role in representing nonlinear relationships. **Answer: A, B, C — Selected: Batch Normalization normalizes activations within a mini-batch to stabilize training.; Dropout randomly deactivates neurons during training to reduce overfitting.; Data augmentation artificially increases dataset diversity and improves generalization.** A: **Correct:** Batch normalization standardizes activations using training-batch statistics. B: **Correct:** Randomly dropping activations during training discourages reliance on particular features. C: **Correct:** Appropriate augmentation broadens observed variation and can improve generalization. D: **Incorrect:** Normalization does not eliminate the need for nonlinear activations.

**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: A symmetric encoder-decoder structure that captures both context and localization.; Skip connections that transfer high-resolution features directly to the decoder.; Upsampling operations that progressively recover spatial resolution.** 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 applies multiple convolution filters of different sizes in parallel within a single module.; It uses 1 × 1 convolutions to reduce channel dimensionality before expensive operations.; It introduces auxiliary classifiers to improve gradient flow during training.** 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. Use local connectivity:** A convolutional neuron reads a small image patch instead of connecting densely to every input pixel. **2. Share the filter:** The same learned weights are reused as the kernel moves across image locations. A convolution has \(k_hk_wC_{in}C_{out}\) weights, independent of image width/height. **3. Explain the efficiency:** Locality limits connections, while sharing avoids a separate weight set at every position. Nonlinearities are retained to model complex patterns. **Answer: A, B — Selected: Local connectivity, where neurons are connected only to a small spatial region of the input.; Parameter sharing, where the same filter is applied across different spatial locations.** A: **Correct:** Local receptive fields reduce connections compared with a dense layer over the whole image. B: **Correct:** Weight reuse across positions is the defining parameter-sharing mechanism. C: **Incorrect:** Independent dense weights at every location lose the convolutional sharing advantage. D: **Incorrect:** Removing activations would reduce nonlinear representational power.

**1. Read the layer defaults:** Conv2d has 3 input channels, 32 output channels, kernel 3, padding 1 and default stride 1. **2. Compute spatial size:** \[H_{out}=W_{out}=\lfloor(64+2-3)/1\rfloor+1=64.\] **3. Assemble the shape:** Batch size remains 1 and 32 filters produce 32 channels. The result is \((1,32,64,64)\). **Answer: B — Selected: Padding preserves spatial dimensions while increasing channel depth to 32.** A: **Incorrect:** Padding compensates for the kernel boundary reduction here. B: **Correct:** Spatial size stays 64-by-64 and channel count becomes 32. C: **Incorrect:** Output channels are determined by the 32 filters. D: **Incorrect:** Convolution preserves the leading batch dimension.

**1. Write a residual block:** Its output is \(y=x+F(x)\) when dimensions match. A projection can align dimensions when necessary. **2. Explain identity learning:** If little transformation is needed, learning \(F(x)\approx0\) lets the block approximate the identity. The skip path also supplies a direct route for gradient propagation. **3. Connect to degradation:** This makes deep networks easier to optimize than forcing every added layer stack to relearn identity behavior through all its parameters. **Answer: B — Selected: By enabling identity mappings that allow gradients to flow more easily through deep networks.** A: **Incorrect:** Residual connections do not generally reduce parameter count by themselves. B: **Correct:** Identity paths improve optimization and gradient flow through deep stacks. C: **Incorrect:** Residual networks still use activation functions. D: **Incorrect:** They retain convolutional feature extraction rather than converting it into dense layers.

**1. State the original YOLO definition:** In YOLO v1, a box-confidence target combines object presence and localization quality: \[C=\Pr(\text{object})\operatorname{IoU}(B,B_{true}).\] **2. Interpret the factors:** Confidence should be low if no object is present or if overlap is poor. It is distinct from the conditional class probabilities predicted for the cell. **3. State the convention:** B is the original YOLO confidence definition intended here. At inference, ground-truth boxes are unavailable; the network predicts this learned score. Modern YOLO variants, including YOLOv8, use different head/score formulations, so this is not a universal formula for every version. **Answer: B — Selected: The probability that an object exists in the box multiplied by the IoU with the ground truth.** A: **Incorrect:** This omits localization/objectness from the original box-confidence definition. B: **Correct:** Original YOLO combines object existence with overlap quality. C: **Incorrect:** IoU alone omits whether an object is present. D: **Incorrect:** Distance from the image center is not the confidence definition.

**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. **Learnability detail:** Pixel Shuffle itself is a parameter-free rearrangement; preceding convolution layers learn the low-resolution features that it rearranges. **Answer: C — Selected: It rearranges channel information into spatial dimensions, enabling learnable and efficient upsampling.** A: **Incorrect:** Pixel Shuffle does not intrinsically specify how many convolution layers are required. B: **Incorrect:** It does not guarantee recovery of information discarded by a downsampling process. C: **Correct:** Learned convolutional channel features are efficiently rearranged into higher spatial resolution. D: **Incorrect:** Global pooling removes spatial layout and is not the Pixel Shuffle operation.

**1. Define anchors:** They are reference boxes at feature-map locations, using several scales/aspect ratios. **2. Learn refinements:** The region proposal network predicts objectness and coordinate offsets relative to anchors. Training compares anchors/proposals with ground-truth objects to set targets and regress better boxes. **3. Separate prediction stages:** Anchors are starting references, not final classifications and not a replacement for the CNN feature extractor. **Answer: A — Selected: They serve as reference boxes that are adjusted (via regression) to match ground truth objects.** A: **Correct:** Regression adjusts reference anchors toward ground-truth object boxes. B: **Incorrect:** Anchors alone are not refined final class predictions. C: **Incorrect:** Anchors are used in target assignment and regression during training as well as prediction. D: **Incorrect:** Anchor predictions depend on extracted convolutional features.

**1. Compare equal-width kernels:** With C input/output channels, one 7-by-7 convolution has \(49C^2\) weights. Three 3-by-3 layers have \(3(9C^2)=27C^2\), ignoring biases. **2. Compare receptive fields:** Three stride-one 3-by-3 layers have effective receptive field \(1+3(3-1)=7\), matching a 7-by-7 spatial extent under the usual conditions. **3. Add nonlinear transformations:** Activations between the smaller convolutions add modeling flexibility. Parameter savings depend on the channel choices; this design does not guarantee absence of overfitting. **Answer: B — Selected: It reduces the total number of parameters while introducing additional non- linearities, improving representational power.** A: **Incorrect:** The stacked design normally adds intermediate activations rather than removing them. B: **Correct:** At comparable widths it reduces parameters and allows additional nonlinear transformations. C: **Incorrect:** Stacked small kernels can achieve the same effective receptive-field size. D: **Incorrect:** Generalization is not guaranteed by kernel size alone.